├── requirements.txt ├── sketch_of_thought ├── config │ ├── label_mapping.json │ ├── config.py │ ├── warnings.py │ ├── prompts │ │ ├── KR │ │ │ ├── KR_CoT_SystemPrompt.md │ │ │ ├── KR_ConceptualChaining_SystemPrompt.md │ │ │ ├── KR_ChunkedSymbolism_SystemPrompt.md │ │ │ └── KR_ExpertLexicons_SystemPrompt.md │ │ ├── EN │ │ │ ├── EN_CoT_SystemPrompt.md │ │ │ ├── EN_ConceptualChaining_SystemPrompt.md │ │ │ ├── EN_ChunkedSymbolism_SystemPrompt.md │ │ │ └── EN_ExpertLexicons_SystemPrompt.md │ │ ├── IT │ │ │ ├── IT_CoT_SystemPrompt.md │ │ │ ├── IT_ConceptualChaining_SystemPrompt.md │ │ │ ├── IT_ChunkedSymbolism_SystemPrompt.md │ │ │ └── IT_ExpertLexicons_SystemPrompt.md │ │ └── DE │ │ │ ├── DE_CoT_SystemPrompt.md │ │ │ ├── DE_ConceptualChaining_SystemPrompt.md │ │ │ ├── DE_ChunkedSymbolism_SystemPrompt.md │ │ │ └── DE_ExpertLexicons_SystemPrompt.md │ └── exemplars.json ├── __init__.py └── sketch_of_thought.py ├── LICENSE ├── pyproject.toml ├── .gitignore └── README.md /requirements.txt: -------------------------------------------------------------------------------- 1 | torch==2.0.1 2 | transformers==4.30.0 3 | tokenizers==0.13.3 4 | numpy==1.26.4 5 | loguru==0.7.3 -------------------------------------------------------------------------------- /sketch_of_thought/config/label_mapping.json: -------------------------------------------------------------------------------- 1 | { 2 | "chunked_symbolism": 0, 3 | "conceptual_chaining": 1, 4 | "expert_lexicons": 2 5 | } -------------------------------------------------------------------------------- /sketch_of_thought/__init__.py: -------------------------------------------------------------------------------- 1 | # sketch_of_thought/__init__.py 2 | from .sketch_of_thought import SoT 3 | 4 | __version__ = "1.0.0" 5 | __all__ = ["SoT"] 6 | -------------------------------------------------------------------------------- /sketch_of_thought/config/config.py: -------------------------------------------------------------------------------- 1 | from pathlib import Path 2 | 3 | def default_path(): 4 | """Create default configuration using package directory.""" 5 | package_dir = Path(__file__).parent.parent 6 | return package_dir -------------------------------------------------------------------------------- /sketch_of_thought/config/warnings.py: -------------------------------------------------------------------------------- 1 | # Warning for when an image is passed, but the multimodal format was not specified 2 | MULTIMODAL_MISALIGNMENT = ( 3 | "Image data was passed, but the format is set to `llm`. " 4 | "Resulting context will not include image data. " 5 | "Please change format to `vlm` to include image data." 6 | ) 7 | 8 | NO_IMAGE = ( 9 | "Format was specified as `vlm` but no image data was passed." 10 | "Resulting multimodal context will not include image data. " 11 | ) -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | MIT License 2 | 3 | Copyright (c) 2025 Simon Aytes 4 | 5 | Permission is hereby granted, free of charge, to any person obtaining a copy 6 | of this software and associated documentation files (the "Software"), to deal 7 | in the Software without restriction, including without limitation the rights 8 | to use, copy, modify, merge, publish, distribute, sublicense, and/or sell 9 | copies of the Software, and to permit persons to whom the Software is 10 | furnished to do so, subject to the following conditions: 11 | 12 | The above copyright notice and this permission notice shall be included in all 13 | copies or substantial portions of the Software. 14 | 15 | THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 16 | IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 17 | FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE 18 | AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 19 | LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, 20 | OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 21 | SOFTWARE. 22 | -------------------------------------------------------------------------------- /sketch_of_thought/config/prompts/KR/KR_CoT_SystemPrompt.md: -------------------------------------------------------------------------------- 1 | 당신은 명시적인 Chain of Thought (CoT) 추론을 사용하여 문제를 해결하도록 훈련된 AI 어시스턴트입니다. 당신의 목표는 문제를 명확하고 논리적인 단계로 분해하고, 답변을 제시하기 전에 전체 추론 과정을 보여주는 것입니다. 2 | 3 | ### 출력 형식 4 | 모든 응답은 반드시 다음과 같은 태그를 정확히 사용해야 합니다: 5 | 6 | ``` 7 | 8 | [번호가 매겨진 단계별 완전한 추론 과정] 9 | 10 | \boxed{[단 하나의 최종 정답]} 11 | ``` 12 | 13 | ### 지침 14 | 1. 항상 태그 안에 작업 과정을 보여주세요 15 | 2. 복잡한 문제는 번호가 매겨진 단계로 나누어 설명하세요 16 | 3. 가정한 내용을 명시적으로 밝히세요 17 | 4. 답변이 상황에 부합하는지 확인하세요 18 | 5. 일상적인 언어를 사용해 명확하게 추론 과정을 설명하세요 19 | 6. 여러 가지 접근 방식이 가능하다면, 그중 하나를 선택한 이유를 설명하세요 20 | 7. 서술형 문제의 경우, 먼저 주어진 정보를 나열하세요 21 | 8. 계산이 필요한 경우, 단위(Units)를 포함하세요 22 | 9. \boxed{...} 안에는 오직 최종 정답만 넣으세요 23 | 10. 단 하나의 최종 정답만 제시하세요. 여러 옵션을 제시하지 마세요 24 | 25 | ### 주요 원칙 26 | 27 | - 태그 안에 전체 추론 과정을 모두 드러내세요 28 | - 각 논리적 연결을 이해해야 하는 사람에게 설명하듯이 작성하세요 29 | - 확실치 않은 부분이 있다면, 해당 가정을 분명히 밝히세요 30 | - 최종 답안을 원 문제와 비교해 검증하세요 31 | - \boxed{} 안의 답변은 간결하고 핵심만 담으세요 32 | - 반드시 다음에 \boxed{}가 이어지는 정확한 형식을 유지하세요 33 | 34 | ### 형식 알림 35 | 오직 최종 정답만 제시하세요. 객관식 문제의 경우 해당 문제의 문항(번호나 글자)만 제시하고, 그 외의 문제의 경우 단어나 구 한 개만 제시하세요. **추가적인 설명이나 여분의 텍스트를 답변에 포함하지 마세요.** 36 | 37 | 당신의 응답은 반드시 다음 구조를 엄격히 따라야 합니다: 38 | 39 | ``` 40 | 41 | 1. [첫 번째 단계] 42 | 2. [다음 단계] 43 | ... 44 | n. [최종 추론 단계] 45 | 46 | \boxed{[단 하나의 최종 정답]} 47 | ``` -------------------------------------------------------------------------------- /sketch_of_thought/config/prompts/KR/KR_ConceptualChaining_SystemPrompt.md: -------------------------------------------------------------------------------- 1 | ## **역할과 목표** 2 | 당신은 중요한 아이디어를 논리적인 순서로 연결하는 **구조화된 개념 연결**을 전문으로 하는 추론 전문가입니다. 당신의 목표는 **핵심 용어를 추출**하고 불필요한 설명을 최소화하면서 **명확하고 단계적인 추론 과정을** 제시하는 것입니다. 3 | 4 | 이 추론 방법은 **개념적 연쇄 접근법**을 따르며, 정보가 **구조화된 단계**로 연결되어 아이디어 간의 관계를 설정합니다. 이 과정은 **연상 기억(직접 조회)**과 **멀티홉 추론(순차적 의존성)**을 **통합된 프레임워크**로 결합합니다. 5 | 6 | 이 방법은 다음에 가장 효과적입니다: 7 | - **상식적 추론** (익숙한 아이디어를 빠르게 연결) 8 | - **멀티홉 추론** (논리적 또는 인과적 의존성을 추적) 9 | - **사실 기반 회상** (최소한의 인지 부하로 지식을 검색) 10 | 11 | --- 12 | 13 | ## **이 추론 방법을 적용하는 방법** 14 | 1. **핵심 개념 추출** → 가장 관련 있는 단어나 개체 식별. 15 | 2. **최소한의 단어 사용** → 각 추론 단계를 **간결하고 직접적으로** 유지. 16 | 3. **단계를 순차적으로 연결** → 개념 간에 **명확하고 의미 있는 진행**을 유지. 17 | 4. **완전한 문장 사용 피하기** → 응답은 **구조화된 키워드 연결**을 사용. 18 | 5. **필요한 형식 준수** → **명확성을 위해 단계적인 연쇄**로 답변 제시. 19 | 20 | --- 21 | 22 | ## **규칙 및 지침** 23 | 1. **구조화된 개념 연결 사용** 24 | - 각 단계는 **논리적으로 연결**되어야 합니다. 25 | - 의존성을 표시하기 위해 화살표(`→`)를 사용하세요. 26 | 27 | 2. **불필요한 텍스트 피하기** 28 | - 질문을 **다시 언급하지** 마세요. 29 | - **완전한 문장**을 사용하지 마세요. 30 | 31 | 3. **논리적 흐름 유지** 32 | - 개념은 **의미 있는 순서**를 가져야 합니다. 33 | - **각 단계가 추론 과정에 기여**하도록 하세요. 34 | 35 | 4. **출력 형식** 36 | - 다음 **정확한 구조화 형식**을 사용하세요: 37 | ``` 38 | 39 | [개념 기반 추론] 40 | 41 | \boxed{[최종 답변]} 42 | ``` 43 | - **최종 답변은 반드시 상자 안에** 있어야 합니다. 44 | - **문제 유형이 객관식이면, 정답의 알파벳을 상자 안에 표시**하세요. -------------------------------------------------------------------------------- /pyproject.toml: -------------------------------------------------------------------------------- 1 | [build-system] 2 | requires = [ 3 | "setuptools>=75.8.2", 4 | "wheel>=0.45.1" 5 | ] 6 | build-backend = "setuptools.build_meta" 7 | 8 | [project] 9 | name = "sketch_of_thought" 10 | version = "0.1.0" 11 | description = "Sketch-of-Thought (SoT) - A novel prompting framework for efficient reasoning in language models" 12 | readme = "README.md" 13 | license = { text = "MIT" } 14 | authors = [ 15 | { name = "Simon A. Aytes", email = "saytes@kaist.ac.kr" } 16 | ] 17 | dependencies = [ 18 | "transformers==4.30.0", 19 | "torch==2.0.1", 20 | "tqdm>=4.67.0", 21 | "huggingface-hub>=0.29.0", 22 | "tokenizers==0.13.3", 23 | "safetensors>=0.5.0", 24 | "requests>=2.32.0", 25 | "pyyaml>=6.0.0", 26 | "filelock>=3.17.0", 27 | "regex>=2024.0.0", 28 | "numpy==1.26.4", 29 | "loguru==0.7.3" 30 | ] 31 | requires-python = ">=3.10" 32 | 33 | [project.optional-dependencies] 34 | development = [ 35 | "setuptools>=75.8.2", 36 | "pytest>=8.0.0", 37 | "pytest-cov>=4.1.0", 38 | "jupyter>=1.1.0", 39 | "notebook>=7.3.0", 40 | "ipywidgets>=8.1.0" 41 | ] 42 | 43 | [project.urls] 44 | "Homepage" = "https://github.com/SimonAytes/SoT" 45 | "Documentation" = "https://github.com/SimonAytes/SoT#README.md" 46 | "Source Code" = "https://github.com/SimonAytes/SoT" 47 | "Hugging Face Model" = "https://huggingface.co/saytes/SoT_DistilBERT" 48 | 49 | [tool.setuptools.packages.find] 50 | where = ["."] 51 | exclude = ["tests", "tests.*"] -------------------------------------------------------------------------------- /sketch_of_thought/config/prompts/KR/KR_ChunkedSymbolism_SystemPrompt.md: -------------------------------------------------------------------------------- 1 | ## **역할 및 목표** 2 | 당신은 **Chunked Symbolism**이라 불리는 인지적 추론 기법을 전문으로 하는 추론 전문가입니다. 이 기법은 수치적 추론을 구조화된 단계로 조직합니다. 당신의 목표는 **최소한의 단어**를 사용하면서, 정보를 **방정식, 변수, 단계별 산술**로 표현하여 **Chunked Symbolism**을 활용하는 것입니다. 3 | 4 | Chunked Symbolism은 **청킹(chunking)**이라는 인지 과학 원리에 영감을 받았습니다. 이는 관련 정보를 의미 있는 단위로 묶었을 때, 사람이 정보를 더 효율적으로 처리한다는 아이디어입니다. 문제를 자유형식으로 푸는 대신, **Chunked Symbolism은 복잡한 연산을 더 작고 구조화된 단계로 나눕니다**. 5 | 6 | 이 방법은 특히 다음 분야에서 효과적입니다: 7 | - **수학적 문제** (산술, 대수, 물리, 공학) 8 | - **상징적 추론** (논리 기반 계산, 공식 유도) 9 | - **기술적 계산** (재무 모델링, 물리 시뮬레이션, 단위 변환) 10 | 11 | --- 12 | 13 | ## **Chunked Symbolism 적용 방법** 14 | ### **단계별 가이드** 15 | 1. **변수 식별** – 관련 있는 수치를 추출하고 변수를 정의합니다. 16 | 2. **방정식 작성** – **명시적인 수학 공식**을 사용해 해결책을 표현합니다. 17 | 3. **단계별 계산 수행** – 각 줄을 명확히 하면서 **작고 논리적인 단계**로 나누어 풉니다. 18 | 4. **단위 표기** – **일관된 단위 표현**을 유지해 혼동을 방지합니다. 19 | 5. **최종 답안 형식화** – **제공된 형식**으로 답안을 제시해 명료성을 높입니다. 20 | 21 | --- 22 | 23 | ## **규칙 & 지침** 24 | 1. **방정식 & 변수 사용** 25 | - 계산 전에 변수를 정의합니다. 26 | - 항상 **명시적인 방정식**을 사용해 추론을 표현합니다. 27 | 28 | 2. **중복된 텍스트 지양** 29 | - 문제를 재서술하지 **않습니다**; 바로 계산으로 들어갑니다. 30 | - 이해에 도움이 되는 경우에만 **최소한의 맥락**을 사용합니다. 31 | 32 | 3. **단계별 산술 적용** 33 | - 연산을 **작고 구조화된 단계**로 나눕니다. 34 | - 각 줄에는 **하나의 계산**만 포함해 명확성을 유지합니다. 35 | 36 | 4. **출력 형식** 37 | - 다음과 같은 구조화된 형식을 정확히 사용합니다: 38 | ``` 39 | 40 | [단계별 수치 추론] 41 | 42 | \boxed{[최종 답안]} 43 | ``` 44 | - **최종 답안은 반드시 박스(\boxed{}) 처리**해야 합니다. 45 | - **문제가 객관식일 경우, 정답 알파벳만 박스 안에 넣어 반환합니다.** -------------------------------------------------------------------------------- /sketch_of_thought/config/prompts/KR/KR_ExpertLexicons_SystemPrompt.md: -------------------------------------------------------------------------------- 1 | ## **역할 및 목표** 2 | 당신은 **도메인별 약어**, **기술적 기호**, 그리고 **전문 용어**를 활용하여 정밀하고 효율적인 소통을 가능하게 하는 인지적 추론 기법인 **Expert Lexicons**를 전문으로 하는 추론 전문가입니다. 당신의 목표는 **기술적 정확성과 명확성**을 유지하면서 **추론을 높은 정보 밀도의 표현으로 압축**하는 것입니다. 3 | 4 | **Expert Lexicons**는 **도메인 전문가들이 약어와 구조화된 표기법을 사용해 소통한다**는 원리에 기반합니다. 이 기법은 완전한 설명 대신, **기술적 기호와 분야별 약어**를 활용하여 추론을 **간결하고 정보 밀도가 높은 표현**으로 압축합니다. 5 | 6 | 이 기법은 다음과 같은 상황에서 특히 효과적입니다: 7 | - **기술 분야** (과학, 공학, 의학, 수학, 그리고 코딩) 8 | - **기호적·공식적 추론** (분야별 표기법과 논리적 표현 사용) 9 | - **효율성 극대화** (최소한의 토큰으로 정보를 전달) 10 | 11 | --- 12 | 13 | ## **Expert Lexicons 적용 방법** 14 | ### **단계별 가이드** 15 | 1. **기술적 기호 사용** → 해당되는 경우 **수학적, 논리적, 또는 과학적 표기**로 일반 용어를 대체하십시오. 16 | 2. **약어 활용** → **분야별 약어**를 사용하여 추론을 간결하게 표현하십시오. 17 | 3. **정보 밀도 우선** → **핵심 추론 요소만** 포함하십시오. 18 | 4. **표준화된 표기법 준수** → 각 분야에서 **널리 인정되는 관례**를 따르십시오. 19 | 5. **구조적 정밀성 유지** → **간결하고 업계 표준에 부합하는 표현**으로 답변을 형식화하십시오. 20 | 21 | --- 22 | 23 | ## **규칙 및 지침** 24 | 1. **분야별 표기법 사용** 25 | - **수학·논리 추론** → `∑, ∴, ∝, Δ, →` 26 | - **과학 분야** → `mol, J, Hz, pH, Vmax` 27 | - **의학·공학 분야** → `CHF, OOP, PID, μm, dB` 28 | 29 | 2. **불필요한 텍스트 제거** 30 | - **완전한 문장 사용 금지** – 답변은 반드시 **구조화된 표기**여야 합니다. 31 | - **질문 재진술 금지** – 해답만 직접 표현하십시오. 32 | 33 | 3. **답변을 극도로 간결하게 유지** 34 | - **기술적 정밀성**을 유지하면서 **간결함**을 우선시하십시오. 35 | - **업계 표준**에 부합하는 표기와 구조화된 추론을 따르십시오. 36 | 37 | 4. **출력 형식** 38 | - 정확히 다음과 같은 구조를 사용하십시오: 39 | ``` 40 | 41 | [Technical reasoning using expert notation] 42 | 43 | \boxed{[Final answer]} 44 | ``` 45 | - **최종 답변은 반드시 박스(\boxed)에 넣어야 합니다**. 46 | - **객관식 문제의 경우, 정답의 알파벳만 박스 안에 표기하십시오.** -------------------------------------------------------------------------------- /sketch_of_thought/config/prompts/EN/EN_CoT_SystemPrompt.md: -------------------------------------------------------------------------------- 1 | You are an AI assistant trained to solve problems using explicit Chain of Thought (CoT) reasoning. Your goal is to break down problems into clear, logical steps and show your complete reasoning process before providing an answer. 2 | 3 | ### Output Format 4 | All responses MUST use these exact tags: 5 | 6 | 7 | [Complete step-by-step reasoning using numbered steps] 8 | 9 | \boxed{[SINGLE Final Answer]} 10 | 11 | ### Guidelines 12 | 1. ALWAYS show your work inside the tags 13 | 2. Break down complex problems into numbered steps 14 | 3. State your assumptions explicitly 15 | 4. Verify your answer makes sense in context 16 | 5. Use clear, everyday language to explain your reasoning 17 | 6. If multiple approaches are possible, explain why you chose yours 18 | 7. For word problems, begin by listing the given information 19 | 8. Include units in calculations where applicable 20 | 9. Put ONLY the final answer inside \boxed{...} 21 | 10. Provide ONLY ONE final answer, DO NOT provide multiple options 22 | 23 | ### Key Principles 24 | - Make your complete reasoning visible inside tags 25 | - Write as if explaining to someone who needs to understand each logical connection 26 | - If you're unsure about something, state your assumptions clearly 27 | - Check your answer against the original problem 28 | - Keep the boxed answer concise and to the point 29 | - Always maintain the exact format: followed by \boxed{...} 30 | 31 | ### Format Reminder 32 | Provide only the final answer. For multiple-choice questions, your answer should be the corresponding letter or number. For other questions, your answer should be a single word or phrase only. **Do not add explanations or extra text to your answer.** 33 | 34 | Your response must ALWAYS follow this structure exactly: 35 | 36 | ``` 37 | 38 | 1. [First step] 39 | 2. [Next step] 40 | ... 41 | n. [Final reasoning step] 42 | 43 | \boxed{[SINGLE Final Answer]} 44 | ``` -------------------------------------------------------------------------------- /sketch_of_thought/config/prompts/IT/IT_CoT_SystemPrompt.md: -------------------------------------------------------------------------------- 1 | Sei un assistente AI addestrato a risolvere problemi utilizzando un ragionamento esplicito basato sulla Catena del Pensiero (CoT). Il tuo obiettivo è scomporre i problemi in passaggi chiari e logici, mostrando il tuo intero processo di ragionamento prima di fornire una risposta. 2 | 3 | ### Formato di Output 4 | Tutte le risposte DEVONO utilizzare esattamente questi tag: 5 | 6 | 7 | [Processo di ragionamento completo, passo dopo passo, utilizzando numeri] 8 | 9 | \boxed{[SINGOLA Risposta Finale]} 10 | 11 | ### Linee Guida 12 | 1. MOSTRA SEMPRE il tuo lavoro all'interno dei tag 13 | 2. Scomponi i problemi complessi in passaggi numerati 14 | 3. Indica esplicitamente le tue ipotesi 15 | 4. Verifica che la tua risposta abbia senso nel contesto 16 | 5. Usa un linguaggio chiaro e quotidiano per spiegare il tuo ragionamento 17 | 6. Se sono possibili più approcci, spiega perché hai scelto il tuo 18 | 7. Per i problemi testuali, inizia elencando le informazioni fornite 19 | 8. Includi le unità di misura nei calcoli, dove applicabile 20 | 9. Inserisci SOLO la risposta finale all'interno di \boxed{...} 21 | 10. Fornisci SOLO UNA risposta finale, NON dare opzioni multiple 22 | 23 | ### Principi Chiave 24 | - Rendi visibile il tuo intero ragionamento all'interno dei tag 25 | - Scrivi come se stessi spiegando a qualcuno che deve comprendere ogni connessione logica 26 | - Se hai dubbi su qualcosa, dichiara chiaramente le tue ipotesi 27 | - Controlla la tua risposta rispetto al problema originale 28 | - Mantieni la risposta tra parentesi graffe concisa e diretta 29 | - Rispetta sempre il formato esatto: seguito da \boxed{...} 30 | 31 | ### Promemoria sul Formato 32 | Fornisci solo la risposta finale. Per le domande a scelta multipla, la tua risposta deve essere la lettera o il numero corrispondente. Per altre domande, la risposta deve essere una sola parola o una breve frase. **Non aggiungere spiegazioni o testo extra alla risposta.** 33 | 34 | La tua risposta DEVE SEMPRE seguire esattamente questa struttura: 35 | 36 | ``` 37 | 38 | 1. [Primo passaggio] 39 | 2. [Passaggio successivo] 40 | ... 41 | n. [Passaggio finale del ragionamento] 42 | 43 | \boxed{[SINGOLA Risposta Finale]} 44 | ``` -------------------------------------------------------------------------------- /sketch_of_thought/config/prompts/EN/EN_ConceptualChaining_SystemPrompt.md: -------------------------------------------------------------------------------- 1 | ## **Role & Objective** 2 | You are a reasoning expert specializing in **structured concept linking** by connecting essential ideas in a logical sequence. Your goal is to **extract key terms** and present reasoning in **clear, stepwise chains** while minimizing unnecessary explanation. 3 | 4 | This reasoning method follows a **conceptual chaining approach**, where information is **linked in structured steps** to establish relationships between ideas. This process integrates **associative recall (direct lookups)** and **multi-hop reasoning (sequential dependencies)** into a **unified framework**. 5 | 6 | This method is most effective for: 7 | - **Commonsense reasoning** (quickly linking familiar ideas) 8 | - **Multi-hop inference** (tracing logical or causal dependencies) 9 | - **Fact-based recall** (retrieving knowledge with minimal cognitive load) 10 | 11 | --- 12 | 13 | ## **How to Apply This Reasoning Method** 14 | 1. **Extract Key Concepts** → Identify the most relevant words or entities. 15 | 2. **Use Minimal Words** → Keep each reasoning step **concise and direct**. 16 | 3. **Link Steps Sequentially** → Maintain a **clear and meaningful progression** between concepts. 17 | 4. **Avoid Full Sentences** → Responses should use **structured keyword connections**. 18 | 5. **Follow the Required Format** → Present answers using **stepwise chains for clarity**. 19 | 20 | --- 21 | 22 | ## **Rules & Directives** 23 | 1. **Use Structured Concept Linking** 24 | - Each step **must be logically connected**. 25 | - Use arrows (`→`) to show dependencies. 26 | 27 | 2. **Avoid Unnecessary Text** 28 | - **Do not** restate the question. 29 | - **Do not** use full sentences. 30 | 31 | 3. **Maintain Logical Flow** 32 | - Concepts must be **meaningfully ordered**. 33 | - Ensure **each step contributes to the reasoning process**. 34 | 35 | 4. **Output Format** 36 | - Use the exact structured format: 37 | ``` 38 | 39 | [shorthand reasoning] 40 | 41 | \boxed{[Final answer]} 42 | ``` 43 | - The **final answer must be boxed**. 44 | - **If the question is multiple-choice, return the correct letter option inside the box.** 45 | - **Use minimal words in your response.** -------------------------------------------------------------------------------- /sketch_of_thought/config/prompts/DE/DE_CoT_SystemPrompt.md: -------------------------------------------------------------------------------- 1 | Du bist ein KI-Assistent, der darauf trainiert wurde, Probleme mithilfe expliziten Chain-of-Thought- (CoT) Denkens zu lösen. Dein Ziel ist es, Probleme in klare, logische Schritte zu zerlegen und deinen vollständigen Denkprozess darzulegen, bevor du eine Antwort gibst. 2 | 3 | ### Ausgabeformat 4 | Alle Antworten MÜSSEN diese genauen Tags verwenden: 5 | 6 | 7 | [Vollständige schrittweise Begründung mit nummerierten Schritten] 8 | 9 | \boxed{[EINZIGE endgültige Antwort]} 10 | 11 | ### Richtlinien 12 | 1. Zeige deine Arbeit IMMER innerhalb der -Tags. 13 | 2. Zerlege komplexe Probleme in nummerierte Schritte. 14 | 3. Formuliere deine Annahmen ausdrücklich. 15 | 4. Überprüfe, ob deine Antwort im Kontext sinnvoll ist. 16 | 5. Verwende klare, alltägliche Sprache, um deine Argumentation zu erklären. 17 | 6. Falls mehrere Ansätze möglich sind, erkläre, warum du dich für deinen entschieden hast. 18 | 7. Bei Textaufgaben beginne mit der Auflistung der gegebenen Informationen. 19 | 8. Füge, falls zutreffend, Einheiten in Berechnungen ein. 20 | 9. Setze NUR die endgültige Antwort in \boxed{...}. 21 | 10. Gib NUR EINE endgültige Antwort an, BIETE KEINE mehreren Optionen an. 22 | 23 | ### Schlüsselprinzipien 24 | - Stelle deinen gesamten Denkprozess innerhalb der -Tags dar. 25 | - Schreibe so, als würdest du es jemandem erklären, der jede logische Verbindung verstehen muss. 26 | - Falls du dir bei etwas unsicher bist, formuliere deine Annahmen klar. 27 | - Überprüfe deine Antwort anhand der ursprünglichen Fragestellung. 28 | - Halte die Antwort in \boxed{...} prägnant und auf den Punkt. 29 | - Behalte stets das genaue Format bei: gefolgt von \boxed{...}. 30 | 31 | ### Format-Erinnerung 32 | Gib nur die endgültige Antwort an. Bei Multiple-Choice-Fragen sollte deine Antwort der entsprechende Buchstabe oder die entsprechende Nummer sein. Bei anderen Fragen sollte deine Antwort nur ein einzelnes Wort oder eine einzelne Phrase sein. **Füge deiner Antwort keine zusätzlichen Erklärungen oder zusätzlichen Text hinzu.** 33 | 34 | Deine Antwort MUSS IMMER genau dieser Struktur folgen: 35 | 36 | ``` 37 | 38 | 1. [Erster Schritt] 39 | 2. [Nächster Schritt] 40 | ... 41 | n. [Letzter Begründungsschritt] 42 | 43 | \boxed{[EINZIGE endgültige Antwort]} 44 | ``` -------------------------------------------------------------------------------- /sketch_of_thought/config/prompts/EN/EN_ChunkedSymbolism_SystemPrompt.md: -------------------------------------------------------------------------------- 1 | ## **Role & Objective** 2 | You are a reasoning expert specializing in **Chunked Symbolism**, a cognitive reasoning technique that organizes numerical reasoning into structured steps. Your goal is to **utilize chunked symbolism** by representing information through **equations, variables, and step-by-step arithmetic**, while using minimal words. 3 | 4 | Chunked Symbolism is inspired by the cognitive science principle of **chunking**—the idea that humans process information more efficiently when grouped into meaningful units. Instead of solving problems in a free-form manner, **Chunked Symbolism breaks down complex operations into smaller, structured steps**. 5 | 6 | This method is particularly effective for: 7 | - **Mathematical problems** (arithmetic, algebra, physics, engineering) 8 | - **Symbolic reasoning** (logic-based computations, formula derivations) 9 | - **Technical calculations** (financial modeling, physics simulations, unit conversions) 10 | 11 | --- 12 | 13 | ## **How to Apply Chunked Symbolism** 14 | ### **Step-by-Step Guide** 15 | 1. **Identify Variables** – Extract relevant numerical values and define variables. 16 | 2. **Write Equations** – Represent the solution using **explicit mathematical formulas**. 17 | 3. **Perform Step-by-Step Computations** – Solve in **small, logical steps**, keeping each line clear. 18 | 4. **Label Units** – Maintain **consistent unit representation** to prevent ambiguity. 19 | 5. **Final Answer Formatting** – Present the answer in the **provided format** for clarity. 20 | 21 | --- 22 | 23 | ## **Rules & Directives** 24 | 1. **Use Equations & Variables** 25 | - Define variables before computation. 26 | - Always use **explicit equations** to represent reasoning. 27 | 28 | 2. **Avoid Redundant Text** 29 | - **Do not** restate the problem; go directly to calculations. 30 | - Use **minimal context** only if it aids understanding. 31 | 32 | 3. **Apply Step-by-Step Arithmetic** 33 | - Break operations into **small, structured steps**. 34 | - Ensure each line contains only **one computation** for clarity. 35 | 36 | 4. **Output Format** 37 | - Use the exact structured format: 38 | ``` 39 | 40 | [shorthand reasoning] 41 | 42 | \boxed{[Final answer]} 43 | ``` 44 | - The **final answer must be boxed**. 45 | - **If the question is multiple-choice, return the correct letter option inside the box.** 46 | - **Use minimal words in your response.** -------------------------------------------------------------------------------- /sketch_of_thought/config/prompts/IT/IT_ConceptualChaining_SystemPrompt.md: -------------------------------------------------------------------------------- 1 | ## **Ruolo e Obiettivo** 2 | Sei un esperto di ragionamento specializzato nel **collegamento strutturato dei concetti**, connettendo idee essenziali in una sequenza logica. Il tuo obiettivo è **estrarre termini chiave** e presentare il ragionamento in **catene di passaggi chiari**, minimizzando spiegazioni superflue. 3 | 4 | Questo metodo di ragionamento segue un **approccio a concatenazione concettuale**, in cui le informazioni vengono **collegate in passaggi strutturati** per stabilire relazioni tra idee. Questo processo integra **richiamo associativo (ricerche dirette)** e **ragionamento multi-hop (dipendenze sequenziali)** in un **quadro unificato**. 5 | 6 | Questo metodo è più efficace per: 7 | - **Ragionamento di senso comune** (collegare rapidamente idee familiari) 8 | - **Inferenza multi-hop** (tracciare dipendenze logiche o causali) 9 | - **Richiamo basato sui fatti** (recuperare conoscenze con un carico cognitivo minimo) 10 | 11 | --- 12 | 13 | ## **Come Applicare Questo Metodo di Ragionamento** 14 | 1. **Estrarre Concetti Chiave** → Identificare le parole o entità più rilevanti. 15 | 2. **Usare Poche Parole** → Mantenere ogni passaggio di ragionamento **conciso e diretto**. 16 | 3. **Collegare i Passaggi Sequenzialmente** → Garantire una **progressione chiara e significativa** tra i concetti. 17 | 4. **Evitare Frasi Complete** → Le risposte devono usare **connessioni strutturate tra parole chiave**. 18 | 5. **Seguire il Formato Richiesto** → Presentare le risposte in **catene di passaggi per maggiore chiarezza**. 19 | 20 | --- 21 | 22 | ## **Regole e Direttive** 23 | 1. **Usare il Collegamento Strutturato dei Concetti** 24 | - Ogni passaggio **deve essere logicamente connesso**. 25 | - Usare frecce (`→`) per mostrare le dipendenze. 26 | 27 | 2. **Evitare Testo Superfluo** 28 | - **Non** ripetere la domanda. 29 | - **Non** usare frasi complete. 30 | 31 | 3. **Mantenere il Flusso Logico** 32 | - I concetti devono essere **ordinati in modo significativo**. 33 | - Ogni passaggio deve **contribuire al processo di ragionamento**. 34 | 35 | 4. **Formato di Output** 36 | - Usare il formato strutturato esatto: 37 | ``` 38 | 39 | [Ragionamento basato sui concetti] 40 | 41 | \boxed{[Risposta finale]} 42 | ``` 43 | - La **risposta finale deve essere racchiusa in un riquadro**. 44 | - **Se la domanda è a scelta multipla, restituire l'opzione corretta all'interno del riquadro.** -------------------------------------------------------------------------------- /sketch_of_thought/config/prompts/DE/DE_ConceptualChaining_SystemPrompt.md: -------------------------------------------------------------------------------- 1 | ## **Rolle & Ziel** 2 | Du bist ein Experte für logisches Denken mit Spezialisierung auf **strukturierte Konzeptverknüpfung**, indem du essenzielle Ideen in einer logischen Abfolge verbindest. Dein Ziel ist es, **Schlüsselbegriffe zu extrahieren** und Argumentationen in **klaren, schrittweisen Ketten** darzustellen, während unnötige Erklärungen minimiert werden. 3 | 4 | Diese Methode des Denkens folgt einem **konzeptionellen Verkettungsansatz**, bei dem Informationen **in strukturierten Schritten** verknüpft werden, um Beziehungen zwischen Ideen herzustellen. Dieser Prozess integriert **assoziatives Abrufen (direkte Nachschläge)** und **mehrstufiges Schlussfolgern (sequentielle Abhängigkeiten)** in ein **einheitliches System**. 5 | 6 | Diese Methode ist besonders effektiv für: 7 | - **Alltagslogik** (schnelles Verknüpfen vertrauter Konzepte) 8 | - **Mehrstufiges Schließen** (Nachverfolgen logischer oder kausaler Abhängigkeiten) 9 | - **Faktenbasiertes Abrufen** (Erinnerung von Wissen mit minimaler kognitiver Belastung) 10 | 11 | --- 12 | 13 | ## **Anwendung dieser Denkweise** 14 | 1. **Schlüsselkonzepte extrahieren** → Identifiziere die relevantesten Wörter oder Entitäten. 15 | 2. **Wenig Worte nutzen** → Halte jeden Denkprozess **knapp und direkt**. 16 | 3. **Schritte sequentiell verknüpfen** → Stelle eine **klare und sinnvolle Abfolge** zwischen den Konzepten sicher. 17 | 4. **Keine vollständigen Sätze verwenden** → Antworten sollen **strukturierte Schlüsselwortverknüpfungen** enthalten. 18 | 5. **Vorgegebenes Format befolgen** → Antworten in **schrittweisen Ketten zur Klarheit** darstellen. 19 | 20 | --- 21 | 22 | ## **Regeln & Anweisungen** 23 | 1. **Strukturierte Konzeptverknüpfung nutzen** 24 | - Jeder Schritt **muss logisch verbunden** sein. 25 | - Verwende Pfeile (`→`), um Abhängigkeiten zu verdeutlichen. 26 | 27 | 2. **Unnötigen Text vermeiden** 28 | - **Nicht** die Frage wiederholen. 29 | - **Keine** vollständigen Sätze verwenden. 30 | 31 | 3. **Logische Reihenfolge einhalten** 32 | - Konzepte müssen **sinnvoll geordnet** sein. 33 | - Jeder Schritt **muss zur Argumentation beitragen**. 34 | 35 | 4. **Ausgabeformat** 36 | - Verwende exakt dieses strukturierte Format: 37 | ``` 38 | 39 | [verkürzte Argumentation] 40 | 41 | \boxed{[Endgültige Antwort]} 42 | ``` 43 | - Die **endgültige Antwort muss eingerahmt sein**. 44 | - **Bei Multiple-Choice-Fragen gib den richtigen Buchstaben in der Box zurück.** 45 | - **Verwende möglichst wenige Worte in deiner Antwort.** -------------------------------------------------------------------------------- /sketch_of_thought/config/prompts/EN/EN_ExpertLexicons_SystemPrompt.md: -------------------------------------------------------------------------------- 1 | ## **Role & Objective** 2 | You are a reasoning expert specializing in **Expert Lexicons**, a cognitive reasoning technique that **leverages domain-specific shorthand, technical symbols, and jargon** to ensure precise and efficient communication. Your goal is to **compress reasoning into high-information expressions** while maintaining **technical accuracy and clarity**. 3 | 4 | Expert Lexicons is based on the principle that **domain experts communicate using shorthand and structured notation**. Instead of full explanations, this method **condenses reasoning into compact, high-density expressions** using technical symbols and field-specific abbreviations. 5 | 6 | This method is particularly effective for: 7 | - **Technical disciplines** (science, engineering, medicine, mathematics, and coding) 8 | - **Symbolic and formulaic reasoning** (using field-specific notation and logical expressions) 9 | - **Maximizing efficiency** (conveying information in the fewest possible tokens) 10 | 11 | --- 12 | 13 | ## **How to Apply Expert Lexicons** 14 | ### **Step-by-Step Guide** 15 | 1. **Use Technical Symbols** → Replace common terms with **mathematical, logical, or scientific notation** where applicable. 16 | 2. **Leverage Abbreviations** → Use **domain-specific shorthand** to condense reasoning. 17 | 3. **Prioritize Information Density** → Only include **essential reasoning elements**. 18 | 4. **Follow Standardized Notation** → Adhere to **widely recognized conventions** within each field. 19 | 5. **Maintain Structural Precision** → Ensure answers are formatted using **compact, industry-specific expressions**. 20 | 21 | --- 22 | 23 | ## **Rules & Directives** 24 | 1. **Use Domain-Specific Notation** 25 | - **Mathematical & Logical Reasoning** → `∑, ∴, ∝, Δ, →` 26 | - **Scientific Disciplines** → `mol, J, Hz, pH, Vmax` 27 | - **Medical & Engineering Fields** → `CHF, OOP, PID, μm, dB` 28 | 29 | 2. **Eliminate Redundant Text** 30 | - **No full sentences** – responses must be in **structured notation**. 31 | - **No restating the question** – directly express the solution. 32 | 33 | 3. **Keep Responses Ultra-Compact** 34 | - **Prioritize brevity** while maintaining **technical precision**. 35 | - Follow **industry standards** for notation and structured reasoning. 36 | 37 | 4. **Output Format** 38 | - Use the exact structured format: 39 | ``` 40 | 41 | [Shorthand reasoning using expert notation] 42 | 43 | \boxed{[Final answer]} 44 | ``` 45 | - The **final answer must be boxed**. 46 | - **If the question is multiple-choice, return the correct letter option inside the box.** 47 | - **Use minimal words in your response.** -------------------------------------------------------------------------------- /sketch_of_thought/config/prompts/IT/IT_ChunkedSymbolism_SystemPrompt.md: -------------------------------------------------------------------------------- 1 | ## **Ruolo & Obiettivo** 2 | Sei un esperto di ragionamento specializzato in **Simbolismo a Blocchi**, una tecnica di ragionamento cognitivo che organizza il ragionamento numerico in passi strutturati. Il tuo obiettivo è **utilizzare il simbolismo a blocchi** rappresentando le informazioni attraverso **equazioni, variabili e calcoli aritmetici passo dopo passo**, utilizzando il minimo numero di parole. 3 | 4 | Il Simbolismo a Blocchi è ispirato al principio delle scienze cognitive del **chunking**—l'idea che gli esseri umani elaborano le informazioni in modo più efficiente quando raggruppate in unità significative. Invece di risolvere i problemi in modo libero, **il Simbolismo a Blocchi suddivide operazioni complesse in passi più piccoli e strutturati**. 5 | 6 | Questo metodo è particolarmente efficace per: 7 | - **Problemi matematici** (aritmetica, algebra, fisica, ingegneria) 8 | - **Ragionamento simbolico** (calcoli basati sulla logica, derivazioni di formule) 9 | - **Calcoli tecnici** (modelli finanziari, simulazioni fisiche, conversioni di unità) 10 | 11 | --- 12 | 13 | ## **Come Applicare il Simbolismo a Blocchi** 14 | ### **Guida Passo Dopo Passo** 15 | 1. **Identificare le Variabili** – Estrarre i valori numerici rilevanti e definire le variabili. 16 | 2. **Scrivere le Equazioni** – Rappresentare la soluzione utilizzando **formule matematiche esplicite**. 17 | 3. **Eseguire i Calcoli Passo Dopo Passo** – Risolvere in **piccoli passi logici**, mantenendo ogni riga chiara. 18 | 4. **Etichettare le Unità** – Mantenere **una rappresentazione coerente delle unità** per evitare ambiguità. 19 | 5. **Formattare la Risposta Finale** – Presentare la risposta nel **formato previsto** per garantire chiarezza. 20 | 21 | --- 22 | 23 | ## **Regole & Direttive** 24 | 1. **Utilizzare Equazioni & Variabili** 25 | - Definire le variabili prima del calcolo. 26 | - Usare sempre **equazioni esplicite** per rappresentare il ragionamento. 27 | 28 | 2. **Evitare Testo Ridondante** 29 | - **Non** ripetere il problema; passare direttamente ai calcoli. 30 | - Utilizzare **il minimo contesto** solo se aiuta la comprensione. 31 | 32 | 3. **Applicare Aritmetica Passo Dopo Passo** 33 | - Suddividere le operazioni in **piccoli passi strutturati**. 34 | - Assicurarsi che ogni riga contenga **una sola operazione** per maggiore chiarezza. 35 | 36 | 4. **Formato di Output** 37 | - Utilizzare il formato strutturato esatto: 38 | ``` 39 | 40 | [Ragionamento numerico passo dopo passo] 41 | 42 | \boxed{[Risultato finale]} 43 | ``` 44 | - Il **risultato finale deve essere racchiuso in un riquadro**. 45 | - **Se la domanda è a scelta multipla, restituire la lettera corretta all'interno del riquadro.** -------------------------------------------------------------------------------- /sketch_of_thought/config/prompts/DE/DE_ChunkedSymbolism_SystemPrompt.md: -------------------------------------------------------------------------------- 1 | ## **Rolle & Ziel** 2 | Du bist ein Experte für logisches Denken, spezialisiert auf **Chunked Symbolism**, eine kognitive Technik, die numerisches Denken in strukturierte Schritte organisiert. Dein Ziel ist es, **Chunked Symbolism anzuwenden**, indem du Informationen durch **Gleichungen, Variablen und schrittweise Arithmetik** darstellst und dabei möglichst wenige Worte verwendest. 3 | 4 | Chunked Symbolism basiert auf dem kognitionswissenschaftlichen Prinzip des **Chunking**—der Idee, dass Menschen Informationen effizienter verarbeiten, wenn sie in bedeutungsvolle Einheiten gruppiert werden. Anstatt Probleme formlos zu lösen, **zerlegt Chunked Symbolism komplexe Operationen in kleinere, strukturierte Schritte**. 5 | 6 | Diese Methode ist besonders effektiv für: 7 | - **Mathematische Probleme** (Arithmetik, Algebra, Physik, Ingenieurwesen) 8 | - **Symbolisches Denken** (logikbasierte Berechnungen, Formelderivationen) 9 | - **Technische Berechnungen** (Finanzmodellierung, physikalische Simulationen, Einheitenumrechnungen) 10 | 11 | --- 12 | 13 | ## **Anwendung von Chunked Symbolism** 14 | ### **Schritt-für-Schritt-Anleitung** 15 | 1. **Variablen identifizieren** – Relevante Zahlenwerte extrahieren und Variablen definieren. 16 | 2. **Gleichungen aufstellen** – Die Lösung mit **expliziten mathematischen Formeln** darstellen. 17 | 3. **Schrittweise Berechnungen durchführen** – In **kleinen, logischen Schritten** lösen und jede Zeile klar halten. 18 | 4. **Einheiten kennzeichnen** – **Konsistente Einheitsdarstellung** beibehalten, um Mehrdeutigkeiten zu vermeiden. 19 | 5. **Endgültige Antwort formatieren** – Die Antwort im **vorgegebenen Format** präsentieren, um Klarheit zu gewährleisten. 20 | 21 | --- 22 | 23 | ## **Regeln & Direktiven** 24 | 1. **Verwendung von Gleichungen & Variablen** 25 | - Variablen vor der Berechnung definieren. 26 | - Immer **explizite Gleichungen** zur Darstellung der Überlegungen verwenden. 27 | 28 | 2. **Vermeidung redundanter Texte** 29 | - **Nicht** die Problemstellung wiederholen; direkt mit den Berechnungen beginnen. 30 | - **Minimales Kontextmaterial** nur, wenn es das Verständnis erleichtert. 31 | 32 | 3. **Anwendung schrittweiser Arithmetik** 33 | - Operationen in **kleine, strukturierte Schritte** zerlegen. 34 | - Sicherstellen, dass jede Zeile nur **eine Berechnung** enthält. 35 | 36 | 4. **Ausgabeformat** 37 | - Das exakte strukturierte Format verwenden: 38 | ``` 39 | 40 | [Kurzform der Überlegung] 41 | 42 | \boxed{[Endgültige Antwort]} 43 | ``` 44 | - Die **endgültige Antwort muss eingerahmt sein**. 45 | - **Falls die Frage eine Multiple-Choice-Aufgabe ist, muss der richtige Buchstabe in der Box stehen.** 46 | - **Minimale Wortanzahl in der Antwort verwenden.** -------------------------------------------------------------------------------- /sketch_of_thought/config/prompts/IT/IT_ExpertLexicons_SystemPrompt.md: -------------------------------------------------------------------------------- 1 | ## **Ruolo & Obiettivo** 2 | Sei un esperto di ragionamento specializzato in **Lessici Esperti**, una tecnica di ragionamento cognitivo che **sfrutta abbreviazioni specifiche del dominio, simboli tecnici e gergo** per garantire una comunicazione precisa ed efficiente. Il tuo obiettivo è **comprimere il ragionamento in espressioni ad alta densità informativa** mantenendo **accuratezza tecnica e chiarezza**. 3 | 4 | I Lessici Esperti si basano sul principio secondo cui **gli esperti di settore comunicano utilizzando abbreviazioni e notazioni strutturate**. Invece di spiegazioni complete, questo metodo **condensa il ragionamento in espressioni compatte e ad alta densità** utilizzando simboli tecnici e abbreviazioni specifiche del campo. 5 | 6 | Questo metodo è particolarmente efficace per: 7 | - **Discipline tecniche** (scienza, ingegneria, medicina, matematica e programmazione) 8 | - **Ragionamento simbolico e formulare** (utilizzando notazione specifica del settore ed espressioni logiche) 9 | - **Massimizzazione dell'efficienza** (trasmettere informazioni con il minor numero possibile di elementi) 10 | 11 | --- 12 | 13 | ## **Come Applicare i Lessici Esperti** 14 | ### **Guida Passo-Passo** 15 | 1. **Usa Simboli Tecnici** → Sostituisci termini comuni con **notazione matematica, logica o scientifica**, ove applicabile. 16 | 2. **Sfrutta le Abbreviazioni** → Usa **shorthand specifico del settore** per condensare il ragionamento. 17 | 3. **Dai Priorità alla Densità Informativa** → Includi solo **gli elementi essenziali del ragionamento**. 18 | 4. **Segui la Notazione Standardizzata** → Attieniti alle **convenzioni ampiamente riconosciute** in ciascun campo. 19 | 5. **Mantieni la Precisione Strutturale** → Assicurati che le risposte siano formattate utilizzando **espressioni compatte e specifiche del settore**. 20 | 21 | --- 22 | 23 | ## **Regole & Direttive** 24 | 1. **Utilizza Notazione Specifica del Settore** 25 | - **Ragionamento Matematico & Logico** → `∑, ∴, ∝, Δ, →` 26 | - **Discipline Scientifiche** → `mol, J, Hz, pH, Vmax` 27 | - **Campi Medici & Ingegneristici** → `CHF, OOP, PID, μm, dB` 28 | 29 | 2. **Elimina Testo Ridondante** 30 | - **Niente frasi complete** – le risposte devono essere in **notazione strutturata**. 31 | - **Niente ripetizione della domanda** – esprimi direttamente la soluzione. 32 | 33 | 3. **Mantieni le Risposte Ultra-Compatte** 34 | - **Dai priorità alla brevità** mantenendo **precisione tecnica**. 35 | - Segui **gli standard di settore** per la notazione e il ragionamento strutturato. 36 | 37 | 4. **Formato di Output** 38 | - Usa esattamente il formato strutturato: 39 | ``` 40 | 41 | [Ragionamento tecnico usando notazione esperta] 42 | 43 | \boxed{[Risultato finale]} 44 | ``` 45 | - Il **risultato finale deve essere incorniciato (boxed)**. 46 | - **Se la domanda è a scelta multipla, restituisci l'opzione corretta all'interno del box.** -------------------------------------------------------------------------------- /sketch_of_thought/config/prompts/DE/DE_ExpertLexicons_SystemPrompt.md: -------------------------------------------------------------------------------- 1 | ## **Rolle & Zielsetzung** 2 | Du bist ein Experte für logisches Denken, spezialisiert auf **Expert Lexicons**, eine kognitive Argumentationstechnik, die **domänenspezifische Kurzschrift, technische Symbole und Fachjargon** nutzt, um präzise und effiziente Kommunikation zu gewährleisten. Dein Ziel ist es, **Argumentationen in hochinformationsdichte Ausdrücke zu komprimieren**, während du **technische Genauigkeit und Klarheit** bewahrst. 3 | 4 | Expert Lexicons basiert auf dem Prinzip, dass **Fachexperten mit Kurzschrift und strukturierter Notation kommunizieren**. Anstatt vollständige Erklärungen zu liefern, **verdichtet diese Methode Argumentationen in kompakte, hochdichte Ausdrücke**, die technische Symbole und feldspezifische Abkürzungen verwenden. 5 | 6 | Diese Methode ist besonders effektiv für: 7 | - **Technische Disziplinen** (Naturwissenschaften, Ingenieurwesen, Medizin, Mathematik und Programmierung) 8 | - **Symbolisches und formelhaftes Denken** (unter Nutzung feldspezifischer Notation und logischer Ausdrücke) 9 | - **Maximierung der Effizienz** (Übermittlung von Informationen mit minimaler Zeichenanzahl) 10 | 11 | --- 12 | 13 | ## **Anwendung von Expert Lexicons** 14 | ### **Schritt-für-Schritt-Anleitung** 15 | 1. **Verwende technische Symbole** → Ersetze allgemeine Begriffe durch **mathematische, logische oder wissenschaftliche Notation**, wo anwendbar. 16 | 2. **Nutze Abkürzungen** → Verwende **domänenspezifische Kurzschrift**, um Argumentationen zu verdichten. 17 | 3. **Priorisiere Informationsdichte** → Behalte nur **essenzielle Argumentationselemente** bei. 18 | 4. **Folge standardisierter Notation** → Halte dich an **weit verbreitete Konventionen** innerhalb des Fachgebiets. 19 | 5. **Erhalte strukturelle Präzision** → Stelle sicher, dass Antworten in **kompakten, branchenspezifischen Ausdrücken** formatiert sind. 20 | 21 | --- 22 | 23 | ## **Regeln & Richtlinien** 24 | 1. **Verwende domänenspezifische Notation** 25 | - **Mathematische & logische Argumentation** → `∑, ∴, ∝, Δ, →` 26 | - **Naturwissenschaftliche Disziplinen** → `mol, J, Hz, pH, Vmax` 27 | - **Medizin & Ingenieurwesen** → `CHF, OOP, PID, μm, dB` 28 | 29 | 2. **Eliminiere redundanten Text** 30 | - **Keine vollständigen Sätze** – Antworten müssen in **strukturierter Notation** erfolgen. 31 | - **Keine Wiederholung der Frage** – direkt die Lösung formulieren. 32 | 33 | 3. **Halte Antworten extrem kompakt** 34 | - **Priorisiere Kürze**, ohne die **technische Präzision** zu verlieren. 35 | - Befolge **Branchennormen** für Notation und strukturierte Argumentation. 36 | 37 | 4. **Ausgabeformat** 38 | - Verwende das exakte strukturierte Format: 39 | ``` 40 | 41 | [Kurzschriftliche Argumentation mit Expertennotation] 42 | 43 | \boxed{[Endgültige Antwort]} 44 | ``` 45 | - Die **endgültige Antwort muss eingerahmt sein**. 46 | - **Bei Multiple-Choice-Fragen wird der richtige Buchstabe innerhalb der Box zurückgegeben.** 47 | - **Verwende minimalen Text in der Antwort.** -------------------------------------------------------------------------------- /.gitignore: -------------------------------------------------------------------------------- 1 | # Byte-compiled / optimized / DLL files 2 | __pycache__/ 3 | *.py[cod] 4 | *$py.class 5 | 6 | # C extensions 7 | *.so 8 | 9 | # Distribution / packaging 10 | .Python 11 | build/ 12 | develop-eggs/ 13 | dist/ 14 | downloads/ 15 | eggs/ 16 | .eggs/ 17 | lib/ 18 | lib64/ 19 | parts/ 20 | sdist/ 21 | var/ 22 | wheels/ 23 | share/python-wheels/ 24 | *.egg-info/ 25 | .installed.cfg 26 | *.egg 27 | MANIFEST 28 | 29 | # PyInstaller 30 | # Usually these files are written by a python script from a template 31 | # before PyInstaller builds the exe, so as to inject date/other infos into it. 32 | *.manifest 33 | *.spec 34 | 35 | # Installer logs 36 | pip-log.txt 37 | pip-delete-this-directory.txt 38 | 39 | # Unit test / coverage reports 40 | htmlcov/ 41 | .tox/ 42 | .nox/ 43 | .coverage 44 | .coverage.* 45 | .cache 46 | nosetests.xml 47 | coverage.xml 48 | *.cover 49 | *.py,cover 50 | .hypothesis/ 51 | .pytest_cache/ 52 | cover/ 53 | 54 | # Translations 55 | *.mo 56 | *.pot 57 | 58 | # Django stuff: 59 | *.log 60 | local_settings.py 61 | db.sqlite3 62 | db.sqlite3-journal 63 | 64 | # Flask stuff: 65 | instance/ 66 | .webassets-cache 67 | 68 | # Scrapy stuff: 69 | .scrapy 70 | 71 | # Sphinx documentation 72 | docs/_build/ 73 | 74 | # PyBuilder 75 | .pybuilder/ 76 | target/ 77 | 78 | # Jupyter Notebook 79 | .ipynb_checkpoints 80 | 81 | # IPython 82 | profile_default/ 83 | ipython_config.py 84 | 85 | # pyenv 86 | # For a library or package, you might want to ignore these files since the code is 87 | # intended to run in multiple environments; otherwise, check them in: 88 | # .python-version 89 | 90 | # pipenv 91 | # According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control. 92 | # However, in case of collaboration, if having platform-specific dependencies or dependencies 93 | # having no cross-platform support, pipenv may install dependencies that don't work, or not 94 | # install all needed dependencies. 95 | #Pipfile.lock 96 | 97 | # UV 98 | # Similar to Pipfile.lock, it is generally recommended to include uv.lock in version control. 99 | # This is especially recommended for binary packages to ensure reproducibility, and is more 100 | # commonly ignored for libraries. 101 | #uv.lock 102 | 103 | # poetry 104 | # Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control. 105 | # This is especially recommended for binary packages to ensure reproducibility, and is more 106 | # commonly ignored for libraries. 107 | # https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control 108 | #poetry.lock 109 | 110 | # pdm 111 | # Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control. 112 | #pdm.lock 113 | # pdm stores project-wide configurations in .pdm.toml, but it is recommended to not include it 114 | # in version control. 115 | # https://pdm.fming.dev/latest/usage/project/#working-with-version-control 116 | .pdm.toml 117 | .pdm-python 118 | .pdm-build/ 119 | 120 | # PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm 121 | __pypackages__/ 122 | 123 | # Celery stuff 124 | celerybeat-schedule 125 | celerybeat.pid 126 | 127 | # SageMath parsed files 128 | *.sage.py 129 | 130 | # Environments 131 | .env 132 | .venv 133 | env/ 134 | venv/ 135 | ENV/ 136 | env.bak/ 137 | venv.bak/ 138 | 139 | # Spyder project settings 140 | .spyderproject 141 | .spyproject 142 | 143 | # Rope project settings 144 | .ropeproject 145 | 146 | # mkdocs documentation 147 | /site 148 | 149 | # mypy 150 | .mypy_cache/ 151 | .dmypy.json 152 | dmypy.json 153 | 154 | # Pyre type checker 155 | .pyre/ 156 | 157 | # pytype static type analyzer 158 | .pytype/ 159 | 160 | # Cython debug symbols 161 | cython_debug/ 162 | 163 | # PyCharm 164 | # JetBrains specific template is maintained in a separate JetBrains.gitignore that can 165 | # be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore 166 | # and can be added to the global gitignore or merged into this file. For a more nuclear 167 | # option (not recommended) you can uncomment the following to ignore the entire idea folder. 168 | #.idea/ 169 | 170 | # PyPI configuration file 171 | .pypirc 172 | -------------------------------------------------------------------------------- /sketch_of_thought/sketch_of_thought.py: -------------------------------------------------------------------------------- 1 | from transformers import DistilBertTokenizer, DistilBertForSequenceClassification 2 | import torch 3 | from .config.config import default_path 4 | from .config.warnings import * 5 | import json 6 | import os 7 | import copy 8 | from loguru import logger 9 | 10 | class SoT: 11 | def __init__(self): 12 | # Load the model from HF 13 | self.__MODEL_PATH = "saytes/SoT_DistilBERT" 14 | self.model = DistilBertForSequenceClassification.from_pretrained(self.__MODEL_PATH) 15 | self.tokenizer = DistilBertTokenizer.from_pretrained(self.__MODEL_PATH) 16 | 17 | # Load the label mapping 18 | self.__LABEL_MAPPING_PATH = os.path.join(str(default_path()), "config/label_mapping.json") 19 | self.__LABEL_MAPPING = json.load(open(self.__LABEL_MAPPING_PATH)) 20 | 21 | # Handle paths 22 | self.__PROMPT_PATH_BASE = os.path.join(str(default_path()), "config/prompts/") 23 | self.__CONTEXT_PATH_BASE = os.path.join(str(default_path()), "config/exemplars.json") 24 | self.__PROMPT_FILENAMES = { 25 | "chunked_symbolism": "ChunkedSymbolism_SystemPrompt.md", 26 | "expert_lexicons": "ExpertLexicons_SystemPrompt.md", 27 | "conceptual_chaining": "ConceptualChaining_SystemPrompt.md", 28 | } 29 | 30 | # Handle data 31 | self.PROMPT_CACHE = {} 32 | self.CONTEXT_CACHE = {} 33 | 34 | # Preload prompts and contexts 35 | self.__preload_contexts() 36 | self.__LANGUAGE_CODES = list(self.CONTEXT_CACHE.keys()) 37 | self.__preload_prompts() 38 | 39 | def __preload_prompts(self): 40 | """ 41 | Loads all available system prompts into memory at startup. 42 | """ 43 | 44 | for lang in self.__LANGUAGE_CODES: 45 | self.PROMPT_CACHE[lang] = {} 46 | for paradigm, filename in self.__PROMPT_FILENAMES.items(): 47 | file_path = os.path.join(self.__PROMPT_PATH_BASE, lang, f"{lang}_{filename}") 48 | if os.path.exists(file_path): 49 | with open(file_path, "r", encoding="utf-8") as file: 50 | self.PROMPT_CACHE[lang][paradigm] = file.read() 51 | 52 | def __preload_contexts(self): 53 | """ 54 | Loads all available contexts into memory at startup. 55 | """ 56 | with open(self.__CONTEXT_PATH_BASE, "r") as f: 57 | self.CONTEXT_CACHE = json.load(f) 58 | 59 | def available_languages(self): 60 | """ 61 | Lists all currently supported languages. 62 | """ 63 | return self.__LANGUAGE_CODES 64 | 65 | def available_paradigms(self): 66 | """ 67 | Returns list of all currently supported paradigms. 68 | """ 69 | return list(self.__PROMPT_FILENAMES.keys()) 70 | 71 | def get_system_prompt(self, paradigm="chunked_symbolism", language_code="EN"): 72 | """ 73 | Retrieves the preloaded system prompt based on the given paradigm and language code. 74 | 75 | :param paradigm: The type of prompt (e.g., "chunked_symbolism", "expert_lexicons", "conceptual_chaining"). 76 | :param language_code: The language code (e.g., "EN" for English, "KR" for Korean, etc.). 77 | :return: The content of the corresponding prompt file or None if not found. 78 | """ 79 | assert paradigm in self.available_paradigms(), f"`{paradigm}` is not a recognized paradigm!" 80 | assert language_code in self.available_languages(), f"`{language_code}` is not a compatible language!" 81 | 82 | return copy.deepcopy(self.PROMPT_CACHE[language_code][paradigm]) 83 | 84 | 85 | def get_initialized_context(self, paradigm, question=None, image_data=None, language_code="EN", include_system_prompt=True, format="llm"): 86 | """ 87 | Retrieves the preloaded conversation context for the given paradigm and language. 88 | Dynamically inserts the user's question and system prompt. 89 | 90 | :param paradigm: The reasoning paradigm ("conceptual_chaining", "chunked_symbolism", "expert_lexicons", "cot"). 91 | :param question: The user's question to be added to the context. If `None` or empty, it will not be added. 92 | :param image_data: The image associated with the user's question. Required `format="vlm"`. 93 | :param language_code: The language code (e.g., "KR" for Korean). 94 | :param include_system_prompt: Whether to add the system prompt to the context. Not available in raw format. 95 | :param format: The format to return. Accepted values are: `llm`, `raw`, or `vlm`. 96 | :return: The full initialized conversation list. 97 | """ 98 | 99 | assert paradigm in self.available_paradigms(), f"`{paradigm}` is not a recognized paradigm!" 100 | assert language_code in self.available_languages(), f"`{language_code}` is not a compatible language!" 101 | 102 | if format.lower() == "llm": 103 | # Warn for multimodal misalignment 104 | if image_data: 105 | logger.warning(MULTIMODAL_MISALIGNMENT) 106 | 107 | exemplars = self.CONTEXT_CACHE[language_code][paradigm] 108 | if include_system_prompt: 109 | context = [{"role": "system", "content": self.get_system_prompt(paradigm=paradigm, language_code=language_code)}] 110 | else: 111 | context = [] 112 | 113 | for ex in exemplars: 114 | context += [ 115 | {"role": "user", "content": ex['question']}, 116 | {"role": "assistant", "content": ex['answer']}, 117 | ] 118 | 119 | # Add user question, if it exists 120 | if question and question != "": 121 | context += [{"role": "user", "content": question}] 122 | 123 | return context 124 | 125 | elif format.lower() == "vlm": 126 | # Warn for missing image 127 | if image_data is None: 128 | logger.warning(NO_IMAGE) 129 | 130 | exemplars = self.CONTEXT_CACHE[language_code][paradigm] 131 | if include_system_prompt: 132 | context = [{"role": "system", "content": [{"type": "text", "text": self.get_system_prompt(paradigm=paradigm, language_code=language_code)}]}] 133 | else: 134 | context = [] 135 | 136 | for ex in exemplars: 137 | context += [ 138 | {"role": "user", "content": [{"type": "text", "text": ex['question']}]}, 139 | {"role": "assistant", "content": [{"type": "text", "text": ex['answer']}]}, 140 | ] 141 | 142 | # Add user question, if it exists 143 | if question and question != "": 144 | context += [{"role": "user", "content": [{"type": "text", "text": question}, {"type": "image", "image": image_data}]}] 145 | return context 146 | 147 | else: # Default case, return raw format 148 | return copy.deepcopy(self.CONTEXT_CACHE[language_code][paradigm]) 149 | 150 | def classify_question(self, question): 151 | """ 152 | Uses the pretrained DistilBERT model to classify the paradigm of a question. 153 | """ 154 | 155 | inputs = self.tokenizer(question, return_tensors="pt", truncation=True, padding=True) 156 | outputs = self.model(**inputs) 157 | predicted_class = torch.argmax(outputs.logits, dim=1).item() 158 | 159 | # Reverse mapping to get the paradigm name 160 | label_mapping_reverse = {v: k for k, v in self.__LABEL_MAPPING.items()} 161 | return label_mapping_reverse[predicted_class] 162 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # Sketch-of-Thought (SoT) 2 | 3 | [![License](https://img.shields.io/badge/License-MIT-blue.svg)](LICENSE) 4 | [![Python](https://img.shields.io/badge/Python-3.10+-blue.svg)](https://www.python.org/downloads/) 5 | [![PyTorch](https://img.shields.io/badge/PyTorch-2.0+-orange.svg)](https://pytorch.org/) 6 | [![Hugging Face](https://img.shields.io/badge/🤗%20Hugging%20Face-Compatible-yellow)](https://huggingface.co/saytes/SoT_DistilBERT) 7 | [![Hugging Face](https://img.shields.io/badge/arxiv-paper-blue)](https://arxiv.org/abs/2503.05179) 8 | 9 | ## Introduction 10 | 11 | Sketch-of-Thought (SoT) is a novel prompting framework for efficient reasoning in language models that combines cognitive-inspired reasoning paradigms with linguistic constraints to minimize output token usage while preserving reasoning accuracy. 12 | 13 | Unlike conventional Chain of Thought (CoT) approaches that produce verbose reasoning chains, SoT implements three distinct reasoning paradigms: 14 | 15 | - **Conceptual Chaining**: Connects essential ideas in logical sequences through structured step links. Effective for commonsense reasoning, multi-hop inference, and fact-based recall tasks. 16 | 17 | - **Chunked Symbolism**: Organizes numerical and symbolic reasoning into structured steps with equations, variables, and arithmetic operations. Excels in mathematical problems and technical calculations. 18 | 19 | - **Expert Lexicons**: Leverages domain-specific shorthand, technical symbols, and jargon for precise and efficient communication. Suited for technical disciplines requiring maximum information density. 20 | 21 | SoT includes a paradigm selection model that automatically determines the optimal reasoning approach for a given query, eliminating the need for manual heuristics. 22 | 23 | ## System Prompts 24 | 25 | Here are the system prompts used in our paper. We offer them in English, Korean, Italian, and German. 26 | 27 | | Language | Available System Prompts | 28 | |----------|------------------------------| 29 | | English (EN) | • [Chunked Symbolism](sketch_of_thought/config/prompts/EN/EN_ChunkedSymbolism_SystemPrompt.md)
• [Conceptual Chaining](sketch_of_thought/config/prompts/EN/EN_ConceptualChaining_SystemPrompt.md)
• [Expert Lexicons](sketch_of_thought/config/prompts/EN/EN_ExpertLexicons_SystemPrompt.md) | 30 | | Korean (KR) | • [Chunked Symbolism](sketch_of_thought/config/prompts/KR/KR_ChunkedSymbolism_SystemPrompt.md)
• [Conceptual Chaining](sketch_of_thought/config/prompts/KR/KR_ConceptualChaining_SystemPrompt.md)
• [Expert Lexicons](sketch_of_thought/config/prompts/KR/KR_ExpertLexicons_SystemPrompt.md) | 31 | | Italian (IT) | • [Chunked Symbolism](sketch_of_thought/config/prompts/IT/IT_ChunkedSymbolism_SystemPrompt.md)
• [Conceptual Chaining](sketch_of_thought/config/prompts/IT/IT_ConceptualChaining_SystemPrompt.md)
• [Expert Lexicons](sketch_of_thought/config/prompts/IT/IT_ExpertLexicons_SystemPrompt.md) | 32 | | German (DE) | • [Chunked Symbolism](sketch_of_thought/config/prompts/DE/DE_ChunkedSymbolism_SystemPrompt.md)
• [Conceptual Chaining](sketch_of_thought/config/prompts/DE/DE_ConceptualChaining_SystemPrompt.md)
• [Expert Lexicons](sketch_of_thought/config/prompts/DE/DE_ExpertLexicons_SystemPrompt.md) | 33 | 34 | ## Installation 35 | 36 | 1. **Clone the Repository** 37 | 38 | ```bash 39 | git clone https://github.com/SimonAytes/SoT.git 40 | cd SoT 41 | ``` 42 | 43 | 2. **Create a Conda Environment (Recommended)** 44 | 45 | ```bash 46 | conda create -n sot python=3.10 -y 47 | conda activate sot 48 | ``` 49 | 50 | 3. **Install Dependencies** 51 | 52 | ```bash 53 | pip install -r requirements.txt 54 | pip install -e . 55 | ``` 56 | 57 | ## Quickstart 58 | 59 | Here's a minimal example showing how to use SoT with any LLM: 60 | 61 | ```python 62 | from sketch_of_thought import SoT 63 | 64 | # Initialize SoT 65 | sot = SoT() 66 | 67 | # Classify a question to determine the best reasoning paradigm 68 | question = "Alice has 5 apples. She gives 3 apples to Bob. How many apples does Alice have?" 69 | paradigm = sot.classify_question(question) 70 | # Returns: 'chunked_symbolism' 71 | 72 | # Get the appropriate system prompt for the paradigm 73 | system_prompt = sot.get_system_prompt(paradigm) 74 | 75 | # Get initialized context with exemplars for the selected paradigm 76 | context = sot.get_initialized_context( 77 | paradigm=paradigm, 78 | question=question, 79 | format="llm", 80 | include_system_prompt=True 81 | ) 82 | 83 | # The context can now be passed to any LLM 84 | ``` 85 | 86 | ## Example with Qwen2.5-7B 87 | 88 | Here's a complete example using Qwen2.5-7B-Instruct: 89 | 90 | ```python 91 | from transformers import AutoModelForCausalLM, AutoTokenizer 92 | from sketch_of_thought import SoT 93 | 94 | # Initialize SoT 95 | sot = SoT() 96 | 97 | # Load Qwen model 98 | model_name = "Qwen/Qwen2.5-7B-Instruct" 99 | model = AutoModelForCausalLM.from_pretrained( 100 | model_name, 101 | torch_dtype="auto", 102 | device_map="auto" 103 | ) 104 | tokenizer = AutoTokenizer.from_pretrained(model_name) 105 | 106 | # Prepare the question 107 | prompt = "Alice has 5 apples. She gives 3 apples to Bob. How many apples does Alice have?" 108 | 109 | # Classify and get appropriate context 110 | paradigm = sot.classify_question(prompt) 111 | messages = sot.get_initialized_context( 112 | paradigm, 113 | prompt, 114 | format="llm", 115 | include_system_prompt=True 116 | ) 117 | 118 | # Format for the model 119 | text = tokenizer.apply_chat_template( 120 | messages, 121 | tokenize=False, 122 | add_generation_prompt=True 123 | ) 124 | model_inputs = tokenizer([text], return_tensors="pt").to(model.device) 125 | 126 | # Generate response 127 | generated_ids = model.generate( 128 | **model_inputs, 129 | max_new_tokens=512 130 | ) 131 | generated_ids = [ 132 | output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids) 133 | ] 134 | 135 | # Decode response 136 | response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0] 137 | print(response) 138 | ``` 139 | 140 | **Output:** 141 | 142 | ``` 143 | 144 | A = 5 145 | A -= 3 146 | A = 2 147 | 148 | 149 | \boxed{2} 150 | ``` 151 | 152 | ## Helper Functions 153 | 154 | SoT provides several utility functions: 155 | 156 | ```python 157 | # List available reasoning paradigms 158 | sot.available_paradigms() 159 | # Returns: ['chunked_symbolism', 'conceptual_chaining', 'expert_lexicons'] 160 | 161 | # List supported languages 162 | sot.available_languages() 163 | # Returns: ['EN', 'KR', 'IT', 'DE'] 164 | 165 | # Get formatted context without a question 166 | context = sot.get_initialized_context(paradigm="conceptual_chaining", format="llm") 167 | 168 | # Get raw exemplars 169 | raw_examples = sot.get_initialized_context(paradigm="chunked_symbolism", format="raw") 170 | ``` 171 | 172 | ## Supported Formats 173 | 174 | Our code supports multiple output formats: 175 | 176 | - `"llm"`: Standard chat format for text-only LLMs 177 | - `"vlm"`: Multimodal format for vision-language models 178 | - `"raw"`: Raw exemplars without formatting 179 | 180 |
181 | What's the difference? 182 | 183 | ### LLM Format 184 | 185 | Standard `messages` format for Large Language Models. 186 | 187 | ```python 188 | [ 189 | { 190 | "role": "system", 191 | "content": "SYSTEM_PROMPT_HERE" 192 | }, 193 | { 194 | "role": "user", 195 | "content": "EXAMPLE_QUESTION_HERE" 196 | }, 197 | { 198 | "role": "assistant", 199 | "content": "EXAMPLE_ANSWER_HERE" 200 | }, 201 | { 202 | "role": "user", 203 | "content": "USER_QUESTION_HERE" 204 | } 205 | ] 206 | ``` 207 | 208 | ### VLM Format 209 | 210 | Standard `messages` format for Large Vision-Language Models. 211 | 212 | ```python 213 | [ 214 | { 215 | "role": "system", 216 | "content": "SYSTEM_PROMPT_HERE" 217 | }, 218 | { 219 | "role": "user", 220 | "content": [{"type": "text", "text": "EXAMPLE_QUESTION_HERE"}] 221 | }, 222 | { 223 | "role": "assistant", 224 | "content": [{"type": "text", "text": "EXAMPLE_ANSWER_HERE"}] 225 | }, 226 | { 227 | "role": "user", 228 | "content": [{"type": "text", "text": "USER_QUESTION_HERE"}] 229 | } 230 | ] 231 | ``` 232 | 233 | ### Raw Format 234 | 235 | Raw exemplar data. Apply your own format! 236 | 237 | ```python 238 | [ 239 | { 240 | "question": "EXAMPLE_QUESTION_HERE", 241 | "answer": "EXAMPLE_ANSWER_HERE" 242 | }, 243 | { 244 | "question": "EXAMPLE_QUESTION_HERE", 245 | "answer": "EXAMPLE_ANSWER_HERE" 246 | } 247 | ] 248 | ``` 249 |
250 | 251 | ## Multilingual Support 252 | 253 | SoT supports multiple languages (depending on your configuration). System prompts and exemplars are automatically loaded in the requested language. 254 | 255 | ## Paradigm Selection Model 256 | 257 | SoT includes a pretrained DistilBERT model for automatic paradigm selection based on the question. The model is available on Hugging Face: [saytes/SoT_DistilBERT](https://huggingface.co/saytes/SoT_DistilBERT) 258 | 259 | ## Datasets 260 | 261 | The SoT_DistilBERT model was evaluated on the following datasets: 262 | 263 | | Dataset | HF ID | Subset | Split | Evaluation Type | 264 | |---------|-------|--------|-------|----------------| 265 | | GSM8K | [gsm8k](https://huggingface.co/datasets/gsm8k) | main | test | numerical | 266 | | SVAMP | [ChilleD/SVAMP](https://huggingface.co/datasets/ChilleD/SVAMP) | - | test | numerical | 267 | | AQUA-RAT | [aqua_rat](https://huggingface.co/datasets/aqua_rat) | - | test | multiple_choice | 268 | | DROP | [drop](https://huggingface.co/datasets/drop) | - | validation | open | 269 | | OpenbookQA | [openbookqa](https://huggingface.co/datasets/openbookqa) | - | test | multiple_choice | 270 | | StrategyQA | [ChilleD/StrategyQA](https://huggingface.co/datasets/ChilleD/StrategyQA) | - | test | yesno | 271 | | LogiQA | [lucasmccabe/logiqa](https://huggingface.co/datasets/lucasmccabe/logiqa) | default | test | multiple_choice | 272 | | Reclor | [metaeval/reclor](https://huggingface.co/datasets/metaeval/reclor) | - | validation | multiple_choice | 273 | | HotPotQA | [hotpot_qa](https://huggingface.co/datasets/hotpot_qa) | distractor | validation | open | 274 | | MuSiQue-Ans | [dgslibisey/MuSiQue](https://huggingface.co/datasets/dgslibisey/MuSiQue) | - | validation | open | 275 | | QASC | [allenai/qasc](https://huggingface.co/datasets/allenai/qasc) | - | validation | multiple_choice | 276 | | Worldtree | [nguyen-brat/worldtree](https://huggingface.co/datasets/nguyen-brat/worldtree) | - | train | multiple_choice | 277 | | PubMedQA | [qiaojin/PubMedQA](https://huggingface.co/datasets/qiaojin/PubMedQA) | pqa_labeled | train | yesno | 278 | | MedQA | [bigbio/med_qa](https://huggingface.co/datasets/bigbio/med_qa) | med_qa_en_source | validation | multiple_choice | 279 | 280 | ## Citation 281 | 282 | If you find our work helpful, please cite: 283 | 284 | ``` 285 | @misc{aytes2025sot, 286 | title={Sketch-of-Thought: Efficient LLM Reasoning with Adaptive Cognitive-Inspired Sketching}, 287 | author={Simon A. Aytes and Jinheon Baek and Sung Ju Hwang}, 288 | year={2025}, 289 | eprint={2503.05179}, 290 | archivePrefix={arXiv}, 291 | primaryClass={cs.CL}, 292 | url={https://arxiv.org/abs/2503.05179}, 293 | } 294 | ``` 295 | 296 | ## License 297 | 298 | This project is licensed under the MIT License - see the [LICENSE](LICENSE) file for details. 299 | -------------------------------------------------------------------------------- /sketch_of_thought/config/exemplars.json: -------------------------------------------------------------------------------- 1 | { 2 | "EN": { 3 | "conceptual_chaining": [ 4 | { 5 | "question": "What is the name of the currency used in Seoul?", 6 | "answer": "\n#Seoul \u2192 #South_Korea \u2192 Won\n\n\\boxed{Korean Won}" 7 | }, 8 | { 9 | "question": "Which planet has the highest surface temperature?\nChoices:\nA) Mercury\nB) Venus\nC) Mars\nD) Jupiter", 10 | "answer": "\n#heat_trap\nMercury \u2192 no atmosphere \u2192 loses heat\nVenus \u2192 thick CO2 \u2192 traps heat \u2192 hottest\nMars \u2192 thin CO2 \u2192 cold\nJupiter \u2192 no solid surface\n\n\\boxed{B}" 11 | }, 12 | { 13 | "question": "Which vitamin is essential for blood clotting?", 14 | "answer": "\n#blood_clotting \u2192 #vitamin_K\n\n\\boxed{Vitamin K}" 15 | } 16 | ], 17 | "chunked_symbolism": [ 18 | { 19 | "question": "A car accelerates at 2.5 m/s^2 for 10 seconds. If its initial velocity was 15 m/s, what is its final velocity?", 20 | "answer": "\na = 2.5 m/s^2\nt = 10 s\nvi = 15 m/s\nvf = 15 + (2.5 \u00d7 10)\nvf = 40 m/s\n\n\\boxed{40}" 21 | }, 22 | { 23 | "question": "If a product costs $120 and there is a 15% discount, what is the final price?\nChoices:\nA) $10\nB) $97\nC) 102", 24 | "answer": "\nop = 120\nd = 15%\ndp = 120 \u00d7 (15 / 100) = 18\nfp = 120 - 18 = 102\n\n\\boxed{C}" 25 | }, 26 | { 27 | "question": "Question: A circuit has a voltage of 12V and a resistance of 4\u03a9. What is the current?", 28 | "answer": "\nV = 12V\nR = 4\u03a9\nI = 12 / 4 = 3A\n\n\\boxed{3}" 29 | } 30 | ], 31 | "expert_lexicons": [ 32 | { 33 | "question": "Context: The discovery of the first interstellar object passing through the Solar System, 1I/2017 U1 ('Oumuamua), provoked intense and continuing interest from the scientific community and the general public.\nQuestion: The interstellar object 1I/2017 U1 ('Oumuamua) exhibited unusual characteristics that led to various hypotheses about its origin. What does the designation \"1I/2017 U1\" signify?\nChoices:\nA) 1st Intergalactic object detected in 2017, classified under category U1\nB) 1st Interstellar object cataloged, detected in 2017, following IAU naming conventions\nC) 1st Independent Unclassified body observed beyond Neptune in 2017", 34 | "answer": "\n1I \u2192 1st interstellar object\n2017 \u2192 Year detected\nU1 \u2192 Sequence ID\nIAU \u2192 Naming rules\n\u2234 1st cataloged interstellar object (2017)\n\n\\boxed{B}" 35 | }, 36 | { 37 | "question": "A patient with STEMI is given MONA therapy. They have a history of being allergic to aspirin. Are they at risk with this treatment?", 38 | "answer": "\nSTEMI \u2192 ST-Elevation MI\nMONA \u2192 {Morphine, O2, Nitrates, Aspirin}\n\u2234 Aspirin \u2208 MONA\n\n\\boxed{Yes}" 39 | }, 40 | { 41 | "question": "What does EBITDA measure?", 42 | "answer": "\nEBITDA \u2192 Earnings Before Interest, Taxes, Depreciation, Amortization\n\u2234 Measures Core Profitability\n\n\\boxed{Core Profitability}" 43 | } 44 | ], 45 | "cot": [ 46 | { 47 | "question": "Question: \"If a train travels 60 miles per hour for 3 hours, how far does it go?\"", 48 | "answer": "\n1. I understand we need to find the total distance traveled by:\n\t- A train moving at 60 miles per hour\n\t- For a duration of 3 hours\n\n2. To calculate the distance, I'll use the formula:\n\t- Distance = Speed \u00d7 Time\n\t- Distance = 60 miles/hour \u00d7 3 hours\n\n3. Now I'll perform the calculation:\n\t- Distance = 60 \u00d7 3 = 180 miles\n\n4. Verification:\n\t- This makes sense because the train moves 60 miles each hour\n\t- After 3 hours, it will have covered 3 times that distance\n\n\\boxed{180}" 49 | }, 50 | { 51 | "question": "Which planet has the highest surface temperature?\nChoices:\nA) Mercury\nB) Venus\nC) Mars\nD) Jupiter", 52 | "answer": "\n1. First, let's recall some basic facts about the planets listed:\n\t- Mercury: Closest to the Sun, no substantial atmosphere.\n\t- Venus: Second planet from the Sun, thick atmosphere with strong greenhouse effect.\n\t- Mars: Fourth planet from the Sun, thin atmosphere.\n\t- Jupiter: Fifth planet from the Sun, gas giant with no solid surface.\n\n2. Consider the factors affecting surface temperature:\n\t- Proximity to the Sun.\n\t- Atmospheric composition and thickness.\n\t- Presence of a solid surface to measure temperature on.\n\n3. Evaluate each planet:\n\t- Mercury: Despite being closest to the Sun, its lack of atmosphere means it experiences extreme temperature variations between day and night.\n\t- Venus: Its thick atmosphere traps heat due to a strong greenhouse effect, leading to very high surface temperatures.\n\t- Mars: Being farther from the Sun and having a thin atmosphere results in lower temperatures.\n\t- Jupiter: As a gas giant, it doesn't have a solid surface, so it doesn't fit the criteria for \"surface\" temperature comparison.\n\n4. Conclusion:\n\t- Among the planets with a solid surface, Venus has the highest surface temperature due to its dense atmosphere trapping heat.\n\n\\boxed{B}" 53 | } 54 | ] 55 | }, 56 | "KR": { 57 | "conceptual_chaining": [ 58 | { 59 | "question": "\uc11c\uc6b8\uc5d0\uc11c \uc0ac\uc6a9\ub418\ub294 \ud1b5\ud654\uc758 \uc774\ub984\uc740 \ubb34\uc5c7\uc778\uac00\uc694?", 60 | "answer": "\n#\uc11c\uc6b8 \u2192 #\ub300\ud55c\ubbfc\uad6d \u2192 \uc6d0\n\n\\boxed{\ub300\ud55c\ubbfc\uad6d \uc6d0}" 61 | }, 62 | { 63 | "question": "\ud45c\uba74 \uc628\ub3c4\uac00 \uac00\uc7a5 \ub192\uc740 \ud589\uc131\uc740 \ubb34\uc5c7\uc778\uac00\uc694?\n\uc120\ud0dd\uc9c0:\nA) \uc218\uc131\nB) \uae08\uc131\nC) \ud654\uc131\nD) \ubaa9\uc131", 64 | "answer": "\n#\uc5f4_\uac00\ub450\uae30\n\uc218\uc131 \u2192 \ub300\uae30 \uc5c6\uc74c \u2192 \uc5f4 \uc190\uc2e4\n\uae08\uc131 \u2192 \ub450\uaebc\uc6b4 CO2 \u2192 \uc5f4 \uac00\ub460 \u2192 \uac00\uc7a5 \ub728\uac70\uc6c0\n\ud654\uc131 \u2192 \uc587\uc740 CO2 \u2192 \ucd94\uc6c0\n\ubaa9\uc131 \u2192 \uace0\uccb4 \ud45c\uba74 \uc5c6\uc74c\n\n\\boxed{B}" 65 | }, 66 | { 67 | "question": "\ud608\uc561 \uc751\uace0\uc5d0 \ud544\uc218\uc801\uc778 \ube44\ud0c0\ubbfc\uc740 \ubb34\uc5c7\uc778\uac00\uc694?", 68 | "answer": "\n#\ud608\uc561_\uc751\uace0 \u2192 #\ube44\ud0c0\ubbfc_K\n\n\\boxed{\ube44\ud0c0\ubbfc K}" 69 | } 70 | ], 71 | "chunked_symbolism": [ 72 | { 73 | "question": "\uc790\ub3d9\ucc28\uac00 2.5 m/s^2\uc758 \uac00\uc18d\ub3c4\ub85c 10\ucd08 \ub3d9\uc548 \uac00\uc18d\ud569\ub2c8\ub2e4. \ucd08\uae30 \uc18d\ub3c4\uac00 15 m/s\uc600\ub2e4\uba74 \ucd5c\uc885 \uc18d\ub3c4\ub294 \uc5bc\ub9c8\uc785\ub2c8\uae4c?", 74 | "answer": "\na = 2.5 m/s^2\nt = 10 s\nvi = 15 m/s\nvf = 15 + (2.5 \u00d7 10)\nvf = 40 m/s\n\n\\boxed{40}" 75 | }, 76 | { 77 | "question": "\uc0c1\ud488\uc758 \uac00\uaca9\uc774 $120\uc774\uace0 15% \ud560\uc778\uc774 \uc801\uc6a9\ub420 \uacbd\uc6b0 \ucd5c\uc885 \uac00\uaca9\uc740 \uc5bc\ub9c8\uc785\ub2c8\uae4c?\n\uc120\ud0dd\uc9c0:\nA) $10\nB) $97\nC) 102", 78 | "answer": "\nop = 120\nd = 15%\ndp = 120 \u00d7 (15 / 100) = 18\nfp = 120 - 18 = 102\n\n\\boxed{C}" 79 | }, 80 | { 81 | "question": "\uc9c8\ubb38: \ud68c\ub85c\uc758 \uc804\uc555\uc774 12V\uc774\uace0 \uc800\ud56d\uc774 4\u03a9\uc77c \ub54c, \uc804\ub958\ub294 \uc5bc\ub9c8\uc785\ub2c8\uae4c?", 82 | "answer": "\nV = 12V\nR = 4\u03a9\nI = 12 / 4 = 3A\n\n\\boxed{3}" 83 | } 84 | ], 85 | "expert_lexicons": [ 86 | { 87 | "question": "\ub9e5\ub77d: \ud0dc\uc591\uacc4\ub97c \ud1b5\uacfc\ud558\ub294 \uccab \ubc88\uc9f8 \uc131\uac04 \ucc9c\uccb4 1I/2017 U1 ('Oumuamua)\uc758 \ubc1c\uacac\uc740 \uacfc\ud559\uacc4\uc640 \ub300\uc911\uc758 \uac15\ud55c \uad00\uc2ec\uc744 \ubd88\ub7ec\uc77c\uc73c\ucf30\uc2b5\ub2c8\ub2e4.\n\uc9c8\ubb38: \uc131\uac04 \ucc9c\uccb4 1I/2017 U1 ('Oumuamua)\uc740 \uae30\uc6d0\uc5d0 \ub300\ud55c \ub2e4\uc591\ud55c \uac00\uc124\uc744 \ucd09\ubc1c\ud55c \ud2b9\uc774\ud55c \ud2b9\uc131\uc744 \ubcf4\uc600\uc2b5\ub2c8\ub2e4. \"1I/2017 U1\"\uc774\ub77c\ub294 \uba85\uce6d\uc740 \ubb34\uc5c7\uc744 \uc758\ubbf8\ud569\ub2c8\uae4c?\n\uc120\ud0dd\uc9c0:\nA) 2017\ub144\uc5d0 \ud0d0\uc9c0\ub41c \ucd5c\ucd08\uc758 \uc740\ud558 \uac04 \ucc9c\uccb4\ub85c, U1 \ubc94\uc8fc\ub85c \ubd84\ub958\ub428\nB) 2017\ub144\uc5d0 \ud0d0\uc9c0\ub418\uc5b4 \uad6d\uc81c\ucc9c\ubb38\uc5f0\ub9f9(IAU) \uba85\uba85\ubc95\uc744 \ub530\ub978 \ucd5c\ucd08\uc758 \uc131\uac04 \ucc9c\uccb4\ub85c \ub4f1\ub85d\ub428\nC) 2017\ub144\uc5d0 \ud574\uc655\uc131 \ub108\uba38\uc5d0\uc11c \uad00\uce21\ub41c \ucd5c\ucd08\uc758 \ub3c5\ub9bd\uc801\uc778 \ubbf8\ubd84\ub958 \ucc9c\uccb4", 88 | "answer": "\n1I \u2192 \ucd5c\ucd08\uc758 \uc131\uac04 \ucc9c\uccb4\n2017 \u2192 \ud0d0\uc9c0 \uc5f0\ub3c4\nU1 \u2192 \uc21c\ubc88 ID\nIAU \u2192 \uba85\uba85 \uaddc\uce59\n\u2234 2017\ub144 \ucd5c\ucd08\ub85c \ub4f1\ub85d\ub41c \uc131\uac04 \ucc9c\uccb4\n\n\\boxed{B}" 89 | }, 90 | { 91 | "question": "STEMI \ud658\uc790\uac00 MONA \uce58\ub8cc\ub97c \ubc1b\uc558\uc2b5\ub2c8\ub2e4. \uadf8\ub7ec\ub098 \uc774 \ud658\uc790\ub294 \uc544\uc2a4\ud53c\ub9b0\uc5d0 \uc54c\ub808\ub974\uae30\uac00 \uc788\ub294 \ubcd1\ub825\uc774 \uc788\uc2b5\ub2c8\ub2e4. \uc774 \uce58\ub8cc\uac00 \uc704\ud5d8\ud560 \uac00\ub2a5\uc131\uc774 \uc788\uc2b5\ub2c8\uae4c?", 92 | "answer": "\nSTEMI \u2192 ST \ubd84\uc808 \uc0c1\uc2b9 \uc2ec\uadfc\uacbd\uc0c9 (ST-Elevation MI)\nMONA \u2192 {\ubaa8\ub974\ud540, \uc0b0\uc18c, \uc9c8\uc0b0\uc5fc, \uc544\uc2a4\ud53c\ub9b0}\n\u2234 \uc544\uc2a4\ud53c\ub9b0 \u2208 MONA\n\n\\boxed{Yes}" 93 | }, 94 | { 95 | "question": "EBITDA\ub294 \ubb34\uc5c7\uc744 \uce21\uc815\ud558\ub098\uc694?", 96 | "answer": "\nEBITDA \u2192 \uc774\uc790, \uc138\uae08, \uac10\uac00\uc0c1\uac01 \ubc0f \ubb34\ud615\uc790\uc0b0\uc0c1\uac01 \uc804 \uc774\uc775\n\u2234 \ud575\uc2ec \uc218\uc775\uc131 \uce21\uc815\n\n\\boxed{\ud575\uc2ec \uc218\uc775\uc131}" 97 | } 98 | ], 99 | "cot": [ 100 | { 101 | "question": "\uc9c8\ubb38: \"\uae30\ucc28\uac00 \uc2dc\uc18d 60\ub9c8\uc77c\ub85c 3\uc2dc\uac04 \ub3d9\uc548 \uc774\ub3d9\ud558\uba74, \ucd1d \uc5bc\ub9c8\ub098 \uc774\ub3d9\ud558\ub294\uac00?\"", 102 | "answer": "\n1. \uc6b0\ub9ac\ub294 \ucd1d \uc774\ub3d9 \uac70\ub9ac\ub97c \uad6c\ud574\uc57c \ud569\ub2c8\ub2e4:\n\t- \uae30\ucc28\ub294 \uc2dc\uc18d 60\ub9c8\uc77c\ub85c \uc774\ub3d9\ud569\ub2c8\ub2e4.\n\t- \uc774\ub3d9 \uc2dc\uac04\uc740 3\uc2dc\uac04\uc785\ub2c8\ub2e4.\n\n2. \uac70\ub9ac\ub97c \uacc4\uc0b0\ud558\uae30 \uc704\ud574 \ub2e4\uc74c \uacf5\uc2dd\uc744 \uc0ac\uc6a9\ud569\ub2c8\ub2e4:\n\t- \uac70\ub9ac = \uc18d\ub825 \u00d7 \uc2dc\uac04\n\t- \uac70\ub9ac = 60\ub9c8\uc77c/\uc2dc\uac04 \u00d7 3\uc2dc\uac04\n\n3. \uc774\uc81c \uacc4\uc0b0\uc744 \uc218\ud589\ud569\ub2c8\ub2e4:\n\t- \uac70\ub9ac = 60 \u00d7 3 = 180\ub9c8\uc77c\n\n4. \uac80\uc99d:\n\t- \uae30\ucc28\ub294 \ub9e4\uc2dc\uac04 60\ub9c8\uc77c\uc744 \uc774\ub3d9\ud569\ub2c8\ub2e4.\n\t- \ub530\ub77c\uc11c 3\uc2dc\uac04 \ud6c4\uc5d0\ub294 3\ubc30\uc758 \uac70\ub9ac\ub97c \uc774\ub3d9\ud558\uac8c \ub429\ub2c8\ub2e4.\n\n\\boxed{180}" 103 | }, 104 | { 105 | "question": "\uc5b4\ub290 \ud589\uc131\uc774 \uac00\uc7a5 \ub192\uc740 \ud45c\uba74 \uc628\ub3c4\ub97c \uac00\uc9c0\uace0 \uc788\ub294\uac00?\n\uc120\ud0dd\uc9c0:\nA) \uc218\uc131\nB) \uae08\uc131\nC) \ud654\uc131\nD) \ubaa9\uc131", 106 | "answer": "\n1. \uba3c\uc800 \uc8fc\uc5b4\uc9c4 \ud589\uc131\ub4e4\uc5d0 \ub300\ud55c \uae30\ubcf8 \uc815\ubcf4\ub97c \ub5a0\uc62c\ub824 \ubd05\uc2dc\ub2e4:\n\t- \uc218\uc131: \ud0dc\uc591\uacfc \uac00\uc7a5 \uac00\uae4c\uc6b4 \ud589\uc131\uc774\uc9c0\ub9cc, \ub300\uae30\uac00 \uac70\uc758 \uc5c6\uc74c.\n\t- \uae08\uc131: \ud0dc\uc591\uc73c\ub85c\ubd80\ud130 \ub450 \ubc88\uc9f8 \ud589\uc131\uc774\uba70, \ub450\uaebc\uc6b4 \ub300\uae30\uc640 \uac15\ud55c \uc628\uc2e4\ud6a8\uacfc\uac00 \uc788\uc74c.\n\t- \ud654\uc131: \ud0dc\uc591\uc73c\ub85c\ubd80\ud130 \ub124 \ubc88\uc9f8 \ud589\uc131\uc774\uba70, \uc587\uc740 \ub300\uae30\ub97c \uac00\uc9d0.\n\t- \ubaa9\uc131: \ud0dc\uc591\uc73c\ub85c\ubd80\ud130 \ub2e4\uc12f \ubc88\uc9f8 \ud589\uc131\uc774\uba70, \uace0\uccb4 \ud45c\uba74\uc774 \uc5c6\ub294 \uac00\uc2a4\ud615 \ud589\uc131.\n\n2. \ud45c\uba74 \uc628\ub3c4\uc5d0 \uc601\ud5a5\uc744 \ubbf8\uce58\ub294 \uc694\uc778\uc744 \uace0\ub824\ud569\ub2c8\ub2e4:\n\t- \ud0dc\uc591\uacfc\uc758 \uac70\ub9ac.\n\t- \ub300\uae30\uc758 \uc870\uc131\uacfc \ub450\uaed8.\n\t- \uce21\uc815 \uac00\ub2a5\ud55c \uace0\uccb4 \ud45c\uba74\uc758 \uc874\uc7ac \uc5ec\ubd80.\n\n3. \uac01 \ud589\uc131\uc744 \ud3c9\uac00\ud569\ub2c8\ub2e4:\n\t- \uc218\uc131: \ud0dc\uc591\uacfc \uac00\uc7a5 \uac00\uae5d\uc9c0\ub9cc, \ub300\uae30\uac00 \uc5c6\uc5b4\uc11c \ub0ae\uacfc \ubc24\uc758 \uc628\ub3c4 \ucc28\uc774\uac00 \uadf9\uc2ec\ud568.\n\t- \uae08\uc131: \ub450\uaebc\uc6b4 \ub300\uae30\uac00 \uac15\ud55c \uc628\uc2e4\ud6a8\uacfc\ub97c \uc77c\uc73c\ucf1c \ub9e4\uc6b0 \ub192\uc740 \ud45c\uba74 \uc628\ub3c4\ub97c \uc720\uc9c0\ud568.\n\t- \ud654\uc131: \ud0dc\uc591\uc5d0\uc11c \uba40\uace0 \ub300\uae30\uac00 \uc587\uc544 \uc628\ub3c4\uac00 \ub0ae\uc74c.\n\t- \ubaa9\uc131: \uac00\uc2a4\ud615 \ud589\uc131\uc774\ubbc0\ub85c \"\ud45c\uba74\" \uc628\ub3c4\ub97c \ube44\uad50\ud558\ub294 \uac83\uc774 \uc801\uc808\ud558\uc9c0 \uc54a\uc74c.\n\n4. \uacb0\ub860:\n\t- \uace0\uccb4 \ud45c\uba74\uc774 \uc788\ub294 \ud589\uc131 \uc911\uc5d0\uc11c, \uae08\uc131\uc774 \ub450\uaebc\uc6b4 \ub300\uae30\ub85c \uc778\ud574 \uac00\uc7a5 \ub192\uc740 \ud45c\uba74 \uc628\ub3c4\ub97c \uac00\uc9d0.\n\n\\boxed{B}" 107 | } 108 | ] 109 | }, 110 | "IT": { 111 | "conceptual_chaining": [ 112 | { 113 | "question": "Qual \u00e8 il nome della valuta utilizzata a Seul?", 114 | "answer": "\n#Seul \u2192 #Corea_del_Sud \u2192 Won\n\n\\boxed{Won sudcoreano}" 115 | }, 116 | { 117 | "question": "Quale pianeta ha la temperatura superficiale pi\u00f9 alta?\nOpzioni:\nA) Mercurio\nB) Venere\nC) Marte\nD) Giove", 118 | "answer": "\n#intrappolamento_del_calore\nMercurio \u2192 nessuna atmosfera \u2192 perde calore\nVenere \u2192 CO2 densa \u2192 intrappola calore \u2192 pi\u00f9 caldo\nMarte \u2192 CO2 sottile \u2192 freddo\nGiove \u2192 nessuna superficie solida\n\n\\boxed{B}" 119 | }, 120 | { 121 | "question": "Quale vitamina \u00e8 essenziale per la coagulazione del sangue?", 122 | "answer": "\n#coagulazione_del_sangue \u2192 #vitamina_K\n\n\\boxed{Vitamina K}" 123 | } 124 | ], 125 | "chunked_symbolism": [ 126 | { 127 | "question": "Un'auto accelera a 2,5 m/s^2 per 10 secondi. Se la sua velocit\u00e0 iniziale era di 15 m/s, qual \u00e8 la sua velocit\u00e0 finale?", 128 | "answer": "\na = 2,5 m/s^2\nt = 10 s\nvi = 15 m/s\nvf = 15 + (2,5 \u00d7 10)\nvf = 40 m/s\n\n\\boxed{40}" 129 | }, 130 | { 131 | "question": "Se un prodotto costa $120 e c'\u00e8 uno sconto del 15%, qual \u00e8 il prezzo finale?\nOpzioni:\nA) $10\nB) $97\nC) 102", 132 | "answer": "\nop = 120\nd = 15%\ndp = 120 \u00d7 (15 / 100) = 18\nfp = 120 - 18 = 102\n\n\\boxed{C}" 133 | }, 134 | { 135 | "question": "Domanda: Un circuito ha una tensione di 12V e una resistenza di 4\u03a9. Qual \u00e8 la corrente?", 136 | "answer": "\nV = 12V\nR = 4\u03a9\nI = 12 / 4 = 3A\n\n\\boxed{3}" 137 | } 138 | ], 139 | "expert_lexicons": [ 140 | { 141 | "question": "Contesto: La scoperta del primo oggetto interstellare in transito attraverso il Sistema Solare, 1I/2017 U1 ('Oumuamua), ha suscitato un intenso e continuo interesse sia nella comunit\u00e0 scientifica che nel pubblico generale.\nDomanda: L'oggetto interstellare 1I/2017 U1 ('Oumuamua) ha mostrato caratteristiche insolite che hanno portato a varie ipotesi sulla sua origine. Cosa significa la designazione \"1I/2017 U1\"?\nScelte:\nA) Primo oggetto intergalattico rilevato nel 2017, classificato nella categoria U1\nB) Primo oggetto interstellare catalogato, rilevato nel 2017, secondo le convenzioni di nomenclatura dell'IAU\nC) Primo corpo indipendente non classificato osservato oltre Nettuno nel 2017", 142 | "answer": "\n1I \u2192 Primo oggetto interstellare\n2017 \u2192 Anno di rilevamento\nU1 \u2192 ID di sequenza\nIAU \u2192 Regole di nomenclatura\n\u2234 Primo oggetto interstellare catalogato (2017)\n\n\\boxed{B}" 143 | }, 144 | { 145 | "question": "A un paziente con STEMI viene somministrata la terapia MONA. Ha una storia di allergia all'aspirina. \u00c8 a rischio con questo trattamento?", 146 | "answer": "\nSTEMI \u2192 Infarto miocardico con sopraslivellamento del tratto ST\nMONA \u2192 {Morfina, O2, Nitrati, Aspirina}\n\u2234 Aspirina \u2208 MONA\n\n\\boxed{S\u00ec}" 147 | }, 148 | { 149 | "question": "Cosa misura l'EBITDA?", 150 | "answer": "\nEBITDA \u2192 Utile prima di interessi, imposte, deprezzamento e ammortamento\n\u2234 Misura la redditivit\u00e0 operativa\n\n\\boxed{Redditivit\u00e0 operativa}" 151 | } 152 | ], 153 | "cot": [ 154 | { 155 | "question": "Domanda: \"Se un treno viaggia a 60 miglia all'ora per 3 ore, quanto lontano arriva?\"", 156 | "answer": "\n1. Capisco che dobbiamo trovare la distanza totale percorsa da:\n\t- Un treno che si muove a 60 miglia all'ora\n\t- Per una durata di 3 ore\n\n2. Per calcolare la distanza, user\u00f2 la formula:\n\t- Distanza = Velocit\u00e0 \u00d7 Tempo\n\t- Distanza = 60 miglia/ora \u00d7 3 ore\n\n3. Ora eseguir\u00f2 il calcolo:\n\t- Distanza = 60 \u00d7 3 = 180 miglia\n\n4. Verifica:\n\t- Ha senso perch\u00e9 il treno percorre 60 miglia ogni ora\n\t- Dopo 3 ore, avr\u00e0 coperto 3 volte quella distanza\n\n\\boxed{180}" 157 | }, 158 | { 159 | "question": "Quale pianeta ha la temperatura superficiale pi\u00f9 alta?\nOpzioni:\nA) Mercurio\nB) Venere\nC) Marte\nD) Giove", 160 | "answer": "\n1. Per prima cosa, ricordiamo alcuni fatti di base sui pianeti elencati:\n\t- Mercurio: Il pi\u00f9 vicino al Sole, senza un'atmosfera significativa.\n\t- Venere: Secondo pianeta dal Sole, atmosfera densa con un forte effetto serra.\n\t- Marte: Quarto pianeta dal Sole, atmosfera sottile.\n\t- Giove: Quinto pianeta dal Sole, gigante gassoso senza superficie solida.\n\n2. Consideriamo i fattori che influenzano la temperatura superficiale:\n\t- Prossimit\u00e0 al Sole.\n\t- Composizione e spessore dell'atmosfera.\n\t- Presenza di una superficie solida su cui misurare la temperatura.\n\n3. Valutiamo ciascun pianeta:\n\t- Mercurio: Nonostante sia il pi\u00f9 vicino al Sole, la mancanza di un'atmosfera significa che subisce variazioni estreme di temperatura tra il giorno e la notte.\n\t- Venere: La sua atmosfera densa intrappola il calore a causa di un forte effetto serra, portando a temperature superficiali molto elevate.\n\t- Marte: Essendo pi\u00f9 lontano dal Sole e avendo un'atmosfera sottile, ha temperature pi\u00f9 basse.\n\t- Giove: Essendo un gigante gassoso, non ha una superficie solida, quindi non rientra nei criteri per il confronto della \"temperatura superficiale\".\n\n4. Conclusione:\n\t- Tra i pianeti con una superficie solida, Venere ha la temperatura superficiale pi\u00f9 alta a causa della sua densa atmosfera che intrappola il calore.\n\n\\boxed{B}" 161 | } 162 | ] 163 | }, 164 | "DE": { 165 | "conceptual_chaining": [ 166 | { 167 | "question": "Wie hei\u00dft die W\u00e4hrung, die in Seoul verwendet wird?", 168 | "answer": "\n#Seoul \u2192 #S\u00fcdkorea \u2192 Won\n\n\\boxed{Koreanischer Won}" 169 | }, 170 | { 171 | "question": "Welcher Planet hat die h\u00f6chste Oberfl\u00e4chentemperatur?\nOptionen:\nA) Merkur\nB) Venus\nC) Mars\nD) Jupiter", 172 | "answer": "\n#Hitzefalle\nMerkur \u2192 keine Atmosph\u00e4re \u2192 verliert W\u00e4rme\nVenus \u2192 dicke CO2-Schicht \u2192 speichert W\u00e4rme \u2192 hei\u00dfester Planet\nMars \u2192 d\u00fcnne CO2-Schicht \u2192 kalt\nJupiter \u2192 keine feste Oberfl\u00e4che\n\n\\boxed{B}" 173 | }, 174 | { 175 | "question": "Welches Vitamin ist essenziell f\u00fcr die Blutgerinnung?", 176 | "answer": "\n#Blutgerinnung \u2192 #Vitamin_K\n\n\\boxed{Vitamin K}" 177 | } 178 | ], 179 | "chunked_symbolism": [ 180 | { 181 | "question": "Ein Auto beschleunigt mit 2,5 m/s^2 f\u00fcr 10 Sekunden. Wenn seine Anfangsgeschwindigkeit 15 m/s betr\u00e4gt, wie hoch ist seine Endgeschwindigkeit?", 182 | "answer": "\na = 2,5 m/s^2\nt = 10 s\nvi = 15 m/s\nvf = 15 + (2,5 \u00d7 10)\nvf = 40 m/s\n\n\\boxed{40}" 183 | }, 184 | { 185 | "question": "Wenn ein Produkt 120 $ kostet und es einen Rabatt von 15 % gibt, wie hoch ist der Endpreis?\nAntwortm\u00f6glichkeiten:\nA) 10 $\nB) 97 $\nC) 102 $", 186 | "answer": "\nop = 120\nd = 15%\ndp = 120 \u00d7 (15 / 100) = 18\nfp = 120 - 18 = 102\n\n\\boxed{C}" 187 | }, 188 | { 189 | "question": "Frage: Ein Stromkreis hat eine Spannung von 12V und einen Widerstand von 4\u03a9. Wie hoch ist der Strom?", 190 | "answer": "\nV = 12V\nR = 4\u03a9\nI = 12 / 4 = 3A\n\n\\boxed{3}" 191 | } 192 | ], 193 | "expert_lexicons": [ 194 | { 195 | "question": "Kontext: Die Entdeckung des ersten interstellaren Objekts, das das Sonnensystem durchquert, 1I/2017 U1 ('Oumuamua), l\u00f6ste gro\u00dfes und anhaltendes Interesse in der wissenschaftlichen Gemeinschaft und der allgemeinen \u00d6ffentlichkeit aus.\nFrage: Das interstellare Objekt 1I/2017 U1 ('Oumuamua) zeigte ungew\u00f6hnliche Eigenschaften, die zu verschiedenen Hypothesen \u00fcber seine Herkunft f\u00fchrten. Was bedeutet die Bezeichnung \"1I/2017 U1\"?\nAntwortm\u00f6glichkeiten:\nA) Erstes intergalaktisches Objekt, das 2017 entdeckt wurde, klassifiziert unter der Kategorie U1\nB) Erstes katalogisiertes interstellares Objekt, entdeckt im Jahr 2017, gem\u00e4\u00df den Benennungsrichtlinien der IAU\nC) Erstes unabh\u00e4ngiges, nicht klassifiziertes Objekt, das jenseits von Neptun im Jahr 2017 beobachtet wurde", 196 | "answer": "\n1I \u2192 Erstes interstellares Objekt\n2017 \u2192 Jahr der Entdeckung\nU1 \u2192 Sequenz-ID\nIAU \u2192 Benennungsrichtlinien\n\u2234 Erstes katalogisiertes interstellares Objekt (2017)\n\n\\boxed{B}" 197 | }, 198 | { 199 | "question": "Ein Patient mit STEMI erh\u00e4lt eine MONA-Therapie. Er hat eine Vorgeschichte von Aspirinallergie. Besteht ein Risiko mit dieser Behandlung?", 200 | "answer": "\nSTEMI \u2192 ST-Hebungsinfarkt\nMONA \u2192 {Morphin, O2, Nitrate, Aspirin}\n\u2234 Aspirin \u2208 MONA\n\n\\boxed{Ja}" 201 | }, 202 | { 203 | "question": "Was misst EBITDA?", 204 | "answer": "\nEBITDA \u2192 Gewinn vor Zinsen, Steuern, Abschreibungen und Amortisation\n\u2234 Misst die Kernprofitabilit\u00e4t\n\n\\boxed{Kernprofitabilit\u00e4t}" 205 | } 206 | ], 207 | "cot": [ 208 | { 209 | "question": "Frage: \"Wenn ein Zug mit 60 Meilen pro Stunde f\u00fcr 3 Stunden f\u00e4hrt, wie weit kommt er?\"", 210 | "answer": "\n1. Ich verstehe, dass wir die gesamte zur\u00fcckgelegte Strecke berechnen m\u00fcssen:\n\t- Ein Zug bewegt sich mit 60 Meilen pro Stunde\n\t- F\u00fcr eine Dauer von 3 Stunden\n\n2. Um die Strecke zu berechnen, verwende ich die Formel:\n\t- Strecke = Geschwindigkeit \u00d7 Zeit\n\t- Strecke = 60 Meilen/Stunde \u00d7 3 Stunden\n\n3. Jetzt f\u00fchre ich die Berechnung durch:\n\t- Strecke = 60 \u00d7 3 = 180 Meilen\n\n4. \u00dcberpr\u00fcfung:\n\t- Das ergibt Sinn, da der Zug jede Stunde 60 Meilen zur\u00fccklegt\n\t- Nach 3 Stunden hat er die dreifache Strecke zur\u00fcckgelegt\n\n\\boxed{180}" 211 | }, 212 | { 213 | "question": "Welcher Planet hat die h\u00f6chste Oberfl\u00e4chentemperatur?\nAntwortm\u00f6glichkeiten:\nA) Merkur\nB) Venus\nC) Mars\nD) Jupiter", 214 | "answer": "\n1. Zuerst rufe ich mir einige grundlegende Fakten \u00fcber die aufgef\u00fchrten Planeten ins Ged\u00e4chtnis:\n\t- Merkur: Am n\u00e4chsten zur Sonne, keine nennenswerte Atmosph\u00e4re.\n\t- Venus: Zweiter Planet von der Sonne, dichte Atmosph\u00e4re mit starkem Treibhauseffekt.\n\t- Mars: Vierter Planet von der Sonne, d\u00fcnne Atmosph\u00e4re.\n\t- Jupiter: F\u00fcnfter Planet von der Sonne, Gasriese ohne feste Oberfl\u00e4che.\n\n2. Faktoren, die die Oberfl\u00e4chentemperatur beeinflussen:\n\t- N\u00e4he zur Sonne.\n\t- Zusammensetzung und Dichte der Atmosph\u00e4re.\n\t- Vorhandensein einer festen Oberfl\u00e4che zur Messung der Temperatur.\n\n3. Bewertung der einzelnen Planeten:\n\t- Merkur: Obwohl er der Sonne am n\u00e4chsten ist, f\u00fchrt das Fehlen einer Atmosph\u00e4re zu extremen Temperaturschwankungen zwischen Tag und Nacht.\n\t- Venus: Ihre dichte Atmosph\u00e4re speichert W\u00e4rme aufgrund eines starken Treibhauseffekts, was zu sehr hohen Oberfl\u00e4chentemperaturen f\u00fchrt.\n\t- Mars: Aufgrund der gr\u00f6\u00dferen Entfernung zur Sonne und der d\u00fcnnen Atmosph\u00e4re sind die Temperaturen niedriger.\n\t- Jupiter: Als Gasriese hat er keine feste Oberfl\u00e4che, sodass er nicht f\u00fcr einen Vergleich der \u201eOberfl\u00e4chen\u201c-Temperaturen geeignet ist.\n\n4. Fazit:\n\t- Unter den Planeten mit einer festen Oberfl\u00e4che hat die Venus die h\u00f6chste Oberfl\u00e4chentemperatur, da ihre dichte Atmosph\u00e4re W\u00e4rme speichert.\n\n\\boxed{B}" 215 | } 216 | ] 217 | } 218 | } --------------------------------------------------------------------------------