├── .github └── workflows │ └── mypy.yml ├── .gitignore ├── Dec-2023.filelist ├── Dec-2023.tar.gz ├── LICENSE ├── Nov-2023.filelist ├── Nov-2023.tar.gz ├── Oct-2023.filelist ├── Oct-2023.tar.gz ├── README.md ├── api_query_cache ├── Anthropic_Eval_claude-3-opus-20240229_temp=0.0_cot=False_fs=5.tar.gz ├── Anthropic_Eval_claude-3-opus-20240229_temp=0.0_cot=True_fs=0.tar.gz ├── Anthropic_Eval_claude-3-opus-20240229_temp=0.0_cot=True_fs=5.tar.gz ├── Mistral_Eval_mistral-large-2402_temp=0.0_cot=False_fs=5.tar.gz ├── Mistral_Eval_mistral-large-2402_temp=0.0_cot=True_fs=0.tar.gz ├── Mistral_Eval_mistral-large-2402_temp=0.0_cot=True_fs=5.tar.gz ├── Mistral_Eval_mistral-medium-2312_temp=0.0_cot=False_fs=5.tar.gz ├── Mistral_Eval_mistral-medium-2312_temp=0.0_cot=True_fs=0.tar.gz ├── Mistral_Eval_mistral-medium-2312_temp=0.0_cot=True_fs=5.tar.gz ├── OAI_Eval_gpt-3.5-turbo_temp=0.0_cot=False_fs=5.tar.gz ├── OAI_Eval_gpt-3.5-turbo_temp=0.0_cot=True_fs=0.tar.gz ├── OAI_Eval_gpt-3.5-turbo_temp=0.0_cot=True_fs=5.tar.gz ├── OAI_Eval_gpt-4_temp=0.0_cot=False_fs=5.tar.gz ├── OAI_Eval_gpt-4_temp=0.0_cot=True_fs=0.tar.gz ├── OAI_Eval_gpt-4_temp=0.0_cot=True_fs=5.tar.gz ├── Together_Eval_Open-Orca_Mistral-7B-OpenOrca_temp=0.0_cot=False_fs=5.tar.gz ├── Together_Eval_Open-Orca_Mistral-7B-OpenOrca_temp=0.0_cot=True_fs=0.tar.gz ├── Together_Eval_Open-Orca_Mistral-7B-OpenOrca_temp=0.0_cot=True_fs=5.tar.gz ├── Together_Eval_Qwen_Qwen1.5-72B_temp=0.0_cot=False_fs=5.tar.gz ├── Together_Eval_Qwen_Qwen1.5-72B_temp=0.0_cot=True_fs=0.tar.gz ├── Together_Eval_Qwen_Qwen1.5-72B_temp=0.0_cot=True_fs=5.tar.gz ├── Together_Eval_allenai_OLMo-7B-Instruct_temp=0.0_cot=False_fs=5.tar.gz ├── Together_Eval_allenai_OLMo-7B-Instruct_temp=0.0_cot=True_fs=0.tar.gz ├── Together_Eval_allenai_OLMo-7B-Instruct_temp=0.0_cot=True_fs=5.tar.gz ├── Together_Eval_google_gemma-7b_temp=0.0_cot=False_fs=5.tar.gz ├── Together_Eval_google_gemma-7b_temp=0.0_cot=True_fs=0.tar.gz ├── Together_Eval_google_gemma-7b_temp=0.0_cot=True_fs=5.tar.gz ├── Together_Eval_meta-llama_Llama-2-70b-hf_temp=0.0_cot=False_fs=5.tar.gz ├── Together_Eval_meta-llama_Llama-2-70b-hf_temp=0.0_cot=True_fs=0.tar.gz ├── Together_Eval_meta-llama_Llama-2-70b-hf_temp=0.0_cot=True_fs=5.tar.gz ├── Together_Eval_microsoft_phi-2_temp=0.0_cot=False_fs=5.tar.gz ├── Together_Eval_microsoft_phi-2_temp=0.0_cot=True_fs=0.tar.gz ├── Together_Eval_microsoft_phi-2_temp=0.0_cot=True_fs=5.tar.gz ├── Together_Eval_togethercomputer_StripedHyena-Nous-7B_temp=0.0_cot=False_fs=5.tar.gz ├── Together_Eval_togethercomputer_StripedHyena-Nous-7B_temp=0.0_cot=True_fs=0.tar.gz ├── Together_Eval_togethercomputer_StripedHyena-Nous-7B_temp=0.0_cot=True_fs=5.tar.gz ├── Together_Eval_zero-one-ai_Yi-34B_temp=0.0_cot=False_fs=5.tar.gz ├── Together_Eval_zero-one-ai_Yi-34B_temp=0.0_cot=True_fs=0.tar.gz └── Together_Eval_zero-one-ai_Yi-34B_temp=0.0_cot=True_fs=5.tar.gz ├── chain_of_thought.py ├── eval ├── __init__.py ├── measure_math.py └── runner.py ├── evaluate.py ├── evaluated_models.json ├── few_shot.py ├── fn_snapshot.py ├── helper_utils.py ├── loader.py ├── math_utils ├── README.md ├── __init__.py ├── math_equivalence.py └── math_helpers.py ├── model_api ├── __init__.py ├── ant.py ├── claude.py ├── closed_api.py ├── file_api.py ├── gpt.py ├── mis.py ├── mistral.py ├── oai.py ├── offline.py ├── oss.py └── tog.py ├── monthly_snapshots.json ├── monthly_snapshots_only_oct.json ├── persist.py ├── query_cache.py ├── report ├── .gitignore ├── Makefile ├── all-coverage.pdf ├── all-levels.pdf ├── all-reasoning-gap.pdf ├── all-static-vs-func-accuracy.pdf ├── all-subjects.pdf ├── data_files.tar.gz ├── gpt4-by-level.pdf ├── gpt4-by-subject.pdf ├── gpt4-k=3-suffices.pdf ├── gpt4-levels.pdf ├── gpt4-reasoning-needed.pdf ├── gpt4-subjects.pdf ├── gpt4-surprise.pdf ├── main.tex ├── oss-by-level.pdf ├── oss-by-subject.pdf ├── oss-k=3-suffices.pdf ├── oss-reasoning-needed.pdf ├── oss-surprise.pdf └── refs.bib ├── requirements.txt ├── summarize_evals.py └── unformatted_llm.py /.github/workflows/mypy.yml: -------------------------------------------------------------------------------- 1 | name: Typecheck fneval 2 | 3 | on: 4 | push: 5 | branches: [main] 6 | paths: 7 | - "**" 8 | - ".github/workflows/mypy.yml" 9 | pull_request: 10 | paths: 11 | - "**" 12 | - ".github/workflows/mypy.yml" 13 | 14 | jobs: 15 | build: 16 | runs-on: ubuntu-22.04 17 | strategy: 18 | matrix: 19 | python-version: [3.9] 20 | 21 | steps: 22 | - uses: actions/checkout@v2 23 | - name: Set up Python ${{ matrix.python-version }} 24 | uses: actions/setup-python@v1 25 | with: 26 | python-version: ${{ matrix.python-version }} 27 | - name: Install dependencies 28 | run: | 29 | python3 -m pip install types-tqdm 30 | python -m pip install --upgrade pip 31 | pip install -r requirements.txt 32 | pip install mypy flake8 33 | - name: Check types 34 | run: | 35 | # flake8 . --ignore E501,W503,W504,W505,E266 36 | mypy --check-untyped-defs --pretty . 37 | -------------------------------------------------------------------------------- /.gitignore: -------------------------------------------------------------------------------- 1 | # save your api keys to this file, for convenient . ./api_keys.sh export 2 | api_keys.sh 3 | 4 | # source static dataset 5 | MATH.tar 6 | MATH/ 7 | 8 | # untarred monthly snapshots 9 | Dec-2023/ 10 | Nov-2023/ 11 | Oct-2023/ 12 | 13 | # caches (the saved snapshots can be checked in as .tar.gz, but no need for the untarred dirs) 14 | # with CoT and Few Shot 15 | api_query_cache/Anthropic_Eval_claude-3-opus-20240229_temp=0.0_cot=False_fs=5 16 | api_query_cache/Anthropic_Eval_claude-3-opus-20240229_temp=0.0_cot=True_fs=0 17 | api_query_cache/Anthropic_Eval_claude-3-opus-20240229_temp=0.0_cot=True_fs=5 18 | api_query_cache/OAI_Eval_gpt-3.5-turbo_temp=0.0_cot=False_fs=5 19 | api_query_cache/OAI_Eval_gpt-3.5-turbo_temp=0.0_cot=True_fs=0 20 | api_query_cache/OAI_Eval_gpt-3.5-turbo_temp=0.0_cot=True_fs=5 21 | api_query_cache/OAI_Eval_gpt-4_temp=0.0_cot=False_fs=5 22 | api_query_cache/OAI_Eval_gpt-4_temp=0.0_cot=True_fs=0 23 | api_query_cache/OAI_Eval_gpt-4_temp=0.0_cot=True_fs=5 24 | api_query_cache/Together_Eval_togethercomputer_StripedHyena-Nous-7B_temp=0.0_cot=False_fs=5 25 | api_query_cache/Together_Eval_togethercomputer_StripedHyena-Nous-7B_temp=0.0_cot=True_fs=0 26 | api_query_cache/Together_Eval_togethercomputer_StripedHyena-Nous-7B_temp=0.0_cot=True_fs=5 27 | api_query_cache/Together_Eval_Open-Orca_Mistral-7B-OpenOrca_temp=0.0_cot=False_fs=5 28 | api_query_cache/Together_Eval_Open-Orca_Mistral-7B-OpenOrca_temp=0.0_cot=True_fs=0 29 | api_query_cache/Together_Eval_Open-Orca_Mistral-7B-OpenOrca_temp=0.0_cot=True_fs=5 30 | api_query_cache/Together_Eval_zero-one-ai_Yi-34B_temp=0.0_cot=False_fs=5 31 | api_query_cache/Together_Eval_zero-one-ai_Yi-34B_temp=0.0_cot=True_fs=0 32 | api_query_cache/Together_Eval_zero-one-ai_Yi-34B_temp=0.0_cot=True_fs=5 33 | api_query_cache/Together_Eval_Qwen_Qwen1.5-72B_temp=0.0_cot=False_fs=5 34 | api_query_cache/Together_Eval_Qwen_Qwen1.5-72B_temp=0.0_cot=True_fs=0 35 | api_query_cache/Together_Eval_Qwen_Qwen1.5-72B_temp=0.0_cot=True_fs=5 36 | api_query_cache/Together_Eval_allenai_OLMo-7B-Instruct_temp=0.0_cot=False_fs=5 37 | api_query_cache/Together_Eval_allenai_OLMo-7B-Instruct_temp=0.0_cot=True_fs=0 38 | api_query_cache/Together_Eval_allenai_OLMo-7B-Instruct_temp=0.0_cot=True_fs=5 39 | api_query_cache/Together_Eval_google_gemma-7b_temp=0.0_cot=False_fs=5 40 | api_query_cache/Together_Eval_google_gemma-7b_temp=0.0_cot=True_fs=0 41 | api_query_cache/Together_Eval_google_gemma-7b_temp=0.0_cot=True_fs=5 42 | api_query_cache/Together_Eval_meta-llama_Llama-2-70b-hf_temp=0.0_cot=False_fs=5 43 | api_query_cache/Together_Eval_meta-llama_Llama-2-70b-hf_temp=0.0_cot=True_fs=0 44 | api_query_cache/Together_Eval_meta-llama_Llama-2-70b-hf_temp=0.0_cot=True_fs=5 45 | api_query_cache/Together_Eval_microsoft_phi-2_temp=0.0_cot=False_fs=5 46 | api_query_cache/Together_Eval_microsoft_phi-2_temp=0.0_cot=True_fs=0 47 | api_query_cache/Together_Eval_microsoft_phi-2_temp=0.0_cot=True_fs=5 48 | api_query_cache/Mistral_Eval_mistral-large-2402_temp=0.0_cot=False_fs=5 49 | api_query_cache/Mistral_Eval_mistral-large-2402_temp=0.0_cot=True_fs=0 50 | api_query_cache/Mistral_Eval_mistral-large-2402_temp=0.0_cot=True_fs=5 51 | api_query_cache/Mistral_Eval_mistral-medium-2312_temp=0.0_cot=False_fs=5 52 | api_query_cache/Mistral_Eval_mistral-medium-2312_temp=0.0_cot=True_fs=5 53 | api_query_cache/Mistral_Eval_mistral-medium-2312_temp=0.0_cot=True_fs=0 54 | 55 | # OLD 56 | # api_query_cache/Anthropic_Eval_claude-2.1/ 57 | # api_query_cache/Mistral_Eval_mistral-medium/ 58 | # api_query_cache/OAI_Eval_davinci/ 59 | # api_query_cache/OAI_Eval_gpt-3.5-turbo/ 60 | # api_query_cache/OAI_Eval_gpt-4/ 61 | # api_query_cache/Together_Eval_EleutherAI_llemma_7b/ 62 | # api_query_cache/Together_Eval_WizardLM_WizardCoder-Python-34B-V1.0/ 63 | # api_query_cache/Together_Eval_mistralai_Mixtral-8x7B-Instruct-v0.1/ 64 | # api_query_cache/Together_Eval_mistralai_Mixtral-8x7B-v0.1/ 65 | # api_query_cache/Together_Eval_togethercomputer_Qwen-7B/ 66 | # api_query_cache/Together_Eval_togethercomputer_StripedHyena-Hessian-7B/ 67 | # api_query_cache/Together_Eval_togethercomputer_StripedHyena-Nous-7B/ 68 | # api_query_cache/Together_Eval_togethercomputer_llama-2-70b/ 69 | # api_query_cache/Together_Eval_zero-one-ai_Yi-34B-Chat/ 70 | # api_query_cache/Together_Eval_zero-one-ai_Yi-34B/ 71 | 72 | # output of individual evals, intermediate from eval/measure_math.py 73 | intermediate 74 | 75 | # individual test results from model evals; output of evaluate.py 76 | *.json.pickle 77 | 78 | # output of summarize_evals.py 79 | *.json.pickle.* 80 | 81 | # mac 82 | .DS_Store 83 | 84 | # pycache 85 | __pycache__ 86 | 87 | # vim swaps 88 | *.swp 89 | -------------------------------------------------------------------------------- /Dec-2023.tar.gz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/Dec-2023.tar.gz -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | GNU GENERAL PUBLIC LICENSE 2 | Version 3, 29 June 2007 3 | 4 | Copyright (C) 2007 Free Software Foundation, Inc. 5 | Everyone is permitted to copy and distribute verbatim copies 6 | of this license document, but changing it is not allowed. 7 | 8 | Preamble 9 | 10 | The GNU General Public License is a free, copyleft license for 11 | software and other kinds of works. 12 | 13 | The licenses for most software and other practical works are designed 14 | to take away your freedom to share and change the works. By contrast, 15 | the GNU General Public License is intended to guarantee your freedom to 16 | share and change all versions of a program--to make sure it remains free 17 | software for all its users. We, the Free Software Foundation, use the 18 | GNU General Public License for most of our software; it applies also to 19 | any other work released this way by its authors. You can apply it to 20 | your programs, too. 21 | 22 | When we speak of free software, we are referring to freedom, not 23 | price. Our General Public Licenses are designed to make sure that you 24 | have the freedom to distribute copies of free software (and charge for 25 | them if you wish), that you receive source code or can get it if you 26 | want it, that you can change the software or use pieces of it in new 27 | free programs, and that you know you can do these things. 28 | 29 | To protect your rights, we need to prevent others from denying you 30 | these rights or asking you to surrender the rights. Therefore, you have 31 | certain responsibilities if you distribute copies of the software, or if 32 | you modify it: responsibilities to respect the freedom of others. 33 | 34 | For example, if you distribute copies of such a program, whether 35 | gratis or for a fee, you must pass on to the recipients the same 36 | freedoms that you received. You must make sure that they, too, receive 37 | or can get the source code. And you must show them these terms so they 38 | know their rights. 39 | 40 | Developers that use the GNU GPL protect your rights with two steps: 41 | (1) assert copyright on the software, and (2) offer you this License 42 | giving you legal permission to copy, distribute and/or modify it. 43 | 44 | For the developers' and authors' protection, the GPL clearly explains 45 | that there is no warranty for this free software. For both users' and 46 | authors' sake, the GPL requires that modified versions be marked as 47 | changed, so that their problems will not be attributed erroneously to 48 | authors of previous versions. 49 | 50 | Some devices are designed to deny users access to install or run 51 | modified versions of the software inside them, although the manufacturer 52 | can do so. This is fundamentally incompatible with the aim of 53 | protecting users' freedom to change the software. The systematic 54 | pattern of such abuse occurs in the area of products for individuals to 55 | use, which is precisely where it is most unacceptable. Therefore, we 56 | have designed this version of the GPL to prohibit the practice for those 57 | products. If such problems arise substantially in other domains, we 58 | stand ready to extend this provision to those domains in future versions 59 | of the GPL, as needed to protect the freedom of users. 60 | 61 | Finally, every program is threatened constantly by software patents. 62 | States should not allow patents to restrict development and use of 63 | software on general-purpose computers, but in those that do, we wish to 64 | avoid the special danger that patents applied to a free program could 65 | make it effectively proprietary. To prevent this, the GPL assures that 66 | patents cannot be used to render the program non-free. 67 | 68 | The precise terms and conditions for copying, distribution and 69 | modification follow. 70 | 71 | TERMS AND CONDITIONS 72 | 73 | 0. Definitions. 74 | 75 | "This License" refers to version 3 of the GNU General Public License. 76 | 77 | "Copyright" also means copyright-like laws that apply to other kinds of 78 | works, such as semiconductor masks. 79 | 80 | "The Program" refers to any copyrightable work licensed under this 81 | License. Each licensee is addressed as "you". "Licensees" and 82 | "recipients" may be individuals or organizations. 83 | 84 | To "modify" a work means to copy from or adapt all or part of the work 85 | in a fashion requiring copyright permission, other than the making of an 86 | exact copy. The resulting work is called a "modified version" of the 87 | earlier work or a work "based on" the earlier work. 88 | 89 | A "covered work" means either the unmodified Program or a work based 90 | on the Program. 91 | 92 | To "propagate" a work means to do anything with it that, without 93 | permission, would make you directly or secondarily liable for 94 | infringement under applicable copyright law, except executing it on a 95 | computer or modifying a private copy. Propagation includes copying, 96 | distribution (with or without modification), making available to the 97 | public, and in some countries other activities as well. 98 | 99 | To "convey" a work means any kind of propagation that enables other 100 | parties to make or receive copies. Mere interaction with a user through 101 | a computer network, with no transfer of a copy, is not conveying. 102 | 103 | An interactive user interface displays "Appropriate Legal Notices" 104 | to the extent that it includes a convenient and prominently visible 105 | feature that (1) displays an appropriate copyright notice, and (2) 106 | tells the user that there is no warranty for the work (except to the 107 | extent that warranties are provided), that licensees may convey the 108 | work under this License, and how to view a copy of this License. If 109 | the interface presents a list of user commands or options, such as a 110 | menu, a prominent item in the list meets this criterion. 111 | 112 | 1. Source Code. 113 | 114 | The "source code" for a work means the preferred form of the work 115 | for making modifications to it. "Object code" means any non-source 116 | form of a work. 117 | 118 | A "Standard Interface" means an interface that either is an official 119 | standard defined by a recognized standards body, or, in the case of 120 | interfaces specified for a particular programming language, one that 121 | is widely used among developers working in that language. 122 | 123 | The "System Libraries" of an executable work include anything, other 124 | than the work as a whole, that (a) is included in the normal form of 125 | packaging a Major Component, but which is not part of that Major 126 | Component, and (b) serves only to enable use of the work with that 127 | Major Component, or to implement a Standard Interface for which an 128 | implementation is available to the public in source code form. A 129 | "Major Component", in this context, means a major essential component 130 | (kernel, window system, and so on) of the specific operating system 131 | (if any) on which the executable work runs, or a compiler used to 132 | produce the work, or an object code interpreter used to run it. 133 | 134 | The "Corresponding Source" for a work in object code form means all 135 | the source code needed to generate, install, and (for an executable 136 | work) run the object code and to modify the work, including scripts to 137 | control those activities. However, it does not include the work's 138 | System Libraries, or general-purpose tools or generally available free 139 | programs which are used unmodified in performing those activities but 140 | which are not part of the work. For example, Corresponding Source 141 | includes interface definition files associated with source files for 142 | the work, and the source code for shared libraries and dynamically 143 | linked subprograms that the work is specifically designed to require, 144 | such as by intimate data communication or control flow between those 145 | subprograms and other parts of the work. 146 | 147 | The Corresponding Source need not include anything that users 148 | can regenerate automatically from other parts of the Corresponding 149 | Source. 150 | 151 | The Corresponding Source for a work in source code form is that 152 | same work. 153 | 154 | 2. Basic Permissions. 155 | 156 | All rights granted under this License are granted for the term of 157 | copyright on the Program, and are irrevocable provided the stated 158 | conditions are met. This License explicitly affirms your unlimited 159 | permission to run the unmodified Program. The output from running a 160 | covered work is covered by this License only if the output, given its 161 | content, constitutes a covered work. This License acknowledges your 162 | rights of fair use or other equivalent, as provided by copyright law. 163 | 164 | You may make, run and propagate covered works that you do not 165 | convey, without conditions so long as your license otherwise remains 166 | in force. You may convey covered works to others for the sole purpose 167 | of having them make modifications exclusively for you, or provide you 168 | with facilities for running those works, provided that you comply with 169 | the terms of this License in conveying all material for which you do 170 | not control copyright. Those thus making or running the covered works 171 | for you must do so exclusively on your behalf, under your direction 172 | and control, on terms that prohibit them from making any copies of 173 | your copyrighted material outside their relationship with you. 174 | 175 | Conveying under any other circumstances is permitted solely under 176 | the conditions stated below. Sublicensing is not allowed; section 10 177 | makes it unnecessary. 178 | 179 | 3. Protecting Users' Legal Rights From Anti-Circumvention Law. 180 | 181 | No covered work shall be deemed part of an effective technological 182 | measure under any applicable law fulfilling obligations under article 183 | 11 of the WIPO copyright treaty adopted on 20 December 1996, or 184 | similar laws prohibiting or restricting circumvention of such 185 | measures. 186 | 187 | When you convey a covered work, you waive any legal power to forbid 188 | circumvention of technological measures to the extent such circumvention 189 | is effected by exercising rights under this License with respect to 190 | the covered work, and you disclaim any intention to limit operation or 191 | modification of the work as a means of enforcing, against the work's 192 | users, your or third parties' legal rights to forbid circumvention of 193 | technological measures. 194 | 195 | 4. Conveying Verbatim Copies. 196 | 197 | You may convey verbatim copies of the Program's source code as you 198 | receive it, in any medium, provided that you conspicuously and 199 | appropriately publish on each copy an appropriate copyright notice; 200 | keep intact all notices stating that this License and any 201 | non-permissive terms added in accord with section 7 apply to the code; 202 | keep intact all notices of the absence of any warranty; and give all 203 | recipients a copy of this License along with the Program. 204 | 205 | You may charge any price or no price for each copy that you convey, 206 | and you may offer support or warranty protection for a fee. 207 | 208 | 5. Conveying Modified Source Versions. 209 | 210 | You may convey a work based on the Program, or the modifications to 211 | produce it from the Program, in the form of source code under the 212 | terms of section 4, provided that you also meet all of these conditions: 213 | 214 | a) The work must carry prominent notices stating that you modified 215 | it, and giving a relevant date. 216 | 217 | b) The work must carry prominent notices stating that it is 218 | released under this License and any conditions added under section 219 | 7. This requirement modifies the requirement in section 4 to 220 | "keep intact all notices". 221 | 222 | c) You must license the entire work, as a whole, under this 223 | License to anyone who comes into possession of a copy. This 224 | License will therefore apply, along with any applicable section 7 225 | additional terms, to the whole of the work, and all its parts, 226 | regardless of how they are packaged. This License gives no 227 | permission to license the work in any other way, but it does not 228 | invalidate such permission if you have separately received it. 229 | 230 | d) If the work has interactive user interfaces, each must display 231 | Appropriate Legal Notices; however, if the Program has interactive 232 | interfaces that do not display Appropriate Legal Notices, your 233 | work need not make them do so. 234 | 235 | A compilation of a covered work with other separate and independent 236 | works, which are not by their nature extensions of the covered work, 237 | and which are not combined with it such as to form a larger program, 238 | in or on a volume of a storage or distribution medium, is called an 239 | "aggregate" if the compilation and its resulting copyright are not 240 | used to limit the access or legal rights of the compilation's users 241 | beyond what the individual works permit. Inclusion of a covered work 242 | in an aggregate does not cause this License to apply to the other 243 | parts of the aggregate. 244 | 245 | 6. Conveying Non-Source Forms. 246 | 247 | You may convey a covered work in object code form under the terms 248 | of sections 4 and 5, provided that you also convey the 249 | machine-readable Corresponding Source under the terms of this License, 250 | in one of these ways: 251 | 252 | a) Convey the object code in, or embodied in, a physical product 253 | (including a physical distribution medium), accompanied by the 254 | Corresponding Source fixed on a durable physical medium 255 | customarily used for software interchange. 256 | 257 | b) Convey the object code in, or embodied in, a physical product 258 | (including a physical distribution medium), accompanied by a 259 | written offer, valid for at least three years and valid for as 260 | long as you offer spare parts or customer support for that product 261 | model, to give anyone who possesses the object code either (1) a 262 | copy of the Corresponding Source for all the software in the 263 | product that is covered by this License, on a durable physical 264 | medium customarily used for software interchange, for a price no 265 | more than your reasonable cost of physically performing this 266 | conveying of source, or (2) access to copy the 267 | Corresponding Source from a network server at no charge. 268 | 269 | c) Convey individual copies of the object code with a copy of the 270 | written offer to provide the Corresponding Source. This 271 | alternative is allowed only occasionally and noncommercially, and 272 | only if you received the object code with such an offer, in accord 273 | with subsection 6b. 274 | 275 | d) Convey the object code by offering access from a designated 276 | place (gratis or for a charge), and offer equivalent access to the 277 | Corresponding Source in the same way through the same place at no 278 | further charge. You need not require recipients to copy the 279 | Corresponding Source along with the object code. If the place to 280 | copy the object code is a network server, the Corresponding Source 281 | may be on a different server (operated by you or a third party) 282 | that supports equivalent copying facilities, provided you maintain 283 | clear directions next to the object code saying where to find the 284 | Corresponding Source. Regardless of what server hosts the 285 | Corresponding Source, you remain obligated to ensure that it is 286 | available for as long as needed to satisfy these requirements. 287 | 288 | e) Convey the object code using peer-to-peer transmission, provided 289 | you inform other peers where the object code and Corresponding 290 | Source of the work are being offered to the general public at no 291 | charge under subsection 6d. 292 | 293 | A separable portion of the object code, whose source code is excluded 294 | from the Corresponding Source as a System Library, need not be 295 | included in conveying the object code work. 296 | 297 | A "User Product" is either (1) a "consumer product", which means any 298 | tangible personal property which is normally used for personal, family, 299 | or household purposes, or (2) anything designed or sold for incorporation 300 | into a dwelling. In determining whether a product is a consumer product, 301 | doubtful cases shall be resolved in favor of coverage. For a particular 302 | product received by a particular user, "normally used" refers to a 303 | typical or common use of that class of product, regardless of the status 304 | of the particular user or of the way in which the particular user 305 | actually uses, or expects or is expected to use, the product. A product 306 | is a consumer product regardless of whether the product has substantial 307 | commercial, industrial or non-consumer uses, unless such uses represent 308 | the only significant mode of use of the product. 309 | 310 | "Installation Information" for a User Product means any methods, 311 | procedures, authorization keys, or other information required to install 312 | and execute modified versions of a covered work in that User Product from 313 | a modified version of its Corresponding Source. The information must 314 | suffice to ensure that the continued functioning of the modified object 315 | code is in no case prevented or interfered with solely because 316 | modification has been made. 317 | 318 | If you convey an object code work under this section in, or with, or 319 | specifically for use in, a User Product, and the conveying occurs as 320 | part of a transaction in which the right of possession and use of the 321 | User Product is transferred to the recipient in perpetuity or for a 322 | fixed term (regardless of how the transaction is characterized), the 323 | Corresponding Source conveyed under this section must be accompanied 324 | by the Installation Information. But this requirement does not apply 325 | if neither you nor any third party retains the ability to install 326 | modified object code on the User Product (for example, the work has 327 | been installed in ROM). 328 | 329 | The requirement to provide Installation Information does not include a 330 | requirement to continue to provide support service, warranty, or updates 331 | for a work that has been modified or installed by the recipient, or for 332 | the User Product in which it has been modified or installed. Access to a 333 | network may be denied when the modification itself materially and 334 | adversely affects the operation of the network or violates the rules and 335 | protocols for communication across the network. 336 | 337 | Corresponding Source conveyed, and Installation Information provided, 338 | in accord with this section must be in a format that is publicly 339 | documented (and with an implementation available to the public in 340 | source code form), and must require no special password or key for 341 | unpacking, reading or copying. 342 | 343 | 7. Additional Terms. 344 | 345 | "Additional permissions" are terms that supplement the terms of this 346 | License by making exceptions from one or more of its conditions. 347 | Additional permissions that are applicable to the entire Program shall 348 | be treated as though they were included in this License, to the extent 349 | that they are valid under applicable law. If additional permissions 350 | apply only to part of the Program, that part may be used separately 351 | under those permissions, but the entire Program remains governed by 352 | this License without regard to the additional permissions. 353 | 354 | When you convey a copy of a covered work, you may at your option 355 | remove any additional permissions from that copy, or from any part of 356 | it. (Additional permissions may be written to require their own 357 | removal in certain cases when you modify the work.) You may place 358 | additional permissions on material, added by you to a covered work, 359 | for which you have or can give appropriate copyright permission. 360 | 361 | Notwithstanding any other provision of this License, for material you 362 | add to a covered work, you may (if authorized by the copyright holders of 363 | that material) supplement the terms of this License with terms: 364 | 365 | a) Disclaiming warranty or limiting liability differently from the 366 | terms of sections 15 and 16 of this License; or 367 | 368 | b) Requiring preservation of specified reasonable legal notices or 369 | author attributions in that material or in the Appropriate Legal 370 | Notices displayed by works containing it; or 371 | 372 | c) Prohibiting misrepresentation of the origin of that material, or 373 | requiring that modified versions of such material be marked in 374 | reasonable ways as different from the original version; or 375 | 376 | d) Limiting the use for publicity purposes of names of licensors or 377 | authors of the material; or 378 | 379 | e) Declining to grant rights under trademark law for use of some 380 | trade names, trademarks, or service marks; or 381 | 382 | f) Requiring indemnification of licensors and authors of that 383 | material by anyone who conveys the material (or modified versions of 384 | it) with contractual assumptions of liability to the recipient, for 385 | any liability that these contractual assumptions directly impose on 386 | those licensors and authors. 387 | 388 | All other non-permissive additional terms are considered "further 389 | restrictions" within the meaning of section 10. If the Program as you 390 | received it, or any part of it, contains a notice stating that it is 391 | governed by this License along with a term that is a further 392 | restriction, you may remove that term. If a license document contains 393 | a further restriction but permits relicensing or conveying under this 394 | License, you may add to a covered work material governed by the terms 395 | of that license document, provided that the further restriction does 396 | not survive such relicensing or conveying. 397 | 398 | If you add terms to a covered work in accord with this section, you 399 | must place, in the relevant source files, a statement of the 400 | additional terms that apply to those files, or a notice indicating 401 | where to find the applicable terms. 402 | 403 | Additional terms, permissive or non-permissive, may be stated in the 404 | form of a separately written license, or stated as exceptions; 405 | the above requirements apply either way. 406 | 407 | 8. Termination. 408 | 409 | You may not propagate or modify a covered work except as expressly 410 | provided under this License. Any attempt otherwise to propagate or 411 | modify it is void, and will automatically terminate your rights under 412 | this License (including any patent licenses granted under the third 413 | paragraph of section 11). 414 | 415 | However, if you cease all violation of this License, then your 416 | license from a particular copyright holder is reinstated (a) 417 | provisionally, unless and until the copyright holder explicitly and 418 | finally terminates your license, and (b) permanently, if the copyright 419 | holder fails to notify you of the violation by some reasonable means 420 | prior to 60 days after the cessation. 421 | 422 | Moreover, your license from a particular copyright holder is 423 | reinstated permanently if the copyright holder notifies you of the 424 | violation by some reasonable means, this is the first time you have 425 | received notice of violation of this License (for any work) from that 426 | copyright holder, and you cure the violation prior to 30 days after 427 | your receipt of the notice. 428 | 429 | Termination of your rights under this section does not terminate the 430 | licenses of parties who have received copies or rights from you under 431 | this License. If your rights have been terminated and not permanently 432 | reinstated, you do not qualify to receive new licenses for the same 433 | material under section 10. 434 | 435 | 9. Acceptance Not Required for Having Copies. 436 | 437 | You are not required to accept this License in order to receive or 438 | run a copy of the Program. Ancillary propagation of a covered work 439 | occurring solely as a consequence of using peer-to-peer transmission 440 | to receive a copy likewise does not require acceptance. However, 441 | nothing other than this License grants you permission to propagate or 442 | modify any covered work. These actions infringe copyright if you do 443 | not accept this License. Therefore, by modifying or propagating a 444 | covered work, you indicate your acceptance of this License to do so. 445 | 446 | 10. Automatic Licensing of Downstream Recipients. 447 | 448 | Each time you convey a covered work, the recipient automatically 449 | receives a license from the original licensors, to run, modify and 450 | propagate that work, subject to this License. You are not responsible 451 | for enforcing compliance by third parties with this License. 452 | 453 | An "entity transaction" is a transaction transferring control of an 454 | organization, or substantially all assets of one, or subdividing an 455 | organization, or merging organizations. If propagation of a covered 456 | work results from an entity transaction, each party to that 457 | transaction who receives a copy of the work also receives whatever 458 | licenses to the work the party's predecessor in interest had or could 459 | give under the previous paragraph, plus a right to possession of the 460 | Corresponding Source of the work from the predecessor in interest, if 461 | the predecessor has it or can get it with reasonable efforts. 462 | 463 | You may not impose any further restrictions on the exercise of the 464 | rights granted or affirmed under this License. For example, you may 465 | not impose a license fee, royalty, or other charge for exercise of 466 | rights granted under this License, and you may not initiate litigation 467 | (including a cross-claim or counterclaim in a lawsuit) alleging that 468 | any patent claim is infringed by making, using, selling, offering for 469 | sale, or importing the Program or any portion of it. 470 | 471 | 11. Patents. 472 | 473 | A "contributor" is a copyright holder who authorizes use under this 474 | License of the Program or a work on which the Program is based. The 475 | work thus licensed is called the contributor's "contributor version". 476 | 477 | A contributor's "essential patent claims" are all patent claims 478 | owned or controlled by the contributor, whether already acquired or 479 | hereafter acquired, that would be infringed by some manner, permitted 480 | by this License, of making, using, or selling its contributor version, 481 | but do not include claims that would be infringed only as a 482 | consequence of further modification of the contributor version. For 483 | purposes of this definition, "control" includes the right to grant 484 | patent sublicenses in a manner consistent with the requirements of 485 | this License. 486 | 487 | Each contributor grants you a non-exclusive, worldwide, royalty-free 488 | patent license under the contributor's essential patent claims, to 489 | make, use, sell, offer for sale, import and otherwise run, modify and 490 | propagate the contents of its contributor version. 491 | 492 | In the following three paragraphs, a "patent license" is any express 493 | agreement or commitment, however denominated, not to enforce a patent 494 | (such as an express permission to practice a patent or covenant not to 495 | sue for patent infringement). To "grant" such a patent license to a 496 | party means to make such an agreement or commitment not to enforce a 497 | patent against the party. 498 | 499 | If you convey a covered work, knowingly relying on a patent license, 500 | and the Corresponding Source of the work is not available for anyone 501 | to copy, free of charge and under the terms of this License, through a 502 | publicly available network server or other readily accessible means, 503 | then you must either (1) cause the Corresponding Source to be so 504 | available, or (2) arrange to deprive yourself of the benefit of the 505 | patent license for this particular work, or (3) arrange, in a manner 506 | consistent with the requirements of this License, to extend the patent 507 | license to downstream recipients. "Knowingly relying" means you have 508 | actual knowledge that, but for the patent license, your conveying the 509 | covered work in a country, or your recipient's use of the covered work 510 | in a country, would infringe one or more identifiable patents in that 511 | country that you have reason to believe are valid. 512 | 513 | If, pursuant to or in connection with a single transaction or 514 | arrangement, you convey, or propagate by procuring conveyance of, a 515 | covered work, and grant a patent license to some of the parties 516 | receiving the covered work authorizing them to use, propagate, modify 517 | or convey a specific copy of the covered work, then the patent license 518 | you grant is automatically extended to all recipients of the covered 519 | work and works based on it. 520 | 521 | A patent license is "discriminatory" if it does not include within 522 | the scope of its coverage, prohibits the exercise of, or is 523 | conditioned on the non-exercise of one or more of the rights that are 524 | specifically granted under this License. You may not convey a covered 525 | work if you are a party to an arrangement with a third party that is 526 | in the business of distributing software, under which you make payment 527 | to the third party based on the extent of your activity of conveying 528 | the work, and under which the third party grants, to any of the 529 | parties who would receive the covered work from you, a discriminatory 530 | patent license (a) in connection with copies of the covered work 531 | conveyed by you (or copies made from those copies), or (b) primarily 532 | for and in connection with specific products or compilations that 533 | contain the covered work, unless you entered into that arrangement, 534 | or that patent license was granted, prior to 28 March 2007. 535 | 536 | Nothing in this License shall be construed as excluding or limiting 537 | any implied license or other defenses to infringement that may 538 | otherwise be available to you under applicable patent law. 539 | 540 | 12. No Surrender of Others' Freedom. 541 | 542 | If conditions are imposed on you (whether by court order, agreement or 543 | otherwise) that contradict the conditions of this License, they do not 544 | excuse you from the conditions of this License. If you cannot convey a 545 | covered work so as to satisfy simultaneously your obligations under this 546 | License and any other pertinent obligations, then as a consequence you may 547 | not convey it at all. For example, if you agree to terms that obligate you 548 | to collect a royalty for further conveying from those to whom you convey 549 | the Program, the only way you could satisfy both those terms and this 550 | License would be to refrain entirely from conveying the Program. 551 | 552 | 13. Use with the GNU Affero General Public License. 553 | 554 | Notwithstanding any other provision of this License, you have 555 | permission to link or combine any covered work with a work licensed 556 | under version 3 of the GNU Affero General Public License into a single 557 | combined work, and to convey the resulting work. The terms of this 558 | License will continue to apply to the part which is the covered work, 559 | but the special requirements of the GNU Affero General Public License, 560 | section 13, concerning interaction through a network will apply to the 561 | combination as such. 562 | 563 | 14. Revised Versions of this License. 564 | 565 | The Free Software Foundation may publish revised and/or new versions of 566 | the GNU General Public License from time to time. Such new versions will 567 | be similar in spirit to the present version, but may differ in detail to 568 | address new problems or concerns. 569 | 570 | Each version is given a distinguishing version number. If the 571 | Program specifies that a certain numbered version of the GNU General 572 | Public License "or any later version" applies to it, you have the 573 | option of following the terms and conditions either of that numbered 574 | version or of any later version published by the Free Software 575 | Foundation. If the Program does not specify a version number of the 576 | GNU General Public License, you may choose any version ever published 577 | by the Free Software Foundation. 578 | 579 | If the Program specifies that a proxy can decide which future 580 | versions of the GNU General Public License can be used, that proxy's 581 | public statement of acceptance of a version permanently authorizes you 582 | to choose that version for the Program. 583 | 584 | Later license versions may give you additional or different 585 | permissions. However, no additional obligations are imposed on any 586 | author or copyright holder as a result of your choosing to follow a 587 | later version. 588 | 589 | 15. Disclaimer of Warranty. 590 | 591 | THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY 592 | APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT 593 | HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY 594 | OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, 595 | THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 596 | PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM 597 | IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF 598 | ALL NECESSARY SERVICING, REPAIR OR CORRECTION. 599 | 600 | 16. Limitation of Liability. 601 | 602 | IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING 603 | WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS 604 | THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY 605 | GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE 606 | USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF 607 | DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD 608 | PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), 609 | EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF 610 | SUCH DAMAGES. 611 | 612 | 17. Interpretation of Sections 15 and 16. 613 | 614 | If the disclaimer of warranty and limitation of liability provided 615 | above cannot be given local legal effect according to their terms, 616 | reviewing courts shall apply local law that most closely approximates 617 | an absolute waiver of all civil liability in connection with the 618 | Program, unless a warranty or assumption of liability accompanies a 619 | copy of the Program in return for a fee. 620 | 621 | END OF TERMS AND CONDITIONS 622 | 623 | How to Apply These Terms to Your New Programs 624 | 625 | If you develop a new program, and you want it to be of the greatest 626 | possible use to the public, the best way to achieve this is to make it 627 | free software which everyone can redistribute and change under these terms. 628 | 629 | To do so, attach the following notices to the program. It is safest 630 | to attach them to the start of each source file to most effectively 631 | state the exclusion of warranty; and each file should have at least 632 | the "copyright" line and a pointer to where the full notice is found. 633 | 634 | 635 | Copyright (C) 636 | 637 | This program is free software: you can redistribute it and/or modify 638 | it under the terms of the GNU General Public License as published by 639 | the Free Software Foundation, either version 3 of the License, or 640 | (at your option) any later version. 641 | 642 | This program is distributed in the hope that it will be useful, 643 | but WITHOUT ANY WARRANTY; without even the implied warranty of 644 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 645 | GNU General Public License for more details. 646 | 647 | You should have received a copy of the GNU General Public License 648 | along with this program. If not, see . 649 | 650 | Also add information on how to contact you by electronic and paper mail. 651 | 652 | If the program does terminal interaction, make it output a short 653 | notice like this when it starts in an interactive mode: 654 | 655 | Copyright (C) 656 | This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'. 657 | This is free software, and you are welcome to redistribute it 658 | under certain conditions; type `show c' for details. 659 | 660 | The hypothetical commands `show w' and `show c' should show the appropriate 661 | parts of the General Public License. Of course, your program's commands 662 | might be different; for a GUI interface, you would use an "about box". 663 | 664 | You should also get your employer (if you work as a programmer) or school, 665 | if any, to sign a "copyright disclaimer" for the program, if necessary. 666 | For more information on this, and how to apply and follow the GNU GPL, see 667 | . 668 | 669 | The GNU General Public License does not permit incorporating your program 670 | into proprietary programs. If your program is a subroutine library, you 671 | may consider it more useful to permit linking proprietary applications with 672 | the library. If this is what you want to do, use the GNU Lesser General 673 | Public License instead of this License. But first, please read 674 | . 675 | -------------------------------------------------------------------------------- /Nov-2023.tar.gz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/Nov-2023.tar.gz -------------------------------------------------------------------------------- /Oct-2023.tar.gz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/Oct-2023.tar.gz -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | 2 | # Functional Benchmarks and Reasoning Gap 3 | 4 | This repository accompanies the paper 5 | [Functional Benchmarks for Robust Evaluation of Reasoning Performance, 6 | and the Reasoning Gap](https://arxiv.org/abs/2402.19450) 7 | 8 | Summary tweet threads: 9 | [Without CoT](https://twitter.com/_saurabh/status/1763626711407816930), 10 | and [with CoT](https://twitter.com/_saurabh/status/1769852207925805080). 11 | 12 | --- 13 | 14 | **Note (Feb'24): This repo and associated paper will not be finalized until the 15 | Q2'24 release. We are working to get 100% coverage over MATH and GSM8K. We are 16 | releasing the first version (Q1'24) for early access to the community.** 17 | 18 | # Overview 19 | We propose a way to evaluate reasoning capabilities of models that could have 20 | been trained on the entire (compressed on uncompressed) text of the internet, 21 | and possibly retrieval augmented. 22 | 23 | We rewrite each benchmark question, in parameterized functional form, such that 24 | the reasoning involved in solving each instance is identical, but the question 25 | and answer are textually distinct each time. We snapshot a new instance of the 26 | entire benchmark each month. A model has to solve the last K instances 27 | correctly to count as correctly reasoning through that question. We notate 28 | these functionalized versions of the benchmarks with "()", e.g., "MATH()" for 29 | the "MATH" benchmark. 30 | 31 | For each model the evaluation script tabulates: 32 | * Static accuracy: accuracy on the static benchmark (e.g., MATH). 33 | * Functional accuracy: accuracy over the last K instances of the functional benchmark 34 | (e.g., MATH()) and the static benchmark. 35 | * Reasoning Gap: Percent drop from static to functional. 36 | * Hallucination [0,100]: Percent times incorrect solution output, instead 37 | of stating no solution. 38 | 39 | # Run (with external API calls) 40 | Export the following keys: 41 | ``` 42 | export OPENAI_ORG="org-..." 43 | export OPENAI_API_KEY="sk-..." 44 | export ANTHROPIC_API_KEY="sk-..." 45 | export TOGETHER_API_KEY="..." 46 | export MISTRAL_API_KEY="..." 47 | export HUGGING_FACE_HUB_TOKEN="hf_..." 48 | ``` 49 | 50 | After exporting the API keys, run: 51 | ``` 52 | python3 -m evaluate 53 | 54 | # compute accuracies for static/functional (or run with just --stat_fn to see all available stats) 55 | python3 -m summarize_evals --stat_fn stat_accuracy 56 | ``` 57 | 58 | # Run (offline, with manual model completions) 59 | If you have an internal model that cannot be called via API, then it might be best to dump out all prompts to JSON, and complete them offline and bring them back into the harness. There is a fake api called `model_api/offline.py` that lets you do that. 60 | 61 | Create two model spec JSONs, one for writing prompts, and one for reading completions 62 | 63 | Writing prompts `evaluated_models_write_qs.json` should look like: 64 | ``` 65 | [ { "name": "your-local-model", "script": "model_api/offline.py", "include": true, "CoT": false, "extra_params": "{\"write_qs\": true}" } ] 66 | ``` 67 | Reading completions `evaluated_models.json` should look like: 68 | ``` 69 | [ { "name": "your-local-model", "script": "model_api/offline.py", "include": true, "CoT": false } ] 70 | ``` 71 | 72 | ``` 73 | # dump prompts 74 | python -m evaluate --model_specs evaluated_models_write_qs.json 75 | 76 | # run your local model and complete each of the prompts in the output JSON 77 | # create a new JSON by adding the completions as a "completion" field in the prompts JSON 78 | 79 | # run completions 80 | python3 -m evaluate --model_specs evaluated_models.json # this step is not fully implemented in the code 81 | 82 | # compute accuracies for static/functional (or run with just --stat_fn to see all available stats) 83 | python3 -m summarize_evals --stat_fn stat_accuracy 84 | ``` 85 | 86 | # FAQs 87 | 88 | 1. Why do the accuracy numbers differ from the best reported for the models? 89 | We have not optimized for getting the best reported accuracy for each model, 90 | and we likely will not do that. The best reported might need undisclosed 91 | prompting, or post-processing. 92 | We conjecture that the overall conclusions about reasoning gap presence 93 | will not change. 94 | 95 | 1. Open question: pass@k and maj@k increase topline performance, but do they 96 | increase or decrease the reasoning gap? 97 | Work-in-progress. Requires 100% MATH coverage. 98 | 99 | 1. Why is the GPT4 number not 78.2% as reported in the [Let's why step by 100 | step](https://arxiv.org/abs/2305.20050) paper? 101 | As mentioned in their repository section [MATH 102 | Splits](https://github.com/openai/prm800k#math-splits) and in the paper, their 103 | evaluation is non-standard: "In order to avoid the risk of over-fitting on the 104 | 7,500 MATH training problems, we expanded the training set to include 4,500 105 | MATH test split problems. We therefore evaluate our models only on the 106 | remaining 500 held-out problems. We selected these 500 test problems uniformly 107 | at random, and we believe they are representative of the test set as a whole." 108 | For the 500 representative problems they pick PRM, ORM, and Majority Voting get 109 | 78.2%, 72.4% and 69.6% when using best-of-1860 (Figure 3). 110 | 111 | # Known Issues 112 | * More liberal output math equivalence matching. E.g., `C(21,7)` instead of `116280` 113 | (Claude non-CoT); `9 choose 2 = 36` instead of `36` (Mixtral non-CoT). 114 | * Functional instantiations can sometimes result in problems or answers that 115 | are too long. E.g., in one of the instantiations in the `Dec-2023` snapshot a 116 | benchmark's output number has more than 40 digits. 117 | 118 | # Citation 119 | 120 | @misc{srivastava2024functional, 121 | title={Functional Benchmarks for Robust Evaluation of Reasoning Performance, and the Reasoning Gap}, 122 | author={Saurabh Srivastava and Annarose M B and Anto P V au2 and Shashank Menon and Ajay Sukumar and Adwaith Samod T and Alan Philipose and Stevin Prince and Sooraj Thomas}, 123 | year={2024}, 124 | eprint={2402.19450}, 125 | archivePrefix={arXiv}, 126 | primaryClass={cs.AI} 127 | } 128 | -------------------------------------------------------------------------------- /api_query_cache/Anthropic_Eval_claude-3-opus-20240229_temp=0.0_cot=False_fs=5.tar.gz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/api_query_cache/Anthropic_Eval_claude-3-opus-20240229_temp=0.0_cot=False_fs=5.tar.gz -------------------------------------------------------------------------------- /api_query_cache/Anthropic_Eval_claude-3-opus-20240229_temp=0.0_cot=True_fs=0.tar.gz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/api_query_cache/Anthropic_Eval_claude-3-opus-20240229_temp=0.0_cot=True_fs=0.tar.gz -------------------------------------------------------------------------------- /api_query_cache/Anthropic_Eval_claude-3-opus-20240229_temp=0.0_cot=True_fs=5.tar.gz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/api_query_cache/Anthropic_Eval_claude-3-opus-20240229_temp=0.0_cot=True_fs=5.tar.gz -------------------------------------------------------------------------------- /api_query_cache/Mistral_Eval_mistral-large-2402_temp=0.0_cot=False_fs=5.tar.gz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/api_query_cache/Mistral_Eval_mistral-large-2402_temp=0.0_cot=False_fs=5.tar.gz -------------------------------------------------------------------------------- /api_query_cache/Mistral_Eval_mistral-large-2402_temp=0.0_cot=True_fs=0.tar.gz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/api_query_cache/Mistral_Eval_mistral-large-2402_temp=0.0_cot=True_fs=0.tar.gz -------------------------------------------------------------------------------- /api_query_cache/Mistral_Eval_mistral-large-2402_temp=0.0_cot=True_fs=5.tar.gz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/api_query_cache/Mistral_Eval_mistral-large-2402_temp=0.0_cot=True_fs=5.tar.gz -------------------------------------------------------------------------------- /api_query_cache/Mistral_Eval_mistral-medium-2312_temp=0.0_cot=False_fs=5.tar.gz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/api_query_cache/Mistral_Eval_mistral-medium-2312_temp=0.0_cot=False_fs=5.tar.gz -------------------------------------------------------------------------------- /api_query_cache/Mistral_Eval_mistral-medium-2312_temp=0.0_cot=True_fs=0.tar.gz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/api_query_cache/Mistral_Eval_mistral-medium-2312_temp=0.0_cot=True_fs=0.tar.gz -------------------------------------------------------------------------------- /api_query_cache/Mistral_Eval_mistral-medium-2312_temp=0.0_cot=True_fs=5.tar.gz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/api_query_cache/Mistral_Eval_mistral-medium-2312_temp=0.0_cot=True_fs=5.tar.gz -------------------------------------------------------------------------------- /api_query_cache/OAI_Eval_gpt-3.5-turbo_temp=0.0_cot=False_fs=5.tar.gz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/api_query_cache/OAI_Eval_gpt-3.5-turbo_temp=0.0_cot=False_fs=5.tar.gz -------------------------------------------------------------------------------- /api_query_cache/OAI_Eval_gpt-3.5-turbo_temp=0.0_cot=True_fs=0.tar.gz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/api_query_cache/OAI_Eval_gpt-3.5-turbo_temp=0.0_cot=True_fs=0.tar.gz -------------------------------------------------------------------------------- /api_query_cache/OAI_Eval_gpt-3.5-turbo_temp=0.0_cot=True_fs=5.tar.gz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/api_query_cache/OAI_Eval_gpt-3.5-turbo_temp=0.0_cot=True_fs=5.tar.gz -------------------------------------------------------------------------------- /api_query_cache/OAI_Eval_gpt-4_temp=0.0_cot=False_fs=5.tar.gz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/api_query_cache/OAI_Eval_gpt-4_temp=0.0_cot=False_fs=5.tar.gz -------------------------------------------------------------------------------- /api_query_cache/OAI_Eval_gpt-4_temp=0.0_cot=True_fs=0.tar.gz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/api_query_cache/OAI_Eval_gpt-4_temp=0.0_cot=True_fs=0.tar.gz -------------------------------------------------------------------------------- /api_query_cache/OAI_Eval_gpt-4_temp=0.0_cot=True_fs=5.tar.gz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/api_query_cache/OAI_Eval_gpt-4_temp=0.0_cot=True_fs=5.tar.gz -------------------------------------------------------------------------------- /api_query_cache/Together_Eval_Open-Orca_Mistral-7B-OpenOrca_temp=0.0_cot=False_fs=5.tar.gz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/api_query_cache/Together_Eval_Open-Orca_Mistral-7B-OpenOrca_temp=0.0_cot=False_fs=5.tar.gz -------------------------------------------------------------------------------- /api_query_cache/Together_Eval_Open-Orca_Mistral-7B-OpenOrca_temp=0.0_cot=True_fs=0.tar.gz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/api_query_cache/Together_Eval_Open-Orca_Mistral-7B-OpenOrca_temp=0.0_cot=True_fs=0.tar.gz -------------------------------------------------------------------------------- /api_query_cache/Together_Eval_Open-Orca_Mistral-7B-OpenOrca_temp=0.0_cot=True_fs=5.tar.gz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/api_query_cache/Together_Eval_Open-Orca_Mistral-7B-OpenOrca_temp=0.0_cot=True_fs=5.tar.gz -------------------------------------------------------------------------------- /api_query_cache/Together_Eval_Qwen_Qwen1.5-72B_temp=0.0_cot=False_fs=5.tar.gz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/api_query_cache/Together_Eval_Qwen_Qwen1.5-72B_temp=0.0_cot=False_fs=5.tar.gz -------------------------------------------------------------------------------- /api_query_cache/Together_Eval_Qwen_Qwen1.5-72B_temp=0.0_cot=True_fs=0.tar.gz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/api_query_cache/Together_Eval_Qwen_Qwen1.5-72B_temp=0.0_cot=True_fs=0.tar.gz -------------------------------------------------------------------------------- /api_query_cache/Together_Eval_Qwen_Qwen1.5-72B_temp=0.0_cot=True_fs=5.tar.gz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/api_query_cache/Together_Eval_Qwen_Qwen1.5-72B_temp=0.0_cot=True_fs=5.tar.gz -------------------------------------------------------------------------------- /api_query_cache/Together_Eval_allenai_OLMo-7B-Instruct_temp=0.0_cot=False_fs=5.tar.gz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/api_query_cache/Together_Eval_allenai_OLMo-7B-Instruct_temp=0.0_cot=False_fs=5.tar.gz -------------------------------------------------------------------------------- /api_query_cache/Together_Eval_allenai_OLMo-7B-Instruct_temp=0.0_cot=True_fs=0.tar.gz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/api_query_cache/Together_Eval_allenai_OLMo-7B-Instruct_temp=0.0_cot=True_fs=0.tar.gz -------------------------------------------------------------------------------- /api_query_cache/Together_Eval_allenai_OLMo-7B-Instruct_temp=0.0_cot=True_fs=5.tar.gz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/api_query_cache/Together_Eval_allenai_OLMo-7B-Instruct_temp=0.0_cot=True_fs=5.tar.gz -------------------------------------------------------------------------------- /api_query_cache/Together_Eval_google_gemma-7b_temp=0.0_cot=False_fs=5.tar.gz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/api_query_cache/Together_Eval_google_gemma-7b_temp=0.0_cot=False_fs=5.tar.gz -------------------------------------------------------------------------------- /api_query_cache/Together_Eval_google_gemma-7b_temp=0.0_cot=True_fs=0.tar.gz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/api_query_cache/Together_Eval_google_gemma-7b_temp=0.0_cot=True_fs=0.tar.gz -------------------------------------------------------------------------------- /api_query_cache/Together_Eval_google_gemma-7b_temp=0.0_cot=True_fs=5.tar.gz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/api_query_cache/Together_Eval_google_gemma-7b_temp=0.0_cot=True_fs=5.tar.gz -------------------------------------------------------------------------------- /api_query_cache/Together_Eval_meta-llama_Llama-2-70b-hf_temp=0.0_cot=False_fs=5.tar.gz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/api_query_cache/Together_Eval_meta-llama_Llama-2-70b-hf_temp=0.0_cot=False_fs=5.tar.gz -------------------------------------------------------------------------------- /api_query_cache/Together_Eval_meta-llama_Llama-2-70b-hf_temp=0.0_cot=True_fs=0.tar.gz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/api_query_cache/Together_Eval_meta-llama_Llama-2-70b-hf_temp=0.0_cot=True_fs=0.tar.gz -------------------------------------------------------------------------------- /api_query_cache/Together_Eval_meta-llama_Llama-2-70b-hf_temp=0.0_cot=True_fs=5.tar.gz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/api_query_cache/Together_Eval_meta-llama_Llama-2-70b-hf_temp=0.0_cot=True_fs=5.tar.gz -------------------------------------------------------------------------------- /api_query_cache/Together_Eval_microsoft_phi-2_temp=0.0_cot=False_fs=5.tar.gz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/api_query_cache/Together_Eval_microsoft_phi-2_temp=0.0_cot=False_fs=5.tar.gz -------------------------------------------------------------------------------- /api_query_cache/Together_Eval_microsoft_phi-2_temp=0.0_cot=True_fs=0.tar.gz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/api_query_cache/Together_Eval_microsoft_phi-2_temp=0.0_cot=True_fs=0.tar.gz -------------------------------------------------------------------------------- /api_query_cache/Together_Eval_microsoft_phi-2_temp=0.0_cot=True_fs=5.tar.gz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/api_query_cache/Together_Eval_microsoft_phi-2_temp=0.0_cot=True_fs=5.tar.gz -------------------------------------------------------------------------------- /api_query_cache/Together_Eval_togethercomputer_StripedHyena-Nous-7B_temp=0.0_cot=False_fs=5.tar.gz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/api_query_cache/Together_Eval_togethercomputer_StripedHyena-Nous-7B_temp=0.0_cot=False_fs=5.tar.gz -------------------------------------------------------------------------------- /api_query_cache/Together_Eval_togethercomputer_StripedHyena-Nous-7B_temp=0.0_cot=True_fs=0.tar.gz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/api_query_cache/Together_Eval_togethercomputer_StripedHyena-Nous-7B_temp=0.0_cot=True_fs=0.tar.gz -------------------------------------------------------------------------------- /api_query_cache/Together_Eval_togethercomputer_StripedHyena-Nous-7B_temp=0.0_cot=True_fs=5.tar.gz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/api_query_cache/Together_Eval_togethercomputer_StripedHyena-Nous-7B_temp=0.0_cot=True_fs=5.tar.gz -------------------------------------------------------------------------------- /api_query_cache/Together_Eval_zero-one-ai_Yi-34B_temp=0.0_cot=False_fs=5.tar.gz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/api_query_cache/Together_Eval_zero-one-ai_Yi-34B_temp=0.0_cot=False_fs=5.tar.gz -------------------------------------------------------------------------------- /api_query_cache/Together_Eval_zero-one-ai_Yi-34B_temp=0.0_cot=True_fs=0.tar.gz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/api_query_cache/Together_Eval_zero-one-ai_Yi-34B_temp=0.0_cot=True_fs=0.tar.gz -------------------------------------------------------------------------------- /api_query_cache/Together_Eval_zero-one-ai_Yi-34B_temp=0.0_cot=True_fs=5.tar.gz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/api_query_cache/Together_Eval_zero-one-ai_Yi-34B_temp=0.0_cot=True_fs=5.tar.gz -------------------------------------------------------------------------------- /chain_of_thought.py: -------------------------------------------------------------------------------- 1 | # Ref: Chain-of-Thought Prompting Elicits Reasoning in Large Language Models https://arxiv.org/abs/2201.11903 2 | # TODO, v2 implement option to use different prompting: 3 | # Ref: Lewkowycz’ 2022 https://arxiv.org/pdf/2206.14858.pdf 4 | # Ref: Section 9.7.6 of the Gemini 1.5 whitepaper: https://arxiv.org/pdf/2403.05530.pdf 5 | 6 | from typing import Dict, List, Tuple 7 | from math_utils.math_helpers import rm_latex_math 8 | from few_shot import CoTFewShotAnswerSamples, FewShotBuilder 9 | from helper_utils import NO_SOLUTION_PREFIX 10 | 11 | PREAMBLE = "Answer the MATH query below. " \ 12 | f"If the answer cannot be computed, or you are not confident, say {NO_SOLUTION_PREFIX}. " 13 | COT_PROMPT = "Think step by step. " 14 | ANSWER_TAG = "The answer is: " 15 | END_WITH_ANSWER = f"End the answer with \"{ANSWER_TAG} \" followed by only the answer you compute and no extra words.\n" 16 | 17 | COT_INSTRUCTION = PREAMBLE + COT_PROMPT + END_WITH_ANSWER 18 | 19 | class ChainOfThought(FewShotBuilder): 20 | def __init__(self): 21 | mk_io = lambda fs: (fs.prb, fs.sol + '\n' + ANSWER_TAG + rm_latex_math(fs.outcome)) 22 | self.few_shot_io: List[Tuple[str, str]] = [ mk_io(fs) for fs in CoTFewShotAnswerSamples ] 23 | 24 | def few_shot_limited(self, count: int) -> Dict[str, str]: 25 | limited_samples = self.few_shot_io[:count] if count != -1 else self.few_shot_io 26 | return dict(limited_samples) 27 | 28 | def extract_answer(self, answer): 29 | # locate the first ANSWER_TAG 30 | ans_index = answer.find(ANSWER_TAG) 31 | extracted = answer[ans_index + len(ANSWER_TAG):] if ans_index != -1 else '' 32 | 33 | # expecting only a single line in the answer 34 | empty_line = extracted.find('\n') 35 | extracted = extracted[:empty_line] if empty_line != -1 else extracted 36 | 37 | # remove any whitespaces, and any ending periods 38 | extracted = extracted.strip() 39 | extracted = extracted[:-1] if extracted.endswith('.') else extracted 40 | 41 | # Most of time this processing works: 42 | # The answer is: $68/125$. 43 | # The answer is: $\frac{\pi}{4}$ 44 | # The answer is: 165. 45 | # The answer is: 6 46 | 47 | # But then there are other models insists on outputting english; or other extraneous tokens: 48 | # The answer is: $n=5$. 49 | # The answer is: The probability that Phil and Sarah get the same number is $\frac{3}{10}$ or 0.3 or 30%. 50 | # The answer is: There are 28 ways for Mary to put the 6 identical basil plants on the 3 window sills 51 | # The answer is: 3/8 or approximately 0.375. 52 | # The answer is: 18/343 53 | # The answer is: The probability is approximately 0.2616 or about 26.16%. 54 | # The answer is: 1 slice has both pepperoni and mushrooms. 55 | # The answer is: 110 students take physics. 56 | 57 | verbose = False 58 | console_confirm = False 59 | 60 | if console_confirm: 61 | print(f'\n---------------\n') 62 | print(f'From:{answer}') 63 | print(f'\n---------------\n') 64 | print(f'Extracted: {extracted}') 65 | print(f'\n---------------\n') 66 | input('Continue? ') 67 | if verbose: 68 | if ' ' in extracted: 69 | print(f'Extracted: {extracted}') 70 | if not extracted: 71 | print(f'Missing extraction; no answer tag; likely exceeded token length.') 72 | 73 | return extracted 74 | -------------------------------------------------------------------------------- /eval/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/eval/__init__.py -------------------------------------------------------------------------------- /eval/measure_math.py: -------------------------------------------------------------------------------- 1 | from __future__ import annotations 2 | import os 3 | import csv 4 | import json 5 | from math_utils.math_helpers import clean_numbers, last_boxed_only # these are unused; why? 6 | from math_utils.math_helpers import last_boxed_only_string, remove_boxed 7 | from math_utils.math_equivalence import is_equiv 8 | from helper_utils import NO_SOLUTION_PREFIX 9 | import numpy as np 10 | from typing import Dict, Any, List, Tuple, Optional 11 | from pathlib import Path 12 | from fn_snapshot import FnSnapshot, root_test 13 | from tqdm import tqdm 14 | 15 | INTERMEDIATES_DIR = "intermediate" 16 | PREFIX_FN_SOLVED = os.path.join(INTERMEDIATES_DIR, "functionals_solved_") 17 | PREFIX_ANSWERS = os.path.join(INTERMEDIATES_DIR, "answers_") 18 | PREFIX_CORRECT = os.path.join(INTERMEDIATES_DIR, "correct_") 19 | 20 | if not os.path.exists(INTERMEDIATES_DIR): 21 | os.makedirs(INTERMEDIATES_DIR) 22 | 23 | # these should match functional-math/functional/regression/instantiate.py 24 | FN_SOLVED_STATIC_UNSOLVED = 'FN_SOLVED_STATIC_UNSOLVED' 25 | FN_SOLVED_STATIC_SOLVED = 'FN_SOLVED_STATIC_SOLVED' 26 | 27 | DELIM = "#" * 80 + "\n" 28 | SUBJECTS = [ 29 | 'Prealgebra', 30 | 'Algebra', 31 | 'Number Theory', 32 | 'Counting & Probability', 33 | 'Geometry', 34 | 'Intermediate Algebra', 35 | 'Precalculus' 36 | ] 37 | 38 | 39 | class TestMetric: 40 | def __init__(self, correct, frac_correct, output, completion, answer, typ, level, is_static, fname): 41 | # evaluation output and result 42 | self.correct = correct 43 | self.fraction_correct = frac_correct 44 | self.output = output 45 | self.answer = answer 46 | self.raw_completion = completion 47 | 48 | # identification data 49 | self.type = typ 50 | self.level = level 51 | self.is_static = is_static 52 | self.fname = os.path.join(*Path(fname).parts[-2:]) # `subject/id.json` 53 | 54 | def __repr__(self): 55 | s = f"CORRECT: {self.correct} | TYPE: {self.type} | LEVEL: {self.level} | OUTPUT: {self.output} | ANSWER: {self.answer} | STATIC: {self.is_static} | FNAME: {self.fname}" 56 | return s 57 | 58 | def __repr_concise__(self): 59 | s = f'{self.correct} | {self.output} | GT: {self.answer} | STATIC: {self.is_static} | {self.fname}' 60 | return s 61 | 62 | @classmethod 63 | def consensus(cls, many: List[TestMetric]) -> TestMetric: 64 | first = many[0] 65 | agree = lambda field: all(field(first) == field(m) for m in many) 66 | assert agree(lambda x: x.type), f'Differs on `type`: {many}' 67 | assert agree(lambda x: x.level), f'Differs on `level`: {many}' 68 | assert agree(lambda x: x.is_static), f'Differs on `is_static`: {many}' 69 | assert agree(lambda x: x.fname), f'Differs on `fname`: {many}' 70 | 71 | t, l, s, f = first.type, first.level, first.is_static, first.fname 72 | 73 | # if all correct then doesn't matter, we'll default to the first of the many 74 | assign = lambda x: (x.output, x.raw_completion, x.answer) 75 | valid, o, c, a = True, *assign(first) 76 | for m in many: 77 | if not m.correct: 78 | valid, o, c, a = False, *assign(m) 79 | break 80 | frac_valid = [m.correct for m in many].count(True) / float(len(many)) 81 | return TestMetric(valid, frac_valid, o, c, a, t, l, s, f) 82 | 83 | 84 | class Metrics: 85 | def __init__(self): 86 | self.individuals: List[TestMetric] = [] 87 | self.name = None 88 | 89 | self.cors: Dict[Any, Any] = {} 90 | self.subject_cors: Dict[Any, Any] = {} 91 | self.level_cors: Dict[Any, Any] = {} 92 | self.correct = 0 93 | self.total = 0 94 | self.total_incorrect = 0 95 | self.hallucinated = 0 96 | 97 | def update_with(self, rslt: TestMetric): 98 | self.individuals.append(rslt) 99 | equiv = rslt.correct 100 | prob_level = rslt.level 101 | prob_type = rslt.type 102 | 103 | # update subject and level corrects 104 | if (prob_level, prob_type) in self.cors: 105 | self.cors[(prob_level, prob_type)].append(equiv) 106 | else: 107 | self.cors[(prob_level, prob_type)] = [equiv] 108 | if prob_level in self.level_cors: 109 | self.level_cors[prob_level].append(equiv) 110 | else: 111 | if prob_level is not None: 112 | self.level_cors[prob_level] = [equiv] 113 | if prob_type in self.subject_cors: 114 | self.subject_cors[prob_type].append(equiv) 115 | else: 116 | if prob_type is not None: 117 | self.subject_cors[prob_type] = [equiv] 118 | 119 | # update correct/total counts 120 | if equiv: 121 | self.correct += 1 122 | self.total += 1 123 | 124 | # update hallucinations count 125 | if not equiv: 126 | lowered_answer = rslt.output.lower().strip() 127 | no_soln_prefix = NO_SOLUTION_PREFIX.lower() 128 | says_dont_know = lowered_answer.startswith(no_soln_prefix) 129 | if not says_dont_know: 130 | self.hallucinated += 1 131 | self.total_incorrect += 1 132 | 133 | def accuracy(self): 134 | return self.correct, self.total 135 | 136 | def hallucinations(self): 137 | return self.hallucinated, self.total_incorrect 138 | 139 | def stats_legend(self): 140 | return "[A]ccuracy, [H]allucination" 141 | 142 | def stats(self): 143 | def stat_str(num: int, dim: int) -> str: 144 | pc = 100.0 * (num / float(dim) if dim else 0.0) 145 | return f'{num:04d}/{dim:04d} = {pc:5.2f}%' 146 | 147 | return f"A: {stat_str(*self.accuracy())}"\ 148 | f" | H: {stat_str(*self.hallucinations())}" 149 | 150 | def save_and_log(self, tested_functionally: Dict[TestMetric, TestMetric], model_name: str, verbose): 151 | console = [] 152 | filelog = [] 153 | 154 | for s, f in tested_functionally.items(): 155 | # default - functional unsolved; static solved; or both unsolved 156 | log: Optional[str] = None 157 | row: Optional[Any] = None 158 | if f.correct: 159 | if s.correct: 160 | log = f'[INFO] TOO SIMPLE? - functional solved: {s.fname} at Level {s.level}' 161 | row = [s.fname, FN_SOLVED_STATIC_SOLVED, s.level] 162 | else: 163 | log = f'[WARN] UNEXPECTED - functional solved; static unsolved: {s.fname}' 164 | row = [s.fname, FN_SOLVED_STATIC_UNSOLVED, s.level] 165 | console.append(log) 166 | filelog.append(row) 167 | 168 | solved_fns_file = f'{PREFIX_FN_SOLVED}{model_name}.csv' 169 | with open(solved_fns_file, 'w') as cf: 170 | csvwriter = csv.writer(cf) 171 | for l in filelog: 172 | if l is not None: 173 | csvwriter.writerow(l) 174 | 175 | if verbose: 176 | console_lines = "\n".join(c for c in console if c is not None) 177 | print(f'{console_lines}') 178 | print(f'written solved functionals to {solved_fns_file}') 179 | 180 | # print stats 181 | cases = [case for _, case, _ in [row for row in filelog if row]] 182 | s_f_correct = cases.count(FN_SOLVED_STATIC_SOLVED) 183 | s_incorrect_f_correct = cases.count(FN_SOLVED_STATIC_UNSOLVED) 184 | print(f'[STATS] tested functionally: {len(tested_functionally)}, correct f+s: {s_f_correct} (counts towards accuracy), correct f+!s: {s_incorrect_f_correct} (does not count)') 185 | 186 | def get_overriden(self, func: Metrics, verbose: bool) -> Dict[TestMetric, Optional[TestMetric]]: 187 | if verbose: 188 | statics = set(f.fname for f in func.individuals if f.is_static) 189 | print(f'[STATS] statics in the functional set: {len(statics)} (if 0 then snapshot might have been taken without include_static = True)') 190 | 191 | matching = lambda ident: [f for f in func.individuals if f.fname == ident and not f.is_static] 192 | overriden = {} 193 | for s in self.individuals: 194 | matches = matching(s.fname) 195 | if len(matches) > 1: 196 | print(f'More than one functional for {s.fname}. Picking first') 197 | overriden[s] = matches[0] if matches else None 198 | return overriden 199 | 200 | def with_functional(self, func: Metrics, model_name: str, verbose: bool) -> Tuple[Metrics, Metrics]: 201 | overriden: Dict[TestMetric, Optional[TestMetric]] = self.get_overriden(func, verbose) 202 | with_fn = Metrics() 203 | of_static_correct_fn_correct = Metrics() 204 | tested_functionally: Dict[TestMetric, TestMetric] = {} 205 | total_correct, functionalized_correct = 0, 0 206 | f_none_count_s_correct = 0 207 | for s, f in overriden.items(): 208 | 209 | if f is None: 210 | # update accumulator: only static available, so keep that 211 | with_fn.update_with(s) 212 | if s.correct: 213 | f_none_count_s_correct += 1 214 | 215 | else: 216 | # update accumulator: both s & f available 217 | # - case: both correct or both false, storing s/f is identical 218 | # - case: one of them is false, keep the false one; since we want to be strict 219 | # summary: if s is incorrect store that, otherwise store whatever f's outcome was 220 | if not s.correct: 221 | with_fn.update_with(s) 222 | else: 223 | with_fn.update_with(f) 224 | 225 | # s is correct, and f is not None 226 | # store this as the functional test for when static correct 227 | of_static_correct_fn_correct.update_with(f) 228 | 229 | # for ones overriden, put them in tested_functionally, so we can log to console and file 230 | tested_functionally[s] = f 231 | 232 | if verbose: 233 | print(f'[STATS] statics correct, for which no functional: {f_none_count_s_correct} (counts towards accuracy)') 234 | 235 | with_fn.name = f'{self.name}\\{func.name}' 236 | self.save_and_log(tested_functionally, model_name, verbose) 237 | 238 | return with_fn, of_static_correct_fn_correct 239 | 240 | @classmethod 241 | def consensus(cls, many: List[Metrics]) -> Metrics: 242 | find = lambda fname, m: next(tm for tm in m.individuals if tm.fname == fname) 243 | idents = lambda i: set(tm.fname for tm in many[i].individuals) 244 | ref_idents = idents(0) 245 | assert all(ref_idents == idents(i) for i in range(1, len(many))), f'idents differ' 246 | by_name: Dict[str, List[TestMetric]] = { f: [find(f, m) for m in many] for f in ref_idents } 247 | cs: Dict[str, TestMetric] = { f: TestMetric.consensus(ltm) for f, ltm in by_name.items() } 248 | 249 | aggr = Metrics() 250 | aggr.name = "+".join([f'{m.name}' for m in many]) 251 | for c in cs.values(): 252 | aggr.update_with(c) 253 | return aggr 254 | 255 | class Eval: 256 | def __init__(self, solve_fn, fn_snapshots, model_name, verbose = False): 257 | self.verbose = verbose 258 | self.problems_tested = [] 259 | self.static_metrics, self.func_metrics = self.get_metrics(solve_fn, fn_snapshots, model_name) 260 | 261 | # report combined metrics 262 | combined, _ = self.static_metrics.with_functional(self.func_metrics, model_name, verbose) 263 | self.report_stats(combined) 264 | 265 | if not os.path.exists(INTERMEDIATES_DIR): 266 | os.makedirs(INTERMEDIATES_DIR) 267 | 268 | def get_metrics(self, solve_fn, fn_snapshots, model_name) -> Tuple[Metrics, Metrics]: 269 | static_metrics = None 270 | multiple_func_metrics = [] 271 | for snapshot in fn_snapshots: 272 | # download and setup directories `ROOT/test/subject/id.json` 273 | root = snapshot.ensure_dataset() 274 | 275 | # location 276 | test_dir = root_test(root) 277 | tag = snapshot.date 278 | out_answers_to = f'{PREFIX_ANSWERS}{model_name}-{tag}' 279 | correct_dir = f'{PREFIX_CORRECT}{model_name}-{tag}' 280 | 281 | # eval over snapshot 282 | metrics = self.run_eval(tag, test_dir, solve_fn, out_answers_to) 283 | 284 | # log to data dirs 285 | self.write_correct_completions(metrics, correct_dir) 286 | 287 | # keep static aside, and put all functionals into list 288 | if snapshot.is_static(): 289 | static_metrics = metrics 290 | else: 291 | multiple_func_metrics.append(metrics) 292 | 293 | func_metrics = Metrics.consensus(multiple_func_metrics) 294 | assert static_metrics, f'No static metrics!' 295 | return static_metrics, func_metrics 296 | 297 | def get_solved(self, only_static: bool = False) -> List[str]: 298 | names = lambda mm: set(tm.fname for tm in mm.individuals if tm.correct) 299 | s, f = names(self.static_metrics), names(self.func_metrics) 300 | return list(s if only_static else s.intersection(f)) 301 | 302 | def report_stats(self, m: Metrics): 303 | print(DELIM) 304 | print(f'Legend: {m.stats_legend()}') 305 | print(f'{m.name}: {m.stats()}') 306 | print(DELIM) 307 | 308 | def run_eval(self, name, root, solve_fn, out_answers_file): 309 | metrics = self.eval_test_files(root, solve_fn) 310 | metrics.name = name 311 | self.write_results_file(metrics, out_answers_file) 312 | return metrics 313 | 314 | def eval_test_files(self, root, solve_fn) -> Metrics: 315 | m = Metrics() 316 | 317 | for subdir, dirs, files in os.walk(root): 318 | enum_files = files if self.verbose else tqdm(files, desc = subdir) 319 | for file in enum_files: # type: ignore 320 | 321 | fname = os.path.join(subdir, file) 322 | with open(fname, 'r') as fp: 323 | self.eval_test_data(fname, fp, m, solve_fn) 324 | return m 325 | 326 | def compare_answer(self, ref, solns: List[Tuple[str, str]]) -> Tuple[bool, str, str]: 327 | def eq(t): 328 | try: 329 | return is_equiv(t, ref) 330 | except: 331 | return False 332 | 333 | def pick_best(many_solns: List[Tuple[bool, str, str]]) -> Tuple[bool, str, str]: 334 | return many_solns[0] 335 | 336 | equivs = [(eq(t), t, c) for t, c in solns] 337 | return pick_best(equivs) 338 | 339 | def eval_test_data(self, fname, fp, m, solve_fn) -> None: 340 | try: 341 | problem_data = json.load(fp) 342 | except Exception as e: 343 | print(f"Error loading JSON from {fname}", e) 344 | raise e 345 | prob_level = problem_data["level"] 346 | prob_type = problem_data["type"] 347 | try: 348 | prob_level = int(prob_level.split("Level ")[1]) 349 | except: 350 | prob_level = None 351 | prb = problem_data["problem"] 352 | soln = problem_data["solution"] 353 | ref_answer = remove_boxed(last_boxed_only_string(soln)) 354 | is_static = problem_data["is_static"] if "is_static" in problem_data else True 355 | 356 | # ask model to solve problem 357 | solns: List[Tuple[str, str]] = solve_fn(prb) 358 | equiv, text_answer, completion = self.compare_answer(ref_answer, solns) 359 | frac_correct = 1.0 if equiv else 0.0 360 | 361 | prb_tested = { 362 | 'problem': prb, 363 | 'reference solution': ref_answer, 364 | 'is_static': is_static, 365 | 'level': prob_level, 366 | 'type': prob_type, 367 | 'fname': fname, 368 | } 369 | self.problems_tested.append(prb_tested) 370 | rslt = TestMetric( 371 | correct = equiv, 372 | frac_correct = frac_correct, 373 | output = text_answer, 374 | completion = completion, 375 | answer = ref_answer, 376 | typ = prob_type, 377 | level = prob_level, 378 | is_static = is_static, 379 | fname = fname, 380 | ) 381 | m.update_with(rslt) 382 | 383 | running = m.stats() # percentages accuracy/hallucinations 384 | test_summary = rslt.__repr_concise__() 385 | if self.verbose: 386 | print(f'{running} | {test_summary}') 387 | 388 | def out(self, f, line): 389 | f.write(line + "\n") 390 | if self.verbose: 391 | print(line) 392 | 393 | def write_each_test_results(self, m, f): 394 | for k, t in enumerate(m.individuals): 395 | line = f"{k} {t}" 396 | self.out(f, line) 397 | 398 | def pc(self, cors_list): 399 | return np.sum(cors_list), len(cors_list), 100.0 * np.mean(cors_list) 400 | 401 | def write_level_accuracies(self, m, f): 402 | self.out(f, DELIM) 403 | for level in sorted(m.level_cors): 404 | if level not in m.level_cors.keys(): 405 | continue 406 | cors_list = m.level_cors[level] 407 | line = "Level {} Accuracy = {}/{} = {:5.2f}%".format(level, *self.pc(cors_list)) 408 | self.out(f, line) 409 | self.out(f, DELIM) 410 | 411 | def write_subject_accuracies(self, m, f): 412 | self.out(f, DELIM) 413 | for subject in SUBJECTS: 414 | if subject not in m.subject_cors.keys(): 415 | continue 416 | cors_list = m.subject_cors[subject] 417 | line = "{} Accuracy = {}/{} = {:5.2f}%".format(subject, *self.pc(cors_list)) 418 | self.out(f, line) 419 | self.out(f, DELIM) 420 | 421 | def write_subject_level_accuracies(self, m, f): 422 | self.out(f, DELIM) 423 | for subject in SUBJECTS: 424 | for level in range(1, 6): 425 | key = (level, subject) 426 | if key not in m.cors.keys(): 427 | continue 428 | cors_list = m.cors[key] 429 | line = "{} Level {} Accuracy = {}/{} = {:5.2f}%".format(subject, level, *self.pc(cors_list)) 430 | self.out(f, line) 431 | self.out(f, DELIM) 432 | 433 | def write_correct_completions(self, m, dirpath: str): 434 | print(f'Writing {m.name} raw+correct completions to {dirpath}') 435 | # make directory if it doesn't exist 436 | if not os.path.exists(dirpath): 437 | os.makedirs(dirpath) 438 | 439 | # for each test outcome, if not correct continue, 440 | # if correct then write raw to file in dump directory 441 | for t in m.individuals: 442 | if not t.correct: 443 | continue 444 | 445 | ident, json = os.path.splitext(os.path.basename(t.fname)) 446 | subject = os.path.split(t.fname)[0] 447 | loc = os.path.join(dirpath, subject) 448 | if not os.path.exists(loc): 449 | os.makedirs(loc) 450 | fname = os.path.join(loc, f'{ident}.txt') 451 | raw_completion = t.raw_completion 452 | with open(fname, 'w') as f: 453 | f.write(raw_completion) 454 | 455 | def write_results_file(self, m: Metrics, out_answers_file: str): 456 | with open(out_answers_file, "w+") as f: 457 | # write all the raw test results 458 | self.write_each_test_results(m, f) 459 | 460 | # write accuracies by subject/level 461 | self.write_subject_level_accuracies(m, f) 462 | self.write_level_accuracies(m, f) 463 | self.write_subject_accuracies(m, f) 464 | 465 | # write overall accuracy 466 | self.out(f, f'Overall {m.name}: {m.stats()}') 467 | 468 | -------------------------------------------------------------------------------- /eval/runner.py: -------------------------------------------------------------------------------- 1 | from __future__ import annotations 2 | from typing import List, Optional, Tuple 3 | from eval.measure_math import Eval 4 | from helper_utils import NO_SOLUTION_BAD_FORMAT 5 | from fn_snapshot import FnSnapshot 6 | 7 | HR = "\n" + '-' * 80 + "\n" 8 | BENCHMARK = "MATH" 9 | 10 | class RunEval: 11 | def __init__(self): 12 | self.model_name = None 13 | self.snapshots = None 14 | self.verbose = True 15 | 16 | def do(self): 17 | return Eval(self.solve_fn, 18 | self.snapshots, 19 | self.model_name, 20 | verbose = self.verbose) 21 | 22 | def solve_fn(self, prb: str) -> List[Tuple[str, str]]: 23 | answers: List[Tuple[str, str]] = self.answers(prb) 24 | # If no answers return indicator that we got no answer section 25 | if not answers: 26 | return [(NO_SOLUTION_BAD_FORMAT, "")] 27 | return answers 28 | 29 | def answers(self, prb: str) -> List[Tuple[str, str]]: # type: ignore [empty-body] 30 | pass 31 | 32 | mk_dir_safe = lambda x: x.replace('/', '_') 33 | 34 | class EvalRunner(RunEval): 35 | def __init__(self, answerer, snapshots_specs, verbose): 36 | print(f'Reading specs from: {snapshots_specs}') 37 | self.answers = answerer.answers # type: ignore [method-assign] 38 | self.model_name = mk_dir_safe(answerer.model.name) 39 | self.snapshots = FnSnapshot.load(BENCHMARK, snapshots_specs) 40 | self.verbose = verbose 41 | -------------------------------------------------------------------------------- /evaluate.py: -------------------------------------------------------------------------------- 1 | from loader import load_mod 2 | from eval.measure_math import Eval 3 | import json 4 | import argparse 5 | from typing import List, Dict, Any, Tuple 6 | from persist import Persist 7 | 8 | 9 | DEFAULT_MODEL_SPECS_FILE = 'evaluated_models.json' 10 | DEFAULT_SNAPSHOTS_FILE = 'monthly_snapshots.json' 11 | DEFAULT_EVAL_PICKLE_FILE = DEFAULT_MODEL_SPECS_FILE + '.pickle' 12 | 13 | DEFAULT_FEW_SHOT_NUM = 5 14 | DEFAULT_TEMPERATURE = 0.0 15 | 16 | SPEC_SCRIPT = 'script' 17 | SPEC_INCLUDE = 'include' 18 | SPEC_NAME = 'name' 19 | SPEC_COT = 'CoT' 20 | SPEC_FEW_SHOT = 'few shot' 21 | SPEC_TEMP = 'temperature' 22 | SPEC_EXTRA_PARAMS = 'extra_params' 23 | SPEC_EXTRA_PARAMS_MODE_WRITE_OFFLINE = 'write_qs' 24 | 25 | 26 | class ModelSpec: 27 | def __init__(self, spec): 28 | self.script = spec[SPEC_SCRIPT] 29 | self.model = spec[SPEC_NAME] 30 | self.chain_of_thought = spec[SPEC_COT] 31 | self.few_shot_num = spec[SPEC_FEW_SHOT] if SPEC_FEW_SHOT in spec else DEFAULT_FEW_SHOT_NUM 32 | self.temperature = spec[SPEC_TEMP] if SPEC_TEMP in spec else DEFAULT_TEMPERATURE 33 | self.extra_params = json.loads(spec[SPEC_EXTRA_PARAMS]) if SPEC_EXTRA_PARAMS in spec else {} 34 | 35 | def ident(self) -> str: 36 | cot_yn = 'yes' if self.chain_of_thought else 'no' 37 | return self.model + f'_cot={cot_yn}_fs={self.few_shot_num}_temp={self.temperature}' 38 | 39 | def __repr__(self) -> str: 40 | return self.ident() 41 | 42 | def spec_csv(self) -> Tuple[str,str]: 43 | hdr = f'model,cot,fewshot,temp' 44 | row = f'{self.model},{self.chain_of_thought},{self.few_shot_num},{self.temperature}' 45 | return hdr,row 46 | 47 | class ModelRunners: 48 | def __init__(self, spec_file: str, snapshots_specs: str, verbose: bool, save_snapshot: bool): 49 | self.verbose = verbose 50 | self.save_snapshot = save_snapshot 51 | self.snapshots_specs = snapshots_specs 52 | 53 | with open(spec_file, 'r') as f: 54 | js = json.load(f) 55 | script_names = set(spec[SPEC_SCRIPT] for spec in js if spec[SPEC_INCLUDE] ) 56 | model_specs = [ ModelSpec(spec) for spec in js if spec[SPEC_INCLUDE] ] 57 | 58 | models_for = lambda sc: [ ms for ms in model_specs if ms.script == sc ] 59 | 60 | self.models: Dict[str, List[ModelSpec]] = { script: models_for(script) for script in script_names } 61 | self.scripts: Dict[str, Any] = { script: load_mod(script) for script in script_names } 62 | 63 | def run(self, ms: ModelSpec) -> Eval: 64 | self.banner(f"Evaluating {ms.ident()}") 65 | runner = self.scripts[ms.script].Answers.run 66 | return runner(ms.model, 67 | self.snapshots_specs, 68 | ms.chain_of_thought, 69 | ms.few_shot_num, 70 | ms.temperature, 71 | self.verbose, 72 | self.save_snapshot, 73 | ms.extra_params) 74 | 75 | def run_models_for(self, script: str) -> Dict[ModelSpec, Eval]: 76 | return { ms: self.run(ms) for ms in self.models[script] } 77 | 78 | def api_script_names(self) -> List[str]: 79 | return list(self.scripts.keys()) 80 | 81 | def banner(self, msg): 82 | hr = '\n' + '*' * 80 + '\n' 83 | print(f'{hr}{msg}{hr}') 84 | 85 | 86 | class Evaluate: 87 | def __init__(self, spec_file: str, snapshots_specs: str, verbose: bool, save_snapshot: bool): 88 | self.model_runners = ModelRunners(spec_file, snapshots_specs, verbose, save_snapshot) 89 | 90 | def build_evals(self) -> Dict[ModelSpec, Eval]: 91 | api_scripts: List[str] = self.model_runners.api_script_names() 92 | evals: List[Dict[ModelSpec, Eval]] = list(map(self.rfn, api_scripts)) 93 | return { ms: e for evd in evals for ms, e in evd.items() } 94 | 95 | def rfn(self, sc) -> Dict[ModelSpec, Eval]: 96 | return self.model_runners.run_models_for(sc) 97 | 98 | 99 | if __name__ == "__main__": 100 | parser = argparse.ArgumentParser() 101 | parser.add_argument("--model_specs", type=str, default=DEFAULT_MODEL_SPECS_FILE, 102 | help = f"JSON of model specs, name + run script + enabled") 103 | parser.add_argument("--snapshots_specs", type=str, default=DEFAULT_SNAPSHOTS_FILE, 104 | help = f"JSON of static and monthly snapshots") 105 | parser.add_argument("--verbose", action='store_true', default=False, 106 | help="Outcomes of each test, one per line, on console") 107 | parser.add_argument("--save_snapshot", action='store_true', default=False, 108 | help="Save API query snapshot as .tar.gz file") 109 | args = parser.parse_args() 110 | 111 | outfile = f'{args.model_specs}.pickle' 112 | e = Evaluate(args.model_specs, args.snapshots_specs, args.verbose, args.save_snapshot) 113 | evals: Dict[ModelSpec, Eval] = e.build_evals() 114 | Persist.save(evals, outfile, force_overwrite = True) 115 | ofile_msg = f'--in_file {outfile}' if outfile != DEFAULT_EVAL_PICKLE_FILE else '' 116 | print(f'Evals written. Run python3 -m summarize_evals {ofile_msg} to see summary.') 117 | -------------------------------------------------------------------------------- /evaluated_models.json: -------------------------------------------------------------------------------- 1 | [ 2 | { "name": "c3-opus", "script": "model_api/ant.py", "include": true, "CoT": false }, 3 | { "name": "c3-opus", "script": "model_api/ant.py", "include": true, "CoT": true, "few shot": 0 }, 4 | { "name": "c3-opus", "script": "model_api/ant.py", "include": true, "CoT": true, "few shot": 5 }, 5 | 6 | { "name": "mistral-large", "script": "model_api/mis.py", "include": true, "CoT": false }, 7 | { "name": "mistral-large", "script": "model_api/mis.py", "include": true, "CoT": true, "few shot": 0 }, 8 | { "name": "mistral-large", "script": "model_api/mis.py", "include": true, "CoT": true, "few shot": 5 }, 9 | 10 | { "name": "mistral-medium", "script": "model_api/mis.py", "include": true, "CoT": false }, 11 | { "name": "mistral-medium", "script": "model_api/mis.py", "include": true, "CoT": true, "few shot": 5 }, 12 | { "name": "mistral-medium", "script": "model_api/mis.py", "include": true, "CoT": true, "few shot": 0 }, 13 | 14 | { "name": "gpt3", "script": "model_api/oai.py", "include": true, "CoT": false }, 15 | { "name": "gpt3", "script": "model_api/oai.py", "include": true, "CoT": true, "few shot": 0 }, 16 | { "name": "gpt3", "script": "model_api/oai.py", "include": true, "CoT": true, "few shot": 5 }, 17 | 18 | { "name": "gpt4", "script": "model_api/oai.py", "include": true, "CoT": false }, 19 | { "name": "gpt4", "script": "model_api/oai.py", "include": true, "CoT": true, "few shot": 0 }, 20 | { "name": "gpt4", "script": "model_api/oai.py", "include": true, "CoT": true, "few shot": 5 }, 21 | 22 | { "name": "togethercomputer/StripedHyena-Nous-7B", "script": "model_api/tog.py", "include": true, "CoT": false }, 23 | { "name": "togethercomputer/StripedHyena-Nous-7B", "script": "model_api/tog.py", "include": true, "CoT": true, "few shot": 0 }, 24 | { "name": "togethercomputer/StripedHyena-Nous-7B", "script": "model_api/tog.py", "include": true, "CoT": true, "few shot": 5 }, 25 | 26 | { "name": "Open-Orca/Mistral-7B-OpenOrca", "script": "model_api/tog.py", "include": true, "CoT": false }, 27 | { "name": "Open-Orca/Mistral-7B-OpenOrca", "script": "model_api/tog.py", "include": true, "CoT": true, "few shot": 0 }, 28 | { "name": "Open-Orca/Mistral-7B-OpenOrca", "script": "model_api/tog.py", "include": true, "CoT": true, "few shot": 5 }, 29 | 30 | { "name": "zero-one-ai/Yi-34B", "script": "model_api/tog.py", "include": true, "CoT": false }, 31 | { "name": "zero-one-ai/Yi-34B", "script": "model_api/tog.py", "include": true, "CoT": true, "few shot": 0 }, 32 | { "name": "zero-one-ai/Yi-34B", "script": "model_api/tog.py", "include": true, "CoT": true, "few shot": 5 }, 33 | 34 | { "name": "Qwen/Qwen1.5-72B", "script": "model_api/tog.py", "include": true, "CoT": false }, 35 | { "name": "Qwen/Qwen1.5-72B", "script": "model_api/tog.py", "include": true, "CoT": true, "few shot": 0 }, 36 | { "name": "Qwen/Qwen1.5-72B", "script": "model_api/tog.py", "include": true, "CoT": true, "few shot": 5 }, 37 | 38 | { "name": "allenai/OLMo-7B-Instruct", "script": "model_api/tog.py", "include": true, "CoT": false }, 39 | { "name": "allenai/OLMo-7B-Instruct", "script": "model_api/tog.py", "include": true, "CoT": true, "few shot": 0 }, 40 | { "name": "allenai/OLMo-7B-Instruct", "script": "model_api/tog.py", "include": true, "CoT": true, "few shot": 5 }, 41 | 42 | { "name": "google/gemma-7b", "script": "model_api/tog.py", "include": true, "CoT": false }, 43 | { "name": "google/gemma-7b", "script": "model_api/tog.py", "include": true, "CoT": true, "few shot": 0 }, 44 | { "name": "google/gemma-7b", "script": "model_api/tog.py", "include": true, "CoT": true, "few shot": 5 }, 45 | 46 | { "name": "meta-llama/Llama-2-70b-hf", "script": "model_api/tog.py", "include": true, "CoT": false }, 47 | { "name": "meta-llama/Llama-2-70b-hf", "script": "model_api/tog.py", "include": true, "CoT": true, "few shot": 0 }, 48 | { "name": "meta-llama/Llama-2-70b-hf", "script": "model_api/tog.py", "include": true, "CoT": true, "few shot": 5 }, 49 | 50 | { "name": "microsoft/phi-2", "script": "model_api/tog.py", "include": true, "CoT": false }, 51 | { "name": "microsoft/phi-2", "script": "model_api/tog.py", "include": true, "CoT": true, "few shot": 0 }, 52 | { "name": "microsoft/phi-2", "script": "model_api/tog.py", "include": true, "CoT": true, "few shot": 5 } 53 | ] 54 | -------------------------------------------------------------------------------- /few_shot.py: -------------------------------------------------------------------------------- 1 | from typing import List, Dict, Optional 2 | 3 | class FewShot: 4 | def __init__(self, prb: str, outcome: str, sol: Optional[str] = None): 5 | self.prb = prb 6 | self.outcome = outcome 7 | self.sol = sol 8 | 9 | 10 | class FewShotBuilder: 11 | def few_shot_limited(self, count: int) -> Dict[str, str]: # type: ignore [empty-body] 12 | pass 13 | 14 | 15 | FewShotAnswerSamples = [ 16 | FewShot( 17 | prb = "What is $\left(\\frac{7}{8}\\right)^3 \cdot \left(\\frac{7}{8}\\right)^{-3}$?", 18 | outcome = "$1$", 19 | ), 20 | FewShot( 21 | prb = "In how many ways can 4 books be selected from a shelf of 6 books if the order in which the books are selected does not matter?", 22 | outcome = "$15$", 23 | ), 24 | FewShot( 25 | prb = "Find the distance between the points $(2,1,-4)$ and $(5,8,-3).$", 26 | outcome = "$\sqrt{59}$", 27 | ), 28 | FewShot( 29 | prb = "The faces of an octahedral die are labeled with digits $1$ through $8$. What is the probability, expressed as a common fraction, of rolling a sum of $15$ with a pair of such octahedral dice?", 30 | outcome = "$\\frac{1}{32}$", 31 | ), 32 | FewShot( 33 | prb = "The first three terms of an arithmetic sequence are 1, 10 and 19, respectively. What is the value of the 21st term?", 34 | outcome = "$181$", 35 | ), 36 | FewShot( 37 | prb = "Calculate $6 \\cdot 8\\frac{1}{3}", 38 | outcome = "$16$", 39 | ), 40 | FewShot( 41 | prb = "When the binary number $100101110010_2$ is divided by 4, what is the remainder (give your answer in base 10)?", 42 | outcome = "$2$", 43 | ), 44 | FewShot( 45 | prb = "How many zeros are at the end of the product 25 $\\times$ 240?", 46 | outcome = "$3$", 47 | ), 48 | ] 49 | 50 | CoTFewShotAnswerSamples = [ 51 | FewShot( 52 | prb = "What is $\left(\\frac{7}{8}\\right)^3 \cdot \left(\\frac{7}{8}\\right)^{-3}$?", 53 | outcome = "$1$", 54 | sol = "Let's think about this step by step. We have the left part of the product raised to the power of 3, and the right part raised to the power of -3. Raising x to a negative power is the same as the 1/x, so we have a product of x and 1/x, which equals 1.", 55 | ), 56 | FewShot( 57 | prb = "In how many ways can 4 books be selected from a shelf of 6 books if the order in which the books are selected does not matter?", 58 | outcome = "$15$", 59 | sol = "Let's think about this step by step. This is a simple combinations problem, and the choice is 6 choose 4, which equals 15.", 60 | ), 61 | FewShot( 62 | prb = "Find the distance between the points $(2,1,-4)$ and $(5,8,-3).$", 63 | outcome = "$\sqrt{59}$", 64 | sol = "Let's think about this step by step. We need to find the distance in 3D space, which is equal to the square of difference in each coordinate, summed, and then sqrt of that. In this case, that is sqrt of (3^2 + 7^2 + (-1)^2) which is sqrt of 9 + 49 + 1, which equals \sqrt{59}.", 65 | ), 66 | FewShot( 67 | prb = "The faces of an octahedral die are labeled with digits $1$ through $8$. What is the probability, expressed as a common fraction, of rolling a sum of $15$ with a pair of such octahedral dice?", 68 | outcome = "$\\frac{1}{32}$", 69 | sol = "Let's think about this step by step. 15 can be made using 7 and 8 rolled. There are 8 times 8 ways of rolling the dies of which only two end up with a 7 and 8 on the face. So the probability is 2/64 which is \frac{1}{32}.", 70 | ), 71 | FewShot( 72 | prb = "The first three terms of an arithmetic sequence are 1, 10 and 19, respectively. What is the value of the 21st term?", 73 | outcome = "$181$", 74 | sol = "Let's think about this step by step. This is an arithmetic sequence with difference 9, so its nth term is given by 1+9(n-1), so its 21st terms is 1+9x20 = 181.", 75 | ), 76 | FewShot( 77 | prb = "Calculate $6 \\cdot 8\\frac{1}{3}", 78 | outcome = "$16$", 79 | sol = "Let's think about this step by step. This equals 6 times 8 divided by 3, which is 2 times 8, or 16.", 80 | ), 81 | FewShot( 82 | prb = "When the binary number $100101110010_2$ is divided by 4, what is the remainder (give your answer in base 10)?", 83 | outcome = "$2$", 84 | sol = "Let's think about this step by step. None of the bits after the 3rd matter in this case, so the binary remainder is 10_2, which is 2 in decimal.", 85 | ), 86 | FewShot( 87 | prb = "How many zeros are at the end of the product 25 $\\times$ 240?", 88 | outcome = "$3$", 89 | sol = "Let's think about this step by step. 25 is 100 by 4, and 4 cleanly divides 24, so two zeros from 100 and one from the other number give us 3 zeros.", 90 | ), 91 | ] 92 | -------------------------------------------------------------------------------- /fn_snapshot.py: -------------------------------------------------------------------------------- 1 | from __future__ import annotations 2 | import os 3 | from datetime import datetime 4 | from pathlib import Path 5 | from helper_utils import download_extract 6 | from typing import Dict, Any, List, Tuple, Optional 7 | import json 8 | 9 | DATE_FMT = '%b-%Y' 10 | STATIC_DATE_TAG = "Jan-1984" 11 | 12 | # MATH/test has the subject dirs within it 13 | root_test = lambda root: os.path.join(root, "test") 14 | 15 | class FnSnapshot: 16 | def __init__(self, date: str, url: str): 17 | self.date = date 18 | self.url = url 19 | 20 | def is_static(self) -> bool: 21 | return self.date == STATIC_DATE_TAG 22 | 23 | def ensure_dataset(self): 24 | # get root "MATH" from "https://people.eecs.berkeley.edu/~hendrycks/MATH.tar" 25 | # or "Dec-2023" from "Dec-2023.tar.gz" 26 | p = Path(self.url) 27 | root = p.name.removesuffix("".join(p.suffixes)) 28 | 29 | downloaded_name = os.path.basename(self.url) 30 | is_gz = downloaded_name.endswith('gz') 31 | there = lambda d: os.path.exists(d) 32 | test_dir = root_test(root) 33 | if not there(root) or not there(test_dir): 34 | download_extract(self.url, root, downloaded_name, root, is_gz) 35 | assert there(root), f'Download/extract failed to create {root} from {self.url}' 36 | assert there(test_dir), f'Download/extract from {self.url} does not have {test_dir}' 37 | return root 38 | 39 | @classmethod 40 | def load(cls, benchmark: str, config: str) -> List[FnSnapshot]: 41 | def load_snaps(fixed_url, fns): 42 | snaps = [FnSnapshot(fn['date'], fn['url']) for fn in fns] 43 | snaps.append(FnSnapshot(STATIC_DATE_TAG, fixed_url)) 44 | snaps.sort(key = lambda x: datetime.strptime(x.date, DATE_FMT)) 45 | return snaps 46 | 47 | with open(config, 'r') as cf: 48 | benchmarks = json.load(cf) 49 | benchmark_meta = None 50 | for meta in benchmarks: 51 | if meta['benchmark'] == benchmark: 52 | benchmark_meta = meta 53 | assert benchmark_meta, f'Did not find {benchmark} in {benchmarks}' 54 | return load_snaps(benchmark_meta['fixed'], benchmark_meta['functionals']) 55 | 56 | -------------------------------------------------------------------------------- /helper_utils.py: -------------------------------------------------------------------------------- 1 | import os 2 | import subprocess 3 | from typing import Tuple, Optional, List 4 | 5 | NO_SOLUTION_PREFIX = "NO_SOLUTION" 6 | NO_SOLUTION_BAD_FORMAT = f"{NO_SOLUTION_PREFIX} [BAD FORMAT]" 7 | NO_SOLUTION_DIMS = f"{NO_SOLUTION_PREFIX} [DIMS]" 8 | NO_SOLUTION_EXCEED_CONTEXT = f"{NO_SOLUTION_PREFIX} [EXCEED CONTEXT]" 9 | 10 | def download_extract(url: str, unzipped_dir: str, downloaded_targz: str, root: str, gz: bool = True): 11 | assert not os.path.exists(unzipped_dir), f'Cannot download extract to {unzipped_dir}. Directory already exists' 12 | tar_flags = '-zxvf' if gz else '-xvf' 13 | get = ['curl', url, '-o', downloaded_targz] if url.startswith("http") else ['cp', url, downloaded_targz] 14 | unzip = ['tar', tar_flags, downloaded_targz] 15 | mv = ['mv', unzipped_dir, f'{root}'] 16 | for cmd in [get, unzip, mv]: 17 | print(f'Running: {cmd}') 18 | subprocess.call(cmd) 19 | 20 | def targz(dirname: str): 21 | tar_name = dirname + ".tar" 22 | targz_name = tar_name + ".gz" 23 | assert not os.path.exists(tar_name), f'Cannot tar to {tar_name}. File already exists' 24 | assert not os.path.exists(targz_name), f'Cannot gz to {targz_name}. File already exists' 25 | tar = ['tar', '-cvf', tar_name, dirname] 26 | gzip = ['gzip', tar_name] 27 | subprocess.call(tar) 28 | subprocess.call(gzip) 29 | 30 | def untargz(targz_name: str, dst: str): 31 | ext = ".tar.gz" 32 | assert targz_name.endswith(ext) 33 | extract_name = targz_name[:-len(ext)] 34 | assert not os.path.exists(extract_name), f'Cannot extract to {extract_name}. Directory already exists' 35 | tar_flags = '-zxvf' 36 | unzip = ['tar', tar_flags, targz_name] 37 | subprocess.call(unzip) 38 | 39 | if extract_name != dst: 40 | print(f'Moving {extract_name} to {dst}') 41 | assert not os.path.exists(dst), f'Cannot mv to {dst}. Directory already exists' 42 | mv = ['mv', extract_name, dst] 43 | subprocess.call(mv) 44 | -------------------------------------------------------------------------------- /loader.py: -------------------------------------------------------------------------------- 1 | import sys 2 | import importlib.util 3 | 4 | def load_mod(fname): 5 | nm = fname 6 | spec = importlib.util.spec_from_file_location(nm, fname) 7 | assert spec, f'Could not load from file: {fname}' 8 | mod = importlib.util.module_from_spec(spec) 9 | sys.modules[nm] = mod 10 | assert spec.loader, f'Spec does not have loader' 11 | spec.loader.exec_module(mod) 12 | return mod 13 | -------------------------------------------------------------------------------- /math_utils/README.md: -------------------------------------------------------------------------------- 1 | original source code from hendrycks MATH: https://github.com/hendrycks/math/ 2 | -------------------------------------------------------------------------------- /math_utils/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/math_utils/__init__.py -------------------------------------------------------------------------------- /math_utils/math_equivalence.py: -------------------------------------------------------------------------------- 1 | # This is a copy of the equivalence helper functions from the MATH repo: 2 | # https://github.com/hendrycks/math/blob/main/modeling/math_equivalence.py 3 | 4 | def _fix_fracs(string): 5 | substrs = string.split("\\frac") 6 | new_str = substrs[0] 7 | if len(substrs) > 1: 8 | substrs = substrs[1:] 9 | for substr in substrs: 10 | new_str += "\\frac" 11 | if substr[0] == "{": 12 | new_str += substr 13 | else: 14 | try: 15 | assert len(substr) >= 2 16 | except: 17 | return string 18 | a = substr[0] 19 | b = substr[1] 20 | if b != "{": 21 | if len(substr) > 2: 22 | post_substr = substr[2:] 23 | new_str += "{" + a + "}{" + b + "}" + post_substr 24 | else: 25 | new_str += "{" + a + "}{" + b + "}" 26 | else: 27 | if len(substr) > 2: 28 | post_substr = substr[2:] 29 | new_str += "{" + a + "}" + b + post_substr 30 | else: 31 | new_str += "{" + a + "}" + b 32 | string = new_str 33 | return string 34 | 35 | def _fix_a_slash_b(string): 36 | if len(string.split("/")) != 2: 37 | return string 38 | a = string.split("/")[0] 39 | b = string.split("/")[1] 40 | try: 41 | a = int(a) 42 | b = int(b) 43 | assert string == "{}/{}".format(a, b) 44 | new_string = "\\frac{" + str(a) + "}{" + str(b) + "}" 45 | return new_string 46 | except: 47 | return string 48 | 49 | def _remove_right_units(string): 50 | # "\\text{ " only ever occurs (at least in the val set) when describing units 51 | if "\\text{ " in string: 52 | splits = string.split("\\text{ ") 53 | assert len(splits) == 2 54 | return splits[0] 55 | else: 56 | return string 57 | 58 | def _fix_sqrt(string): 59 | if "\\sqrt" not in string: 60 | return string 61 | splits = string.split("\\sqrt") 62 | new_string = splits[0] 63 | for split in splits[1:]: 64 | if split[0] != "{": 65 | a = split[0] 66 | new_substr = "\\sqrt{" + a + "}" + split[1:] 67 | else: 68 | new_substr = "\\sqrt" + split 69 | new_string += new_substr 70 | return new_string 71 | 72 | def _strip_string(string): 73 | # linebreaks 74 | string = string.replace("\n", "") 75 | #print(string) 76 | 77 | # remove inverse spaces 78 | string = string.replace("\\!", "") 79 | #print(string) 80 | 81 | # replace \\ with \ 82 | string = string.replace("\\\\", "\\") 83 | #print(string) 84 | 85 | # replace tfrac and dfrac with frac 86 | string = string.replace("tfrac", "frac") 87 | string = string.replace("dfrac", "frac") 88 | #print(string) 89 | 90 | # remove \left and \right 91 | string = string.replace("\\left", "") 92 | string = string.replace("\\right", "") 93 | #print(string) 94 | 95 | # Remove circ (degrees) 96 | string = string.replace("^{\\circ}", "") 97 | string = string.replace("^\\circ", "") 98 | 99 | # remove dollar signs 100 | string = string.replace("\\$", "") 101 | 102 | # remove units (on the right) 103 | string = _remove_right_units(string) 104 | 105 | # remove percentage 106 | string = string.replace("\\%", "") 107 | string = string.replace("\%", "") 108 | 109 | # " 0." equivalent to " ." and "{0." equivalent to "{." Alternatively, add "0" if "." is the start of the string 110 | string = string.replace(" .", " 0.") 111 | string = string.replace("{.", "{0.") 112 | # if empty, return empty string 113 | if len(string) == 0: 114 | return string 115 | if string[0] == ".": 116 | string = "0" + string 117 | 118 | # to consider: get rid of e.g. "k = " or "q = " at beginning 119 | if len(string.split("=")) == 2: 120 | if len(string.split("=")[0]) <= 2: 121 | string = string.split("=")[1] 122 | 123 | # fix sqrt3 --> sqrt{3} 124 | string = _fix_sqrt(string) 125 | 126 | # remove spaces 127 | string = string.replace(" ", "") 128 | 129 | # \frac1b or \frac12 --> \frac{1}{b} and \frac{1}{2}, etc. Even works with \frac1{72} (but not \frac{72}1). Also does a/b --> \\frac{a}{b} 130 | string = _fix_fracs(string) 131 | 132 | # manually change 0.5 --> \frac{1}{2} 133 | if string == "0.5": 134 | string = "\\frac{1}{2}" 135 | 136 | # NOTE: X/Y changed to \frac{X}{Y} in dataset, but in simple cases fix in case the model output is X/Y 137 | string = _fix_a_slash_b(string) 138 | 139 | return string 140 | 141 | def is_equiv(str1, str2, verbose=False): 142 | if str1 is None and str2 is None: 143 | print("WARNING: Both None") 144 | return True 145 | if str1 is None or str2 is None: 146 | return False 147 | 148 | try: 149 | ss1 = _strip_string(str1) 150 | ss2 = _strip_string(str2) 151 | if verbose: 152 | print(ss1, ss2) 153 | return ss1 == ss2 154 | except: 155 | return str1 == str2 156 | -------------------------------------------------------------------------------- /math_utils/math_helpers.py: -------------------------------------------------------------------------------- 1 | 2 | def rm_latex_math(expr: str) -> str: 3 | def wrapped_latex_math(expr: str) -> bool: 4 | expr = expr.strip() 5 | return len(expr) >=2 and expr[0] == '$' and expr[-1] == '$' 6 | 7 | expr = expr.strip() 8 | return expr[1:-1] if wrapped_latex_math(expr) else expr 9 | 10 | 11 | # The below is a copy of the helper functions in the MATH repo: 12 | # https://github.com/hendrycks/math/blob/main/modeling/eval_math_gpt.py 13 | 14 | def remove_boxed(s): 15 | left = "\\boxed{" 16 | try: 17 | assert s[:len(left)] == left 18 | assert s[-1] == "}" 19 | return s[len(left):-1] 20 | except: 21 | return None 22 | 23 | 24 | # The below is a copy of the helper functions in the MATH repo: 25 | # https://github.com/hendrycks/math/blob/main/modeling/dataset/util.py 26 | 27 | def last_boxed_only(sample): 28 | """ 29 | Given a (q,a) sample, filter the answers so that they only contain 30 | the last \boxed{...} or \fbox{...} element 31 | """ 32 | q, a = sample 33 | a = last_boxed_only_string(a) 34 | if a == None: 35 | return None 36 | return (q, a) 37 | 38 | def last_boxed_only_string(string): 39 | idx = string.rfind("\\boxed") 40 | if idx < 0: 41 | idx = string.rfind("\\fbox") 42 | if idx < 0: 43 | return None 44 | 45 | i = idx 46 | right_brace_idx = None 47 | num_left_braces_open = 0 48 | while i < len(string): 49 | if string[i] == "{": 50 | num_left_braces_open += 1 51 | if string[i] == "}": 52 | num_left_braces_open -= 1 53 | if num_left_braces_open == 0: 54 | right_brace_idx = i 55 | break 56 | i += 1 57 | 58 | if right_brace_idx == None: 59 | retval = None 60 | else: 61 | assert right_brace_idx is not None 62 | retval = string[idx:right_brace_idx + 1] 63 | 64 | return retval 65 | 66 | def only_until_first_boxed_from_tokens(string, tokens): 67 | idx = string.find("\\boxed") 68 | if idx < 0: 69 | idx = string.find("\\fbox") 70 | if idx < 0: 71 | return None 72 | 73 | cum_length = 0 74 | for i, t in enumerate(tokens): 75 | cum_length += len(t) 76 | if cum_length >= idx: 77 | break 78 | 79 | return tokens[:i] 80 | 81 | 82 | 83 | def clean_numbers(sample): 84 | if not sample: 85 | return None 86 | new_sample = list() 87 | for s in sample: 88 | new_sample.append(_clean_numbers(s)) 89 | 90 | return tuple(new_sample) 91 | 92 | def _clean_numbers(string): 93 | """ 94 | Clean Numbers in the given string 95 | 96 | >>> _clean_numbers(None, "Hello 123") 97 | 'Hello 123' 98 | >>> _clean_numbers(None, "Hello 1234") 99 | 'Hello 1,234' 100 | >>> _clean_numbers(None, "Hello 1234324asdasd") 101 | 'Hello 1,234,324asdasd' 102 | """ 103 | num_prev_digits = 0 104 | new_string = "" 105 | for i, c in enumerate(string): 106 | # isdigit() doesnt work here because of weird unicode chars. 107 | if c in {'1', '2', '3', '4', '5', '6', '7', '8', '9', '0'}: 108 | num_prev_digits += 1 109 | else: 110 | if num_prev_digits > 3: 111 | # Some fixing 112 | string_number = new_string[-num_prev_digits:] 113 | new_string = new_string[:-num_prev_digits] + "{0:,}".format(int(string_number)) 114 | num_prev_digits = 0 115 | new_string += c 116 | 117 | if num_prev_digits > 3: 118 | # Some fixing 119 | string_number = new_string[-num_prev_digits:] 120 | new_string = new_string[:-num_prev_digits] + "{0:,}".format(int(string_number)) 121 | 122 | return new_string 123 | -------------------------------------------------------------------------------- /model_api/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/model_api/__init__.py -------------------------------------------------------------------------------- /model_api/ant.py: -------------------------------------------------------------------------------- 1 | from eval.runner import EvalRunner, mk_dir_safe 2 | import argparse 3 | import sys 4 | from typing import Dict, List, Tuple, Any 5 | from math_utils.math_helpers import rm_latex_math 6 | from unformatted_llm import UnformattedLLM 7 | from model_api.closed_api import KnownModel, ParameterizedModel, Runner 8 | from model_api.claude import Claude, ANTHROPIC_PARAMS 9 | from evaluate import DEFAULT_SNAPSHOTS_FILE, DEFAULT_FEW_SHOT_NUM, DEFAULT_TEMPERATURE 10 | from chain_of_thought import COT_INSTRUCTION, ChainOfThought 11 | 12 | class Answers(Runner): 13 | def __init__(self, model: KnownModel, use_cot: bool, few_shot_num: int, temperature: float): 14 | self.model = model 15 | self.use_cot = use_cot 16 | self.cot = ChainOfThought() 17 | unformatted = UnformattedLLM() 18 | 19 | few_shot_builder = self.cot if use_cot else unformatted 20 | instruction = COT_INSTRUCTION if use_cot else unformatted.INSTRUCTION 21 | few_shot: Dict[str, str] = few_shot_builder.few_shot_limited(few_shot_num) 22 | cache_infer_params = f'temp={temperature}_cot={use_cot}_fs={few_shot_num}' 23 | cache_model_name = mk_dir_safe(model.name) 24 | params = ParameterizedModel( 25 | temp = temperature, 26 | instruction = instruction, 27 | who_are_you = "", 28 | few_shot = few_shot, 29 | model = model, 30 | agent_name = f"Anthropic_Eval_{cache_model_name}_{cache_infer_params}", 31 | ) 32 | self.claude = Claude(params) 33 | 34 | def answers(self, prb: str) -> List[Tuple[str, str]]: 35 | completion = self.claude.query([prb], rate_limited = False)[0] 36 | answer = self.claude.extract_answer(completion) 37 | if self.use_cot: 38 | answer = self.cot.extract_answer(answer) 39 | return [(answer, completion)] 40 | 41 | @classmethod 42 | def run(cls, name: str, snapshots_specs: str, 43 | cot: bool, few_shot_num: int, temperature: float, 44 | verbose: bool = False, save_snaphot: bool = False, extra_params: Dict[str, Any] = {}): 45 | known_model = ANTHROPIC_PARAMS.SHORT_NAMES[name] 46 | answerer = Answers(known_model, cot, few_shot_num, temperature) 47 | e = EvalRunner(answerer, snapshots_specs, verbose).do() 48 | if save_snaphot: 49 | answerer.claude.snapshot_api_query_cache() 50 | return e 51 | 52 | 53 | if __name__ == "__main__": 54 | parser = argparse.ArgumentParser() 55 | parser.add_argument("--model", required=True, 56 | help = f"Model has to be one of {list(ANTHROPIC_PARAMS.SHORT_NAMES.keys())}") 57 | parser.add_argument("--snapshots_specs", type=str, default=DEFAULT_SNAPSHOTS_FILE, 58 | help = f"JSON of static and monthly snapshots") 59 | parser.add_argument("--verbose", action='store_true', 60 | help="Outcomes of each test, one per line, on console") 61 | parser.add_argument("--use_chain_of_thought", action='store_true', 62 | help="Use chain of thought instruction and postprocessing") 63 | parser.add_argument("--few_shot_num", type=str, default=DEFAULT_FEW_SHOT_NUM, 64 | help=f"Default few shot count: {DEFAULT_FEW_SHOT_NUM}") 65 | parser.add_argument("--temperature", type=float, default=DEFAULT_TEMPERATURE, 66 | help=f"Default temperature: {DEFAULT_TEMPERATURE}") 67 | parser.add_argument("--save_snapshot", action='store_true', 68 | help="Save API query snapshot as .tar.gz file") 69 | args = parser.parse_args() 70 | 71 | Answers.run(args.model, args.snapshots_specs, 72 | args.use_chain_of_thought, args.few_shot_num, args.temperature, 73 | args.verbose, args.save_snapshot) 74 | -------------------------------------------------------------------------------- /model_api/claude.py: -------------------------------------------------------------------------------- 1 | import os 2 | from model_api.closed_api import ClosedAPI, RateLimited, KnownModel, PERIODS_PER_MIN, ParameterizedModel 3 | from typing import Any, Awaitable, Dict, Optional 4 | 5 | from anthropic import AsyncAnthropic, HUMAN_PROMPT, AI_PROMPT 6 | 7 | CLAUDE_API_KEY = os.getenv("ANTHROPIC_API_KEY") 8 | END_TAG = "\n~~~\n" 9 | MAX_TOKENS = 1024 10 | 11 | mk_model = lambda name: KnownModel( 12 | name, 13 | is_chat = False, 14 | api_org = None, 15 | api_key = CLAUDE_API_KEY, 16 | api_base = "") 17 | 18 | class ANTHROPIC_PARAMS: 19 | NAMES = { 20 | 'c2-2.1': 'claude-2.1', 21 | 'c3-opus': 'claude-3-opus-20240229', 22 | 'c3-sonnet': 'claude-3-sonnet-20240229' 23 | } 24 | SHORT_NAMES = { short: mk_model(model) for short, model in NAMES.items() } 25 | 26 | class Claude(ClosedAPI): 27 | def __init__(self, params: ParameterizedModel): 28 | super().__init__(params) 29 | 30 | self.params = params 31 | self.prelude = self.few_shot_format(params.few_shot) 32 | self.anthropic = AsyncAnthropic(api_key = params.model.api_key) 33 | 34 | def deobject(self, message: Any) -> str: 35 | # message is a Message( 36 | # id = 'msg_..', 37 | # content = [ContentBlock(text='', type='text')], 38 | # model='claude-3-opus-20240229', 39 | # role='assistant', 40 | # stop_reason='end_turn', 41 | # stop_sequence=None, 42 | # type='message', 43 | # usage=Usage(input_tokens=84, output_tokens=409)) 44 | content_list = message.content 45 | 46 | # From docs: https://docs.anthropic.com/claude/reference/messages_post 47 | # [{"type": "text", "text": "Hi, I'm Claude."}] 48 | 49 | # Actual output.. 50 | # [ContentBlock(text='To solve this, we ca.......9647\n\nThe answer is: 0.9647', type='text')] 51 | assert len(content_list) == 1, f'Expecting a single content response, got: {content_list}' 52 | single_content = content_list[0] 53 | assert single_content.type == 'text', f'Only expecting text content, got: {single_content}' 54 | completion = single_content.text 55 | 56 | assert isinstance(completion, str), f'claude: model output is not str: {completion}' 57 | return completion 58 | 59 | def extract_answer(self, response) -> str: 60 | completion = response.lstrip() 61 | end = completion.find(END_TAG) 62 | answer = completion[:end] if end != -1 else completion 63 | answer = answer.strip() 64 | return answer 65 | 66 | def to_prompt_task(self, task: str) -> Any: 67 | return self.prelude + self.params.instruction + self.format(task, None) 68 | 69 | def reset_library(self, model: KnownModel) -> None: 70 | return 71 | 72 | def model_ask(self, prompt: str) -> Awaitable[Any]: 73 | # print(f'Sending query: {prompt}') 74 | message = self.anthropic.messages.create( 75 | model = self.params.model.name, 76 | max_tokens = MAX_TOKENS, 77 | temperature = self.params.temperature, 78 | system = "You are a MATH expert.", 79 | messages = [{ 80 | 'role': 'user', 81 | 'content': prompt, 82 | }] 83 | ) 84 | return message 85 | 86 | def spent(self, response) -> float: 87 | return 0.0 88 | 89 | def format(self, inp: str, out: Optional[str]) -> str: 90 | fmt = f"{HUMAN_PROMPT} {inp}{AI_PROMPT}" 91 | extra = out if out else "" 92 | return fmt + extra 93 | 94 | def few_shot_format(self, ios: Dict[str, str]) -> str: 95 | fs = "" 96 | for inp, out in ios.items(): 97 | fs += self.format(inp, out) + END_TAG 98 | return fs 99 | 100 | -------------------------------------------------------------------------------- /model_api/closed_api.py: -------------------------------------------------------------------------------- 1 | from typing import Dict, List, Tuple, Callable, Any, Awaitable, Optional 2 | import time 3 | import sys 4 | import os 5 | import asyncio 6 | import math 7 | import functools 8 | from codetiming import Timer # type: ignore 9 | from tqdm import tqdm # type: ignore 10 | from query_cache import QueryCache 11 | from persist import Persist 12 | from enum import Enum 13 | 14 | # temperature threshold under which we expect deterministic answers 15 | # the queries will be cached and we won't make API calls to OAI 16 | # If temp is above the threshold then every time we will API to OAI 17 | DETERMINISTIC_THRESHOLD = 1.0 18 | 19 | NOW_MS = lambda: round(time.time() * 1000) 20 | PERIODS_PER_MIN = 4 21 | ONE_PERIOD_MS = 1000 * 60 / PERIODS_PER_MIN 22 | 23 | # If ServiceUnavailableError, or Timeout etc 24 | BACKOFF_NUM_RETRIES = 3 25 | BACKOFF_TIME_SECONDS = 120 26 | hr = "=" * 100 27 | BACKOFF_MSG = lambda e: f'\n\n{hr}\nService unavailable.\nError = {e}\nPausing for {BACKOFF_TIME_SECONDS} seconds.\n{hr}\n\n' 28 | BACKOFF_ABORT_MSG = lambda e: f'\n\n{hr}\nService unavailable after {BACKOFF_NUM_RETRIES} attempts.\nError = {e}\n{hr}\nAborting!\n\n' 29 | 30 | OFFLINE_DEFER = "DEFERRED; MAGIC; NOTHING TO SEE HERE!" 31 | 32 | class KnownModel: 33 | def __init__(self, 34 | name: str, 35 | is_chat: bool, 36 | api_org: Optional[str], 37 | api_key: Optional[str], 38 | api_base: str, 39 | stops: Optional[List[str]] = None, 40 | prompt_format: Optional[str] = None, 41 | server_path: Optional[str] = None, 42 | batch_size: int = 200, 43 | num_parallel: int = 50, 44 | cps: float = 0.0, 45 | ): 46 | self.name = name 47 | self.is_chat = is_chat 48 | self.api_org = api_org 49 | self.api_key = api_key 50 | self.api_base = api_base 51 | self.server_path = server_path 52 | 53 | self.stops = stops 54 | self.prompt_format = prompt_format 55 | 56 | self.batch_size = batch_size 57 | self.num_parallel = num_parallel 58 | self.cost_per_token = cps 59 | 60 | def __hash__(self): 61 | hashes = [hash(v) for v in vars(self)] 62 | init = 42 63 | return functools.reduce(lambda a, b: a ^ b, hashes, init) 64 | 65 | def __eq__(self, other): 66 | eq_vars = vars(self) == vars(other) 67 | return (isinstance(other, self.__class__) and eq_vars) 68 | 69 | class QStat: 70 | def __init__(self, cost, time_ms): 71 | self.cost = cost 72 | self.time_ms = time_ms 73 | self.timestamp = NOW_MS() 74 | 75 | class QueryStats: 76 | def __init__(self): 77 | self.stats: List[QStat] = [] 78 | 79 | def log_queries(self, qs: List[QStat]): 80 | self.stats += qs 81 | 82 | def __repr__(self) -> str: 83 | total = len(self.stats) 84 | per_query_spend = 0.0 85 | per_query_ms = 0.0 86 | queries_min = 0.0 87 | 88 | if total > 0: 89 | spend = sum(q.cost for q in self.stats) 90 | ms = sum(q.time_ms for q in self.stats) 91 | per_query_spend = float(spend) / total 92 | per_query_ms = float(ms) / total 93 | 94 | # estimate q/min 95 | max_ts = max(q.timestamp for q in self.stats) 96 | min_ts = min(q.timestamp for q in self.stats) 97 | minutes = (max_ts - min_ts) / (60 * 1000.0) 98 | ms_per_query = ( max_ts - min_ts ) / total 99 | queries_min = PERIODS_PER_MIN * ONE_PERIOD_MS / ms_per_query if ms_per_query else 0.0 100 | 101 | return f'{total} queries; {minutes:.2f} mins total time; per query = ${per_query_spend:.4f} {per_query_ms / 1000.0:.2f}s; rate = {queries_min:.0f}queries/min, 0 time/cost implies cache lookup]' 102 | 103 | ResponseStats = Tuple[str, QStat] 104 | WorkFn = Callable[[Any], Awaitable[Tuple[ResponseStats, bool]]] 105 | 106 | class Workers: 107 | def __init__(self, async_fn: WorkFn, num_parallel): 108 | self.async_fn = async_fn 109 | self.num_parallel = num_parallel 110 | 111 | def do(self, tasks) -> Tuple[List[ResponseStats], int]: 112 | return asyncio.run(self.parallel(tasks)) 113 | 114 | async def work(self, worker_ident, in_q, out_q): 115 | while not in_q.empty(): 116 | tid, task = await in_q.get() 117 | rslt, is_expensive = await self.async_fn(task) 118 | await out_q.put((tid, is_expensive, rslt)) 119 | 120 | async def progress(self, out_q, sz): 121 | pbar = tqdm(total = sz) 122 | out_sz = 0 123 | while out_sz < sz: 124 | new_sz = out_q.qsize() 125 | update = new_sz - out_sz 126 | out_sz = new_sz 127 | pbar.update(update) 128 | await asyncio.sleep(0.5) 129 | pbar.close() 130 | 131 | async def parallel(self, tasks) -> Tuple[List[ResponseStats], int]: 132 | # Create the queue of work 133 | in_q: asyncio.Queue[Tuple[int, Any]] = asyncio.Queue() 134 | out_q: asyncio.Queue[Tuple[int, bool, Any]] = asyncio.Queue() 135 | 136 | # Put some work in the queue 137 | for tid, t in enumerate(tasks): 138 | await in_q.put((tid, t)) 139 | 140 | # create the deferred workers 141 | create = lambda x: asyncio.create_task(self.work(x, in_q, out_q)) 142 | workers = [create(i) for i in range(self.num_parallel)] 143 | workers.append(asyncio.create_task(self.progress(out_q, len(tasks)))) 144 | 145 | # Run the tasks 146 | await asyncio.gather(*workers) 147 | 148 | results = [] 149 | while not out_q.empty(): 150 | results.append(await out_q.get()) 151 | results.sort(key = lambda x: x[0]) 152 | ids, expensive_flags, done = list(zip(*results)) 153 | assert list(ids) == sorted(list(ids)) 154 | expensed = expensive_flags.count(True) 155 | return list(done), expensed 156 | 157 | class LogLevel(Enum): 158 | ERROR = 3 159 | WARN = 2 160 | INFO = 1 161 | 162 | LOWEST = 1 163 | 164 | class RateLimited: 165 | def __init__(self, async_fn: WorkFn, batch_size: int, num_parallel: int, console_log_lvl: LogLevel): 166 | self.workers = Workers(async_fn, num_parallel) 167 | self.overall_stats = QueryStats() 168 | self.batch_size = batch_size 169 | self.console_log_lvl = console_log_lvl 170 | 171 | def do(self, tasks) -> List[str]: 172 | with Timer(text="Total elapsed time: {:.1f}"): 173 | chunks = self.chunks(tasks, self.batch_size) 174 | num_chunks = math.ceil(len(tasks) / self.batch_size) 175 | responses: List[ResponseStats] = self.do_one_per_period(chunks, num_chunks, self.workers.do) 176 | answers, stats = list(zip(*responses)) 177 | self.overall_stats.log_queries(list(stats)) 178 | self.console_log(LogLevel.INFO, f'stats: {self.overall_stats}') 179 | return list(answers) 180 | 181 | def do_one_per_period(self, per_period_chunks, num_chunks, 182 | per_period_fn: Callable[[Any], Tuple[List[ResponseStats], int]]) -> List[ResponseStats]: 183 | responses: List[ResponseStats] = [] 184 | for count, chunk in enumerate(per_period_chunks): 185 | with Timer(text="Expected to spend 60s if API called. Elapsed: {:.1f}s. (< 60s ok, if cache lookup)"): 186 | self.console_log(LogLevel.INFO, f'Sending {len(chunk)} queries to OAI in parallel.') 187 | start = NOW_MS() 188 | resp_list, num_expensive = self.with_unavailable_backoff(per_period_fn, chunk) 189 | responses += resp_list 190 | elapsed = NOW_MS() - start 191 | frac_api = float(num_expensive) / len(chunk) 192 | frac_cache = float(len(chunk) - num_expensive) / len(chunk) 193 | self.console_log(LogLevel.INFO, f'Done {count + 1}/{num_chunks} ({frac_api * 100:.2f}% API calls, {frac_cache * 100:.2f}% cache lookup)') 194 | 195 | # sleep for 100ms more than what is needed to make it one min 196 | pause_ms = frac_api * (ONE_PERIOD_MS - elapsed + 100) 197 | if pause_ms > 0: 198 | self.console_log(LogLevel.WARN, f'Pausing for {pause_ms / 1000.0}s to stay within API rate budget') 199 | time.sleep(pause_ms / 1000.0) 200 | return responses 201 | 202 | def console_log(self, lvl: LogLevel, msg): 203 | if lvl.value >= self.console_log_lvl.value: 204 | return 205 | print(msg) 206 | 207 | def with_unavailable_backoff(self, per_period_fn, chunk): 208 | tries = 0 209 | while tries < BACKOFF_NUM_RETRIES: 210 | try: 211 | return per_period_fn(chunk) 212 | except Exception as e: 213 | self.console_log(LogLevel.WARN, BACKOFF_MSG(e)) 214 | time.sleep(BACKOFF_TIME_SECONDS) 215 | ex = e 216 | tries += 1 217 | # if failed after enough tries, terminate process 218 | # restarting will pull any data already downloaded from cache 219 | # so no destructive loss. we just revert to user to manually restart 220 | self.console_log(LogLevel.WARN, BACKOFF_ABORT_MSG(ex)) 221 | raise ex 222 | 223 | def chunks(self, lst, batch_size): 224 | for i in range(0, len(lst), batch_size): 225 | yield lst[i:i + batch_size] 226 | 227 | 228 | class ParameterizedModel: 229 | def __init__(self, 230 | temp: float, 231 | instruction: str, 232 | who_are_you: str, 233 | few_shot: Dict[str, str], 234 | model: KnownModel, 235 | agent_name: str, 236 | mk_tasks: bool = False, 237 | batch_sz: int = 200, 238 | num_parallel: int = 50, 239 | ): 240 | self.pull_from_cache = temp < DETERMINISTIC_THRESHOLD 241 | 242 | self.temperature = temp 243 | self.instruction = instruction 244 | self.who_are_you = who_are_you 245 | self.few_shot = few_shot 246 | self.model = model 247 | 248 | self.agent_name = agent_name 249 | self.batch_sz = batch_sz 250 | self.num_parallel = num_parallel 251 | 252 | self.mk_tasks = mk_tasks 253 | 254 | def __hash__(self): 255 | hashes = [hash(v) for v in vars(self)] 256 | init = 42 257 | return functools.reduce(lambda a, b: a ^ b, hashes, init) 258 | 259 | def __eq__(self, other): 260 | eq_vars = vars(self) == vars(other) 261 | return (isinstance(other, self.__class__) and eq_vars) 262 | 263 | def __repr__(self): 264 | return f'PM({self.model.name}, {self.agent_name}, task_maker = {self.mk_tasks})' 265 | 266 | class TasksForModel: 267 | def __init__(self, tasks: List[str], params: ParameterizedModel): 268 | self.tasks = tasks 269 | self.params = params 270 | 271 | class OracleTaskQ: 272 | TASKS_DIR = "oracle_task_queue" 273 | TASK_FILE_SUFFIX = ".pickle" 274 | 275 | @classmethod 276 | def write_tasks(cls, tasks: List[str], params: ParameterizedModel) -> None: 277 | cls.ensure_tq_dir_exists() 278 | existing: List[int] = cls.idents_in_task_dir() 279 | ident = max(existing) + 1 if existing else 0 280 | fname = cls.task_file(ident) 281 | tm = TasksForModel(tasks, params) 282 | Persist.save(tm, fname) 283 | 284 | @classmethod 285 | def task_file(cls, ident: int) -> str: 286 | cls.ensure_tq_dir_exists() 287 | return os.path.join(cls.TASKS_DIR, f"{ident}{cls.TASK_FILE_SUFFIX}") 288 | 289 | @classmethod 290 | def idents_in_task_dir(cls) -> List[int]: 291 | cls.ensure_tq_dir_exists() 292 | extract = lambda fname: int(fname[:-len(cls.TASK_FILE_SUFFIX)]) 293 | return sorted([extract(f) for f in os.listdir(cls.TASKS_DIR)]) 294 | 295 | @classmethod 296 | def purge_tasks(cls, idents: List[int]) -> None: 297 | cls.ensure_tq_dir_exists() 298 | for i in idents: 299 | fname = cls.task_file(i) 300 | os.remove(fname) 301 | 302 | @classmethod 303 | def ensure_tq_dir_exists(cls) -> None: 304 | if not os.path.exists(cls.TASKS_DIR): 305 | os.makedirs(cls.TASKS_DIR, exist_ok = True) 306 | 307 | class ClosedAPI: 308 | MAKER_PLACEHOLDER = "PLACEHOLDER_FILL_LATER" 309 | 310 | def __init__(self, params: ParameterizedModel, console_loglevel: LogLevel = LogLevel.LOWEST): 311 | self.params = params 312 | self.cacher = QueryCache(params.agent_name) 313 | self.rate_limited_runner = RateLimited(self.cached_api_call, params.batch_sz, params.num_parallel, console_loglevel) 314 | self.TASK_MAKER_MODE = params.mk_tasks 315 | 316 | # to be overriden 317 | def deobject(self, obj_resp: Any) -> str: # type: ignore [empty-body] 318 | pass 319 | 320 | # to be overriden 321 | def extract_answer(self, resp: str) -> str: # type: ignore [empty-body] 322 | pass 323 | 324 | # to be overriden 325 | def to_prompt_task(self, task: str) -> Any: # type: ignore [empty-body] 326 | pass 327 | 328 | # to be overriden 329 | def reset_library(self, model: KnownModel) -> None: # type: ignore [empty-body] 330 | pass 331 | 332 | # to be overriden 333 | def model_ask(self, prompt: str) -> Awaitable[Any]: # type: ignore [empty-body] 334 | pass 335 | 336 | # to be overriden 337 | def spent(self, response) -> float: # type: ignore [empty-body] 338 | pass 339 | 340 | def sync_solver(self, task: str): 341 | ans_stat, api_call_not_cache = asyncio.run(self.cached_api_call(task)) 342 | ans, stat = ans_stat 343 | return ans 344 | 345 | def maker_placeholder(self, sz: int) -> List[str]: 346 | return [ ClosedAPI.MAKER_PLACEHOLDER ] * sz 347 | 348 | def make_tasks(self, tasks: List[str]) -> List[str]: 349 | OracleTaskQ.write_tasks(tasks, self.params) 350 | placeholders = self.maker_placeholder(len(tasks)) 351 | return placeholders 352 | 353 | def query(self, tasks: List[str], rate_limited: bool = True) -> List[str]: 354 | if self.TASK_MAKER_MODE: 355 | return self.make_tasks(tasks) 356 | 357 | # actual solving mode, call API 358 | self.reset_library(self.params.model) 359 | 360 | if rate_limited: 361 | solved = self.rate_limited_runner.do(tasks) 362 | else: 363 | solved = [self.sync_solver(s) for s in tasks] 364 | return solved 365 | 366 | async def cached_api_call(self, task: str) -> Tuple[ResponseStats, bool]: 367 | # if in cache pull from cache 368 | if self.params.pull_from_cache: 369 | cached = self.cacher.get(task) 370 | if cached: 371 | return (cached, QStat(0, 0)), False 372 | 373 | # not cached, make api call 374 | answer, stat = await self.api_call(task) 375 | 376 | if answer != OFFLINE_DEFER: 377 | # write to cache even in the case when we don't pull from it 378 | self.cacher.put(task, answer) 379 | 380 | return (answer, stat), True 381 | 382 | async def api_call(self, task: str) -> ResponseStats: 383 | 384 | start_ms = NOW_MS() 385 | response = await self.model_ask(self.to_prompt_task(task)) 386 | elapsed = NOW_MS() - start_ms 387 | 388 | if response == OFFLINE_DEFER: 389 | return OFFLINE_DEFER, QStat(0, 0) 390 | 391 | cost = self.spent(response) 392 | txt_response = self.deobject(response) 393 | return txt_response, QStat(cost, elapsed) 394 | 395 | def snapshot_api_query_cache(self): 396 | self.cacher.save_snapshot() 397 | 398 | class Runner: 399 | @classmethod 400 | def run(cls, name: str, snapshots_specs: str, 401 | cot: bool, few_shot_num: int, temperature: float, 402 | verbose: bool = False, save_snaphot: bool = False): # type: ignore [empty-body] 403 | pass 404 | -------------------------------------------------------------------------------- /model_api/file_api.py: -------------------------------------------------------------------------------- 1 | import os 2 | from model_api.closed_api import ClosedAPI, RateLimited, KnownModel, PERIODS_PER_MIN, ParameterizedModel, OFFLINE_DEFER 3 | from typing import Any, Awaitable, List, Dict, Optional 4 | 5 | END_TAG = "\n~~~\n" 6 | MAX_TOKENS = 1024 7 | DEFAULT_PROMPT_FORMAT = "Problem:\n{prompt}\nSolution:\n{response}" 8 | 9 | mk_model = lambda name: KnownModel( 10 | name, 11 | is_chat = False, 12 | api_org = None, 13 | api_key = "", 14 | api_base = "", 15 | prompt_format = DEFAULT_PROMPT_FORMAT) 16 | 17 | class OFFLINE_PARAMS: 18 | NAMES = { 19 | 'gemini-ultra': 'gemini-ultra', 20 | 'gemini-pro': 'gemini-pro', 21 | 'gemini-nano': 'gemini-nano', 22 | } 23 | SHORT_NAMES = { short: mk_model(model) for short, model in NAMES.items() } 24 | 25 | class FileAPI(ClosedAPI): 26 | def __init__(self, params: ParameterizedModel, mode_write_questions: bool): 27 | super().__init__(params) 28 | 29 | self.params = params 30 | self.prelude = self.few_shot_format(params.few_shot) 31 | 32 | self.mode_write_questions = mode_write_questions 33 | self.offline_prompts: List[Dict[str, Any]] = [] 34 | self.offline_answers = self.read_offline(params.model.name) if not mode_write_questions else None 35 | 36 | def read_offline(self, model_name: str) -> Dict[str, str]: 37 | assert False, f'unimplemented' 38 | 39 | def deobject(self, message: Any) -> str: 40 | completion = message 41 | assert isinstance(completion, str), f'model output is not str: {completion}' 42 | return completion 43 | 44 | def extract_answer(self, response) -> str: 45 | completion = response.lstrip() 46 | end = completion.find(END_TAG) 47 | answer = completion[:end] if end != -1 else completion 48 | answer = answer.strip() 49 | return answer 50 | 51 | def to_prompt_task(self, task: str) -> Any: 52 | return self.prelude + self.params.instruction + self.format(task, None) 53 | 54 | def reset_library(self, model: KnownModel) -> None: 55 | return 56 | 57 | def model_ask(self, prompt: str) -> Awaitable[Any]: 58 | if self.mode_write_questions: 59 | # question mode 60 | return self.async_write_prompt(prompt, 61 | model = self.params.model.name, 62 | max_tokens = MAX_TOKENS, 63 | temperature = self.params.temperature) 64 | else: 65 | # answer mode 66 | return self.async_read_response(prompt) 67 | 68 | async def async_write_prompt(self, prompt, 69 | model: str, 70 | max_tokens: int, 71 | temperature: float) -> str: 72 | prompt_json = { 73 | 'prompt': prompt, 74 | 'model': model, 75 | 'max_tokens': max_tokens, 76 | 'temperature': temperature 77 | } 78 | self.offline_prompts.append(prompt_json) 79 | return OFFLINE_DEFER 80 | 81 | async def async_read_response(self, prompt) -> str: 82 | assert self.offline_answers, f'No offline answers available' 83 | assert prompt in self.offline_answers, f'Prompt not completed: {prompt}' 84 | message = self.offline_answers[prompt] 85 | return message 86 | 87 | def spent(self, response) -> float: 88 | return 0.0 89 | 90 | def format(self, inp: str, out: Optional[str]) -> str: 91 | resp = out if out else "" 92 | assert self.params.model.prompt_format, f'prompt format not set' 93 | formatted = self.params.model.prompt_format.format(prompt = inp, response = resp) 94 | return formatted 95 | 96 | def few_shot_format(self, ios: Dict[str, str]) -> str: 97 | fs = "" 98 | for inp, out in ios.items(): 99 | fs += self.format(inp, out) + END_TAG 100 | return fs 101 | 102 | -------------------------------------------------------------------------------- /model_api/gpt.py: -------------------------------------------------------------------------------- 1 | import sys 2 | import os 3 | from copy import deepcopy 4 | from typing import Dict, List, Tuple, Callable, Any, Awaitable, Optional 5 | 6 | import openai # type: ignore 7 | 8 | from model_api.closed_api import ClosedAPI, RateLimited, KnownModel, PERIODS_PER_MIN, ParameterizedModel 9 | from model_api.closed_api import OracleTaskQ 10 | from unformatted_llm import UnformattedLLM 11 | from persist import Persist 12 | 13 | # In debug mode will only make limited API queries 14 | DEBUG_MODE_NUM_LIMIT = None # 200 # or None for non-debug mode 15 | 16 | # end few shot instances with special tag 17 | END_TAG = "\n~~~\n" 18 | 19 | # how much buffer on being nice with API calls; e.g., if 100/min allowed then we do 50/min if factor = 2 20 | BATCH_SZ_FACTOR_PER_MIN = 8 21 | # extra nice on API calls, by reducing batch size "permitted" based on API rate, divided by this factor 22 | EXTRA_NICE = 4 23 | 24 | # find these keys at https://platform.openai.com/account/api-keys 25 | # ... org-... at url above Organization->Settings 26 | # ... sk- at url above User->API Keys 27 | API_ORG = os.getenv("OPENAI_ORG") 28 | API_KEY = os.getenv("OPENAI_API_KEY") 29 | 30 | def ensure_keys(): 31 | if API_KEY and API_ORG: 32 | return 33 | print(f'OAI keys missing. export OPENAI_API_KEY and OPENAI_ORG. Aborting.') 34 | print(f'[WARN] You might have to purge the oracle task queue at {OracleTaskQ.TASKS_DIR}') 35 | print(f'[WARN] Some no-key tasks might have been written to queue') 36 | exit(-1) 37 | 38 | class OAI_PARAMS: 39 | # gpt rate limits: https://platform.openai.com/account/rate-limits 40 | QUERY_PER_MIN_LIMIT = { 41 | "gpt4": 200, 42 | "gpt3": 3500, 43 | } 44 | if isinstance(DEBUG_MODE_NUM_LIMIT, int): 45 | QUERY_PER_MIN_LIMIT = { 46 | "gpt4": min(200, DEBUG_MODE_NUM_LIMIT), 47 | "gpt3": min(3500, DEBUG_MODE_NUM_LIMIT), 48 | } 49 | 50 | # https://openai.com/pricing#language-models 51 | MODEL_COST_PER_TOKEN = { 52 | "gpt4": 0.03 / 1000, 53 | "gpt3": 0.0015 / 1000, 54 | } 55 | 56 | # Note, legacy text models, e.g., DAVINCI are being deprecated Jan 2024 57 | # https://platform.openai.com/docs/deprecations/base-gpt-models 58 | @classmethod 59 | def cps(cls, m): 60 | return cls.MODEL_COST_PER_TOKEN[m] 61 | @classmethod 62 | def bsz(cls, m): 63 | return int(cls.QUERY_PER_MIN_LIMIT[m] / (PERIODS_PER_MIN * BATCH_SZ_FACTOR_PER_MIN)) 64 | @classmethod 65 | def np(cls, m): 66 | return int(cls.bsz(m) / EXTRA_NICE) if DEBUG_MODE_NUM_LIMIT is None else 10 67 | @classmethod 68 | def rates(cls, m): 69 | return { "batch_size": cls.bsz(m), "num_parallel": cls.np(m), "cps": cls.cps(m) } 70 | LOC = { "api_org": API_ORG, "api_key": API_KEY, "api_base": "https://api.openai.com/v1" } 71 | 72 | class OAI_MODELS: 73 | GPT4 = KnownModel("gpt-4", is_chat = True, **OAI_PARAMS.LOC, **OAI_PARAMS.rates("gpt4")) # type: ignore 74 | GPT3 = KnownModel("gpt-3.5-turbo", is_chat = True, **OAI_PARAMS.LOC, **OAI_PARAMS.rates("gpt3")) # type: ignore 75 | 76 | SHORT_NAMES = {"gpt4": GPT4, "gpt3": GPT3} 77 | 78 | DEFAULT_MODEL = GPT3 79 | 80 | class GPT(ClosedAPI): 81 | def __init__(self, params: ParameterizedModel): 82 | self.params = params 83 | super().__init__(params) 84 | 85 | a, t, c, p = GPT.model_fns(params.model, params.temperature, params.few_shot, params.instruction, params.who_are_you) 86 | self.model_ask = a # type: ignore [method-assign] 87 | self.to_prompt_task = t # type: ignore [method-assign] 88 | self.deobject = c # type: ignore [method-assign] 89 | self.prelude = p 90 | 91 | def extract_answer(self, response) -> str: 92 | completion = response.lstrip() 93 | end = completion.find(END_TAG) 94 | answer = completion[:end] if end != -1 else completion 95 | answer = answer.strip() 96 | return answer 97 | 98 | def reset_library(self, model_params): 99 | # reset openai library 100 | openai.organization = model_params.api_org 101 | openai.api_key = model_params.api_key 102 | openai.api_base = model_params.api_base 103 | 104 | def spent(self, response) -> float: 105 | return response['usage']['total_tokens'] * self.params.model.cost_per_token 106 | 107 | @classmethod 108 | def model_fns(cls, model, temp, few_shot, instruction, who_are_you) -> Tuple[Any, Any, Any, Any]: 109 | # note `acreate` instead of `create` for async 110 | # docs: https://github.com/openai/openai-python#async-api 111 | 112 | # chat ask, raw api query 113 | def chat_ask(msgs): 114 | # print(f'Sending query: {msgs}') 115 | return openai.ChatCompletion.acreate( 116 | model = model.name, 117 | messages = msgs, 118 | temperature = temp, 119 | ) 120 | # prelude added to each query 121 | who = {"role": "system", "content": who_are_you} 122 | chat_prelude = [who] 123 | for inp, out in few_shot.items(): 124 | que = {"role": "user", "content": instruction + inp} 125 | chat_prelude.append(que) 126 | ans = {"role": "assistant", "content": out + END_TAG} 127 | chat_prelude.append(ans) 128 | # extract 129 | chat_txt_from_obj = lambda response: response['choices'][0]['message']['content'] 130 | 131 | def chat_task(task: str): 132 | messages = deepcopy(chat_prelude) 133 | que = {"role": "user", "content": instruction + task} 134 | messages.append(que) 135 | return messages 136 | 137 | # pick appropriate functions based on model type 138 | callfn = chat_ask 139 | promptfn = chat_task 140 | deobject = chat_txt_from_obj 141 | prelude = chat_prelude 142 | return callfn, promptfn, deobject, prelude 143 | 144 | def check_model_access(self): 145 | # check that we have access to this model 146 | self.reset_library(self.params.model) 147 | models = openai.Model.list()["data"] 148 | assert list(filter(lambda x: x["id"] == self.params.model.name, models)), "No access to {self.params.model.name}" 149 | -------------------------------------------------------------------------------- /model_api/mis.py: -------------------------------------------------------------------------------- 1 | from eval.runner import EvalRunner, mk_dir_safe 2 | import argparse 3 | import sys 4 | from typing import Dict, List, Tuple, Any 5 | from math_utils.math_helpers import rm_latex_math 6 | from unformatted_llm import UnformattedLLM 7 | from model_api.closed_api import KnownModel, ParameterizedModel, Runner 8 | from model_api.mistral import Mistral, MISTRAL_PARAMS 9 | from evaluate import DEFAULT_SNAPSHOTS_FILE, DEFAULT_FEW_SHOT_NUM, DEFAULT_TEMPERATURE 10 | from chain_of_thought import COT_INSTRUCTION, ChainOfThought 11 | 12 | _MISTRAL = MISTRAL_PARAMS() 13 | 14 | class Answers(Runner): 15 | def __init__(self, model: KnownModel, use_cot: bool, few_shot_num: int, temperature: float): 16 | self.model = model 17 | self.use_cot = use_cot 18 | self.cot = ChainOfThought() 19 | unformatted = UnformattedLLM() 20 | 21 | few_shot_builder = self.cot if use_cot else unformatted 22 | instruction = COT_INSTRUCTION if use_cot else unformatted.INSTRUCTION 23 | few_shot: Dict[str, str] = few_shot_builder.few_shot_limited(few_shot_num) 24 | cache_infer_params = f'temp={temperature}_cot={use_cot}_fs={few_shot_num}' 25 | cache_model_name = mk_dir_safe(model.name) 26 | params = ParameterizedModel( 27 | temp = temperature, 28 | instruction = instruction, 29 | who_are_you = "", 30 | few_shot = few_shot, 31 | model = model, 32 | agent_name = f"Mistral_Eval_{cache_model_name}_{cache_infer_params}", 33 | ) 34 | self.mistral = Mistral(params) 35 | 36 | def answers(self, prb: str) -> List[Tuple[str, str]]: 37 | completion = self.mistral.query([prb], rate_limited = False)[0] 38 | answer = self.mistral.extract_answer(completion) 39 | if self.use_cot: 40 | answer = self.cot.extract_answer(answer) 41 | return [(answer, completion)] 42 | 43 | @classmethod 44 | def run(cls, name: str, snapshots_specs: str, 45 | cot: bool, few_shot_num: int, temperature: float, 46 | verbose: bool = False, save_snaphot: bool = False, extra_params: Dict[str, Any] = {}): 47 | known_model = _MISTRAL.models[name] 48 | answerer = Answers(known_model, cot, few_shot_num, temperature) 49 | e = EvalRunner(answerer, snapshots_specs, verbose).do() 50 | if save_snaphot: 51 | answerer.mistral.snapshot_api_query_cache() 52 | return e 53 | 54 | 55 | if __name__ == "__main__": 56 | parser = argparse.ArgumentParser() 57 | parser.add_argument("--model", required=True, 58 | help = f"Model has to be one of {list(_MISTRAL.models.keys())}") 59 | parser.add_argument("--snapshots_specs", type=str, default=DEFAULT_SNAPSHOTS_FILE, 60 | help = f"JSON of static and monthly snapshots") 61 | parser.add_argument("--verbose", action='store_true', 62 | help="Outcomes of each test, one per line, on console") 63 | parser.add_argument("--use_chain_of_thought", action='store_true', 64 | help="Use chain of thought instruction and postprocessing") 65 | parser.add_argument("--few_shot_num", type=str, default=DEFAULT_FEW_SHOT_NUM, 66 | help=f"Default few shot count: {DEFAULT_FEW_SHOT_NUM}") 67 | parser.add_argument("--temperature", type=float, default=DEFAULT_TEMPERATURE, 68 | help=f"Default temperature: {DEFAULT_TEMPERATURE}") 69 | parser.add_argument("--save_snapshot", action='store_true', 70 | help="Save API query snapshot as .tar.gz file") 71 | args = parser.parse_args() 72 | 73 | Answers.run(args.model, args.snapshots_specs, 74 | args.use_chain_of_thought, args.few_shot_num, args.temperature, 75 | args.verbose, args.save_snapshot) 76 | -------------------------------------------------------------------------------- /model_api/mistral.py: -------------------------------------------------------------------------------- 1 | from mistralai.client import MistralClient 2 | from mistralai.models.chat_completion import ChatMessage 3 | import os 4 | from model_api.closed_api import ClosedAPI, RateLimited, KnownModel, PERIODS_PER_MIN, ParameterizedModel 5 | from unformatted_llm import UnformattedLLM 6 | from typing import Any, Awaitable, Dict, Optional, List, Tuple 7 | from tqdm import tqdm # type: ignore 8 | 9 | MISTRAL_API_KEY = os.getenv("MISTRAL_API_KEY") 10 | END_TAG = "\n~~~\n" 11 | DEFAULT_END_TAGS = [END_TAG, ''] 12 | DEFAULT_PROMPT_FORMAT = ": {prompt}\n: {response}" 13 | MAX_TOKENS = 1024 14 | 15 | class MISTRAL_PARAMS: 16 | def __init__(self): 17 | # https://docs.mistral.ai/platform/endpoints/ 18 | # small = 8x7b 19 | # medium = unknown but 8x32b rumoured 20 | # large = unknown, but apparently does well on evals: https://mistral.ai/news/mistral-large/ 21 | models_names: Dict[str, str] = { 22 | "mistral-small": "mistral-small-2402", 23 | "mistral-medium": "mistral-medium-2312", 24 | "mistral-large": "mistral-large-2402", 25 | } 26 | self.models: Dict[str, KnownModel] = { n: self.known_from_config(endpoint) for n, endpoint in models_names.items() } 27 | 28 | def known_from_config(self, name: str) -> KnownModel: 29 | is_chat = True 30 | stops: List[str] = DEFAULT_END_TAGS 31 | prompt_format: str = DEFAULT_PROMPT_FORMAT 32 | 33 | return KnownModel(name, 34 | is_chat, 35 | api_org = None, 36 | api_key = MISTRAL_API_KEY, 37 | api_base = "", 38 | stops = stops, 39 | prompt_format = prompt_format) 40 | 41 | 42 | class Mistral(ClosedAPI): 43 | def __init__(self, params: ParameterizedModel): 44 | super().__init__(params) 45 | 46 | self.params = params 47 | self.prelude = self.few_shot_format(params.few_shot) 48 | self.client = MistralClient(api_key = params.model.api_key) 49 | 50 | def deobject(self, obj_resp: Any) -> str: 51 | extract = lambda resp: resp.choices[0].message.content 52 | completion = obj_resp if isinstance(obj_resp, str) else extract(obj_resp) 53 | assert isinstance(completion, str), f'model output is not str: {obj_resp}' 54 | return completion 55 | 56 | def extract_answer(self, response) -> str: 57 | completion = response.lstrip() 58 | assert self.params.model.stops, f'stop tokens not set' 59 | ends = [completion.find(s) for s in self.params.model.stops if completion.find(s) != -1] 60 | answer = completion[:ends[0]] if ends else completion 61 | answer = answer.strip() 62 | return answer 63 | 64 | async def async_ask(self, prompt: str): 65 | # print(f'Sending query: {prompt}') 66 | return self.client.chat( 67 | model = self.params.model.name, 68 | messages = [ 69 | ChatMessage(role="user", content=prompt) 70 | ], 71 | max_tokens = MAX_TOKENS, 72 | temperature = self.params.temperature, 73 | ) 74 | 75 | def to_prompt_task(self, task: str) -> Any: 76 | return self.prelude + self.params.instruction + self.format(task, None) 77 | 78 | def reset_library(self, model: KnownModel) -> None: 79 | return 80 | 81 | def model_ask(self, prompt: str) -> Awaitable[Any]: 82 | return self.async_ask(prompt) 83 | 84 | def spent(self, response) -> float: 85 | return 0.0 86 | 87 | def format(self, inp: str, out: Optional[str]) -> str: 88 | resp = out if out else "" 89 | assert self.params.model.prompt_format, f'prompt format not set' 90 | formatted = self.params.model.prompt_format.format(prompt = inp, response = resp) 91 | return formatted 92 | 93 | def few_shot_format(self, ios: Dict[str, str]) -> str: 94 | fs = "" 95 | for inp, out in ios.items(): 96 | assert self.params.model.stops, f'stop tokens not set' 97 | fs += self.format(inp, out) + self.params.model.stops[0] 98 | return fs 99 | -------------------------------------------------------------------------------- /model_api/oai.py: -------------------------------------------------------------------------------- 1 | from eval.runner import EvalRunner, mk_dir_safe 2 | import argparse 3 | import sys 4 | from typing import Dict, List, Tuple, Any 5 | from math_utils.math_helpers import rm_latex_math 6 | from unformatted_llm import UnformattedLLM 7 | from model_api.gpt import GPT, OAI_MODELS 8 | from model_api.closed_api import KnownModel, ParameterizedModel, Runner 9 | from evaluate import DEFAULT_SNAPSHOTS_FILE, DEFAULT_FEW_SHOT_NUM, DEFAULT_TEMPERATURE 10 | from chain_of_thought import COT_INSTRUCTION, ChainOfThought 11 | 12 | WHO_ARE_YOU = "You are a expert in mathematics, logic, and formal reasoning. " 13 | 14 | class Answers(Runner): 15 | def __init__(self, model: KnownModel, use_cot: bool, few_shot_num: int, temperature: float): 16 | self.model = model 17 | self.use_cot = use_cot 18 | self.cot = ChainOfThought() 19 | unformatted = UnformattedLLM() 20 | 21 | few_shot_builder = self.cot if use_cot else unformatted 22 | instruction = COT_INSTRUCTION if use_cot else unformatted.INSTRUCTION 23 | few_shot: Dict[str, str] = few_shot_builder.few_shot_limited(few_shot_num) 24 | cache_infer_params = f'temp={temperature}_cot={use_cot}_fs={few_shot_num}' 25 | cache_model_name = mk_dir_safe(model.name) 26 | params = ParameterizedModel( 27 | temp = temperature, 28 | instruction = instruction, 29 | who_are_you = WHO_ARE_YOU, 30 | few_shot = few_shot, 31 | model = model, 32 | agent_name = f"OAI_Eval_{cache_model_name}_{cache_infer_params}", 33 | ) 34 | self.gpt = GPT(params) 35 | 36 | def answers(self, prb: str) -> List[Tuple[str, str]]: 37 | try: 38 | completion = self.gpt.query([prb], rate_limited = False)[0] 39 | except Exception as e: 40 | if self.known_error(e): 41 | completion = "" 42 | else: 43 | raise e 44 | answer = self.gpt.extract_answer(completion) 45 | if self.use_cot: 46 | answer = self.cot.extract_answer(answer) 47 | return [(answer, completion)] 48 | 49 | def known_error(self, e) -> bool: 50 | KNOWN = [ 51 | "Sorry! We've encountered an issue with repetitive patterns in your prompt. Please try again with a different prompt." 52 | ] 53 | print(f'API ERROR: {str(e)}') 54 | return str(e) in KNOWN 55 | 56 | @classmethod 57 | def run(cls, name: str, snapshots_specs: str, 58 | cot: bool, few_shot_num: int, temperature: float, 59 | verbose: bool = False, save_snaphot: bool = False, extra_params: Dict[str, Any] = {}): 60 | known_model = OAI_MODELS.SHORT_NAMES[name] 61 | answerer = Answers(known_model, cot, few_shot_num, temperature) 62 | e = EvalRunner(answerer, snapshots_specs, verbose).do() 63 | if save_snaphot: 64 | answerer.gpt.snapshot_api_query_cache() 65 | return e 66 | 67 | 68 | if __name__ == "__main__": 69 | parser = argparse.ArgumentParser() 70 | parser.add_argument("--model", required=True, 71 | help = f"Model has to be one of {list(OAI_MODELS.SHORT_NAMES.keys())}") 72 | parser.add_argument("--snapshots_specs", type=str, default=DEFAULT_SNAPSHOTS_FILE, 73 | help = f"JSON of static and monthly snapshots") 74 | parser.add_argument("--verbose", action='store_true', 75 | help="Outcomes of each test, one per line, on console") 76 | parser.add_argument("--use_chain_of_thought", action='store_true', 77 | help="Use chain of thought instruction and postprocessing") 78 | parser.add_argument("--few_shot_num", type=str, default=DEFAULT_FEW_SHOT_NUM, 79 | help=f"Default few shot count: {DEFAULT_FEW_SHOT_NUM}") 80 | parser.add_argument("--temperature", type=float, default=DEFAULT_TEMPERATURE, 81 | help=f"Default temperature: {DEFAULT_TEMPERATURE}") 82 | parser.add_argument("--save_snapshot", action='store_true', 83 | help="Save API query snapshot as .tar.gz file") 84 | args = parser.parse_args() 85 | 86 | Answers.run(args.model, args.snapshots_specs, 87 | args.use_chain_of_thought, args.few_shot_num, args.temperature, 88 | args.verbose, args.save_snapshot) 89 | -------------------------------------------------------------------------------- /model_api/offline.py: -------------------------------------------------------------------------------- 1 | from eval.runner import EvalRunner, mk_dir_safe 2 | import argparse 3 | import sys 4 | import json 5 | from typing import Dict, List, Tuple, Optional, Any 6 | from math_utils.math_helpers import rm_latex_math 7 | from unformatted_llm import UnformattedLLM 8 | from model_api.closed_api import KnownModel, ParameterizedModel 9 | from model_api.file_api import FileAPI, OFFLINE_PARAMS 10 | from evaluate import DEFAULT_SNAPSHOTS_FILE, DEFAULT_FEW_SHOT_NUM, DEFAULT_TEMPERATURE, SPEC_EXTRA_PARAMS_MODE_WRITE_OFFLINE 11 | from chain_of_thought import COT_INSTRUCTION, ChainOfThought 12 | 13 | class Answers: 14 | def __init__(self, model: KnownModel, use_cot: bool, few_shot_num: int, temperature: float, mode_write_questions: bool): 15 | self.model = model 16 | self.use_cot = use_cot 17 | self.cot = ChainOfThought() 18 | unformatted = UnformattedLLM() 19 | 20 | few_shot_builder = self.cot if use_cot else unformatted 21 | instruction = COT_INSTRUCTION if use_cot else unformatted.INSTRUCTION 22 | few_shot: Dict[str, str] = few_shot_builder.few_shot_limited(few_shot_num) 23 | cache_infer_params = f'temp={temperature}_cot={use_cot}_fs={few_shot_num}' 24 | cache_model_name = mk_dir_safe(model.name) 25 | params = ParameterizedModel( 26 | temp = temperature, 27 | instruction = instruction, 28 | who_are_you = "", 29 | few_shot = few_shot, 30 | model = model, 31 | agent_name = f"Offline_Eval_{cache_model_name}_{cache_infer_params}", 32 | ) 33 | self.file_api = FileAPI(params, mode_write_questions) 34 | 35 | def answers(self, prb: str) -> List[Tuple[str, str]]: 36 | completion = self.file_api.query([prb], rate_limited = False)[0] 37 | answer = self.file_api.extract_answer(completion) 38 | if self.use_cot: 39 | answer = self.cot.extract_answer(answer) 40 | return [(answer, completion)] 41 | 42 | def write_questions(self, e): 43 | def write(data): 44 | fname = self.file_api.params.agent_name + '.json' 45 | with open(fname, "w") as jfile: 46 | json.dump(data, jfile, indent=4) 47 | 48 | duplicates = 0 49 | for ref in e.problems_tested: 50 | matched = None 51 | for prompt in self.file_api.offline_prompts: 52 | if ref['problem'] in prompt['prompt']: 53 | if matched and not ref['is_static']: 54 | duplicates += 1 55 | matched = prompt 56 | assert matched, f'Failed to match: {prompt}' 57 | for k, v in matched.items(): 58 | ref[k] = v 59 | print(f'Duplicates found: {duplicates}. Investigate!') 60 | 61 | write(e.problems_tested) 62 | 63 | @classmethod 64 | def run(cls, name: str, snapshots_specs: str, 65 | cot: bool, few_shot_num: int, temperature: float, 66 | verbose: bool = False, save_snaphot: bool = False, extra_params: Dict[str, Any] = {}): 67 | mode_write_questions = extra_params[SPEC_EXTRA_PARAMS_MODE_WRITE_OFFLINE] if SPEC_EXTRA_PARAMS_MODE_WRITE_OFFLINE in extra_params else False 68 | known_model = OFFLINE_PARAMS.SHORT_NAMES[name] 69 | answerer = Answers(known_model, cot, few_shot_num, temperature, mode_write_questions) 70 | e = EvalRunner(answerer, snapshots_specs, verbose).do() 71 | if mode_write_questions: 72 | answerer.write_questions(e) 73 | return None 74 | if save_snaphot: 75 | answerer.file_api.snapshot_api_query_cache() 76 | return e 77 | 78 | 79 | if __name__ == "__main__": 80 | parser = argparse.ArgumentParser() 81 | parser.add_argument("--model", required=True, 82 | help = f"Model has to be one of {list(OFFLINE_PARAMS.SHORT_NAMES.keys())}") 83 | parser.add_argument("--snapshots_specs", type=str, default=DEFAULT_SNAPSHOTS_FILE, 84 | help = f"JSON of static and monthly snapshots") 85 | parser.add_argument("--verbose", action='store_true', 86 | help="Outcomes of each test, one per line, on console") 87 | parser.add_argument("--use_chain_of_thought", action='store_true', 88 | help="Use chain of thought instruction and postprocessing") 89 | parser.add_argument("--few_shot_num", type=str, default=DEFAULT_FEW_SHOT_NUM, 90 | help=f"Default few shot count: {DEFAULT_FEW_SHOT_NUM}") 91 | parser.add_argument("--temperature", type=float, default=DEFAULT_TEMPERATURE, 92 | help=f"Default temperature: {DEFAULT_TEMPERATURE}") 93 | parser.add_argument("--save_snapshot", action='store_true', 94 | help="Save API query snapshot as .tar.gz file") 95 | parser.add_argument("--mode_write_questions", action='store_true', 96 | help="Write questions to output dump, to be solved offline") 97 | args = parser.parse_args() 98 | 99 | extra_params = { SPEC_EXTRA_PARAMS_MODE_WRITE_OFFLINE: args.mode_write_questions } 100 | Answers.run(args.model, args.snapshots_specs, 101 | args.use_chain_of_thought, args.few_shot_num, args.temperature, 102 | args.verbose, args.save_snapshot, extra_params) 103 | -------------------------------------------------------------------------------- /model_api/oss.py: -------------------------------------------------------------------------------- 1 | import together # type: ignore 2 | import os 3 | from model_api.closed_api import ClosedAPI, RateLimited, KnownModel, PERIODS_PER_MIN, ParameterizedModel 4 | from unformatted_llm import UnformattedLLM 5 | from typing import Any, Awaitable, Dict, Optional, List, Tuple 6 | from tqdm import tqdm # type: ignore 7 | 8 | TOGETHER_API_KEY = os.getenv("TOGETHER_API_KEY") 9 | END_TAG = "\n~~~\n" 10 | DEFAULT_END_TAGS = [END_TAG, ''] 11 | DEFAULT_PROMPT_FORMAT = ": {prompt}\n: {response}" 12 | MAX_TOKENS = 1024 13 | 14 | # Model list: https://docs.together.ai/docs/inference-models 15 | WORTH_IT_LANG = [ 16 | "EleutherAI/llemma_7b", 17 | "zero-one-ai/Yi-34B", 18 | "Open-Orca/Mistral-7B-OpenOrca", 19 | "zero-one-ai/Yi-34B", 20 | "meta-llama/Llama-2-70b-hf", 21 | "Qwen/Qwen1.5-72B", 22 | "microsoft/phi-2", 23 | "google/gemma-7b", 24 | "togethercomputer/RedPajama-INCITE-7B-Instruct", 25 | "allenai/OLMo-7B-Instruct", 26 | 27 | "mistralai/Mixtral-8x7B-v0.1", 28 | "togethercomputer/StripedHyena-Hessian-7B", 29 | "WizardLM/WizardLM-70B-V1.0", 30 | ] 31 | 32 | 33 | WORTH_IT_CODE = [ 34 | "togethercomputer/CodeLlama-34b-Python", 35 | "togethercomputer/CodeLlama-34b", 36 | "WizardLM/WizardCoder-Python-34B-V1.0", 37 | ] 38 | 39 | WORTH_IT_CHAT = [ 40 | "zero-one-ai/Yi-34B-Chat", 41 | "togethercomputer/CodeLlama-34b-Instruct", 42 | "togethercomputer/llama-2-70b-chat", 43 | "mistralai/Mixtral-8x7B-Instruct-v0.1", 44 | "Open-Orca/Mistral-7B-OpenOrca", 45 | "togethercomputer/Pythia-Chat-Base-7B-v0.16", 46 | "togethercomputer/Qwen-7B-Chat", 47 | "togethercomputer/StripedHyena-Nous-7B", 48 | "lmsys/vicuna-13b-v1.5", 49 | "togethercomputer/CodeLlama-34b-Instruct", 50 | "upstage/SOLAR-0-70b-16bit", 51 | ] 52 | 53 | def worth_it_exclude(names: List[str]) -> Tuple[List[str], List[str]]: 54 | keep, exclude = [], [] 55 | for n in names: 56 | if n in WORTH_IT_CHAT or n in WORTH_IT_LANG or n in WORTH_IT_CODE: 57 | keep.append(n) 58 | else: 59 | exclude.append(n) 60 | return keep, exclude 61 | 62 | class TOGETHER_PARAMS: 63 | def __init__(self): 64 | models_names: List[str] = [m['name'] for m in together.Models.list()] 65 | print(f'## Available models: {len(models_names)}\n') 66 | 67 | # keep the "excluded" print, it'll show whats available on together as new models get added 68 | worth_it, excluded = worth_it_exclude(models_names) 69 | print(f'## Worth trying: {worth_it}\n') 70 | print(f'## Excluded: {excluded}\n') 71 | 72 | self.models: Dict[str, KnownModel] = { n: self.known_from_config(n) for n in tqdm(worth_it, desc = 'Getting info for "worth it" models') } 73 | 74 | def known_from_config(self, name: str) -> KnownModel: 75 | info = together.Models.info(name) 76 | 77 | # https://docs.together.ai/docs 78 | # >>> together.Models.info(model='mistralai/Mixtral-8x7B-v0.1') 79 | # info: {'modelInstanceConfig': {'appearsIn': [], 'order': 0}, '_id': '6577bf1034e6c1e2bb5283d9', 'name': 'mistralai/Mixtral-8x7B-v0.1', 'display_name': 'Mixtral-8x7B', 'display_type': 'language', 'description': 'The Mixtral-8x7B Large Language Model (LLM) is a pretrained generative Sparse Mixture of Experts.', 'license': 'apache-2.0', 'link': 'https://huggingface.co/mistralai/Mixtral-8x7B-v0.1', 'creator_organization': 'mistralai', 'pricing_tier': 'Featured', 'access': 'open', 'num_parameters': '56000000000', 'show_in_playground': True, 'isFeaturedModel': True, 'context_length': 32768, 'pricing': {'input': 150, 'output': 150, 'hourly': 0}, 'created_at': '2023-12-12T02:01:52.674Z', 'update_at': '2023-12-12T02:01:52.674Z', 'instances': [{'avzone': 'us-east-1a', 'cluster': 'happypiglet'}], 'renamed': 'mistralai/mixtral-8x7b-32kseqlen', 'hardware_label': '', 'descriptionLink': '', 'depth': {'num_asks': 1, 'num_bids': 0, 'num_running': 0, 'qps': 0, 'throughput_in': 0, 'throughput_out': 0, 'error_rate': 0, 'retry_rate': 0, 'stats': [{'avzone': 'us-east-1a', 'cluster': 'happypiglet', 'capacity': 0, 'qps': 0, 'throughput_in': 0, 'throughput_out': 0, 'error_rate': 0, 'retry_rate': 0}]}} 80 | 81 | is_chat = info["display_type"] == 'chat' 82 | 83 | stops: List[str] = DEFAULT_END_TAGS 84 | prompt_format: str = DEFAULT_PROMPT_FORMAT 85 | # override the stops and prompt_format if they are specified (which is the case for chat models) 86 | if is_chat: 87 | # >>> together.Models.info(model='mistralai/Mixtral-8x7B-Instruct-v0.1')["config"] 88 | # {'stop': ['', '[INST]'], 'prompt_format': '[INST] {prompt} [/INST]', 'chat_template_name': 'llama'} 89 | stops = info["config"]["stop"] 90 | prompt_format = info["config"]["prompt_format"] + "{response}" 91 | 92 | return KnownModel(name, 93 | is_chat, 94 | api_org = None, 95 | api_key = TOGETHER_API_KEY, 96 | api_base = "https://api.together.xyz", 97 | stops = stops, 98 | prompt_format = prompt_format) 99 | 100 | 101 | class OSS(ClosedAPI): 102 | def __init__(self, params: ParameterizedModel): 103 | super().__init__(params) 104 | 105 | self.params = params 106 | self.prelude = self.few_shot_format(params.few_shot) 107 | 108 | def deobject(self, obj_resp: Any) -> str: 109 | extract = lambda output: output['output']['choices'][0]['text'] 110 | completion = obj_resp if isinstance(obj_resp, str) else extract(obj_resp) 111 | assert isinstance(completion, str), f'model output is not str: {obj_resp}' 112 | return completion 113 | 114 | def extract_answer(self, response) -> str: 115 | completion = response.lstrip() 116 | assert self.params.model.stops, f'stop tokens not set' 117 | ends = [completion.find(s) for s in self.params.model.stops if completion.find(s) != -1] 118 | answer = completion[:ends[0]] if ends else completion 119 | answer = answer.strip() 120 | return answer 121 | 122 | def to_prompt_task(self, task: str) -> Any: 123 | return self.prelude + self.params.instruction + self.format(task, None) 124 | 125 | def reset_library(self, model: KnownModel) -> None: 126 | return 127 | 128 | def model_ask(self, prompt: str) -> Awaitable[Any]: 129 | return self.async_ask(prompt) 130 | 131 | async def async_ask(self, prompt: str): 132 | # print(f'Sending query: {prompt}') 133 | return together.Complete.create( 134 | model = self.params.model.name, 135 | max_tokens = MAX_TOKENS, 136 | temperature = self.params.temperature, 137 | prompt = prompt, 138 | stop = self.params.model.stops 139 | ) 140 | 141 | def spent(self, response) -> float: 142 | return 0.0 143 | 144 | def format(self, inp: str, out: Optional[str]) -> str: 145 | resp = out if out else "" 146 | assert self.params.model.prompt_format, f'prompt format not set' 147 | formatted = self.params.model.prompt_format.format(prompt = inp, response = resp) 148 | return formatted 149 | 150 | def few_shot_format(self, ios: Dict[str, str]) -> str: 151 | fs = "" 152 | for inp, out in ios.items(): 153 | assert self.params.model.stops, f'stop tokens not set' 154 | fs += self.format(inp, out) + '\n' + self.params.model.stops[0] 155 | return fs 156 | 157 | -------------------------------------------------------------------------------- /model_api/tog.py: -------------------------------------------------------------------------------- 1 | from eval.runner import EvalRunner, mk_dir_safe 2 | import argparse 3 | import sys 4 | from typing import Dict, List, Tuple, Any 5 | from math_utils.math_helpers import rm_latex_math 6 | from unformatted_llm import UnformattedLLM 7 | from model_api.closed_api import KnownModel, ParameterizedModel 8 | from model_api.oss import OSS, TOGETHER_PARAMS 9 | from evaluate import DEFAULT_SNAPSHOTS_FILE, DEFAULT_FEW_SHOT_NUM, DEFAULT_TEMPERATURE 10 | from chain_of_thought import COT_INSTRUCTION, ChainOfThought 11 | 12 | _TOGETHER = TOGETHER_PARAMS() 13 | 14 | KNOWN_CRASH="400 Client Error: Bad Request for url: https://api.together.xyz/api/inference" 15 | 16 | class Answers: 17 | def __init__(self, model: KnownModel, use_cot: bool, few_shot_num: int, temperature: float): 18 | self.model = model 19 | self.use_cot = use_cot 20 | self.cot = ChainOfThought() 21 | unformatted = UnformattedLLM() 22 | 23 | few_shot_builder = self.cot if use_cot else unformatted 24 | instruction = COT_INSTRUCTION if use_cot else unformatted.INSTRUCTION 25 | few_shot: Dict[str, str] = few_shot_builder.few_shot_limited(few_shot_num) 26 | cache_infer_params = f'temp={temperature}_cot={use_cot}_fs={few_shot_num}' 27 | cache_model_name = mk_dir_safe(model.name) 28 | params = ParameterizedModel( 29 | temp = temperature, 30 | instruction = instruction, 31 | who_are_you = "", 32 | few_shot = few_shot, 33 | model = model, 34 | agent_name = f"Together_Eval_{cache_model_name}_{cache_infer_params}", 35 | ) 36 | self.oss = OSS(params) 37 | 38 | def answers(self, prb: str) -> List[Tuple[str, str]]: 39 | try: 40 | completion = self.oss.query([prb], rate_limited = False)[0] 41 | except Exception as he: 42 | if f'{he}' == KNOWN_CRASH: 43 | # print(f'[WARN] CRASH on prb below\n-----\n{prb}\n-----\nIgnoring and Continue!') 44 | completion = "" 45 | else: 46 | raise he 47 | answer = self.oss.extract_answer(completion) 48 | if self.use_cot: 49 | answer = self.cot.extract_answer(answer) 50 | return [(answer, completion)] 51 | 52 | @classmethod 53 | def run(cls, name: str, snapshots_specs: str, 54 | cot: bool, few_shot_num: int, temperature: float, 55 | verbose: bool = False, save_snaphot: bool = False, extra_params: Dict[str, Any] = {}): 56 | known_model = _TOGETHER.models[name] 57 | answerer = Answers(known_model, cot, few_shot_num, temperature) 58 | e = EvalRunner(answerer, snapshots_specs, verbose).do() 59 | if save_snaphot: 60 | answerer.oss.snapshot_api_query_cache() 61 | return e 62 | 63 | 64 | if __name__ == "__main__": 65 | parser = argparse.ArgumentParser() 66 | parser.add_argument("--model", required=True, 67 | help = f"Model has to be one of {list(_TOGETHER.models.keys())}") 68 | parser.add_argument("--snapshots_specs", type=str, default=DEFAULT_SNAPSHOTS_FILE, 69 | help = f"JSON of static and monthly snapshots") 70 | parser.add_argument("--verbose", action='store_true', 71 | help="Outcomes of each test, one per line, on console") 72 | parser.add_argument("--use_chain_of_thought", action='store_true', 73 | help="Use chain of thought instruction and postprocessing") 74 | parser.add_argument("--few_shot_num", type=str, default=DEFAULT_FEW_SHOT_NUM, 75 | help=f"Default few shot count: {DEFAULT_FEW_SHOT_NUM}") 76 | parser.add_argument("--temperature", type=float, default=DEFAULT_TEMPERATURE, 77 | help=f"Default temperature: {DEFAULT_TEMPERATURE}") 78 | parser.add_argument("--save_snapshot", action='store_true', 79 | help="Save API query snapshot as .tar.gz file") 80 | args = parser.parse_args() 81 | 82 | Answers.run(args.model, args.snapshots_specs, 83 | args.use_chain_of_thought, args.few_shot_num, args.temperature, 84 | args.verbose, args.save_snapshot) 85 | -------------------------------------------------------------------------------- /monthly_snapshots.json: -------------------------------------------------------------------------------- 1 | [ 2 | { 3 | "benchmark": "MATH", 4 | "fixed": "https://people.eecs.berkeley.edu/~hendrycks/MATH.tar", 5 | "functionals": [ 6 | { 7 | "date": "Dec-2023", 8 | "url": "Dec-2023.tar.gz" 9 | }, 10 | { 11 | "date": "Nov-2023", 12 | "url": "Nov-2023.tar.gz" 13 | }, 14 | { 15 | "date": "Oct-2023", 16 | "url": "Oct-2023.tar.gz" 17 | } 18 | ] 19 | }, 20 | { 21 | "benchmark": "GSM8k", 22 | "static": "https://", 23 | "functionals": [] 24 | } 25 | ] 26 | -------------------------------------------------------------------------------- /monthly_snapshots_only_oct.json: -------------------------------------------------------------------------------- 1 | [ 2 | { 3 | "benchmark": "MATH", 4 | "fixed": "https://people.eecs.berkeley.edu/~hendrycks/MATH.tar", 5 | "functionals": [ 6 | { 7 | "date": "Oct-2023", 8 | "url": "Oct-2023.tar.gz" 9 | } 10 | ] 11 | }, 12 | { 13 | "benchmark": "GSM8k", 14 | "static": "https://", 15 | "functionals": [] 16 | } 17 | ] 18 | -------------------------------------------------------------------------------- /persist.py: -------------------------------------------------------------------------------- 1 | import pickle 2 | import os 3 | 4 | class Persist: 5 | @classmethod 6 | def overwrite(cls, fname) -> bool: 7 | ask = lambda: input(f'{fname} exists. Overwrite? [y/n]') 8 | if os.path.exists(fname): 9 | yn = ask() 10 | while not (yn == 'y' or yn == 'n'): 11 | yn = ask() 12 | return yn == 'y' 13 | return True 14 | 15 | @classmethod 16 | def save(cls, obj, fname, force_overwrite = False): 17 | if not force_overwrite: 18 | # ask; if file already present and we need to overwrite 19 | if not Persist.overwrite(fname): 20 | print(f'Will not overwrite {fname}. Save failed!') 21 | return 22 | with open(fname, 'wb') as handle: 23 | pickle.dump(obj, handle, protocol=pickle.HIGHEST_PROTOCOL) 24 | 25 | @classmethod 26 | def load(cls, fname): 27 | with open(fname, 'rb') as handle: 28 | obj = pickle.load(handle) 29 | return obj 30 | -------------------------------------------------------------------------------- /query_cache.py: -------------------------------------------------------------------------------- 1 | import uuid 2 | import os 3 | from typing import Optional, List 4 | import time 5 | from helper_utils import targz, untargz 6 | 7 | CACHE_LOC = "api_query_cache" 8 | 9 | 10 | class UniqId: 11 | def __init__(self, txt: List[str]): 12 | s = "\n".join(txt) 13 | # uuid3 does a MD5 hash; use uuid5 if you need SHA1 14 | self.uuid = uuid.uuid3(uuid.NAMESPACE_OID, s) 15 | 16 | def __repr__(self) -> str: 17 | return str(self.uuid) 18 | 19 | 20 | class QueryCache: 21 | def __init__(self, name): 22 | self.dir = os.path.join(CACHE_LOC, name) 23 | self.ensure_cache_dir() 24 | 25 | def ensure_cache_dir(self): 26 | self.ensure_exists(CACHE_LOC) 27 | self.load_snapshot_or_empty() 28 | 29 | def load_snapshot_or_empty(self): 30 | snapshot = self.snapshot_filename() 31 | if not os.path.exists(self.dir): 32 | if os.path.exists(snapshot) and self.confirm(f'Cache snapshot {snapshot} exists. Initialize from it (highly recommended)'): 33 | self.load_snapshot() 34 | else: 35 | self.ensure_exists(self.dir) 36 | return 37 | 38 | def save_snapshot(self): 39 | save_loc = self.snapshot_filename() 40 | do_save = self.confirm(f"Save query snapshot to {save_loc}") 41 | if not do_save: 42 | return 43 | targz(self.dir) 44 | assert os.path.exists(save_loc), f'Save should have created {save_loc}. Did not.' 45 | 46 | def load_snapshot(self): 47 | untargz(self.snapshot_filename(), self.dir) 48 | assert os.path.exists(self.dir), f'Load should have created {self.dir}. Did not.' 49 | 50 | def confirm(self, msg: str) -> bool: 51 | do = input(msg + "[y]/n? ") 52 | return do != "n" 53 | 54 | def snapshot_filename(self): 55 | return str(self.dir) + ".tar.gz" 56 | 57 | def ensure_exists(self, d): 58 | if not os.path.exists(d): 59 | os.mkdir(d) 60 | 61 | def put(self, q: str, a: str) -> bool: 62 | old = self.get(q) 63 | if old is not None and old != a: 64 | print(f'[WARN] Cache already has map for "{q[:20]}.." -> "{old[:20]}..", but asking to overwrite with different {a[:20]}.. Overwrite ignored!') 65 | return False 66 | uuid = UniqId([q]) 67 | self.cache_write(uuid, a) 68 | return True 69 | 70 | def get(self, q: str) -> Optional[str]: 71 | uuid = UniqId([q]) 72 | return self.cache_get(uuid) if self.cache_exists(uuid) else None 73 | 74 | def cache_get(self, uuid: UniqId) -> str: 75 | with open(self.fname(uuid), 'r') as f: 76 | return f.read() 77 | 78 | def cache_write(self, uuid: UniqId, val: str) -> None: 79 | with open(self.fname(uuid), 'w') as f: 80 | f.write(val) 81 | 82 | def cache_exists(self, uuid: UniqId) -> bool: 83 | return os.path.exists(self.fname(uuid)) 84 | 85 | def fname(self, uuid: UniqId) -> str: 86 | return os.path.join(self.dir, f'{uuid}') 87 | 88 | def purge(self): 89 | for c in os.listdir(self.dir): 90 | cache_file = os.path.join(self.dir, c) 91 | print(f"Purging {cache_file}") 92 | time.sleep(1) 93 | os.remove(cache_file) 94 | 95 | 96 | def test_cache_write() -> None: 97 | q12 = "question 1 2" 98 | q23 = "question 2 3" 99 | a12 = "answer 1 2" 100 | a23 = "answer 2 3" 101 | cache = QueryCache("TEST") 102 | 103 | cache.purge() 104 | 105 | # empty cache 106 | assert cache.get(q12) == None 107 | assert cache.get(q23) == None 108 | 109 | # put a value and retrieve it 110 | assert cache.put(q12, a12) == True 111 | assert cache.get(q12) == a12 112 | 113 | # put new cache value, make sure old value did not change 114 | assert cache.put(q23, a23) == True 115 | assert cache.get(q23) == a23 116 | assert cache.get(q12) == a12 117 | 118 | # fail to overwrite 119 | assert cache.put(q23, a12) == False 120 | assert cache.get(q23) == a23 121 | -------------------------------------------------------------------------------- /report/.gitignore: -------------------------------------------------------------------------------- 1 | *aux 2 | *blg 3 | *bbl 4 | *log 5 | main.pdf 6 | -------------------------------------------------------------------------------- /report/Makefile: -------------------------------------------------------------------------------- 1 | main.pdf: main.tex refs.bib 2 | pdflatex main 3 | bibtex main 4 | pdflatex main 5 | bibtex main 6 | pdflatex main 7 | 8 | clean: 9 | rm main.pdf main.aux main.log main.bbl main.blg 10 | rm missfont.log 11 | rm texput.log 12 | -------------------------------------------------------------------------------- /report/all-coverage.pdf: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/report/all-coverage.pdf -------------------------------------------------------------------------------- /report/all-levels.pdf: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/report/all-levels.pdf -------------------------------------------------------------------------------- /report/all-reasoning-gap.pdf: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/report/all-reasoning-gap.pdf -------------------------------------------------------------------------------- /report/all-static-vs-func-accuracy.pdf: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/report/all-static-vs-func-accuracy.pdf -------------------------------------------------------------------------------- /report/all-subjects.pdf: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/report/all-subjects.pdf -------------------------------------------------------------------------------- /report/data_files.tar.gz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/report/data_files.tar.gz -------------------------------------------------------------------------------- /report/gpt4-by-level.pdf: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/report/gpt4-by-level.pdf -------------------------------------------------------------------------------- /report/gpt4-by-subject.pdf: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/report/gpt4-by-subject.pdf -------------------------------------------------------------------------------- /report/gpt4-k=3-suffices.pdf: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/report/gpt4-k=3-suffices.pdf -------------------------------------------------------------------------------- /report/gpt4-levels.pdf: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/report/gpt4-levels.pdf -------------------------------------------------------------------------------- /report/gpt4-reasoning-needed.pdf: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/report/gpt4-reasoning-needed.pdf -------------------------------------------------------------------------------- /report/gpt4-subjects.pdf: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/report/gpt4-subjects.pdf -------------------------------------------------------------------------------- /report/gpt4-surprise.pdf: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/report/gpt4-surprise.pdf -------------------------------------------------------------------------------- /report/oss-by-level.pdf: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/report/oss-by-level.pdf -------------------------------------------------------------------------------- /report/oss-by-subject.pdf: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/report/oss-by-subject.pdf -------------------------------------------------------------------------------- /report/oss-k=3-suffices.pdf: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/report/oss-k=3-suffices.pdf -------------------------------------------------------------------------------- /report/oss-reasoning-needed.pdf: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/report/oss-reasoning-needed.pdf -------------------------------------------------------------------------------- /report/oss-surprise.pdf: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ConsequentAI/fneval/a89e561466dac1d0a376aaad9d84fa366a9f8e93/report/oss-surprise.pdf -------------------------------------------------------------------------------- /requirements.txt: -------------------------------------------------------------------------------- 1 | pytest>=7.3.1 2 | codetiming 3 | openai==0.28.1 4 | anthropic==0.18.1 5 | numpy 6 | together 7 | mistralai 8 | -------------------------------------------------------------------------------- /summarize_evals.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | import inspect 3 | import sys 4 | import os 5 | import json 6 | import shutil 7 | from persist import Persist 8 | from typing import List, Dict, Any, Set, Tuple 9 | from eval.measure_math import Eval, Metrics, SUBJECTS 10 | from eval.runner import mk_dir_safe 11 | 12 | from evaluate import DEFAULT_EVAL_PICKLE_FILE, ModelSpec 13 | 14 | COVERAGE_PC_TOO_LOW_IF_BELOW = 70 15 | PC = lambda num, dim: 100.0 * (num / float(dim) if dim else 0.0) 16 | 17 | 18 | def write_lines(lines: List[Any], prefix, ext = "txt"): 19 | frm = inspect.stack()[1][3] 20 | outfile = f"{prefix}.{frm}.{ext}" 21 | with open(outfile, 'w') as f: 22 | f.write("\n".join(lines)) 23 | print(f'Written to {outfile}') 24 | 25 | 26 | def solved_by_models_across_snapshots(frac_models, es, prefix): 27 | model_solved: Dict[str, List[str]] = { model: e.get_solved(only_static = False) for model, e in es.items() } 28 | solved: List[List[str]] = [s for _, s in model_solved.items()] 29 | 30 | counts: Dict[str, int] = {} 31 | def incr(s, d): 32 | if s not in d: 33 | d[s] = 0 34 | d[s] += 1 35 | for solns in solved: 36 | for soln in solns: 37 | incr(soln, counts) 38 | 39 | threshold = len(es) * frac_models 40 | above_threshold = [s for s, times in counts.items() if times >= threshold] 41 | 42 | solved_solns = sorted(above_threshold) 43 | write_lines(solved_solns, prefix + f'_{frac_models}') 44 | 45 | 46 | def fraction_functionally_tested(static, func) -> Tuple[int, int]: 47 | static_correct = [tm.fname for tm in static.individuals if tm.correct] 48 | func_tested = [tm.fname for tm in func.individuals if tm.fname in static_correct] 49 | num, dim = len(func_tested), len(static_correct) 50 | return num, dim 51 | 52 | 53 | def reasoning_gap(static, combined) -> Tuple[int, int]: 54 | st = PC(*static.accuracy()) 55 | cb = PC(*combined.accuracy()) 56 | return st-cb, st 57 | 58 | 59 | def model_specs_to_str(evals_spec: Dict[ModelSpec, Eval]) -> Dict[str, Eval]: 60 | to_str = lambda ms: ms.spec_csv()[1] 61 | evals: Dict[str, Eval] = { to_str(ms): e for ms, e in evals_spec.items() } 62 | return evals 63 | 64 | 65 | def stat_solved_by_majority_models_across_snapshots(evaluations, prefix, extra): 66 | es = model_specs_to_str(evaluations) 67 | solved_by_models_across_snapshots(0.5, es, prefix) 68 | 69 | 70 | def stat_solved_by_all_models_across_snapshots(evaluations, prefix, extra): 71 | es = model_specs_to_str(evaluations) 72 | solved_by_models_across_snapshots(1.0, es, prefix) 73 | 74 | 75 | def stat_solved_statics(evaluations, prefix, extra): 76 | es = model_specs_to_str(evaluations) 77 | solved: Dict[str, List[str]] = { model: e.get_solved(only_static = True) for model, e in es.items() } 78 | solved_solns = sorted(list(set(s for _, solns in solved.items() for s in solns))) 79 | write_lines(solved_solns, prefix) 80 | 81 | 82 | def static_func_combined(model, evals) -> Tuple[Any, Any, Any, Any, str, str, bool]: 83 | static = evals.static_metrics 84 | func = evals.func_metrics 85 | verbose = False 86 | model_name = mk_dir_safe(model) 87 | combined, func_of_static = static.with_functional(func, model_name, verbose) 88 | gap = PC(*reasoning_gap(static, combined)) 89 | gap_tag = f'{gap:.2f}%' 90 | pc_fn = PC(*fraction_functionally_tested(static, func)) 91 | pc_fn_tag = f"{pc_fn:.2f}%" 92 | cover_too_low = pc_fn <= COVERAGE_PC_TOO_LOW_IF_BELOW 93 | warn_low_cover = pc_fn_tag + (" (too low)" if cover_too_low else "") 94 | return static, func, func_of_static, combined, gap_tag, warn_low_cover, cover_too_low 95 | 96 | 97 | def stat_accuracy(evaluations, prefix, extra): 98 | es = model_specs_to_str(evaluations) 99 | hr = '-' * 100 100 | accuracies: List[str] = [] 101 | for model, evals in es.items(): 102 | if not accuracies: 103 | legend = evals.static_metrics.stats_legend() 104 | accuracies.append(legend) 105 | static, func, func_sub_static, combined, gap, warn_low_cover, cover_too_low = static_func_combined(model, evals) 106 | if cover_too_low: 107 | print(f'Coverage too low ({warn_low_cover}) for {model}. Still writing raw data to file.') 108 | gap_warn = ' (please dont cite; coverage too low)' if cover_too_low else '' 109 | accuracies += [ 110 | hr, model, hr, 111 | 'static:', static.stats(), 112 | 'functional:', func.stats(), 113 | 'functional correct amongst static correct:', func_sub_static.stats(), 114 | 'combined:', combined.stats(), 115 | 'reasoning gap:', gap + gap_warn, 116 | 'pc functionally tested:', warn_low_cover, 117 | 'coverage low:', str(cover_too_low), 118 | hr] 119 | write_lines(accuracies, prefix) 120 | 121 | 122 | class ModelAccuracies: 123 | def __init__(self, model, evals): 124 | static, func, func_sub_static, combined, gap, warn_low, cover_too_low = static_func_combined(model, evals) 125 | self.st = PC(*static.accuracy()) 126 | self.fn = PC(*combined.accuracy()) 127 | self.pc_fn = PC(*fraction_functionally_tested(static, func)) 128 | self.st_h = PC(*static.hallucinations()) 129 | self.fn_h = PC(*combined.hallucinations()) 130 | self.fn_correct = PC(*func_sub_static.accuracy()) 131 | self.gap = gap 132 | self.warn_low = warn_low 133 | self.cover_too_low = cover_too_low 134 | 135 | HDRS = 'Static,Func,Frac Func Tested,Static Hall,Func Hall,Gap,Fn Coverage,Correct Amongst Fn Tested' 136 | 137 | def __repr__(self): 138 | return f'{self.st:.2f}%,{self.fn:.2f}%,{self.pc_fn:.2f}%,{self.st_h:.2f}%,{self.fn_h:.2f}%,{self.gap},{self.warn_low},{self.fn_correct:.2f}%' 139 | 140 | def stat_csv_static_func(evaluations, prefix, extra): 141 | es = model_specs_to_str(evaluations) 142 | force_write_low_covers = 'FORCE_WRITE_LOW_COVERS' in extra 143 | 144 | model_spec = next(iter(evaluations)).spec_csv()[0] 145 | accuracies: List[str] = [f'{model_spec},{ModelAccuracies.HDRS}'] 146 | some_skipped = False 147 | 148 | for model, evals in es.items(): 149 | model_accs = ModelAccuracies(model, evals) 150 | if model_accs.cover_too_low: 151 | print(f'Coverage too low ({model_accs.warn_low}) for {model}.') 152 | if not force_write_low_covers: 153 | some_skipped = True 154 | continue 155 | accuracies.append(f'{model},{model_accs}') 156 | 157 | if some_skipped: 158 | print('Low coverage values skipped; add extra flag to include them: --extra \'{"FORCE_WRITE_LOW_COVERS": "True"}\'') 159 | write_lines(accuracies, prefix, ext = "csv") 160 | 161 | 162 | def stat_effect_of_cot(evaluations, prefix, extra): 163 | without_cot = {} 164 | with_cot = {} 165 | for model_spec, evals in evaluations.items(): 166 | if model_spec.chain_of_thought: 167 | with_cot[model_spec] = evals 168 | else: 169 | without_cot[model_spec] = evals 170 | 171 | def find_with_cot(ms: ModelSpec): 172 | for ms_with, evals in with_cot.items(): 173 | if ms_with.model == ms.model and \ 174 | ms_with.few_shot_num == ms.few_shot_num and \ 175 | ms_with.temperature == ms.temperature: 176 | return ms_with, evals 177 | return None, None 178 | 179 | to_str = lambda ms: ms.spec_csv()[1] 180 | hdrs = ModelAccuracies.HDRS 181 | with_and_without = [f'model,fs,temp,{hdrs},{hdrs}'] 182 | for ms, evals in without_cot.items(): 183 | with_ms, with_evals = find_with_cot(ms) 184 | if not with_ms: 185 | continue 186 | 187 | model_desc = f'{ms.model},{ms.few_shot_num},{ms.temperature}' 188 | with_nums = ModelAccuracies(to_str(with_ms), with_evals) 189 | without_nums = ModelAccuracies(to_str(ms), evals) 190 | 191 | line = f'{model_desc},{without_nums},{with_nums}' 192 | with_and_without.append(line) 193 | 194 | write_lines(with_and_without, prefix, ext = "csv") 195 | 196 | 197 | def stat_effect_of_fewshot(evaluations, prefix, extra): 198 | without_fs = {} 199 | with_fs = {} 200 | for model_spec, evals in evaluations.items(): 201 | if model_spec.few_shot_num == 0: 202 | without_fs[model_spec] = evals 203 | else: 204 | with_fs[model_spec] = evals 205 | 206 | def find_with_fs(ms: ModelSpec): 207 | for ms_with, evals in with_fs.items(): 208 | if ms_with.model == ms.model and \ 209 | ms_with.chain_of_thought == ms.chain_of_thought and \ 210 | ms_with.temperature == ms.temperature: 211 | return ms_with, evals 212 | return None, None 213 | 214 | to_str = lambda ms: ms.spec_csv()[1] 215 | hdrs = ModelAccuracies.HDRS 216 | with_and_without = [f'model,cot,temp,{hdrs},{hdrs}'] 217 | for ms, evals in without_fs.items(): 218 | with_ms, with_evals = find_with_fs(ms) 219 | if not with_ms: 220 | continue 221 | 222 | model_desc = f'{ms.model},{ms.chain_of_thought},{ms.temperature}' 223 | with_nums = ModelAccuracies(to_str(with_ms), with_evals) 224 | without_nums = ModelAccuracies(to_str(ms), evals) 225 | 226 | line = f'{model_desc},{without_nums},{with_nums}' 227 | with_and_without.append(line) 228 | 229 | write_lines(with_and_without, prefix, ext = "csv") 230 | 231 | 232 | def stat_csv_subject_level(evaluations, prefix, extra): 233 | es = model_specs_to_str(evaluations) 234 | # model -> typ (subject or level) -> (static, func) 235 | ModelTypStatFunc = Dict[str, Dict[Any, Tuple[float, float]]] 236 | 237 | def acc_fn(outcomes: List[bool]): 238 | return PC(outcomes.count(True), len(outcomes)) 239 | 240 | def acc_s(m: Metrics, subject: str): 241 | return acc_fn(m.subject_cors[subject]) 242 | 243 | def acc_l(m: Metrics, lvl: int): 244 | return acc_fn(m.level_cors[lvl]) 245 | 246 | def collapse_models(metrics: ModelTypStatFunc, fn) -> ModelTypStatFunc: 247 | def collapse_tuple(l: List[Tuple[float, float]]) -> Tuple[float, float]: 248 | statics, funcs = list(zip(*l)) 249 | return fn(statics), fn(funcs) 250 | models = list(m for m in metrics) 251 | types = list(t for t in metrics[models[0]]) 252 | return { "all": { t: collapse_tuple(list(metrics[model][t] for model in models)) for t in types } } 253 | 254 | 255 | sub_accs: ModelTypStatFunc = {} 256 | lvl_accs: ModelTypStatFunc = {} 257 | for model, evals in es.items(): 258 | static, _, _, comb, _, _, _ = static_func_combined(model, evals) 259 | s_accs = { sub: (acc_s(static, sub), acc_s(comb, sub)) for sub in SUBJECTS } 260 | l_accs = { lvl: (acc_l(static, lvl), acc_l(comb, lvl)) for lvl in range(1,6) } 261 | sub_accs[model] = s_accs 262 | lvl_accs[model] = l_accs 263 | 264 | def write(tag: str, accs: ModelTypStatFunc): 265 | lines: List[str] = [f"Model,{tag},Static,W Func,Delta"] 266 | for model, accuracies in accs.items(): 267 | for typ, (stat, func) in accuracies.items(): 268 | delta = PC(stat - func, stat) 269 | lines.append(f'{model},{typ},{stat:.2f}%,{func:.2f}%,{delta:.2f}%') 270 | write_lines(lines, prefix + f".{tag}", ext = "csv") 271 | 272 | write("Subject", sub_accs) 273 | write("Level", lvl_accs) 274 | 275 | avg = lambda xs: sum(xs)/len(xs) 276 | write("AllModels.Subject", collapse_models(sub_accs, avg)) 277 | write("AllModels.Level", collapse_models(lvl_accs, avg)) 278 | 279 | 280 | def stat_dropoff(evaluations, prefix, extra): 281 | es = model_specs_to_str(evaluations) 282 | tab = '\t' 283 | def summary_line(s, f): 284 | jsn = f'functional-math/prbs/{s.fname}' 285 | subject, base = s.fname.split('/') 286 | pyid = base[:-len('.json')] 287 | fnpy = f'functional-math/functional/benchmarks/{subject}/m{pyid}.py' 288 | lines = [ 289 | s.answer, 290 | f.answer, 291 | f.output, 292 | f'{f.fraction_correct:.2f}', 293 | f'{s.level}', 294 | jsn, 295 | fnpy 296 | ] 297 | return tab.join(lines) 298 | hdr = tab.join([ 299 | 'static got (= correct)', 300 | 'func correct ref', 301 | 'func got (incorrect)', 302 | 'fraction correct', 303 | 'level', 304 | 'json', 305 | 'functional' 306 | ]) 307 | 308 | def get_static_for(f, static): 309 | return next(tm for tm in static.individuals if tm.fname == f.fname) 310 | 311 | for model, evals in es.items(): 312 | static, func, _, _, _, _, _ = static_func_combined(model, evals) 313 | dropped = [hdr] 314 | for fn in func.individuals: 315 | if not fn.correct: 316 | assert fn.output != fn.answer, 'aggregation failure? some functional incorrect, but diff outs not captured?' 317 | st = get_static_for(fn, static) 318 | if st.correct: 319 | dropped.append(summary_line(st, fn)) 320 | fname = f'dropped.{mk_dir_safe(model)}' 321 | write_lines(dropped, fname, ext = 'tsv') 322 | 323 | 324 | def stat_model_counts(evaluations, prefix, extra): 325 | es = model_specs_to_str(evaluations) 326 | corrects = {} 327 | for model, evals in es.items(): 328 | static, func, _, _, _, _, _ = static_func_combined(model, evals) 329 | corrects[model] = [s.fname for s in static.individuals if s.correct] 330 | cummulative: Dict[str, List[str]] = {} 331 | for model in es: 332 | for c in corrects[model]: 333 | if c not in cummulative: 334 | cummulative[c] = [] 335 | cummulative[c].append(model) 336 | 337 | exclusives = [f'Across all models: {len(cummulative)}'] 338 | only_in_model = {} 339 | for model in es: 340 | only_here = [c for c in cummulative if cummulative[c] == [model]] 341 | only_in_model[model] = only_here 342 | exclusives += [f'Exclusive to {model}: {len(only_here)}'] 343 | write_lines(exclusives, prefix + '.exclusives') 344 | 345 | overlaps = [f'{c},{len(cummulative[c])}' for c in cummulative] 346 | write_lines(overlaps, prefix + '.count_models', ext='csv') 347 | 348 | 349 | def stat_solved_statics_not_functionalized(evaluations, prefix, extra): 350 | es = model_specs_to_str(evaluations) 351 | # root of already functionalized problems 352 | prb_root = extra['ROOT_FUNCMATH_PRBS'] 353 | 354 | def read_functional_math_prbs() -> List[str]: 355 | files = [] 356 | subjects = os.listdir(prb_root) 357 | for s in subjects: 358 | d = os.path.join(prb_root, s) 359 | for f in os.listdir(d): 360 | files.append(os.path.join(s, f)) 361 | return files 362 | 363 | def copy_from_MATH(ufn: str, root: str): 364 | subj, ident = ufn.split('/') 365 | src = os.path.join('MATH', 'test', subj, ident) 366 | dst_dir = os.path.join(root, subj) 367 | dst = os.path.join(dst_dir, ident) 368 | if not os.path.exists(dst_dir): 369 | os.makedirs(dst_dir) 370 | shutil.copyfile(src, dst) 371 | 372 | solved: Dict[str, List[str]] = { model: e.get_solved(only_static = True) for model, e in es.items() } 373 | solved_solns: List[str] = sorted(list(set(s for _, solns in solved.items() for s in solns))) 374 | prbs: List[str] = read_functional_math_prbs() 375 | unfunctionalized = [s for s in solved_solns if not s in prbs] 376 | 377 | # write output /prbs data to this directory 378 | outdir = prefix + '.not_yet_func' 379 | # read jsons and copy to directory structure with root of fn name 380 | ufn_root = os.path.join(outdir, 'prbs') 381 | if not os.path.exists(ufn_root): 382 | os.makedirs(ufn_root) 383 | for ufn in unfunctionalized: 384 | copy_from_MATH(ufn, ufn_root) 385 | print(f'Unfunctionalized MATH test jsons written to: {ufn_root}') 386 | 387 | 388 | if __name__ == "__main__": 389 | parser = argparse.ArgumentParser() 390 | parser.add_argument("--in_file", 391 | help = f"Read pickle file of [ model_name -> eval ]; default = {DEFAULT_EVAL_PICKLE_FILE}") 392 | parser.add_argument("--stat_fn", 393 | help = f"Name of stat fn to run") 394 | parser.add_argument("--extra", 395 | help = f"Opt dict of other parameters") 396 | args = parser.parse_args() 397 | 398 | if not args.stat_fn: 399 | stat_names = [s for s in dir(sys.modules[__name__]) if s.startswith('stat')] 400 | print(f'--stat_fn can take: {stat_names}') 401 | exit(1) 402 | 403 | infile = args.in_file if args.in_file else DEFAULT_EVAL_PICKLE_FILE 404 | 405 | evals: Dict[ModelSpec, Eval] = Persist.load(infile) 406 | print(f'Processing eval data from: {list(evals.keys())}') 407 | stat_fn = getattr(sys.modules[__name__], args.stat_fn) 408 | prefix = infile 409 | extra = json.loads(args.extra) if args.extra else {} 410 | 411 | # run the requested stat fn 412 | stat_fn(evals, prefix, extra) 413 | -------------------------------------------------------------------------------- /unformatted_llm.py: -------------------------------------------------------------------------------- 1 | from typing import Dict, List, Tuple, Optional 2 | from few_shot import FewShotAnswerSamples, FewShotBuilder 3 | from math_utils.math_helpers import rm_latex_math 4 | from helper_utils import NO_SOLUTION_PREFIX 5 | 6 | HDR = "###" 7 | PRB_TAG = f"{HDR} Problem" 8 | ANS_TAG = f"{HDR} Answer" 9 | END_TAG = "\n\n" 10 | 11 | class UnformattedLLM(FewShotBuilder): 12 | INSTRUCTION = f"Given a mathematics problem, determine the answer. "\ 13 | f"Simplify your answer as much as possible. "\ 14 | f"If the answer cannot be computed, or you are not confident, say {NO_SOLUTION_PREFIX}. " 15 | 16 | def few_shot_dict(self, count: int = -1) -> Dict[str, str]: 17 | ios: Dict[str, str] = self.few_shot_limited(count) 18 | return ios 19 | 20 | def few_shot_format(self, ios: Dict[str, str]) -> str: 21 | return "".join([self.format(i, o) for i, o in ios.items()]) 22 | 23 | def prompt_for(self, prb: str, solns: List[str] = []) -> str: 24 | assert solns == [], f'Unformatted does not format soln steps to model prompt' 25 | return self.format(prb) 26 | 27 | def extract_answer(self, completion: str) -> str: 28 | completion = completion.lstrip() 29 | end = completion.find(END_TAG) 30 | answer = completion[:end] if end != -1 else completion 31 | answer = answer.strip() 32 | return answer 33 | 34 | def __init__(self): 35 | mk_io = lambda fs: (fs.prb, rm_latex_math(fs.outcome)) 36 | self.few_shot_io: List[Tuple[str, str]] = [ mk_io(fs) for fs in FewShotAnswerSamples ] 37 | 38 | def few_shot_limited(self, count: int) -> Dict[str, str]: 39 | limited_samples = self.few_shot_io[:count] if count != -1 else self.few_shot_io 40 | return dict(limited_samples) 41 | 42 | def format_fn(self, prb: str, solns: List[str], outcome: str, explain: Optional[List[str]]): 43 | assert solns == [] and explain == None 44 | return self.format(prb, outcome) 45 | 46 | def format(self, inp: str, out: Optional[str] = None) -> str: 47 | lines = [ 48 | UnformattedLLM.INSTRUCTION, 49 | PRB_TAG, 50 | inp, 51 | ANS_TAG, 52 | ] 53 | if out: 54 | lines += [out, END_TAG] 55 | return "\n".join(lines) 56 | 57 | 58 | --------------------------------------------------------------------------------