Note The high school math contest consists of questions covering several branches of mathematics such as algebra, geometry, probability and number theory. Train: 7500 Test: 5000 https://github.com/hendrycks/math

Note 500 questions selected from the MATH benchmark Test: 500 https://github.com/openai/prm800k

Note A variety of elementary math word problem sets (grade school) Train: 200K https://arxiv.org/pdf/2402.14830

Note Grade school math word problems (easy) Train: 7473 Test: 1319 https://arxiv.org/abs/2110.14168

Note AIME 22, AIME 23, and AIME 24 Train: 90 https://artofproblemsolving.com/wiki/index.php/AIME_Problems_and_Solutions

Note 30 problems from the 2024 AIME I and AIME II tests

Note AMC12 2022, AMC12 2023 (Elementary math competition for students in grade 12 and below, examining the more basic high school math knowledge) Train: 83 https://artofproblemsolving.com/wiki/index.php/AMC_12_Problems_and_Solutions

Note AIME 25 Train: 15

Note Decrease generation length, while preserving accuracy across domains such as mathematics, coding, science, and general knowledge. Sky-T1-32B-Preview + PRM800K (12K questions) Train: 10K https://novasky-ai.github.io/posts/reduce-overthinking/

Note Each question has ten multiple-choice options. Train: 12032

Note Train: 12000 Test: 500 https://github.com/openai/prm800k/tree/main

Note Graduate-level multiple-choice questions written by domain experts in biology, physics, and chemistry GPQA Diamond (Rein et al., 2023) consists of 198 PhD level science questions from Biology, Chemistry and Physics. Test: 448 https://github.com/idavidrein/gpqa

Note a continuously updated code benchmark from contests across: LeetCode, AtCoder, and CodeForces Test: 1. release_v1: 400 2. release_v2: 511 3. release_v3: 612 4. release_v4: 713 5. release_v5: 880 https://github.com/LiveCodeBench/LiveCodeBench

Note Competition-level math problems with CoT manner solutions Train: 900K

Note 4428 competition-level problems Test: 4428 https://github.com/KbsdJames/Omni-MATH

Note 11,163 bilingual problems across both text-only and interleaved text-image modalities from 62 distinct Olympic competitions with 13 answer types Test: 11163 https://github.com/GAIR-NLP/OlympicArena

Note Code generation benchmark Train: 10000

Note Math eval benchmark Test: 500

Note Math eval benchmark Test: 272

Note Olympiad-level bilingual multimodal scientific benchmark (math + physics) Test: 8,476 problems from Olympiad-level mathematics and physics competitions

Note TACO is a benchmark for code generation with 26,443 problems. It can be used to evaluate the ability of language models to generate code from natural language specifications. Train: 26443

Note https://github.com/GAIR-NLP/O1-Journey

Note Curated mathematical reasoning data from NuminaMath-CoT, AIME, MATH Train: 817 https://github.com/GAIR-NLP/LIMO

Note 1,000 questions as in s1K but with traces instead generated by DeepSeek r1. Train: 1000 https://github.com/simplescaling/s1

Note Full 59K questions of S1: NuminaMATH, MATH, OlympicArena, OmniMath, AGIEval, xword, OlympiadBench, AIME (1983-2023), TheoremQA, USACO, JEEBench, GPQA, SciEval, s1-prob (128 Stanford statistics qualifying exams), LiveCodeBench, s1-teasers (23 interview questions for quantitative trading positions. Each sample consists of a problem and solution taken from PuzzledQuant (https: //www.puzzledquant.com/). We only take examples with the highest difficulty level ("Hard").) Train: 59029

Note STILL-2 Train: 5K https://github.com/RUCAIBox/Slow_Thinking_with_LLMs

Note STILL-3: MATH, NuminaMathCoT, and AIME 1983-2023 Train: 30K https://github.com/RUCAIBox/Slow_Thinking_with_LLMs

Note Improved the Berkeley Sky-T1 data pipeline using SFT distillation data from DeepSeek-R1 to create Bespoke-Stratos-17k

Note 5k coding data from APPs and TACO, and 10k math data from AIME, MATH, and Olympiads subsets of the NuminaMATH dataset. In addition, we maintain 1k science and puzzle data from STILL-2. Train: 17K https://novasky-ai.github.io/posts/sky-t1/

Note 114k high-quality examples covering math, science, code, and puzzles distilled from DeepSeek-R1 Code: 1. BAAI/TACO 2. codeparrot/apps 3. deepmind/code_contests 4. MatrixStudio/Codeforces-Python-Submissions Math: 1. AI-MO/NuminaMath-CoT Science: 1. camel-ai/chemistry 2. camel-ai/biology 3. camel-ai/physics Puzzle: 1. INK-USC/riddle_sense Train: 113957 https://github.com/open-thoughts/open-thoughts

Note 400k problems from NuminaMath 1.5 distills from DeepSeek R1. Train: 220K default: 94k problems and that achieves the best performance after SFT. extended: 131k samples where we add data sources like cn_k12, and SFT performance is lower.

Note Advanced medical CoT reasoning distils from GPT-4o Train: 25.4K https://github.com/FreedomIntelligence/HuatuoGPT-o1

Note The math subset of OpenThoughts-114k with extra metadata Train: 89120 Of those, 56730/89120 (63%) have correct answers, as checked by Math-Verify

Note High-quality undergraduate to doctoral-le content filtered from 6.69 million web-crawled academic documents, (physics, chemistry, and biology) with solution distilled from o1-mini and QwQ-32B-preview, along with validation flags. Train: 116,756 https://github.com/AQA6666/SCP-116K-open/tree/main

Note Unique mathematics problem-answer pairs from: AIME (American Invitational Mathematics Examination) problems (1984-2023) AMC (American Mathematics Competition) problems (prior to 2023) Omni-MATH dataset Still dataset Train: 40,000

Note English and Chinese multiple-choice mathematical competition from primary,junior high to high school level. Train: 3K Test: 2K

Note 10M Math & Sci related Instruction data from the web (This one is partial data coming mostly from the forums like StackExchange) Train: 2.34M

Note STEM theorem-based reasoning benchmark, covering 350+ theorems spanning across Math, EE&CS, Physics and Finance. Test: 800

Note This dataset contains 30 problems from the 2025 AIME tests, including: AIME I: 15 problems AIME II: 15 problems

Note 1,389 selected questions from MATH (level 3-5) Train: 1389

Note Multimodal reasoning data Generated by GPT4o with reasoning paths and verifiable answers, based on Math360K and Geo170K Train: 8K

Note Collections of open-source datasets of high-quality mathematical problems (with heavy filter) Uniquely verifiable solutions; Open-ended problem formulations; Closed-form solutions Extra 47,000 problems, Big-Math-Reformulated, reformulated open-ended questions from multiple-choice formats. Train: 251K

Note 10k CodeForces problems Train: 10K

Note International Olympiad in Informatics (IOI) 2020-2024 Train: 229 Test: 41

Note fully-synthetic open-source dataset providing verifiable solutions and tests for coding tasks. Train: 444K

Note natural sciences, humanities, social sciences, and general conversations. Train: 323K

Note 360k multimodal problems with diverse domains, including arithmetic, geometry, calculus, science, and more Train: 360K

Note 22mil+ general reasoning synthetic dataset (not verified the reasoning traces and answers for accuracy) Train: 22.2 M

Note general reasoning dataset with ground reference answers and distilled reasoning responses. from pretraining corpora DCLM and FineMath Train: 1.15 M