---
license: cc-by-nc-sa-4.0
language:
- zh
task_categories:
- text-generation
tags:
- math
- reasoning
- DPO
- preference
- translation
- R1
size_categories:
- 1K<n<10K
---
# Dataset Card for HuggingFaceH4-MATH-R1-dpo

<!-- Provide a quick summary of the dataset. -->
本資料集為 [HuggingFaceH4/MATH-500](https://huggingface.co/datasets/HuggingFaceH4/MATH-500) 等 MATH 評測資料延伸的 **R1-style DPO** 偏好集，並翻譯為繁體中文：每筆樣本針對一道數學題，提供「具完整 R1 風格 `<think>` 推理過程的較佳解（chosen）」與「推理品質較差的解（rejected）」。

## Dataset Details

### Dataset Description

<!-- Provide a longer summary of what this dataset is. -->
資料以 MATH 風格題目為基礎（代數、幾何、組合、機率等），題幹翻譯為繁中，並使用 [DeepSeek R1](https://huggingface.co/deepseek-ai/DeepSeek-R1) 等具 R1 風格之模型產出 chosen，配合較弱模型／簡化 prompt 的版本作為 rejected，以建構偏好對。

格式採 LLaMA-Factory 之 `conversations` + `chosen` + `rejected` 風格，gpt 回答前段為 `<think>...</think>` 思考段落，後段為最終解答（多以 `\boxed{...}` 收尾）。

- **Curated by:** [Huang Liang Hsun](https://www.linkedin.com/in/lianghsunhuang/?locale=en_US)
- **Language(s) (NLP):** Traditional Chinese
- **License:** cc-by-nc-sa-4.0

### Dataset Sources

<!-- Provide the basic links for the dataset. -->

- **Repository:** [lianghsun/HuggingFaceH4-MATH-R1-dpo](https://huggingface.co/datasets/lianghsun/HuggingFaceH4-MATH-R1-dpo)
- **Source dataset:** [HuggingFaceH4/MATH-500](https://huggingface.co/datasets/HuggingFaceH4/MATH-500)
- **Paper:** TBA

## Uses

### Direct Use

<!-- This section describes suitable use cases for the dataset. -->
* 對繁中數學 reasoning 模型做 DPO 偏好對齊。
* 與其他 reasoning DPO 集（[Natural-Reasoning-R1-10k-dpo](https://huggingface.co/datasets/lianghsun/Natural-Reasoning-R1-10k-dpo)、[tw-textbook-dpo](https://huggingface.co/datasets/lianghsun/tw-textbook-dpo)）混用以平衡題型分布。

### Out-of-Scope Use

<!-- This section addresses misuse, malicious use, and uses that the dataset will not work well for. -->
* 不適用於完全未經 SFT 的基底模型；DPO 假設模型已具基本解題能力。
* 不適用於 MATH-500 評測；題目部分重疊會造成評測泄漏，請避免在自家評估時直接使用。

## Dataset Structure

<!-- This section provides a description of the dataset fields, and additional information about the dataset structure such as criteria used to create the splits, relationships between data points, etc. -->
```jsonl
{
  "conversations": [{"from": "user", "value": "反射矩陣對於一條通過原點的直線 $\\ell$ 是： $\\begin{pmatrix} ... \\end{pmatrix}$ ..."}],
  "chosen":   {"from": "assistant", "value": "<think>嗯，好的，我現在要解決這個問題...</think>\n\n..."},
  "rejected": {"from": "assistant", "value": "..."}
}
```

## Dataset Creation

### Curation Rationale

<!-- Motivation for the creation of this dataset. -->
英文 MATH/R1 資料對訓練繁中數學 reasoning 模型有用，但語言不一致會造成思考連貫性問題。本資料集將題目與解答翻譯為繁中，讓繁中模型可在自然語言下學習 R1 風格的推理結構。

### Source Data

#### Data Collection and Processing

<!-- This section describes the data collection and processing process such as data selection criteria, filtering and normalization methods, tools and libraries used, etc. -->
1. 取得 HuggingFaceH4 MATH 系列之題目。
2. 將題幹翻譯為繁中，保留 LaTeX 等數學格式。
3. 由 R1 風格模型產出 chosen；由較弱模型產出 rejected。
4. 過濾翻譯失敗或答案明顯不符合的樣本。

#### Who are the source data producers?

<!-- This section describes the people or systems who originally created the data. It should also include self-reported demographic or identity information for the source data creators if this information is available. -->
* 原題目來自 MATH benchmark；HuggingFaceH4 整理並發布。
* 解答內容由 LLM（含 R1 系列）自動生成。

### Annotations

#### Annotation process

<!-- This section describes the annotation process such as annotation tools used in the process, the amount of data annotated, annotation guidelines provided to the annotators, interannotator statistics, annotation validation, etc. -->
無人工標註。

#### Who are the annotators?

<!-- This section describes the people or systems who created the annotations. -->
無人工標註者。

#### Personal and Sensitive Information

<!-- State whether the dataset contains data that might be considered personal, sensitive, or private (e.g., data that reveals addresses, uniquely identifiable names or aliases, racial or ethnic origins, sexual orientations, religious beliefs, political opinions, financial or health data, etc.). If efforts were made to anonymize the data, describe the anonymization process. -->
題目以數學為主，不含 PII。

## Bias, Risks, and Limitations

<!-- This section is meant to convey both technical and sociotechnical limitations. -->
* 翻譯後的數學題可能因語言用詞差異而與原題語意略有偏差。
* chosen 解答仍由 LLM 生成，部分樣本最終答案仍可能錯誤，建議搭配自動評分（boxed 抽取後與標準答案比對）做進一步過濾。
* 與 MATH-500 評測有題目重疊風險，請在評測時務必比對 ID。

### Recommendations

<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
建議在 DPO 前做答案比對，確保 chosen 的最終答案與 ground truth 一致再用於訓練。

## Citation

<!-- If there is a paper or blog post introducing the dataset, the APA and Bibtex information for that should go in this section. -->
```bibtex
@misc{huggingfaceh4_math_r1_dpo,
  title        = {HuggingFaceH4-MATH-R1-dpo: Traditional Chinese R1-style DPO Dataset for MATH},
  author       = {Huang, Liang Hsun},
  year         = {2025},
  howpublished = {\url{https://huggingface.co/datasets/lianghsun/HuggingFaceH4-MATH-R1-dpo}},
  note         = {Translated and re-derived from HuggingFaceH4/MATH-500.}
}
```

## Dataset Card Authors

[Huang Liang Hsun](https://www.linkedin.com/in/lianghsunhuang/?locale=en_US)

## Dataset Card Contact

[Huang Liang Hsun](https://www.linkedin.com/in/lianghsunhuang/?locale=en_US)