CoolFace
Datasetpublic

amphora/researchmath-glm5_23-distill-12k

ResearchMath GLM-5.2/5.3 distill (12k) 12,487 completed reasoning traces on research-level mathematics questions extracted from arXiv papers. For details, see this paper. This dataset is a further extension of the work. Columns column meaning paper_url source arXiv paper signal_type how the question was identified: open_problem, natural_extension, limitation, conjecture, announced_forthcoming source_quote the passage in the paper the question was… See the full description on the dataset page: https://huggingface.co/datasets/amphora/researchmath-glm5_23-distill-12k.

sourceHugging Facemitupdated 22d agoView on Hugging Face
1likes97downloads
Dataset Card

ResearchMath GLM-5.2/5.3 distill (12k)

12,487 completed reasoning traces on research-level mathematics questions extracted from arXiv papers. For details, see this paper.

This dataset is a further extension of the work.

Columns

columnmeaning
paper_urlsource arXiv paper
signal_typehow the question was identified: open_problem, natural_extension, limitation, conjecture, announced_forthcoming
source_quotethe passage in the paper the question was derived from
questionthe refined, self-contained question
reasoningfull chain-of-thought (the <think> block)
responsethe final answer after </think>
modelwhich model generated the trace

Generation

modelprecisionmax tokensrows
zai-org/GLM-5.3-FlashFP8128k5,101
nvidia/GLM-5.2-NVFP4NVFP464k4,801
LibertAIDAI/GLM-5.3-Flash-NVFP4NVFP4128k2,585

Served with SGLang (tp=8, ep=8) on 8x H200 and 8x H100. GLM-5.3 runs used reasoning_effort=high, temperature=1.0, top_p=0.95; the GLM-5.2 run used temperature=0.6 and predates that control.

All runs share one prompt, instructing the model to solve the exact question as posed, to build an explicit construction, not to answer "this is open", and never to cite journal/volume/year/pages.

Collaborations

I'm interested in creating larger datasets to train open models for research-level math. If you are interested let me know. (guijin.son@snu.ac.kr)

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}
}