ResearchMath-Filtered

# ResearchMath-Filtered ResearchMath-Filtered is a quality-filtered collection of **129,927 long-form reasoning traces and solutions for research-level mathematical problems**, released alongside [ResearchMath-14k](https://huggingface.co/datasets/amphora/ResearchMath-14k) as part of the same paper. It is a cleaned subset of [ResearchMath-Reasoning-194K](https://huggingface.co/datasets/amphora/ResearchMath-Reasoning-194K): each record holds a self-contained problem statement, a long chain-of-thought reasoning trace, and a final response, with low-quality and non-solving generations removed. Paper: [ResearchMath-14K: Scaling Research-Level Mathematics via Agents](https://arxiv.org/abs/2605.28003) ## Load ```python from datasets import load_dataset ds = load_dataset("amphora/ResearchMath-Filtered", split="train") print(ds[0]) ``` ## Composition | Teacher model (`model`) | Rows | |---|---:| | `openai/gpt-oss-120b` | 44,171 | | `Qwen/Qwen3-30B-A3B` | 85,756 | | **Total** | **129,927** | ## Schema | Column | Type | Description | |---|---|---| | `problem` | string | Self-contained research-level math problem statement | | `reasoning` | string | Long-form chain-of-thought reasoning trace | | `response` | string | Final answer / solution write-up | | `model` | string | Teacher model that generated the reasoning and response | ## Filtering Starting from the 193,938 records in ResearchMath-Reasoning-194K, the following were removed: 1. **Very short responses** (< 200 characters). 2. **Non-solving / punt responses** — generations that decline to solve, e.g. statements that the problem is an open problem, "not known", "unsolved", "out of reach", "we do not know", etc. 3. **Hallucinated references** — responses containing external links (URLs, arXiv, DOI) or a References/Bibliography section. 4. **Citation-only answers** — responses that merely attribute the answer to a named existing result (e.g. "this is exactly the X theorem") instead of deriving it. 5. **Truncated responses** — generations cut off mid-sentence (no terminal punctuation). 6. **Proof sketches** — responses offering only a sketch rather than a worked solution. 7. **Source/paper leakage** — responses referring to "the paper", "the problem set", "the prompt", etc. 8. **Knowledge-cutoff disclaimers and TODO/placeholder text.** 9. **Degenerate generations** — repetition loops and responses with unbalanced LaTeX environments. ## Citation If you use this dataset, please cite the paper: ``` @article{son2026researchmath, title={ResearchMath-14K: Scaling Research-Level Mathematics via Agents}, author={Son, Guijin and Yi, Seungyeop and Gwak, Minju and Ko, Hyunwoo and Jang, Wongi and Yu, Youngjae}, journal={arXiv preprint arXiv:2605.28003}, year={2026} } ```