dicta-il/MathCOT-oss-vs-DeepSeek

---
configs:
- config_name: default
data_files:
- split: gpt_oss_120b
path: gpt-oss-120b.jsonl.gz
- split: DeepSeek_R1_0528
path: DeepSeek-R1-0528.jsonl.gz
license: cc-by-4.0
task_categories:
- text-generation
language:
- en
---

# Learning to Reason: Training LLMs with GPT-OSS or DeepSeek R1 Reasoning Traces

This dataset is the one used from the paper, available [here](arxiv-link) 📄

This dataset consists of 242k math questions, with the verified generated answer (with reasoning) by both DeepSeek-R1-0528 and gpt-oss-120b.

The original prompts and the DeepSeek-R1-0528 traces were taken from NVIDIA's [Nemotron-Post-Training-Dataset-v1](nvidia/Nemotron-Post-Training-Dataset-v1).

# Citation

If you found this dataset useful, please cite the paper below:

```bibtext
@misc{shmidman2025learningreasontrainingllms,
title={Learning to Reason: Training LLMs with GPT-OSS or DeepSeek R1 Reasoning Traces},
author={Shaltiel Shmidman and Asher Fredman and Oleg Sudakov and Meriem Bendris},
year={2025},
eprint={2511.19333},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2511.19333},
}
```

配置项：
- 配置名称：默认
数据文件：
- 拆分集：gpt_oss_120b
路径：gpt-oss-120b.jsonl.gz
- 拆分集：DeepSeek_R1_0528
路径：DeepSeek-R1-0528.jsonl.gz
许可证：CC BY 4.0
任务类别：
- 文本生成
语言：
- 英语

# 学会推理：使用GPT-OSS或DeepSeek R1推理轨迹训练大语言模型（Large Language Models）

本数据集即该论文所使用的配套数据集，论文原文可通过[此处](arxiv-link)获取 📄

本数据集包含24.2万道数学题目，附带由DeepSeek-R1-0528与gpt-oss-120b生成并经过验证的带完整推理过程的答案。

原始提示词与DeepSeek-R1-0528的推理轨迹均源自NVIDIA发布的[Nemotron-Post-Training-Dataset-v1](nvidia/Nemotron-Post-Training-Dataset-v1)数据集。

## 引用
若本数据集对你的研究有所助益，请引用以下论文：

bibtext
@misc{shmidman2025learningreasontrainingllms,
title={Learning to Reason: Training LLMs with GPT-OSS or DeepSeek R1 Reasoning Traces},
author={Shaltiel Shmidman and Asher Fredman and Oleg Sudakov and Meriem Bendris},
year={2025},
eprint={2511.19333},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2511.19333},
}