├── .github ├── pull_request_template.md └── workflows │ └── schema-validation.yml ├── .gitignore ├── .vscode └── settings.json ├── CONTRIBUTING.md ├── LICENSE.md ├── README.md ├── data ├── .github │ └── CODEOWNERS ├── benchmarks │ ├── aa-index.json │ ├── acebench.json │ ├── activitynet.json │ ├── agieval.json │ ├── ai2-reasoning-challenge-(arc).json │ ├── ai2d.json │ ├── aider-polyglot-edit.json │ ├── aider-polyglot.json │ ├── aider.json │ ├── aime-2024.json │ ├── aime-2025.json │ ├── aime.json │ ├── aitz-em.json │ ├── alignbench.json │ ├── alpacaeval-2.0.json │ ├── amc-2022-23.json │ ├── android-control-high-em.json │ ├── android-control-low-em.json │ ├── androidworld-sr.json │ ├── api-bank.json │ ├── arc-agi-v2.json │ ├── arc-agi.json │ ├── arc-c.json │ ├── arc-e.json │ ├── arc.json │ ├── arena-hard-v2.json │ ├── arena-hard.json │ ├── attaq.json │ ├── autologi.json │ ├── bbh.json │ ├── bfcl-v2.json │ ├── bfcl-v3-multiturn.json │ ├── bfcl-v3.json │ ├── bfcl.json │ ├── big-bench-extra-hard.json │ ├── big-bench-hard.json │ ├── big-bench.json │ ├── bigcodebench-full.json │ ├── bigcodebench-hard.json │ ├── bigcodebench.json │ ├── bird-sql-(dev).json │ ├── blink.json │ ├── boolq.json │ ├── browsecomp-long-128k.json │ ├── browsecomp-long-256k.json │ ├── browsecomp-zh.json │ ├── browsecomp.json │ ├── c-eval.json │ ├── cbnsl.json │ ├── cc-ocr.json │ ├── cfeval.json │ ├── charadessta.json │ ├── chartqa.json │ ├── charxiv-d.json │ ├── charxiv-r.json │ ├── chexpert-cxr.json │ ├── cluewsc.json │ ├── cmmlu.json │ ├── cnmo-2024.json │ ├── codeforces.json │ ├── codegolf-v2.2.json │ ├── collie.json │ ├── common-voice-15.json │ ├── commonsenseqa.json │ ├── complexfuncbench.json │ ├── covost2-en-zh.json │ ├── covost2.json │ ├── crag.json │ ├── creative-writing-v3.json │ ├── crperelation.json │ ├── crux-o.json │ ├── cruxeval-input-cot.json │ ├── cruxeval-o.json │ ├── cruxeval-output-cot.json │ ├── csimpleqa.json │ ├── cybersecurity-ctfs.json │ ├── dermmcqa.json │ ├── docvqa.json │ ├── docvqatest.json │ ├── drop.json │ ├── ds-arena-code.json │ ├── ds-fim-eval.json │ ├── eclektic.json │ ├── egoschema.json │ ├── erqa.json │ ├── evalplus.json │ ├── facts-grounding.json │ ├── factscore.json │ ├── finqa.json │ ├── flenqa.json │ ├── fleurs.json │ ├── frames.json │ ├── french-mmlu.json │ ├── frontiermath.json │ ├── functionalmath.json │ ├── giantsteps-tempo.json │ ├── global-mmlu-lite.json │ ├── global-mmlu.json │ ├── gorilla-benchmark-api-bench.json │ ├── govreport.json │ ├── gpqa-biology.json │ ├── gpqa-chemistry.json │ ├── gpqa-physics.json │ ├── gpqa.json │ ├── graphwalks-bfs-%3C128k.json │ ├── graphwalks-bfs-%3E128k.json │ ├── graphwalks-parents-%3C128k.json │ ├── graphwalks-parents-%3E128k.json │ ├── groundui-1k.json │ ├── gsm-8k-(cot).json │ ├── gsm8k-chat.json │ ├── gsm8k.json │ ├── hallusion-bench.json │ ├── healthbench-hard.json │ ├── healthbench.json │ ├── hellaswag.json │ ├── hiddenmath.json │ ├── hle.json │ ├── hmmt-2025.json │ ├── hmmt25.json │ ├── humaneval+.json │ ├── humaneval-average.json │ ├── humaneval-er.json │ ├── humaneval-mul.json │ ├── humaneval-plus.json │ ├── humaneval.json │ ├── humanevalfim-average.json │ ├── humanity's-last-exam.json │ ├── if.json │ ├── ifeval.json │ ├── include.json │ ├── infinitebench-en.mc.json │ ├── infinitebench-en.qa.json │ ├── infographicsqa.json │ ├── infovqa.json │ ├── infovqatest.json │ ├── instruct-humaneval.json │ ├── intergps.json │ ├── internal-api-instruction-following-(hard).json │ ├── lbpp-(v2).json │ ├── livebench-20241125.json │ ├── livebench.json │ ├── livecodebench(01-09).json │ ├── livecodebench-v5-24.12-25.2.json │ ├── livecodebench-v5.json │ ├── livecodebench-v6.json │ ├── livecodebench.json │ ├── longbench-v2.json │ ├── longfact-concepts.json │ ├── longfact-objects.json │ ├── longvideobench.json │ ├── lsat.json │ ├── lvbench.json │ ├── math-(cot).json │ ├── math-500.json │ ├── math.json │ ├── mathvision.json │ ├── mathvista-mini.json │ ├── mathvista.json │ ├── mbpp+.json │ ├── mbpp-++-base-version.json │ ├── mbpp-evalplus-(base).json │ ├── mbpp-evalplus.json │ ├── mbpp-pass@1.json │ ├── mbpp-plus.json │ ├── mbpp.json │ ├── medxpertqa.json │ ├── mega-mlqa.json │ ├── mega-tydi-qa.json │ ├── mega-udpos.json │ ├── mega-xcopa.json │ ├── mega-xstorycloze.json │ ├── meld.json │ ├── mgsm.json │ ├── mimic-cxr.json │ ├── mlvu-m.json │ ├── mlvu.json │ ├── mm-if-eval.json │ ├── mm-mind2web.json │ ├── mm-mt-bench.json │ ├── mmau-music.json │ ├── mmau-sound.json │ ├── mmau-speech.json │ ├── mmau.json │ ├── mmbench-test.json │ ├── mmbench-v1.1.json │ ├── mmbench-video.json │ ├── mmbench.json │ ├── mme-realworld.json │ ├── mme.json │ ├── mmlu-(cot).json │ ├── mmlu-base.json │ ├── mmlu-chat.json │ ├── mmlu-french.json │ ├── mmlu-pro.json │ ├── mmlu-prox.json │ ├── mmlu-redux-2.0.json │ ├── mmlu-redux.json │ ├── mmlu-stem.json │ ├── mmlu.json │ ├── mmmlu.json │ ├── mmmu-(val).json │ ├── mmmu-(validation).json │ ├── mmmu-pro.json │ ├── mmmu.json │ ├── mmmuval.json │ ├── mmstar.json │ ├── mmt-bench.json │ ├── mmvet.json │ ├── mmvetgpt4turbo.json │ ├── mobileminiwob++-sr.json │ ├── mrcr-1m-(pointwise).json │ ├── mrcr-1m.json │ ├── mrcr-v2-(8-needle).json │ ├── mrcr-v2.json │ ├── mrcr.json │ ├── mt-bench.json │ ├── mtvqa.json │ ├── muirbench.json │ ├── multi-if.json │ ├── multi-swe-bench.json │ ├── multichallenge-(o3-mini-grader).json │ ├── multichallenge.json │ ├── multilf.json │ ├── multilingual-mgsm-(cot).json │ ├── multilingual-mmlu.json │ ├── multipl-e-humaneval.json │ ├── multipl-e-mbpp.json │ ├── multipl-e.json │ ├── musiccaps.json │ ├── musr.json │ ├── mvbench.json │ ├── natural-questions.json │ ├── natural2code.json │ ├── nexus.json │ ├── nih-multi-needle.json │ ├── nmos.json │ ├── nq.json │ ├── ocrbench-v2-(en).json │ ├── ocrbench-v2-(zh).json │ ├── ocrbench-v2.json │ ├── ocrbench.json │ ├── odinw.json │ ├── ojbench.json │ ├── olympiadbench.json │ ├── omnibench-music.json │ ├── omnibench.json │ ├── omnimath.json │ ├── open-rewrite.json │ ├── openai-mmlu.json │ ├── openai-mrcr%3A-2-needle-128k.json │ ├── openai-mrcr%3A-2-needle-1m.json │ ├── openai-mrcr%3A-2-needle-256k.json │ ├── openbookqa.json │ ├── osworld-extended.json │ ├── osworld-screenshot-only.json │ ├── osworld.json │ ├── pathmcqa.json │ ├── perceptiontest.json │ ├── phibench.json │ ├── physicsfinals.json │ ├── piqa.json │ ├── pointgrounding.json │ ├── polymath-en.json │ ├── polymath.json │ ├── pope.json │ ├── popqa.json │ ├── qasper.json │ ├── qmsum.json │ ├── realworldqa.json │ ├── repobench.json │ ├── repoqa.json │ ├── ruler.json │ ├── sat-math.json │ ├── scale-multichallenge.json │ ├── scicode.json │ ├── scienceqa-visual.json │ ├── scienceqa.json │ ├── screenspot-pro.json │ ├── screenspot.json │ ├── simpleqa.json │ ├── slakevqa.json │ ├── social-iqa.json │ ├── spider.json │ ├── squality.json │ ├── stem.json │ ├── summscreenfd.json │ ├── superglue.json │ ├── supergpqa.json │ ├── swe-bench-multilingual.json │ ├── swe-bench-verified-(agentic-coding).json │ ├── swe-bench-verified-(agentless).json │ ├── swe-bench-verified-(multiple-attempts).json │ ├── swe-bench-verified.json │ ├── swe-dev.json │ ├── swe-lancer-(ic-diamond-subset).json │ ├── swe-lancer.json │ ├── tau-bench-airline.json │ ├── tau-bench-retail.json │ ├── tau-bench.json │ ├── tau2-airline.json │ ├── tau2-retail.json │ ├── tau2-telecom.json │ ├── tempcompass.json │ ├── terminal-bench.json │ ├── terminus.json │ ├── textvqa.json │ ├── theoremqa.json │ ├── tldr9+-(test).json │ ├── translation-en-to-set1-comet22.json │ ├── translation-en-to-set1-spbleu.json │ ├── translation-set1-to-en-comet22.json │ ├── translation-set1-to-en-spbleu.json │ ├── triviaqa.json │ ├── truthfulqa.json │ ├── tydiqa.json │ ├── uniform-bar-exam.json │ ├── usamo25.json │ ├── vatex.json │ ├── vcr-en-easy.json │ ├── vibe-eval.json │ ├── video-mme-(long,-no-subtitles).json │ ├── video-mme.json │ ├── video-mmew-sub.json │ ├── videomme-w-o-sub..json │ ├── videomme-w-sub..json │ ├── videommmu.json │ ├── visualwebbench.json │ ├── vocalsound.json │ ├── voicebench-avg.json │ ├── vqa-rad.json │ ├── vqav2-(test).json │ ├── vqav2-(val).json │ ├── vqav2.json │ ├── wild-bench.json │ ├── winogrande.json │ ├── wmt23.json │ ├── wmt24++.json │ ├── writingbench.json │ ├── xlsum-english.json │ ├── xstest.json │ └── zebralogic.json ├── licenses │ ├── apache_2_0.json │ ├── cc_by_nc.json │ ├── creative_commons_attribution_4_0_license.json │ ├── deepseek.json │ ├── gemma.json │ ├── health_ai_developer_foundations_terms_of_use.json │ ├── jamba_open_model_license.json │ ├── llama3_2.json │ ├── llama_3_1_community_license.json │ ├── llama_3_2_community_license.json │ ├── llama_3_3_community_license_agreement.json │ ├── llama_4_community_license_agreement.json │ ├── mistral_research_license.json │ ├── mistral_research_license_(mrl)_for_research;_mistral_commercial_license_for_commercial_use.json │ ├── mit.json │ ├── mit_+_model_license_(commercial_use_allowed).json │ ├── mit_license.json │ ├── mnpl_0_1.json │ ├── modified_mit_license.json │ ├── nvidia_open_model_license_agreement.json │ ├── proprietary.json │ ├── qwen.json │ ├── tongyi_qianwen.json │ └── unknown.json ├── organizations │ ├── ai21 │ │ ├── models │ │ │ ├── jamba-1.5-large │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ └── jamba-1.5-mini │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ └── organization.json │ ├── amazon │ │ ├── models │ │ │ ├── nova-lite │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── nova-micro │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ └── nova-pro │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ └── organization.json │ ├── anthropic │ │ ├── models │ │ │ ├── claude-3-5-haiku-20241022 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── claude-3-5-sonnet-20240620 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── claude-3-5-sonnet-20241022 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── claude-3-7-sonnet-20250219 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── claude-3-haiku-20240307 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── claude-3-opus-20240229 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── claude-3-sonnet-20240229 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── claude-haiku-4-5-20251015 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── claude-opus-4-1-20250805 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── claude-opus-4-20250514 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── claude-sonnet-4-20250514 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ └── claude-sonnet-4-5-20250929 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ └── organization.json │ ├── cohere │ │ ├── models │ │ │ └── command-r-plus-04-2024 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ └── organization.json │ ├── deepseek │ │ ├── models │ │ │ ├── deepseek-r1-0528 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── deepseek-r1-distill-llama-70b │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── deepseek-r1-distill-llama-8b │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── deepseek-r1-distill-qwen-1.5b │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── deepseek-r1-distill-qwen-14b │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── deepseek-r1-distill-qwen-32b │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── deepseek-r1-distill-qwen-7b │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── deepseek-r1-zero │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── deepseek-r1 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── deepseek-v2.5 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── deepseek-v3-0324 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── deepseek-v3.1 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── deepseek-v3.2-exp │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── deepseek-v3 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── deepseek-vl2-small │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── deepseek-vl2-tiny │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ └── deepseek-vl2 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ └── organization.json │ ├── google │ │ ├── models │ │ │ ├── gemini-1.0-pro │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── gemini-1.5-flash-8b │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── gemini-1.5-flash │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── gemini-1.5-pro │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── gemini-2.0-flash-lite │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── gemini-2.0-flash-thinking │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── gemini-2.0-flash │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── gemini-2.5-flash-lite │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── gemini-2.5-flash │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── gemini-2.5-pro-preview-06-05 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── gemini-2.5-pro │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── gemini-diffusion │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── gemma-2-27b-it │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── gemma-2-9b-it │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── gemma-3-12b-it │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── gemma-3-1b-it │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── gemma-3-27b-it │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── gemma-3-4b-it │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── gemma-3n-e2b-it-litert-preview │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── gemma-3n-e2b-it │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── gemma-3n-e2b │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── gemma-3n-e4b-it-litert-preview │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── gemma-3n-e4b-it │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── gemma-3n-e4b │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ └── medgemma-4b-it │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ └── organization.json │ ├── ibm │ │ ├── models │ │ │ ├── granite-3.3-8b-base │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── granite-3.3-8b-instruct │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ └── granite-4.0-tiny-preview │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ └── organization.json │ ├── meta │ │ ├── models │ │ │ ├── llama-3.1-405b-instruct │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── llama-3.1-70b-instruct │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── llama-3.1-8b-instruct │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── llama-3.2-11b-instruct │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── llama-3.2-3b-instruct │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── llama-3.2-90b-instruct │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── llama-3.3-70b-instruct │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── llama-4-maverick │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ └── llama-4-scout │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ └── organization.json │ ├── microsoft │ │ ├── models │ │ │ ├── phi-3.5-mini-instruct │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── phi-3.5-moe-instruct │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── phi-3.5-vision-instruct │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── phi-4-mini-reasoning │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── phi-4-mini │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── phi-4-multimodal-instruct │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── phi-4-reasoning-plus │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── phi-4-reasoning │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ └── phi-4 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ └── organization.json │ ├── mistral │ │ ├── models │ │ │ ├── codestral-22b │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── devstral-medium-2507 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── devstral-small-2507 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── magistral-medium │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── magistral-small-2506 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── ministral-8b-instruct-2410 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── mistral-large-2-2407 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── mistral-nemo-instruct-2407 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── mistral-small-2409 │ │ │ │ └── model.json │ │ │ ├── mistral-small-24b-base-2501 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── mistral-small-24b-instruct-2501 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── mistral-small-3.1-24b-base-2503 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── mistral-small-3.1-24b-instruct-2503 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── mistral-small-3.2-24b-instruct-2506 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── pixtral-12b-2409 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ └── pixtral-large │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ └── organization.json │ ├── moonshotai │ │ ├── models │ │ │ ├── kimi-k1.5 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── kimi-k2-0905 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── kimi-k2-base │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── kimi-k2-instruct-0905 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ └── kimi-k2-instruct │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ └── organization.json │ ├── nvidia │ │ ├── models │ │ │ ├── llama-3.1-nemotron-70b-instruct │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── llama-3.1-nemotron-nano-8b-v1 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── llama-3.1-nemotron-ultra-253b-v1 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── llama-3.3-nemotron-super-49b-v1 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ └── nemotron-nano-9b-v2 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ └── organization.json │ ├── openai │ │ ├── models │ │ │ ├── gpt-3.5-turbo-0125 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── gpt-4-0613 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── gpt-4-turbo-2024-04-09 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── gpt-4.1-2025-04-14 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── gpt-4.1-mini-2025-04-14 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── gpt-4.1-nano-2025-04-14 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── gpt-4.5 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── gpt-4o-2024-05-13 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── gpt-4o-2024-08-06 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── gpt-4o-mini-2024-07-18 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── gpt-5-2025-08-07 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── gpt-5-codex-2025-09-15 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── gpt-5-mini-2025-08-07 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── gpt-5-nano-2025-08-07 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── gpt-oss-120b │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── gpt-oss-20b │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── o1-2024-12-17 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── o1-mini │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── o1-preview │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── o1-pro │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── o3-2025-04-16 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── o3-mini │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── o3-pro-2025-06-10 │ │ │ │ └── model.json │ │ │ └── o4-mini │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ └── organization.json │ ├── qwen │ │ ├── models │ │ │ ├── qvq-72b-preview │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── qwen-2.5-14b-instruct │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── qwen-2.5-32b-instruct │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── qwen-2.5-72b-instruct │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── qwen-2.5-7b-instruct │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── qwen-2.5-coder-32b-instruct │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── qwen-2.5-coder-7b-instruct │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── qwen2-72b-instruct │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── qwen2-7b-instruct │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── qwen2-vl-72b │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── qwen2.5-omni-7b │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── qwen2.5-vl-32b │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── qwen2.5-vl-72b │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── qwen2.5-vl-7b │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── qwen3-235b-a22b-instruct-2507 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── qwen3-235b-a22b-thinking-2507 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── qwen3-235b-a22b │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── qwen3-30b-a3b │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── qwen3-32b │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── qwen3-next-80b-a3b-base │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── qwen3-next-80b-a3b-instruct │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── qwen3-next-80b-a3b-thinking │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── qwq-32b-preview │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ └── qwq-32b │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ └── organization.json │ ├── unknown │ │ └── organization.json │ ├── xai │ │ ├── models │ │ │ ├── grok-1.5 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── grok-1.5v │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── grok-2-mini │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── grok-2 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── grok-3-mini │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── grok-3 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── grok-4-fast │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── grok-4-heavy │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ ├── grok-4 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ │ └── grok-code-fast-1 │ │ │ │ ├── benchmarks.json │ │ │ │ └── model.json │ │ └── organization.json │ └── zai-org │ │ ├── models │ │ ├── glm-4.5-air │ │ │ ├── benchmarks.json │ │ │ └── model.json │ │ ├── glm-4.5 │ │ │ ├── benchmarks.json │ │ │ └── model.json │ │ ├── glm-4.5v │ │ │ ├── benchmarks.json │ │ │ └── model.json │ │ └── glm-4.6 │ │ │ ├── benchmarks.json │ │ │ └── model.json │ │ └── organization.json └── providers │ ├── anthropic │ ├── models.json │ └── provider.json │ ├── azure │ ├── models.json │ └── provider.json │ ├── bedrock │ ├── models.json │ └── provider.json │ ├── cerebras │ ├── models.json │ └── provider.json │ ├── cohere │ ├── models.json │ └── provider.json │ ├── deepinfra │ ├── models.json │ └── provider.json │ ├── deepseek │ ├── models.json │ └── provider.json │ ├── fireworks │ ├── models.json │ └── provider.json │ ├── google │ ├── models.json │ └── provider.json │ ├── groq │ ├── models.json │ └── provider.json │ ├── hyperbolic │ ├── models.json │ └── provider.json │ ├── lambda │ ├── models.json │ └── provider.json │ ├── mistral │ ├── models.json │ └── provider.json │ ├── novita │ ├── models.json │ └── provider.json │ ├── openai │ ├── models.json │ └── provider.json │ ├── replicate │ ├── models.json │ └── provider.json │ ├── sambanova │ ├── models.json │ └── provider.json │ ├── together │ ├── models.json │ └── provider.json │ ├── xai │ ├── models.json │ └── provider.json │ └── zeroeval │ ├── models.json │ └── provider.json ├── package.json └── schemas ├── README.md ├── benchmark-results.schema.json ├── benchmark.schema.json ├── integrity-validator.js ├── license.schema.json ├── model.schema.json ├── organization.schema.json ├── provider-models.schema.json ├── provider.schema.json └── validator.js /.github/pull_request_template.md: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/.github/pull_request_template.md -------------------------------------------------------------------------------- /.github/workflows/schema-validation.yml: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/.github/workflows/schema-validation.yml -------------------------------------------------------------------------------- /.gitignore: -------------------------------------------------------------------------------- 1 | /node_modules 2 | -------------------------------------------------------------------------------- /.vscode/settings.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/.vscode/settings.json -------------------------------------------------------------------------------- /CONTRIBUTING.md: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/CONTRIBUTING.md -------------------------------------------------------------------------------- /LICENSE.md: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/LICENSE.md -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/README.md -------------------------------------------------------------------------------- /data/.github/CODEOWNERS: -------------------------------------------------------------------------------- 1 | * @JonathanChavezTamales 2 | * @sebastiancrossa 3 | -------------------------------------------------------------------------------- /data/benchmarks/aa-index.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/aa-index.json -------------------------------------------------------------------------------- /data/benchmarks/acebench.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/acebench.json -------------------------------------------------------------------------------- /data/benchmarks/activitynet.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/activitynet.json -------------------------------------------------------------------------------- /data/benchmarks/agieval.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/agieval.json -------------------------------------------------------------------------------- /data/benchmarks/ai2-reasoning-challenge-(arc).json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/ai2-reasoning-challenge-(arc).json -------------------------------------------------------------------------------- /data/benchmarks/ai2d.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/ai2d.json -------------------------------------------------------------------------------- /data/benchmarks/aider-polyglot-edit.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/aider-polyglot-edit.json -------------------------------------------------------------------------------- /data/benchmarks/aider-polyglot.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/aider-polyglot.json -------------------------------------------------------------------------------- /data/benchmarks/aider.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/aider.json -------------------------------------------------------------------------------- /data/benchmarks/aime-2024.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/aime-2024.json -------------------------------------------------------------------------------- /data/benchmarks/aime-2025.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/aime-2025.json -------------------------------------------------------------------------------- /data/benchmarks/aime.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/aime.json -------------------------------------------------------------------------------- /data/benchmarks/aitz-em.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/aitz-em.json -------------------------------------------------------------------------------- /data/benchmarks/alignbench.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/alignbench.json -------------------------------------------------------------------------------- /data/benchmarks/alpacaeval-2.0.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/alpacaeval-2.0.json -------------------------------------------------------------------------------- /data/benchmarks/amc-2022-23.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/amc-2022-23.json -------------------------------------------------------------------------------- /data/benchmarks/android-control-high-em.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/android-control-high-em.json -------------------------------------------------------------------------------- /data/benchmarks/android-control-low-em.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/android-control-low-em.json -------------------------------------------------------------------------------- /data/benchmarks/androidworld-sr.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/androidworld-sr.json -------------------------------------------------------------------------------- /data/benchmarks/api-bank.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/api-bank.json -------------------------------------------------------------------------------- /data/benchmarks/arc-agi-v2.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/arc-agi-v2.json -------------------------------------------------------------------------------- /data/benchmarks/arc-agi.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/arc-agi.json -------------------------------------------------------------------------------- /data/benchmarks/arc-c.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/arc-c.json -------------------------------------------------------------------------------- /data/benchmarks/arc-e.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/arc-e.json -------------------------------------------------------------------------------- /data/benchmarks/arc.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/arc.json -------------------------------------------------------------------------------- /data/benchmarks/arena-hard-v2.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/arena-hard-v2.json -------------------------------------------------------------------------------- /data/benchmarks/arena-hard.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/arena-hard.json -------------------------------------------------------------------------------- /data/benchmarks/attaq.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/attaq.json -------------------------------------------------------------------------------- /data/benchmarks/autologi.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/autologi.json -------------------------------------------------------------------------------- /data/benchmarks/bbh.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/bbh.json -------------------------------------------------------------------------------- /data/benchmarks/bfcl-v2.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/bfcl-v2.json -------------------------------------------------------------------------------- /data/benchmarks/bfcl-v3-multiturn.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/bfcl-v3-multiturn.json -------------------------------------------------------------------------------- /data/benchmarks/bfcl-v3.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/bfcl-v3.json -------------------------------------------------------------------------------- /data/benchmarks/bfcl.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/bfcl.json -------------------------------------------------------------------------------- /data/benchmarks/big-bench-extra-hard.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/big-bench-extra-hard.json -------------------------------------------------------------------------------- /data/benchmarks/big-bench-hard.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/big-bench-hard.json -------------------------------------------------------------------------------- /data/benchmarks/big-bench.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/big-bench.json -------------------------------------------------------------------------------- /data/benchmarks/bigcodebench-full.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/bigcodebench-full.json -------------------------------------------------------------------------------- /data/benchmarks/bigcodebench-hard.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/bigcodebench-hard.json -------------------------------------------------------------------------------- /data/benchmarks/bigcodebench.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/bigcodebench.json -------------------------------------------------------------------------------- /data/benchmarks/bird-sql-(dev).json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/bird-sql-(dev).json -------------------------------------------------------------------------------- /data/benchmarks/blink.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/blink.json -------------------------------------------------------------------------------- /data/benchmarks/boolq.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/boolq.json -------------------------------------------------------------------------------- /data/benchmarks/browsecomp-long-128k.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/browsecomp-long-128k.json -------------------------------------------------------------------------------- /data/benchmarks/browsecomp-long-256k.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/browsecomp-long-256k.json -------------------------------------------------------------------------------- /data/benchmarks/browsecomp-zh.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/browsecomp-zh.json -------------------------------------------------------------------------------- /data/benchmarks/browsecomp.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/browsecomp.json -------------------------------------------------------------------------------- /data/benchmarks/c-eval.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/c-eval.json -------------------------------------------------------------------------------- /data/benchmarks/cbnsl.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/cbnsl.json -------------------------------------------------------------------------------- /data/benchmarks/cc-ocr.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/cc-ocr.json -------------------------------------------------------------------------------- /data/benchmarks/cfeval.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/cfeval.json -------------------------------------------------------------------------------- /data/benchmarks/charadessta.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/charadessta.json -------------------------------------------------------------------------------- /data/benchmarks/chartqa.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/chartqa.json -------------------------------------------------------------------------------- /data/benchmarks/charxiv-d.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/charxiv-d.json -------------------------------------------------------------------------------- /data/benchmarks/charxiv-r.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/charxiv-r.json -------------------------------------------------------------------------------- /data/benchmarks/chexpert-cxr.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/chexpert-cxr.json -------------------------------------------------------------------------------- /data/benchmarks/cluewsc.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/cluewsc.json -------------------------------------------------------------------------------- /data/benchmarks/cmmlu.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/cmmlu.json -------------------------------------------------------------------------------- /data/benchmarks/cnmo-2024.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/cnmo-2024.json -------------------------------------------------------------------------------- /data/benchmarks/codeforces.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/codeforces.json -------------------------------------------------------------------------------- /data/benchmarks/codegolf-v2.2.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/codegolf-v2.2.json -------------------------------------------------------------------------------- /data/benchmarks/collie.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/collie.json -------------------------------------------------------------------------------- /data/benchmarks/common-voice-15.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/common-voice-15.json -------------------------------------------------------------------------------- /data/benchmarks/commonsenseqa.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/commonsenseqa.json -------------------------------------------------------------------------------- /data/benchmarks/complexfuncbench.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/complexfuncbench.json -------------------------------------------------------------------------------- /data/benchmarks/covost2-en-zh.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/covost2-en-zh.json -------------------------------------------------------------------------------- /data/benchmarks/covost2.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/covost2.json -------------------------------------------------------------------------------- /data/benchmarks/crag.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/crag.json -------------------------------------------------------------------------------- /data/benchmarks/creative-writing-v3.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/creative-writing-v3.json -------------------------------------------------------------------------------- /data/benchmarks/crperelation.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/crperelation.json -------------------------------------------------------------------------------- /data/benchmarks/crux-o.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/crux-o.json -------------------------------------------------------------------------------- /data/benchmarks/cruxeval-input-cot.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/cruxeval-input-cot.json -------------------------------------------------------------------------------- /data/benchmarks/cruxeval-o.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/cruxeval-o.json -------------------------------------------------------------------------------- /data/benchmarks/cruxeval-output-cot.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/cruxeval-output-cot.json -------------------------------------------------------------------------------- /data/benchmarks/csimpleqa.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/csimpleqa.json -------------------------------------------------------------------------------- /data/benchmarks/cybersecurity-ctfs.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/cybersecurity-ctfs.json -------------------------------------------------------------------------------- /data/benchmarks/dermmcqa.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/dermmcqa.json -------------------------------------------------------------------------------- /data/benchmarks/docvqa.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/docvqa.json -------------------------------------------------------------------------------- /data/benchmarks/docvqatest.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/docvqatest.json -------------------------------------------------------------------------------- /data/benchmarks/drop.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/drop.json -------------------------------------------------------------------------------- /data/benchmarks/ds-arena-code.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/ds-arena-code.json -------------------------------------------------------------------------------- /data/benchmarks/ds-fim-eval.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/ds-fim-eval.json -------------------------------------------------------------------------------- /data/benchmarks/eclektic.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/eclektic.json -------------------------------------------------------------------------------- /data/benchmarks/egoschema.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/egoschema.json -------------------------------------------------------------------------------- /data/benchmarks/erqa.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/erqa.json -------------------------------------------------------------------------------- /data/benchmarks/evalplus.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/evalplus.json -------------------------------------------------------------------------------- /data/benchmarks/facts-grounding.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/facts-grounding.json -------------------------------------------------------------------------------- /data/benchmarks/factscore.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/factscore.json -------------------------------------------------------------------------------- /data/benchmarks/finqa.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/finqa.json -------------------------------------------------------------------------------- /data/benchmarks/flenqa.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/flenqa.json -------------------------------------------------------------------------------- /data/benchmarks/fleurs.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/fleurs.json -------------------------------------------------------------------------------- /data/benchmarks/frames.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/frames.json -------------------------------------------------------------------------------- /data/benchmarks/french-mmlu.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/french-mmlu.json -------------------------------------------------------------------------------- /data/benchmarks/frontiermath.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/frontiermath.json -------------------------------------------------------------------------------- /data/benchmarks/functionalmath.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/functionalmath.json -------------------------------------------------------------------------------- /data/benchmarks/giantsteps-tempo.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/giantsteps-tempo.json -------------------------------------------------------------------------------- /data/benchmarks/global-mmlu-lite.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/global-mmlu-lite.json -------------------------------------------------------------------------------- /data/benchmarks/global-mmlu.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/global-mmlu.json -------------------------------------------------------------------------------- /data/benchmarks/gorilla-benchmark-api-bench.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/gorilla-benchmark-api-bench.json -------------------------------------------------------------------------------- /data/benchmarks/govreport.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/govreport.json -------------------------------------------------------------------------------- /data/benchmarks/gpqa-biology.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/gpqa-biology.json -------------------------------------------------------------------------------- /data/benchmarks/gpqa-chemistry.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/gpqa-chemistry.json -------------------------------------------------------------------------------- /data/benchmarks/gpqa-physics.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/gpqa-physics.json -------------------------------------------------------------------------------- /data/benchmarks/gpqa.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/gpqa.json -------------------------------------------------------------------------------- /data/benchmarks/graphwalks-bfs-%3C128k.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/graphwalks-bfs-%3C128k.json -------------------------------------------------------------------------------- /data/benchmarks/graphwalks-bfs-%3E128k.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/graphwalks-bfs-%3E128k.json -------------------------------------------------------------------------------- /data/benchmarks/graphwalks-parents-%3C128k.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/graphwalks-parents-%3C128k.json -------------------------------------------------------------------------------- /data/benchmarks/graphwalks-parents-%3E128k.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/graphwalks-parents-%3E128k.json -------------------------------------------------------------------------------- /data/benchmarks/groundui-1k.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/groundui-1k.json -------------------------------------------------------------------------------- /data/benchmarks/gsm-8k-(cot).json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/gsm-8k-(cot).json -------------------------------------------------------------------------------- /data/benchmarks/gsm8k-chat.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/gsm8k-chat.json -------------------------------------------------------------------------------- /data/benchmarks/gsm8k.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/gsm8k.json -------------------------------------------------------------------------------- /data/benchmarks/hallusion-bench.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/hallusion-bench.json -------------------------------------------------------------------------------- /data/benchmarks/healthbench-hard.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/healthbench-hard.json -------------------------------------------------------------------------------- /data/benchmarks/healthbench.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/healthbench.json -------------------------------------------------------------------------------- /data/benchmarks/hellaswag.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/hellaswag.json -------------------------------------------------------------------------------- /data/benchmarks/hiddenmath.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/hiddenmath.json -------------------------------------------------------------------------------- /data/benchmarks/hle.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/hle.json -------------------------------------------------------------------------------- /data/benchmarks/hmmt-2025.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/hmmt-2025.json -------------------------------------------------------------------------------- /data/benchmarks/hmmt25.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/hmmt25.json -------------------------------------------------------------------------------- /data/benchmarks/humaneval+.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/humaneval+.json -------------------------------------------------------------------------------- /data/benchmarks/humaneval-average.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/humaneval-average.json -------------------------------------------------------------------------------- /data/benchmarks/humaneval-er.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/humaneval-er.json -------------------------------------------------------------------------------- /data/benchmarks/humaneval-mul.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/humaneval-mul.json -------------------------------------------------------------------------------- /data/benchmarks/humaneval-plus.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/humaneval-plus.json -------------------------------------------------------------------------------- /data/benchmarks/humaneval.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/humaneval.json -------------------------------------------------------------------------------- /data/benchmarks/humanevalfim-average.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/humanevalfim-average.json -------------------------------------------------------------------------------- /data/benchmarks/humanity's-last-exam.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/humanity's-last-exam.json -------------------------------------------------------------------------------- /data/benchmarks/if.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/if.json -------------------------------------------------------------------------------- /data/benchmarks/ifeval.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/ifeval.json -------------------------------------------------------------------------------- /data/benchmarks/include.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/include.json -------------------------------------------------------------------------------- /data/benchmarks/infinitebench-en.mc.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/infinitebench-en.mc.json -------------------------------------------------------------------------------- /data/benchmarks/infinitebench-en.qa.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/infinitebench-en.qa.json -------------------------------------------------------------------------------- /data/benchmarks/infographicsqa.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/infographicsqa.json -------------------------------------------------------------------------------- /data/benchmarks/infovqa.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/infovqa.json -------------------------------------------------------------------------------- /data/benchmarks/infovqatest.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/infovqatest.json -------------------------------------------------------------------------------- /data/benchmarks/instruct-humaneval.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/instruct-humaneval.json -------------------------------------------------------------------------------- /data/benchmarks/intergps.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/intergps.json -------------------------------------------------------------------------------- /data/benchmarks/internal-api-instruction-following-(hard).json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/internal-api-instruction-following-(hard).json -------------------------------------------------------------------------------- /data/benchmarks/lbpp-(v2).json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/lbpp-(v2).json -------------------------------------------------------------------------------- /data/benchmarks/livebench-20241125.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/livebench-20241125.json -------------------------------------------------------------------------------- /data/benchmarks/livebench.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/livebench.json -------------------------------------------------------------------------------- /data/benchmarks/livecodebench(01-09).json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/livecodebench(01-09).json -------------------------------------------------------------------------------- /data/benchmarks/livecodebench-v5-24.12-25.2.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/livecodebench-v5-24.12-25.2.json -------------------------------------------------------------------------------- /data/benchmarks/livecodebench-v5.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/livecodebench-v5.json -------------------------------------------------------------------------------- /data/benchmarks/livecodebench-v6.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/livecodebench-v6.json -------------------------------------------------------------------------------- /data/benchmarks/livecodebench.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/livecodebench.json -------------------------------------------------------------------------------- /data/benchmarks/longbench-v2.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/longbench-v2.json -------------------------------------------------------------------------------- /data/benchmarks/longfact-concepts.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/longfact-concepts.json -------------------------------------------------------------------------------- /data/benchmarks/longfact-objects.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/longfact-objects.json -------------------------------------------------------------------------------- /data/benchmarks/longvideobench.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/longvideobench.json -------------------------------------------------------------------------------- /data/benchmarks/lsat.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/lsat.json -------------------------------------------------------------------------------- /data/benchmarks/lvbench.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/lvbench.json -------------------------------------------------------------------------------- /data/benchmarks/math-(cot).json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/math-(cot).json -------------------------------------------------------------------------------- /data/benchmarks/math-500.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/math-500.json -------------------------------------------------------------------------------- /data/benchmarks/math.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/math.json -------------------------------------------------------------------------------- /data/benchmarks/mathvision.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mathvision.json -------------------------------------------------------------------------------- /data/benchmarks/mathvista-mini.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mathvista-mini.json -------------------------------------------------------------------------------- /data/benchmarks/mathvista.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mathvista.json -------------------------------------------------------------------------------- /data/benchmarks/mbpp+.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mbpp+.json -------------------------------------------------------------------------------- /data/benchmarks/mbpp-++-base-version.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mbpp-++-base-version.json -------------------------------------------------------------------------------- /data/benchmarks/mbpp-evalplus-(base).json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mbpp-evalplus-(base).json -------------------------------------------------------------------------------- /data/benchmarks/mbpp-evalplus.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mbpp-evalplus.json -------------------------------------------------------------------------------- /data/benchmarks/mbpp-pass@1.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mbpp-pass@1.json -------------------------------------------------------------------------------- /data/benchmarks/mbpp-plus.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mbpp-plus.json -------------------------------------------------------------------------------- /data/benchmarks/mbpp.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mbpp.json -------------------------------------------------------------------------------- /data/benchmarks/medxpertqa.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/medxpertqa.json -------------------------------------------------------------------------------- /data/benchmarks/mega-mlqa.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mega-mlqa.json -------------------------------------------------------------------------------- /data/benchmarks/mega-tydi-qa.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mega-tydi-qa.json -------------------------------------------------------------------------------- /data/benchmarks/mega-udpos.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mega-udpos.json -------------------------------------------------------------------------------- /data/benchmarks/mega-xcopa.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mega-xcopa.json -------------------------------------------------------------------------------- /data/benchmarks/mega-xstorycloze.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mega-xstorycloze.json -------------------------------------------------------------------------------- /data/benchmarks/meld.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/meld.json -------------------------------------------------------------------------------- /data/benchmarks/mgsm.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mgsm.json -------------------------------------------------------------------------------- /data/benchmarks/mimic-cxr.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mimic-cxr.json -------------------------------------------------------------------------------- /data/benchmarks/mlvu-m.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mlvu-m.json -------------------------------------------------------------------------------- /data/benchmarks/mlvu.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mlvu.json -------------------------------------------------------------------------------- /data/benchmarks/mm-if-eval.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mm-if-eval.json -------------------------------------------------------------------------------- /data/benchmarks/mm-mind2web.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mm-mind2web.json -------------------------------------------------------------------------------- /data/benchmarks/mm-mt-bench.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mm-mt-bench.json -------------------------------------------------------------------------------- /data/benchmarks/mmau-music.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mmau-music.json -------------------------------------------------------------------------------- /data/benchmarks/mmau-sound.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mmau-sound.json -------------------------------------------------------------------------------- /data/benchmarks/mmau-speech.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mmau-speech.json -------------------------------------------------------------------------------- /data/benchmarks/mmau.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mmau.json -------------------------------------------------------------------------------- /data/benchmarks/mmbench-test.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mmbench-test.json -------------------------------------------------------------------------------- /data/benchmarks/mmbench-v1.1.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mmbench-v1.1.json -------------------------------------------------------------------------------- /data/benchmarks/mmbench-video.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mmbench-video.json -------------------------------------------------------------------------------- /data/benchmarks/mmbench.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mmbench.json -------------------------------------------------------------------------------- /data/benchmarks/mme-realworld.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mme-realworld.json -------------------------------------------------------------------------------- /data/benchmarks/mme.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mme.json -------------------------------------------------------------------------------- /data/benchmarks/mmlu-(cot).json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mmlu-(cot).json -------------------------------------------------------------------------------- /data/benchmarks/mmlu-base.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mmlu-base.json -------------------------------------------------------------------------------- /data/benchmarks/mmlu-chat.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mmlu-chat.json -------------------------------------------------------------------------------- /data/benchmarks/mmlu-french.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mmlu-french.json -------------------------------------------------------------------------------- /data/benchmarks/mmlu-pro.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mmlu-pro.json -------------------------------------------------------------------------------- /data/benchmarks/mmlu-prox.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mmlu-prox.json -------------------------------------------------------------------------------- /data/benchmarks/mmlu-redux-2.0.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mmlu-redux-2.0.json -------------------------------------------------------------------------------- /data/benchmarks/mmlu-redux.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mmlu-redux.json -------------------------------------------------------------------------------- /data/benchmarks/mmlu-stem.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mmlu-stem.json -------------------------------------------------------------------------------- /data/benchmarks/mmlu.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mmlu.json -------------------------------------------------------------------------------- /data/benchmarks/mmmlu.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mmmlu.json -------------------------------------------------------------------------------- /data/benchmarks/mmmu-(val).json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mmmu-(val).json -------------------------------------------------------------------------------- /data/benchmarks/mmmu-(validation).json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mmmu-(validation).json -------------------------------------------------------------------------------- /data/benchmarks/mmmu-pro.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mmmu-pro.json -------------------------------------------------------------------------------- /data/benchmarks/mmmu.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mmmu.json -------------------------------------------------------------------------------- /data/benchmarks/mmmuval.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mmmuval.json -------------------------------------------------------------------------------- /data/benchmarks/mmstar.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mmstar.json -------------------------------------------------------------------------------- /data/benchmarks/mmt-bench.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mmt-bench.json -------------------------------------------------------------------------------- /data/benchmarks/mmvet.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mmvet.json -------------------------------------------------------------------------------- /data/benchmarks/mmvetgpt4turbo.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mmvetgpt4turbo.json -------------------------------------------------------------------------------- /data/benchmarks/mobileminiwob++-sr.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mobileminiwob++-sr.json -------------------------------------------------------------------------------- /data/benchmarks/mrcr-1m-(pointwise).json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mrcr-1m-(pointwise).json -------------------------------------------------------------------------------- /data/benchmarks/mrcr-1m.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mrcr-1m.json -------------------------------------------------------------------------------- /data/benchmarks/mrcr-v2-(8-needle).json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mrcr-v2-(8-needle).json -------------------------------------------------------------------------------- /data/benchmarks/mrcr-v2.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mrcr-v2.json -------------------------------------------------------------------------------- /data/benchmarks/mrcr.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mrcr.json -------------------------------------------------------------------------------- /data/benchmarks/mt-bench.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mt-bench.json -------------------------------------------------------------------------------- /data/benchmarks/mtvqa.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mtvqa.json -------------------------------------------------------------------------------- /data/benchmarks/muirbench.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/muirbench.json -------------------------------------------------------------------------------- /data/benchmarks/multi-if.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/multi-if.json -------------------------------------------------------------------------------- /data/benchmarks/multi-swe-bench.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/multi-swe-bench.json -------------------------------------------------------------------------------- /data/benchmarks/multichallenge-(o3-mini-grader).json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/multichallenge-(o3-mini-grader).json -------------------------------------------------------------------------------- /data/benchmarks/multichallenge.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/multichallenge.json -------------------------------------------------------------------------------- /data/benchmarks/multilf.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/multilf.json -------------------------------------------------------------------------------- /data/benchmarks/multilingual-mgsm-(cot).json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/multilingual-mgsm-(cot).json -------------------------------------------------------------------------------- /data/benchmarks/multilingual-mmlu.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/multilingual-mmlu.json -------------------------------------------------------------------------------- /data/benchmarks/multipl-e-humaneval.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/multipl-e-humaneval.json -------------------------------------------------------------------------------- /data/benchmarks/multipl-e-mbpp.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/multipl-e-mbpp.json -------------------------------------------------------------------------------- /data/benchmarks/multipl-e.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/multipl-e.json -------------------------------------------------------------------------------- /data/benchmarks/musiccaps.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/musiccaps.json -------------------------------------------------------------------------------- /data/benchmarks/musr.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/musr.json -------------------------------------------------------------------------------- /data/benchmarks/mvbench.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/mvbench.json -------------------------------------------------------------------------------- /data/benchmarks/natural-questions.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/natural-questions.json -------------------------------------------------------------------------------- /data/benchmarks/natural2code.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/natural2code.json -------------------------------------------------------------------------------- /data/benchmarks/nexus.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/nexus.json -------------------------------------------------------------------------------- /data/benchmarks/nih-multi-needle.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/nih-multi-needle.json -------------------------------------------------------------------------------- /data/benchmarks/nmos.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/nmos.json -------------------------------------------------------------------------------- /data/benchmarks/nq.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/nq.json -------------------------------------------------------------------------------- /data/benchmarks/ocrbench-v2-(en).json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/ocrbench-v2-(en).json -------------------------------------------------------------------------------- /data/benchmarks/ocrbench-v2-(zh).json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/ocrbench-v2-(zh).json -------------------------------------------------------------------------------- /data/benchmarks/ocrbench-v2.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/ocrbench-v2.json -------------------------------------------------------------------------------- /data/benchmarks/ocrbench.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/ocrbench.json -------------------------------------------------------------------------------- /data/benchmarks/odinw.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/odinw.json -------------------------------------------------------------------------------- /data/benchmarks/ojbench.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/ojbench.json -------------------------------------------------------------------------------- /data/benchmarks/olympiadbench.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/olympiadbench.json -------------------------------------------------------------------------------- /data/benchmarks/omnibench-music.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/omnibench-music.json -------------------------------------------------------------------------------- /data/benchmarks/omnibench.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/omnibench.json -------------------------------------------------------------------------------- /data/benchmarks/omnimath.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/omnimath.json -------------------------------------------------------------------------------- /data/benchmarks/open-rewrite.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/open-rewrite.json -------------------------------------------------------------------------------- /data/benchmarks/openai-mmlu.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/openai-mmlu.json -------------------------------------------------------------------------------- /data/benchmarks/openai-mrcr%3A-2-needle-128k.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/openai-mrcr%3A-2-needle-128k.json -------------------------------------------------------------------------------- /data/benchmarks/openai-mrcr%3A-2-needle-1m.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/openai-mrcr%3A-2-needle-1m.json -------------------------------------------------------------------------------- /data/benchmarks/openai-mrcr%3A-2-needle-256k.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/openai-mrcr%3A-2-needle-256k.json -------------------------------------------------------------------------------- /data/benchmarks/openbookqa.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/openbookqa.json -------------------------------------------------------------------------------- /data/benchmarks/osworld-extended.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/osworld-extended.json -------------------------------------------------------------------------------- /data/benchmarks/osworld-screenshot-only.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/osworld-screenshot-only.json -------------------------------------------------------------------------------- /data/benchmarks/osworld.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/osworld.json -------------------------------------------------------------------------------- /data/benchmarks/pathmcqa.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/pathmcqa.json -------------------------------------------------------------------------------- /data/benchmarks/perceptiontest.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/perceptiontest.json -------------------------------------------------------------------------------- /data/benchmarks/phibench.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/phibench.json -------------------------------------------------------------------------------- /data/benchmarks/physicsfinals.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/physicsfinals.json -------------------------------------------------------------------------------- /data/benchmarks/piqa.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/piqa.json -------------------------------------------------------------------------------- /data/benchmarks/pointgrounding.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/pointgrounding.json -------------------------------------------------------------------------------- /data/benchmarks/polymath-en.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/polymath-en.json -------------------------------------------------------------------------------- /data/benchmarks/polymath.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/polymath.json -------------------------------------------------------------------------------- /data/benchmarks/pope.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/pope.json -------------------------------------------------------------------------------- /data/benchmarks/popqa.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/popqa.json -------------------------------------------------------------------------------- /data/benchmarks/qasper.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/qasper.json -------------------------------------------------------------------------------- /data/benchmarks/qmsum.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/qmsum.json -------------------------------------------------------------------------------- /data/benchmarks/realworldqa.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/realworldqa.json -------------------------------------------------------------------------------- /data/benchmarks/repobench.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/repobench.json -------------------------------------------------------------------------------- /data/benchmarks/repoqa.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/repoqa.json -------------------------------------------------------------------------------- /data/benchmarks/ruler.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/ruler.json -------------------------------------------------------------------------------- /data/benchmarks/sat-math.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/sat-math.json -------------------------------------------------------------------------------- /data/benchmarks/scale-multichallenge.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/scale-multichallenge.json -------------------------------------------------------------------------------- /data/benchmarks/scicode.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/scicode.json -------------------------------------------------------------------------------- /data/benchmarks/scienceqa-visual.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/scienceqa-visual.json -------------------------------------------------------------------------------- /data/benchmarks/scienceqa.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/scienceqa.json -------------------------------------------------------------------------------- /data/benchmarks/screenspot-pro.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/screenspot-pro.json -------------------------------------------------------------------------------- /data/benchmarks/screenspot.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/screenspot.json -------------------------------------------------------------------------------- /data/benchmarks/simpleqa.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/simpleqa.json -------------------------------------------------------------------------------- /data/benchmarks/slakevqa.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/slakevqa.json -------------------------------------------------------------------------------- /data/benchmarks/social-iqa.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/social-iqa.json -------------------------------------------------------------------------------- /data/benchmarks/spider.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/spider.json -------------------------------------------------------------------------------- /data/benchmarks/squality.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/squality.json -------------------------------------------------------------------------------- /data/benchmarks/stem.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/stem.json -------------------------------------------------------------------------------- /data/benchmarks/summscreenfd.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/summscreenfd.json -------------------------------------------------------------------------------- /data/benchmarks/superglue.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/superglue.json -------------------------------------------------------------------------------- /data/benchmarks/supergpqa.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/supergpqa.json -------------------------------------------------------------------------------- /data/benchmarks/swe-bench-multilingual.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/swe-bench-multilingual.json -------------------------------------------------------------------------------- /data/benchmarks/swe-bench-verified-(agentic-coding).json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/swe-bench-verified-(agentic-coding).json -------------------------------------------------------------------------------- /data/benchmarks/swe-bench-verified-(agentless).json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/swe-bench-verified-(agentless).json -------------------------------------------------------------------------------- /data/benchmarks/swe-bench-verified-(multiple-attempts).json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/swe-bench-verified-(multiple-attempts).json -------------------------------------------------------------------------------- /data/benchmarks/swe-bench-verified.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/swe-bench-verified.json -------------------------------------------------------------------------------- /data/benchmarks/swe-dev.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/swe-dev.json -------------------------------------------------------------------------------- /data/benchmarks/swe-lancer-(ic-diamond-subset).json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/swe-lancer-(ic-diamond-subset).json -------------------------------------------------------------------------------- /data/benchmarks/swe-lancer.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/swe-lancer.json -------------------------------------------------------------------------------- /data/benchmarks/tau-bench-airline.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/tau-bench-airline.json -------------------------------------------------------------------------------- /data/benchmarks/tau-bench-retail.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/tau-bench-retail.json -------------------------------------------------------------------------------- /data/benchmarks/tau-bench.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/tau-bench.json -------------------------------------------------------------------------------- /data/benchmarks/tau2-airline.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/tau2-airline.json -------------------------------------------------------------------------------- /data/benchmarks/tau2-retail.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/tau2-retail.json -------------------------------------------------------------------------------- /data/benchmarks/tau2-telecom.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/tau2-telecom.json -------------------------------------------------------------------------------- /data/benchmarks/tempcompass.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/tempcompass.json -------------------------------------------------------------------------------- /data/benchmarks/terminal-bench.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/terminal-bench.json -------------------------------------------------------------------------------- /data/benchmarks/terminus.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/terminus.json -------------------------------------------------------------------------------- /data/benchmarks/textvqa.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/textvqa.json -------------------------------------------------------------------------------- /data/benchmarks/theoremqa.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/theoremqa.json -------------------------------------------------------------------------------- /data/benchmarks/tldr9+-(test).json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/tldr9+-(test).json -------------------------------------------------------------------------------- /data/benchmarks/translation-en-to-set1-comet22.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/translation-en-to-set1-comet22.json -------------------------------------------------------------------------------- /data/benchmarks/translation-en-to-set1-spbleu.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/translation-en-to-set1-spbleu.json -------------------------------------------------------------------------------- /data/benchmarks/translation-set1-to-en-comet22.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/translation-set1-to-en-comet22.json -------------------------------------------------------------------------------- /data/benchmarks/translation-set1-to-en-spbleu.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/translation-set1-to-en-spbleu.json -------------------------------------------------------------------------------- /data/benchmarks/triviaqa.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/triviaqa.json -------------------------------------------------------------------------------- /data/benchmarks/truthfulqa.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/truthfulqa.json -------------------------------------------------------------------------------- /data/benchmarks/tydiqa.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/tydiqa.json -------------------------------------------------------------------------------- /data/benchmarks/uniform-bar-exam.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/uniform-bar-exam.json -------------------------------------------------------------------------------- /data/benchmarks/usamo25.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/usamo25.json -------------------------------------------------------------------------------- /data/benchmarks/vatex.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/vatex.json -------------------------------------------------------------------------------- /data/benchmarks/vcr-en-easy.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/vcr-en-easy.json -------------------------------------------------------------------------------- /data/benchmarks/vibe-eval.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/vibe-eval.json -------------------------------------------------------------------------------- /data/benchmarks/video-mme-(long,-no-subtitles).json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/video-mme-(long,-no-subtitles).json -------------------------------------------------------------------------------- /data/benchmarks/video-mme.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/video-mme.json -------------------------------------------------------------------------------- /data/benchmarks/video-mmew-sub.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/video-mmew-sub.json -------------------------------------------------------------------------------- /data/benchmarks/videomme-w-o-sub..json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/videomme-w-o-sub..json -------------------------------------------------------------------------------- /data/benchmarks/videomme-w-sub..json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/videomme-w-sub..json -------------------------------------------------------------------------------- /data/benchmarks/videommmu.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/videommmu.json -------------------------------------------------------------------------------- /data/benchmarks/visualwebbench.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/visualwebbench.json -------------------------------------------------------------------------------- /data/benchmarks/vocalsound.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/vocalsound.json -------------------------------------------------------------------------------- /data/benchmarks/voicebench-avg.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/voicebench-avg.json -------------------------------------------------------------------------------- /data/benchmarks/vqa-rad.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/vqa-rad.json -------------------------------------------------------------------------------- /data/benchmarks/vqav2-(test).json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/vqav2-(test).json -------------------------------------------------------------------------------- /data/benchmarks/vqav2-(val).json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/vqav2-(val).json -------------------------------------------------------------------------------- /data/benchmarks/vqav2.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/vqav2.json -------------------------------------------------------------------------------- /data/benchmarks/wild-bench.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/wild-bench.json -------------------------------------------------------------------------------- /data/benchmarks/winogrande.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/winogrande.json -------------------------------------------------------------------------------- /data/benchmarks/wmt23.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/wmt23.json -------------------------------------------------------------------------------- /data/benchmarks/wmt24++.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/wmt24++.json -------------------------------------------------------------------------------- /data/benchmarks/writingbench.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/writingbench.json -------------------------------------------------------------------------------- /data/benchmarks/xlsum-english.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/xlsum-english.json -------------------------------------------------------------------------------- /data/benchmarks/xstest.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/xstest.json -------------------------------------------------------------------------------- /data/benchmarks/zebralogic.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/benchmarks/zebralogic.json -------------------------------------------------------------------------------- /data/licenses/apache_2_0.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/licenses/apache_2_0.json -------------------------------------------------------------------------------- /data/licenses/cc_by_nc.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/licenses/cc_by_nc.json -------------------------------------------------------------------------------- /data/licenses/creative_commons_attribution_4_0_license.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/licenses/creative_commons_attribution_4_0_license.json -------------------------------------------------------------------------------- /data/licenses/deepseek.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/licenses/deepseek.json -------------------------------------------------------------------------------- /data/licenses/gemma.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/licenses/gemma.json -------------------------------------------------------------------------------- /data/licenses/health_ai_developer_foundations_terms_of_use.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/licenses/health_ai_developer_foundations_terms_of_use.json -------------------------------------------------------------------------------- /data/licenses/jamba_open_model_license.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/licenses/jamba_open_model_license.json -------------------------------------------------------------------------------- /data/licenses/llama3_2.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/licenses/llama3_2.json -------------------------------------------------------------------------------- /data/licenses/llama_3_1_community_license.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/licenses/llama_3_1_community_license.json -------------------------------------------------------------------------------- /data/licenses/llama_3_2_community_license.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/licenses/llama_3_2_community_license.json -------------------------------------------------------------------------------- /data/licenses/llama_3_3_community_license_agreement.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/licenses/llama_3_3_community_license_agreement.json -------------------------------------------------------------------------------- /data/licenses/llama_4_community_license_agreement.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/licenses/llama_4_community_license_agreement.json -------------------------------------------------------------------------------- /data/licenses/mistral_research_license.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/licenses/mistral_research_license.json -------------------------------------------------------------------------------- /data/licenses/mit.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/licenses/mit.json -------------------------------------------------------------------------------- /data/licenses/mit_+_model_license_(commercial_use_allowed).json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/licenses/mit_+_model_license_(commercial_use_allowed).json -------------------------------------------------------------------------------- /data/licenses/mit_license.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/licenses/mit_license.json -------------------------------------------------------------------------------- /data/licenses/mnpl_0_1.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/licenses/mnpl_0_1.json -------------------------------------------------------------------------------- /data/licenses/modified_mit_license.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/licenses/modified_mit_license.json -------------------------------------------------------------------------------- /data/licenses/nvidia_open_model_license_agreement.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/licenses/nvidia_open_model_license_agreement.json -------------------------------------------------------------------------------- /data/licenses/proprietary.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/licenses/proprietary.json -------------------------------------------------------------------------------- /data/licenses/qwen.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/licenses/qwen.json -------------------------------------------------------------------------------- /data/licenses/tongyi_qianwen.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/licenses/tongyi_qianwen.json -------------------------------------------------------------------------------- /data/licenses/unknown.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/licenses/unknown.json -------------------------------------------------------------------------------- /data/organizations/ai21/models/jamba-1.5-large/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/ai21/models/jamba-1.5-large/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/ai21/models/jamba-1.5-large/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/ai21/models/jamba-1.5-large/model.json -------------------------------------------------------------------------------- /data/organizations/ai21/models/jamba-1.5-mini/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/ai21/models/jamba-1.5-mini/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/ai21/models/jamba-1.5-mini/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/ai21/models/jamba-1.5-mini/model.json -------------------------------------------------------------------------------- /data/organizations/ai21/organization.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/ai21/organization.json -------------------------------------------------------------------------------- /data/organizations/amazon/models/nova-lite/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/amazon/models/nova-lite/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/amazon/models/nova-lite/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/amazon/models/nova-lite/model.json -------------------------------------------------------------------------------- /data/organizations/amazon/models/nova-micro/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/amazon/models/nova-micro/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/amazon/models/nova-micro/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/amazon/models/nova-micro/model.json -------------------------------------------------------------------------------- /data/organizations/amazon/models/nova-pro/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/amazon/models/nova-pro/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/amazon/models/nova-pro/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/amazon/models/nova-pro/model.json -------------------------------------------------------------------------------- /data/organizations/amazon/organization.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/amazon/organization.json -------------------------------------------------------------------------------- /data/organizations/anthropic/models/claude-3-5-haiku-20241022/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/anthropic/models/claude-3-5-haiku-20241022/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/anthropic/models/claude-3-5-haiku-20241022/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/anthropic/models/claude-3-5-haiku-20241022/model.json -------------------------------------------------------------------------------- /data/organizations/anthropic/models/claude-3-5-sonnet-20240620/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/anthropic/models/claude-3-5-sonnet-20240620/model.json -------------------------------------------------------------------------------- /data/organizations/anthropic/models/claude-3-5-sonnet-20241022/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/anthropic/models/claude-3-5-sonnet-20241022/model.json -------------------------------------------------------------------------------- /data/organizations/anthropic/models/claude-3-7-sonnet-20250219/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/anthropic/models/claude-3-7-sonnet-20250219/model.json -------------------------------------------------------------------------------- /data/organizations/anthropic/models/claude-3-haiku-20240307/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/anthropic/models/claude-3-haiku-20240307/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/anthropic/models/claude-3-haiku-20240307/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/anthropic/models/claude-3-haiku-20240307/model.json -------------------------------------------------------------------------------- /data/organizations/anthropic/models/claude-3-opus-20240229/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/anthropic/models/claude-3-opus-20240229/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/anthropic/models/claude-3-opus-20240229/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/anthropic/models/claude-3-opus-20240229/model.json -------------------------------------------------------------------------------- /data/organizations/anthropic/models/claude-3-sonnet-20240229/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/anthropic/models/claude-3-sonnet-20240229/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/anthropic/models/claude-3-sonnet-20240229/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/anthropic/models/claude-3-sonnet-20240229/model.json -------------------------------------------------------------------------------- /data/organizations/anthropic/models/claude-haiku-4-5-20251015/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/anthropic/models/claude-haiku-4-5-20251015/model.json -------------------------------------------------------------------------------- /data/organizations/anthropic/models/claude-opus-4-1-20250805/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/anthropic/models/claude-opus-4-1-20250805/model.json -------------------------------------------------------------------------------- /data/organizations/anthropic/models/claude-opus-4-20250514/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/anthropic/models/claude-opus-4-20250514/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/anthropic/models/claude-opus-4-20250514/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/anthropic/models/claude-opus-4-20250514/model.json -------------------------------------------------------------------------------- /data/organizations/anthropic/models/claude-sonnet-4-20250514/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/anthropic/models/claude-sonnet-4-20250514/model.json -------------------------------------------------------------------------------- /data/organizations/anthropic/models/claude-sonnet-4-5-20250929/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/anthropic/models/claude-sonnet-4-5-20250929/model.json -------------------------------------------------------------------------------- /data/organizations/anthropic/organization.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/anthropic/organization.json -------------------------------------------------------------------------------- /data/organizations/cohere/models/command-r-plus-04-2024/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/cohere/models/command-r-plus-04-2024/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/cohere/models/command-r-plus-04-2024/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/cohere/models/command-r-plus-04-2024/model.json -------------------------------------------------------------------------------- /data/organizations/cohere/organization.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/cohere/organization.json -------------------------------------------------------------------------------- /data/organizations/deepseek/models/deepseek-r1-0528/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/deepseek/models/deepseek-r1-0528/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/deepseek/models/deepseek-r1-0528/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/deepseek/models/deepseek-r1-0528/model.json -------------------------------------------------------------------------------- /data/organizations/deepseek/models/deepseek-r1-distill-llama-70b/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/deepseek/models/deepseek-r1-distill-llama-70b/model.json -------------------------------------------------------------------------------- /data/organizations/deepseek/models/deepseek-r1-distill-llama-8b/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/deepseek/models/deepseek-r1-distill-llama-8b/model.json -------------------------------------------------------------------------------- /data/organizations/deepseek/models/deepseek-r1-distill-qwen-1.5b/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/deepseek/models/deepseek-r1-distill-qwen-1.5b/model.json -------------------------------------------------------------------------------- /data/organizations/deepseek/models/deepseek-r1-distill-qwen-14b/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/deepseek/models/deepseek-r1-distill-qwen-14b/model.json -------------------------------------------------------------------------------- /data/organizations/deepseek/models/deepseek-r1-distill-qwen-32b/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/deepseek/models/deepseek-r1-distill-qwen-32b/model.json -------------------------------------------------------------------------------- /data/organizations/deepseek/models/deepseek-r1-distill-qwen-7b/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/deepseek/models/deepseek-r1-distill-qwen-7b/model.json -------------------------------------------------------------------------------- /data/organizations/deepseek/models/deepseek-r1-zero/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/deepseek/models/deepseek-r1-zero/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/deepseek/models/deepseek-r1-zero/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/deepseek/models/deepseek-r1-zero/model.json -------------------------------------------------------------------------------- /data/organizations/deepseek/models/deepseek-r1/benchmarks.json: -------------------------------------------------------------------------------- 1 | [] 2 | -------------------------------------------------------------------------------- /data/organizations/deepseek/models/deepseek-r1/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/deepseek/models/deepseek-r1/model.json -------------------------------------------------------------------------------- /data/organizations/deepseek/models/deepseek-v2.5/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/deepseek/models/deepseek-v2.5/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/deepseek/models/deepseek-v2.5/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/deepseek/models/deepseek-v2.5/model.json -------------------------------------------------------------------------------- /data/organizations/deepseek/models/deepseek-v3-0324/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/deepseek/models/deepseek-v3-0324/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/deepseek/models/deepseek-v3-0324/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/deepseek/models/deepseek-v3-0324/model.json -------------------------------------------------------------------------------- /data/organizations/deepseek/models/deepseek-v3.1/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/deepseek/models/deepseek-v3.1/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/deepseek/models/deepseek-v3.1/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/deepseek/models/deepseek-v3.1/model.json -------------------------------------------------------------------------------- /data/organizations/deepseek/models/deepseek-v3.2-exp/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/deepseek/models/deepseek-v3.2-exp/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/deepseek/models/deepseek-v3.2-exp/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/deepseek/models/deepseek-v3.2-exp/model.json -------------------------------------------------------------------------------- /data/organizations/deepseek/models/deepseek-v3/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/deepseek/models/deepseek-v3/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/deepseek/models/deepseek-v3/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/deepseek/models/deepseek-v3/model.json -------------------------------------------------------------------------------- /data/organizations/deepseek/models/deepseek-vl2-small/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/deepseek/models/deepseek-vl2-small/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/deepseek/models/deepseek-vl2-small/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/deepseek/models/deepseek-vl2-small/model.json -------------------------------------------------------------------------------- /data/organizations/deepseek/models/deepseek-vl2-tiny/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/deepseek/models/deepseek-vl2-tiny/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/deepseek/models/deepseek-vl2-tiny/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/deepseek/models/deepseek-vl2-tiny/model.json -------------------------------------------------------------------------------- /data/organizations/deepseek/models/deepseek-vl2/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/deepseek/models/deepseek-vl2/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/deepseek/models/deepseek-vl2/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/deepseek/models/deepseek-vl2/model.json -------------------------------------------------------------------------------- /data/organizations/deepseek/organization.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/deepseek/organization.json -------------------------------------------------------------------------------- /data/organizations/google/models/gemini-1.0-pro/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/google/models/gemini-1.0-pro/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/google/models/gemini-1.0-pro/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/google/models/gemini-1.0-pro/model.json -------------------------------------------------------------------------------- /data/organizations/google/models/gemini-1.5-flash-8b/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/google/models/gemini-1.5-flash-8b/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/google/models/gemini-1.5-flash-8b/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/google/models/gemini-1.5-flash-8b/model.json -------------------------------------------------------------------------------- /data/organizations/google/models/gemini-1.5-flash/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/google/models/gemini-1.5-flash/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/google/models/gemini-1.5-flash/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/google/models/gemini-1.5-flash/model.json -------------------------------------------------------------------------------- /data/organizations/google/models/gemini-1.5-pro/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/google/models/gemini-1.5-pro/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/google/models/gemini-1.5-pro/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/google/models/gemini-1.5-pro/model.json -------------------------------------------------------------------------------- /data/organizations/google/models/gemini-2.0-flash-lite/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/google/models/gemini-2.0-flash-lite/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/google/models/gemini-2.0-flash-lite/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/google/models/gemini-2.0-flash-lite/model.json -------------------------------------------------------------------------------- /data/organizations/google/models/gemini-2.0-flash-thinking/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/google/models/gemini-2.0-flash-thinking/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/google/models/gemini-2.0-flash-thinking/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/google/models/gemini-2.0-flash-thinking/model.json -------------------------------------------------------------------------------- /data/organizations/google/models/gemini-2.0-flash/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/google/models/gemini-2.0-flash/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/google/models/gemini-2.0-flash/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/google/models/gemini-2.0-flash/model.json -------------------------------------------------------------------------------- /data/organizations/google/models/gemini-2.5-flash-lite/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/google/models/gemini-2.5-flash-lite/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/google/models/gemini-2.5-flash-lite/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/google/models/gemini-2.5-flash-lite/model.json -------------------------------------------------------------------------------- /data/organizations/google/models/gemini-2.5-flash/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/google/models/gemini-2.5-flash/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/google/models/gemini-2.5-flash/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/google/models/gemini-2.5-flash/model.json -------------------------------------------------------------------------------- /data/organizations/google/models/gemini-2.5-pro-preview-06-05/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/google/models/gemini-2.5-pro-preview-06-05/model.json -------------------------------------------------------------------------------- /data/organizations/google/models/gemini-2.5-pro/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/google/models/gemini-2.5-pro/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/google/models/gemini-2.5-pro/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/google/models/gemini-2.5-pro/model.json -------------------------------------------------------------------------------- /data/organizations/google/models/gemini-diffusion/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/google/models/gemini-diffusion/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/google/models/gemini-diffusion/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/google/models/gemini-diffusion/model.json -------------------------------------------------------------------------------- /data/organizations/google/models/gemma-2-27b-it/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/google/models/gemma-2-27b-it/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/google/models/gemma-2-27b-it/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/google/models/gemma-2-27b-it/model.json -------------------------------------------------------------------------------- /data/organizations/google/models/gemma-2-9b-it/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/google/models/gemma-2-9b-it/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/google/models/gemma-2-9b-it/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/google/models/gemma-2-9b-it/model.json -------------------------------------------------------------------------------- /data/organizations/google/models/gemma-3-12b-it/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/google/models/gemma-3-12b-it/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/google/models/gemma-3-12b-it/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/google/models/gemma-3-12b-it/model.json -------------------------------------------------------------------------------- /data/organizations/google/models/gemma-3-1b-it/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/google/models/gemma-3-1b-it/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/google/models/gemma-3-1b-it/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/google/models/gemma-3-1b-it/model.json -------------------------------------------------------------------------------- /data/organizations/google/models/gemma-3-27b-it/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/google/models/gemma-3-27b-it/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/google/models/gemma-3-27b-it/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/google/models/gemma-3-27b-it/model.json -------------------------------------------------------------------------------- /data/organizations/google/models/gemma-3-4b-it/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/google/models/gemma-3-4b-it/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/google/models/gemma-3-4b-it/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/google/models/gemma-3-4b-it/model.json -------------------------------------------------------------------------------- /data/organizations/google/models/gemma-3n-e2b-it-litert-preview/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/google/models/gemma-3n-e2b-it-litert-preview/model.json -------------------------------------------------------------------------------- /data/organizations/google/models/gemma-3n-e2b-it/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/google/models/gemma-3n-e2b-it/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/google/models/gemma-3n-e2b-it/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/google/models/gemma-3n-e2b-it/model.json -------------------------------------------------------------------------------- /data/organizations/google/models/gemma-3n-e2b/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/google/models/gemma-3n-e2b/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/google/models/gemma-3n-e2b/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/google/models/gemma-3n-e2b/model.json -------------------------------------------------------------------------------- /data/organizations/google/models/gemma-3n-e4b-it-litert-preview/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/google/models/gemma-3n-e4b-it-litert-preview/model.json -------------------------------------------------------------------------------- /data/organizations/google/models/gemma-3n-e4b-it/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/google/models/gemma-3n-e4b-it/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/google/models/gemma-3n-e4b-it/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/google/models/gemma-3n-e4b-it/model.json -------------------------------------------------------------------------------- /data/organizations/google/models/gemma-3n-e4b/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/google/models/gemma-3n-e4b/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/google/models/gemma-3n-e4b/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/google/models/gemma-3n-e4b/model.json -------------------------------------------------------------------------------- /data/organizations/google/models/medgemma-4b-it/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/google/models/medgemma-4b-it/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/google/models/medgemma-4b-it/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/google/models/medgemma-4b-it/model.json -------------------------------------------------------------------------------- /data/organizations/google/organization.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/google/organization.json -------------------------------------------------------------------------------- /data/organizations/ibm/models/granite-3.3-8b-base/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/ibm/models/granite-3.3-8b-base/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/ibm/models/granite-3.3-8b-base/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/ibm/models/granite-3.3-8b-base/model.json -------------------------------------------------------------------------------- /data/organizations/ibm/models/granite-3.3-8b-instruct/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/ibm/models/granite-3.3-8b-instruct/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/ibm/models/granite-3.3-8b-instruct/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/ibm/models/granite-3.3-8b-instruct/model.json -------------------------------------------------------------------------------- /data/organizations/ibm/models/granite-4.0-tiny-preview/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/ibm/models/granite-4.0-tiny-preview/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/ibm/models/granite-4.0-tiny-preview/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/ibm/models/granite-4.0-tiny-preview/model.json -------------------------------------------------------------------------------- /data/organizations/ibm/organization.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/ibm/organization.json -------------------------------------------------------------------------------- /data/organizations/meta/models/llama-3.1-405b-instruct/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/meta/models/llama-3.1-405b-instruct/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/meta/models/llama-3.1-405b-instruct/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/meta/models/llama-3.1-405b-instruct/model.json -------------------------------------------------------------------------------- /data/organizations/meta/models/llama-3.1-70b-instruct/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/meta/models/llama-3.1-70b-instruct/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/meta/models/llama-3.1-70b-instruct/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/meta/models/llama-3.1-70b-instruct/model.json -------------------------------------------------------------------------------- /data/organizations/meta/models/llama-3.1-8b-instruct/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/meta/models/llama-3.1-8b-instruct/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/meta/models/llama-3.1-8b-instruct/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/meta/models/llama-3.1-8b-instruct/model.json -------------------------------------------------------------------------------- /data/organizations/meta/models/llama-3.2-11b-instruct/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/meta/models/llama-3.2-11b-instruct/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/meta/models/llama-3.2-11b-instruct/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/meta/models/llama-3.2-11b-instruct/model.json -------------------------------------------------------------------------------- /data/organizations/meta/models/llama-3.2-3b-instruct/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/meta/models/llama-3.2-3b-instruct/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/meta/models/llama-3.2-3b-instruct/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/meta/models/llama-3.2-3b-instruct/model.json -------------------------------------------------------------------------------- /data/organizations/meta/models/llama-3.2-90b-instruct/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/meta/models/llama-3.2-90b-instruct/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/meta/models/llama-3.2-90b-instruct/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/meta/models/llama-3.2-90b-instruct/model.json -------------------------------------------------------------------------------- /data/organizations/meta/models/llama-3.3-70b-instruct/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/meta/models/llama-3.3-70b-instruct/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/meta/models/llama-3.3-70b-instruct/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/meta/models/llama-3.3-70b-instruct/model.json -------------------------------------------------------------------------------- /data/organizations/meta/models/llama-4-maverick/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/meta/models/llama-4-maverick/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/meta/models/llama-4-maverick/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/meta/models/llama-4-maverick/model.json -------------------------------------------------------------------------------- /data/organizations/meta/models/llama-4-scout/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/meta/models/llama-4-scout/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/meta/models/llama-4-scout/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/meta/models/llama-4-scout/model.json -------------------------------------------------------------------------------- /data/organizations/meta/organization.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/meta/organization.json -------------------------------------------------------------------------------- /data/organizations/microsoft/models/phi-3.5-mini-instruct/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/microsoft/models/phi-3.5-mini-instruct/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/microsoft/models/phi-3.5-mini-instruct/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/microsoft/models/phi-3.5-mini-instruct/model.json -------------------------------------------------------------------------------- /data/organizations/microsoft/models/phi-3.5-moe-instruct/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/microsoft/models/phi-3.5-moe-instruct/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/microsoft/models/phi-3.5-moe-instruct/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/microsoft/models/phi-3.5-moe-instruct/model.json -------------------------------------------------------------------------------- /data/organizations/microsoft/models/phi-3.5-vision-instruct/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/microsoft/models/phi-3.5-vision-instruct/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/microsoft/models/phi-3.5-vision-instruct/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/microsoft/models/phi-3.5-vision-instruct/model.json -------------------------------------------------------------------------------- /data/organizations/microsoft/models/phi-4-mini-reasoning/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/microsoft/models/phi-4-mini-reasoning/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/microsoft/models/phi-4-mini-reasoning/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/microsoft/models/phi-4-mini-reasoning/model.json -------------------------------------------------------------------------------- /data/organizations/microsoft/models/phi-4-mini/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/microsoft/models/phi-4-mini/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/microsoft/models/phi-4-mini/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/microsoft/models/phi-4-mini/model.json -------------------------------------------------------------------------------- /data/organizations/microsoft/models/phi-4-multimodal-instruct/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/microsoft/models/phi-4-multimodal-instruct/model.json -------------------------------------------------------------------------------- /data/organizations/microsoft/models/phi-4-reasoning-plus/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/microsoft/models/phi-4-reasoning-plus/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/microsoft/models/phi-4-reasoning-plus/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/microsoft/models/phi-4-reasoning-plus/model.json -------------------------------------------------------------------------------- /data/organizations/microsoft/models/phi-4-reasoning/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/microsoft/models/phi-4-reasoning/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/microsoft/models/phi-4-reasoning/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/microsoft/models/phi-4-reasoning/model.json -------------------------------------------------------------------------------- /data/organizations/microsoft/models/phi-4/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/microsoft/models/phi-4/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/microsoft/models/phi-4/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/microsoft/models/phi-4/model.json -------------------------------------------------------------------------------- /data/organizations/microsoft/organization.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/microsoft/organization.json -------------------------------------------------------------------------------- /data/organizations/mistral/models/codestral-22b/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/mistral/models/codestral-22b/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/mistral/models/codestral-22b/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/mistral/models/codestral-22b/model.json -------------------------------------------------------------------------------- /data/organizations/mistral/models/devstral-medium-2507/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/mistral/models/devstral-medium-2507/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/mistral/models/devstral-medium-2507/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/mistral/models/devstral-medium-2507/model.json -------------------------------------------------------------------------------- /data/organizations/mistral/models/devstral-small-2507/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/mistral/models/devstral-small-2507/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/mistral/models/devstral-small-2507/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/mistral/models/devstral-small-2507/model.json -------------------------------------------------------------------------------- /data/organizations/mistral/models/magistral-medium/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/mistral/models/magistral-medium/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/mistral/models/magistral-medium/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/mistral/models/magistral-medium/model.json -------------------------------------------------------------------------------- /data/organizations/mistral/models/magistral-small-2506/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/mistral/models/magistral-small-2506/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/mistral/models/magistral-small-2506/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/mistral/models/magistral-small-2506/model.json -------------------------------------------------------------------------------- /data/organizations/mistral/models/ministral-8b-instruct-2410/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/mistral/models/ministral-8b-instruct-2410/model.json -------------------------------------------------------------------------------- /data/organizations/mistral/models/mistral-large-2-2407/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/mistral/models/mistral-large-2-2407/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/mistral/models/mistral-large-2-2407/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/mistral/models/mistral-large-2-2407/model.json -------------------------------------------------------------------------------- /data/organizations/mistral/models/mistral-nemo-instruct-2407/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/mistral/models/mistral-nemo-instruct-2407/model.json -------------------------------------------------------------------------------- /data/organizations/mistral/models/mistral-small-2409/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/mistral/models/mistral-small-2409/model.json -------------------------------------------------------------------------------- /data/organizations/mistral/models/mistral-small-24b-base-2501/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/mistral/models/mistral-small-24b-base-2501/model.json -------------------------------------------------------------------------------- /data/organizations/mistral/models/pixtral-12b-2409/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/mistral/models/pixtral-12b-2409/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/mistral/models/pixtral-12b-2409/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/mistral/models/pixtral-12b-2409/model.json -------------------------------------------------------------------------------- /data/organizations/mistral/models/pixtral-large/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/mistral/models/pixtral-large/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/mistral/models/pixtral-large/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/mistral/models/pixtral-large/model.json -------------------------------------------------------------------------------- /data/organizations/mistral/organization.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/mistral/organization.json -------------------------------------------------------------------------------- /data/organizations/moonshotai/models/kimi-k1.5/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/moonshotai/models/kimi-k1.5/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/moonshotai/models/kimi-k1.5/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/moonshotai/models/kimi-k1.5/model.json -------------------------------------------------------------------------------- /data/organizations/moonshotai/models/kimi-k2-0905/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/moonshotai/models/kimi-k2-0905/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/moonshotai/models/kimi-k2-0905/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/moonshotai/models/kimi-k2-0905/model.json -------------------------------------------------------------------------------- /data/organizations/moonshotai/models/kimi-k2-base/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/moonshotai/models/kimi-k2-base/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/moonshotai/models/kimi-k2-base/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/moonshotai/models/kimi-k2-base/model.json -------------------------------------------------------------------------------- /data/organizations/moonshotai/models/kimi-k2-instruct-0905/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/moonshotai/models/kimi-k2-instruct-0905/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/moonshotai/models/kimi-k2-instruct-0905/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/moonshotai/models/kimi-k2-instruct-0905/model.json -------------------------------------------------------------------------------- /data/organizations/moonshotai/models/kimi-k2-instruct/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/moonshotai/models/kimi-k2-instruct/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/moonshotai/models/kimi-k2-instruct/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/moonshotai/models/kimi-k2-instruct/model.json -------------------------------------------------------------------------------- /data/organizations/moonshotai/organization.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/moonshotai/organization.json -------------------------------------------------------------------------------- /data/organizations/nvidia/models/llama-3.1-nemotron-70b-instruct/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/nvidia/models/llama-3.1-nemotron-70b-instruct/model.json -------------------------------------------------------------------------------- /data/organizations/nvidia/models/llama-3.1-nemotron-nano-8b-v1/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/nvidia/models/llama-3.1-nemotron-nano-8b-v1/model.json -------------------------------------------------------------------------------- /data/organizations/nvidia/models/llama-3.3-nemotron-super-49b-v1/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/nvidia/models/llama-3.3-nemotron-super-49b-v1/model.json -------------------------------------------------------------------------------- /data/organizations/nvidia/models/nemotron-nano-9b-v2/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/nvidia/models/nemotron-nano-9b-v2/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/nvidia/models/nemotron-nano-9b-v2/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/nvidia/models/nemotron-nano-9b-v2/model.json -------------------------------------------------------------------------------- /data/organizations/nvidia/organization.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/nvidia/organization.json -------------------------------------------------------------------------------- /data/organizations/openai/models/gpt-3.5-turbo-0125/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/openai/models/gpt-3.5-turbo-0125/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/openai/models/gpt-3.5-turbo-0125/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/openai/models/gpt-3.5-turbo-0125/model.json -------------------------------------------------------------------------------- /data/organizations/openai/models/gpt-4-0613/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/openai/models/gpt-4-0613/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/openai/models/gpt-4-0613/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/openai/models/gpt-4-0613/model.json -------------------------------------------------------------------------------- /data/organizations/openai/models/gpt-4-turbo-2024-04-09/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/openai/models/gpt-4-turbo-2024-04-09/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/openai/models/gpt-4-turbo-2024-04-09/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/openai/models/gpt-4-turbo-2024-04-09/model.json -------------------------------------------------------------------------------- /data/organizations/openai/models/gpt-4.1-2025-04-14/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/openai/models/gpt-4.1-2025-04-14/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/openai/models/gpt-4.1-2025-04-14/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/openai/models/gpt-4.1-2025-04-14/model.json -------------------------------------------------------------------------------- /data/organizations/openai/models/gpt-4.1-mini-2025-04-14/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/openai/models/gpt-4.1-mini-2025-04-14/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/openai/models/gpt-4.1-mini-2025-04-14/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/openai/models/gpt-4.1-mini-2025-04-14/model.json -------------------------------------------------------------------------------- /data/organizations/openai/models/gpt-4.1-nano-2025-04-14/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/openai/models/gpt-4.1-nano-2025-04-14/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/openai/models/gpt-4.1-nano-2025-04-14/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/openai/models/gpt-4.1-nano-2025-04-14/model.json -------------------------------------------------------------------------------- /data/organizations/openai/models/gpt-4.5/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/openai/models/gpt-4.5/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/openai/models/gpt-4.5/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/openai/models/gpt-4.5/model.json -------------------------------------------------------------------------------- /data/organizations/openai/models/gpt-4o-2024-05-13/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/openai/models/gpt-4o-2024-05-13/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/openai/models/gpt-4o-2024-05-13/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/openai/models/gpt-4o-2024-05-13/model.json -------------------------------------------------------------------------------- /data/organizations/openai/models/gpt-4o-2024-08-06/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/openai/models/gpt-4o-2024-08-06/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/openai/models/gpt-4o-2024-08-06/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/openai/models/gpt-4o-2024-08-06/model.json -------------------------------------------------------------------------------- /data/organizations/openai/models/gpt-4o-mini-2024-07-18/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/openai/models/gpt-4o-mini-2024-07-18/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/openai/models/gpt-4o-mini-2024-07-18/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/openai/models/gpt-4o-mini-2024-07-18/model.json -------------------------------------------------------------------------------- /data/organizations/openai/models/gpt-5-2025-08-07/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/openai/models/gpt-5-2025-08-07/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/openai/models/gpt-5-2025-08-07/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/openai/models/gpt-5-2025-08-07/model.json -------------------------------------------------------------------------------- /data/organizations/openai/models/gpt-5-codex-2025-09-15/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/openai/models/gpt-5-codex-2025-09-15/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/openai/models/gpt-5-codex-2025-09-15/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/openai/models/gpt-5-codex-2025-09-15/model.json -------------------------------------------------------------------------------- /data/organizations/openai/models/gpt-5-mini-2025-08-07/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/openai/models/gpt-5-mini-2025-08-07/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/openai/models/gpt-5-mini-2025-08-07/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/openai/models/gpt-5-mini-2025-08-07/model.json -------------------------------------------------------------------------------- /data/organizations/openai/models/gpt-5-nano-2025-08-07/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/openai/models/gpt-5-nano-2025-08-07/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/openai/models/gpt-5-nano-2025-08-07/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/openai/models/gpt-5-nano-2025-08-07/model.json -------------------------------------------------------------------------------- /data/organizations/openai/models/gpt-oss-120b/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/openai/models/gpt-oss-120b/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/openai/models/gpt-oss-120b/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/openai/models/gpt-oss-120b/model.json -------------------------------------------------------------------------------- /data/organizations/openai/models/gpt-oss-20b/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/openai/models/gpt-oss-20b/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/openai/models/gpt-oss-20b/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/openai/models/gpt-oss-20b/model.json -------------------------------------------------------------------------------- /data/organizations/openai/models/o1-2024-12-17/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/openai/models/o1-2024-12-17/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/openai/models/o1-2024-12-17/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/openai/models/o1-2024-12-17/model.json -------------------------------------------------------------------------------- /data/organizations/openai/models/o1-mini/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/openai/models/o1-mini/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/openai/models/o1-mini/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/openai/models/o1-mini/model.json -------------------------------------------------------------------------------- /data/organizations/openai/models/o1-preview/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/openai/models/o1-preview/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/openai/models/o1-preview/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/openai/models/o1-preview/model.json -------------------------------------------------------------------------------- /data/organizations/openai/models/o1-pro/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/openai/models/o1-pro/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/openai/models/o1-pro/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/openai/models/o1-pro/model.json -------------------------------------------------------------------------------- /data/organizations/openai/models/o3-2025-04-16/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/openai/models/o3-2025-04-16/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/openai/models/o3-2025-04-16/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/openai/models/o3-2025-04-16/model.json -------------------------------------------------------------------------------- /data/organizations/openai/models/o3-mini/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/openai/models/o3-mini/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/openai/models/o3-mini/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/openai/models/o3-mini/model.json -------------------------------------------------------------------------------- /data/organizations/openai/models/o3-pro-2025-06-10/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/openai/models/o3-pro-2025-06-10/model.json -------------------------------------------------------------------------------- /data/organizations/openai/models/o4-mini/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/openai/models/o4-mini/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/openai/models/o4-mini/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/openai/models/o4-mini/model.json -------------------------------------------------------------------------------- /data/organizations/openai/organization.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/openai/organization.json -------------------------------------------------------------------------------- /data/organizations/qwen/models/qvq-72b-preview/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/qwen/models/qvq-72b-preview/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/qwen/models/qvq-72b-preview/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/qwen/models/qvq-72b-preview/model.json -------------------------------------------------------------------------------- /data/organizations/qwen/models/qwen-2.5-14b-instruct/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/qwen/models/qwen-2.5-14b-instruct/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/qwen/models/qwen-2.5-14b-instruct/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/qwen/models/qwen-2.5-14b-instruct/model.json -------------------------------------------------------------------------------- /data/organizations/qwen/models/qwen-2.5-32b-instruct/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/qwen/models/qwen-2.5-32b-instruct/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/qwen/models/qwen-2.5-32b-instruct/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/qwen/models/qwen-2.5-32b-instruct/model.json -------------------------------------------------------------------------------- /data/organizations/qwen/models/qwen-2.5-72b-instruct/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/qwen/models/qwen-2.5-72b-instruct/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/qwen/models/qwen-2.5-72b-instruct/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/qwen/models/qwen-2.5-72b-instruct/model.json -------------------------------------------------------------------------------- /data/organizations/qwen/models/qwen-2.5-7b-instruct/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/qwen/models/qwen-2.5-7b-instruct/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/qwen/models/qwen-2.5-7b-instruct/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/qwen/models/qwen-2.5-7b-instruct/model.json -------------------------------------------------------------------------------- /data/organizations/qwen/models/qwen-2.5-coder-32b-instruct/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/qwen/models/qwen-2.5-coder-32b-instruct/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/qwen/models/qwen-2.5-coder-32b-instruct/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/qwen/models/qwen-2.5-coder-32b-instruct/model.json -------------------------------------------------------------------------------- /data/organizations/qwen/models/qwen-2.5-coder-7b-instruct/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/qwen/models/qwen-2.5-coder-7b-instruct/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/qwen/models/qwen-2.5-coder-7b-instruct/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/qwen/models/qwen-2.5-coder-7b-instruct/model.json -------------------------------------------------------------------------------- /data/organizations/qwen/models/qwen2-72b-instruct/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/qwen/models/qwen2-72b-instruct/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/qwen/models/qwen2-72b-instruct/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/qwen/models/qwen2-72b-instruct/model.json -------------------------------------------------------------------------------- /data/organizations/qwen/models/qwen2-7b-instruct/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/qwen/models/qwen2-7b-instruct/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/qwen/models/qwen2-7b-instruct/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/qwen/models/qwen2-7b-instruct/model.json -------------------------------------------------------------------------------- /data/organizations/qwen/models/qwen2-vl-72b/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/qwen/models/qwen2-vl-72b/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/qwen/models/qwen2-vl-72b/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/qwen/models/qwen2-vl-72b/model.json -------------------------------------------------------------------------------- /data/organizations/qwen/models/qwen2.5-omni-7b/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/qwen/models/qwen2.5-omni-7b/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/qwen/models/qwen2.5-omni-7b/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/qwen/models/qwen2.5-omni-7b/model.json -------------------------------------------------------------------------------- /data/organizations/qwen/models/qwen2.5-vl-32b/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/qwen/models/qwen2.5-vl-32b/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/qwen/models/qwen2.5-vl-32b/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/qwen/models/qwen2.5-vl-32b/model.json -------------------------------------------------------------------------------- /data/organizations/qwen/models/qwen2.5-vl-72b/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/qwen/models/qwen2.5-vl-72b/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/qwen/models/qwen2.5-vl-72b/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/qwen/models/qwen2.5-vl-72b/model.json -------------------------------------------------------------------------------- /data/organizations/qwen/models/qwen2.5-vl-7b/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/qwen/models/qwen2.5-vl-7b/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/qwen/models/qwen2.5-vl-7b/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/qwen/models/qwen2.5-vl-7b/model.json -------------------------------------------------------------------------------- /data/organizations/qwen/models/qwen3-235b-a22b-instruct-2507/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/qwen/models/qwen3-235b-a22b-instruct-2507/model.json -------------------------------------------------------------------------------- /data/organizations/qwen/models/qwen3-235b-a22b-thinking-2507/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/qwen/models/qwen3-235b-a22b-thinking-2507/model.json -------------------------------------------------------------------------------- /data/organizations/qwen/models/qwen3-235b-a22b/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/qwen/models/qwen3-235b-a22b/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/qwen/models/qwen3-235b-a22b/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/qwen/models/qwen3-235b-a22b/model.json -------------------------------------------------------------------------------- /data/organizations/qwen/models/qwen3-30b-a3b/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/qwen/models/qwen3-30b-a3b/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/qwen/models/qwen3-30b-a3b/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/qwen/models/qwen3-30b-a3b/model.json -------------------------------------------------------------------------------- /data/organizations/qwen/models/qwen3-32b/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/qwen/models/qwen3-32b/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/qwen/models/qwen3-32b/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/qwen/models/qwen3-32b/model.json -------------------------------------------------------------------------------- /data/organizations/qwen/models/qwen3-next-80b-a3b-base/benchmarks.json: -------------------------------------------------------------------------------- 1 | [] 2 | -------------------------------------------------------------------------------- /data/organizations/qwen/models/qwen3-next-80b-a3b-base/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/qwen/models/qwen3-next-80b-a3b-base/model.json -------------------------------------------------------------------------------- /data/organizations/qwen/models/qwen3-next-80b-a3b-instruct/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/qwen/models/qwen3-next-80b-a3b-instruct/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/qwen/models/qwen3-next-80b-a3b-instruct/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/qwen/models/qwen3-next-80b-a3b-instruct/model.json -------------------------------------------------------------------------------- /data/organizations/qwen/models/qwen3-next-80b-a3b-thinking/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/qwen/models/qwen3-next-80b-a3b-thinking/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/qwen/models/qwen3-next-80b-a3b-thinking/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/qwen/models/qwen3-next-80b-a3b-thinking/model.json -------------------------------------------------------------------------------- /data/organizations/qwen/models/qwq-32b-preview/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/qwen/models/qwq-32b-preview/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/qwen/models/qwq-32b-preview/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/qwen/models/qwq-32b-preview/model.json -------------------------------------------------------------------------------- /data/organizations/qwen/models/qwq-32b/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/qwen/models/qwq-32b/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/qwen/models/qwq-32b/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/qwen/models/qwq-32b/model.json -------------------------------------------------------------------------------- /data/organizations/qwen/organization.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/qwen/organization.json -------------------------------------------------------------------------------- /data/organizations/unknown/organization.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/unknown/organization.json -------------------------------------------------------------------------------- /data/organizations/xai/models/grok-1.5/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/xai/models/grok-1.5/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/xai/models/grok-1.5/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/xai/models/grok-1.5/model.json -------------------------------------------------------------------------------- /data/organizations/xai/models/grok-1.5v/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/xai/models/grok-1.5v/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/xai/models/grok-1.5v/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/xai/models/grok-1.5v/model.json -------------------------------------------------------------------------------- /data/organizations/xai/models/grok-2-mini/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/xai/models/grok-2-mini/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/xai/models/grok-2-mini/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/xai/models/grok-2-mini/model.json -------------------------------------------------------------------------------- /data/organizations/xai/models/grok-2/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/xai/models/grok-2/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/xai/models/grok-2/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/xai/models/grok-2/model.json -------------------------------------------------------------------------------- /data/organizations/xai/models/grok-3-mini/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/xai/models/grok-3-mini/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/xai/models/grok-3-mini/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/xai/models/grok-3-mini/model.json -------------------------------------------------------------------------------- /data/organizations/xai/models/grok-3/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/xai/models/grok-3/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/xai/models/grok-3/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/xai/models/grok-3/model.json -------------------------------------------------------------------------------- /data/organizations/xai/models/grok-4-fast/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/xai/models/grok-4-fast/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/xai/models/grok-4-fast/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/xai/models/grok-4-fast/model.json -------------------------------------------------------------------------------- /data/organizations/xai/models/grok-4-heavy/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/xai/models/grok-4-heavy/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/xai/models/grok-4-heavy/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/xai/models/grok-4-heavy/model.json -------------------------------------------------------------------------------- /data/organizations/xai/models/grok-4/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/xai/models/grok-4/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/xai/models/grok-4/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/xai/models/grok-4/model.json -------------------------------------------------------------------------------- /data/organizations/xai/models/grok-code-fast-1/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/xai/models/grok-code-fast-1/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/xai/models/grok-code-fast-1/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/xai/models/grok-code-fast-1/model.json -------------------------------------------------------------------------------- /data/organizations/xai/organization.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/xai/organization.json -------------------------------------------------------------------------------- /data/organizations/zai-org/models/glm-4.5-air/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/zai-org/models/glm-4.5-air/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/zai-org/models/glm-4.5-air/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/zai-org/models/glm-4.5-air/model.json -------------------------------------------------------------------------------- /data/organizations/zai-org/models/glm-4.5/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/zai-org/models/glm-4.5/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/zai-org/models/glm-4.5/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/zai-org/models/glm-4.5/model.json -------------------------------------------------------------------------------- /data/organizations/zai-org/models/glm-4.5v/benchmarks.json: -------------------------------------------------------------------------------- 1 | [] 2 | -------------------------------------------------------------------------------- /data/organizations/zai-org/models/glm-4.5v/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/zai-org/models/glm-4.5v/model.json -------------------------------------------------------------------------------- /data/organizations/zai-org/models/glm-4.6/benchmarks.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/zai-org/models/glm-4.6/benchmarks.json -------------------------------------------------------------------------------- /data/organizations/zai-org/models/glm-4.6/model.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/zai-org/models/glm-4.6/model.json -------------------------------------------------------------------------------- /data/organizations/zai-org/organization.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/organizations/zai-org/organization.json -------------------------------------------------------------------------------- /data/providers/anthropic/models.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/providers/anthropic/models.json -------------------------------------------------------------------------------- /data/providers/anthropic/provider.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/providers/anthropic/provider.json -------------------------------------------------------------------------------- /data/providers/azure/models.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/providers/azure/models.json -------------------------------------------------------------------------------- /data/providers/azure/provider.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/providers/azure/provider.json -------------------------------------------------------------------------------- /data/providers/bedrock/models.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/providers/bedrock/models.json -------------------------------------------------------------------------------- /data/providers/bedrock/provider.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/providers/bedrock/provider.json -------------------------------------------------------------------------------- /data/providers/cerebras/models.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/providers/cerebras/models.json -------------------------------------------------------------------------------- /data/providers/cerebras/provider.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/providers/cerebras/provider.json -------------------------------------------------------------------------------- /data/providers/cohere/models.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/providers/cohere/models.json -------------------------------------------------------------------------------- /data/providers/cohere/provider.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/providers/cohere/provider.json -------------------------------------------------------------------------------- /data/providers/deepinfra/models.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/providers/deepinfra/models.json -------------------------------------------------------------------------------- /data/providers/deepinfra/provider.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/providers/deepinfra/provider.json -------------------------------------------------------------------------------- /data/providers/deepseek/models.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/providers/deepseek/models.json -------------------------------------------------------------------------------- /data/providers/deepseek/provider.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/providers/deepseek/provider.json -------------------------------------------------------------------------------- /data/providers/fireworks/models.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/providers/fireworks/models.json -------------------------------------------------------------------------------- /data/providers/fireworks/provider.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/providers/fireworks/provider.json -------------------------------------------------------------------------------- /data/providers/google/models.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/providers/google/models.json -------------------------------------------------------------------------------- /data/providers/google/provider.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/providers/google/provider.json -------------------------------------------------------------------------------- /data/providers/groq/models.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/providers/groq/models.json -------------------------------------------------------------------------------- /data/providers/groq/provider.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/providers/groq/provider.json -------------------------------------------------------------------------------- /data/providers/hyperbolic/models.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/providers/hyperbolic/models.json -------------------------------------------------------------------------------- /data/providers/hyperbolic/provider.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/providers/hyperbolic/provider.json -------------------------------------------------------------------------------- /data/providers/lambda/models.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/providers/lambda/models.json -------------------------------------------------------------------------------- /data/providers/lambda/provider.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/providers/lambda/provider.json -------------------------------------------------------------------------------- /data/providers/mistral/models.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/providers/mistral/models.json -------------------------------------------------------------------------------- /data/providers/mistral/provider.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/providers/mistral/provider.json -------------------------------------------------------------------------------- /data/providers/novita/models.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/providers/novita/models.json -------------------------------------------------------------------------------- /data/providers/novita/provider.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/providers/novita/provider.json -------------------------------------------------------------------------------- /data/providers/openai/models.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/providers/openai/models.json -------------------------------------------------------------------------------- /data/providers/openai/provider.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/providers/openai/provider.json -------------------------------------------------------------------------------- /data/providers/replicate/models.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/providers/replicate/models.json -------------------------------------------------------------------------------- /data/providers/replicate/provider.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/providers/replicate/provider.json -------------------------------------------------------------------------------- /data/providers/sambanova/models.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/providers/sambanova/models.json -------------------------------------------------------------------------------- /data/providers/sambanova/provider.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/providers/sambanova/provider.json -------------------------------------------------------------------------------- /data/providers/together/models.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/providers/together/models.json -------------------------------------------------------------------------------- /data/providers/together/provider.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/providers/together/provider.json -------------------------------------------------------------------------------- /data/providers/xai/models.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/providers/xai/models.json -------------------------------------------------------------------------------- /data/providers/xai/provider.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/providers/xai/provider.json -------------------------------------------------------------------------------- /data/providers/zeroeval/models.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/providers/zeroeval/models.json -------------------------------------------------------------------------------- /data/providers/zeroeval/provider.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/data/providers/zeroeval/provider.json -------------------------------------------------------------------------------- /package.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/package.json -------------------------------------------------------------------------------- /schemas/README.md: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/schemas/README.md -------------------------------------------------------------------------------- /schemas/benchmark-results.schema.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/schemas/benchmark-results.schema.json -------------------------------------------------------------------------------- /schemas/benchmark.schema.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/schemas/benchmark.schema.json -------------------------------------------------------------------------------- /schemas/integrity-validator.js: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/schemas/integrity-validator.js -------------------------------------------------------------------------------- /schemas/license.schema.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/schemas/license.schema.json -------------------------------------------------------------------------------- /schemas/model.schema.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/schemas/model.schema.json -------------------------------------------------------------------------------- /schemas/organization.schema.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/schemas/organization.schema.json -------------------------------------------------------------------------------- /schemas/provider-models.schema.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/schemas/provider-models.schema.json -------------------------------------------------------------------------------- /schemas/provider.schema.json: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/schemas/provider.schema.json -------------------------------------------------------------------------------- /schemas/validator.js: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/JonathanChavezTamales/llm-leaderboard/HEAD/schemas/validator.js --------------------------------------------------------------------------------