{"paper":{"title":"DeepMath-103K: A Large-Scale, Challenging, Decontaminated, and Verifiable Mathematical Dataset for Advancing Reasoning","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"DeepMath-103K supplies 103K hard, clean math problems that let reinforcement learning reach state-of-the-art reasoning performance.","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Dian Yu, Dong Yu, Haitao Mi, Jiahao Xu, Linfeng Song, Qiuzhi Liu, Rui Wang, Tian Liang, Wenxuan Wang, Xingyu Chen, Yue Wang, Zhaopeng Tu, Zhenwen Liang, Zhiwei He, Zhuosheng Zhang","submitted_at":"2025-04-15T17:59:51Z","abstract_excerpt":"Reinforcement learning (RL) with large language models shows promise in complex reasoning. However, its progress is hindered by the lack of large-scale training data that is sufficiently challenging, contamination-free and verifiable. To this end, we introduce DeepMath-103K, a large-scale mathematical dataset designed with high difficulty (primarily levels 5-9), rigorous decontamination against numerous benchmarks, and verifiable answers for rule-based RL reward. It further includes three distinct R1 solutions adaptable for diverse training paradigms such as supervised fine-tuning (SFT). Spann"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"models trained on DeepMath-103K achieve state-of-the-art results on challenging mathematical benchmarks and demonstrate generalization beyond math such as biology, physics and chemistry","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The decontamination process fully removes overlap with numerous benchmarks and the selected problems remain sufficiently challenging and verifiable to produce genuine gains in reasoning capability.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"DeepMath-103K is a new 103K-problem mathematical dataset with high difficulty, rigorous decontamination, and verifiable answers to support RL training of language-model reasoning.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"DeepMath-103K supplies 103K hard, clean math problems that let reinforcement learning reach state-of-the-art reasoning performance.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"dd32e69dd7a3534871410530c6b7658000018536302e28df13b4762017b619b8"},"source":{"id":"2504.11456","kind":"arxiv","version":2},"verdict":{"id":"ee96a8e9-1767-4fb8-93f7-e4cf7534a95e","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-16T10:27:45.145056Z","strongest_claim":"models trained on DeepMath-103K achieve state-of-the-art results on challenging mathematical benchmarks and demonstrate generalization beyond math such as biology, physics and chemistry","one_line_summary":"DeepMath-103K is a new 103K-problem mathematical dataset with high difficulty, rigorous decontamination, and verifiable answers to support RL training of language-model reasoning.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The decontamination process fully removes overlap with numerous benchmarks and the selected problems remain sufficiently challenging and verifiable to produce genuine gains in reasoning capability.","pith_extraction_headline":"DeepMath-103K supplies 103K hard, clean math problems that let reinforcement learning reach state-of-the-art reasoning performance."},"references":{"count":22,"sample":[{"doi":"","year":null,"title":"Marthe Ballon, Brecht Verbeken, Vincent Ginis, and Andres Algaba","work_id":"00edb0c5-3c62-421d-a532-03e2594f1dba","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"SmolLM2: When Smol Goes Big -- Data-Centric Training of a Small Language Model","work_id":"8472f581-14d4-40f8-8189-62ed8b470c4e","ref_index":2,"cited_arxiv_id":"2502.02737","is_internal_anchor":true},{"doi":"10.18653/v1/2023.emnlp-main.468","year":2023,"title":"doi: 10.18653/v1/2023.emnlp-main.468","work_id":"8d54fadf-29e9-49ff-aab2-e9bc4679ef97","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Do NOT Think That Much for 2+3=? On the Overthinking of o1-Like LLMs","work_id":"d79f8b7d-560d-4fbd-bd70-4d084f6d9ad6","ref_index":4,"cited_arxiv_id":"2412.21187","is_internal_anchor":true},{"doi":"","year":null,"title":"Training Verifiers to Solve Math Word Problems","work_id":"acab1aa8-b4d6-40e0-a3ee-25341701dca2","ref_index":5,"cited_arxiv_id":"2110.14168","is_internal_anchor":true}],"resolved_work":22,"snapshot_sha256":"1a1ffd0dff706fbc80115a2d8f45ae0744dca908e7641f8ffa625ad4cf6fe554","internal_anchors":11},"formal_canon":{"evidence_count":1,"snapshot_sha256":"cfecafb5d9f5c9af3dc6964d79e60f52a9401c53c81e8262383061d23716aa6f"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}