scholar


---
## Dataset Details

This dataset was created to address the lack of high-quality scientific datasets in French. It is based on Baccalauréat and Classes Préparatoires (CPGE) exam questions and their detailed solutions, covering a wide range of subjects, primarily mathematics, physics and chemistry and computer science.

The dataset includes 30.3K annotated samples designed to support both educational and research applications in French-language NLP.

It was also used to train [**Luth-0.6B-Instruct**](https://huggingface.co/kurakurai/Luth-0.6B-Instruct) and [**Luth-1.7B-Instruct**](https://huggingface.co/kurakurai/Luth-1.7B-Instruct).

Our data scripts are available on [GitHub](https://github.com/kurakurai/Luth).

## Datasat Subject Distribution

The dataset covers a diverse set of subjects, as illustrated by the distribution below:


## Data Collection & Processing Pipeline

The data was primarily sourced from:

- [Prépas.org – Sujet CPGE Archive](https://prepas.org/index.php?module=Sujets)
- [Sujetdebac.fr – French Baccalauréat Exams](https://www.sujetdebac.fr/)

A total of approximately 14,000 PDFs were collected from these sources.

## Dataset Construction Steps

1. **Filtering and Pairing**
-Remove outdated PDFs (before 1990) and low-quality files where content could not be reliably extracted.
-Match each exam question PDF with its corresponding correction PDF, and eliminate duplicates.

2. **Extraction**
For each (question, correction) pair, extract:
- A structured list of questions
- The corresponding list of answers

3. **Contextualization**
Associate each question with its full context, including:
- The subject instructions
- Any relevant preceding questions and answers

4. **Refinement using Gemini 2.5**
Use Gemini 2.5 to:
- Correct LaTeX formatting issues
- Fix structural errors
- Reformat and standardize answer quality

5. **Sanity Checks and Cleaning**
Remove samples with:
- Missing data
- Formatting errors
- Misalignments between questions and answers

## Citation

```bibtex
@misc{luth2025kurakurai,
title = {Luth: Efficient French Specialization for Small Language Models and Cross-Lingual Transfer},
author = {Lasbordes, Maxence and Gad, Sinoué},
year = {2025},
howpublished = {\url{https://arxiv.org/abs/2510.05846}},
note = {arXiv:2510.05846}
}
```


---

## 数据集详情
本数据集旨在弥补法语高质量科学数据集的空白。其数据源自法国中学毕业会考（Baccalauréat）与大学预科班（Classes Préparatoires，简称CPGE）的考题及详细解答，覆盖多门学科，主要涵盖数学、物理、化学与计算机科学。

本数据集包含30.3千个带标注样本，可支持法语自然语言处理（NLP）领域的教育与研究应用。该数据集还被用于训练**Luth-0.6B-Instruct**与**Luth-1.7B-Instruct**两款模型，相关模型链接如下：
- https://huggingface.co/kurakurai/Luth-0.6B-Instruct
- https://huggingface.co/kurakurai/Luth-1.7B-Instruct

本数据集的处理脚本已开源至[GitHub](https://github.com/kurakurai/Luth)。

## 数据集学科分布
本数据集覆盖多元学科，具体分布如下方饼图所示：

## 数据采集与处理流程
本数据集主要源自以下渠道：
- [Prépas.org – CPGE考题档案库](https://prepas.org/index.php?module=Sujets)
- [Sujetdebac.fr – 法国中学毕业会考真题平台](https://www.sujetdebac.fr/)

从上述渠道共采集约14000份PDF文档。

## 数据集构建步骤
1. **筛选与配对**
- 移除1990年之前的老旧PDF，以及无法可靠提取内容的低质量文件
- 将每份考题PDF与对应的解答PDF进行匹配，并移除重复样本
2. **内容提取**
针对每一组（考题，解答）配对，提取以下内容：
- 结构化的考题列表
- 对应的解答列表
3. **上下文补全**
为每份考题补充完整上下文信息，包括：
- 学科考试说明
- 相关的前置考题与解答
4. **基于Gemini 2.5的优化打磨**
使用Gemini 2.5完成以下操作：
- 修正LaTeX格式错误
- 修复结构错误
- 重新格式化内容并统一解答质量标准
5. **合理性校验与数据清洗**
移除存在以下问题的样本：
- 数据缺失
- 格式错误
- 考题与解答不匹配

## 引用格式
bibtex
@misc{luth2025kurakurai,
title = {Luth：面向小语言模型的高效法语专精与跨语言迁移},
author = {Lasbordes, Maxence and Gad, Sinoué},
year = {2025},
howpublished = {url{https://arxiv.org/abs/2510.05846}},
note = {arXiv:2510.05846}
}