bashqort-task
收藏资源简介:
Bashqort Topic Classification 是一个用于巴什基尔语(Bashkir)新闻标题的主题分类数据集。该数据集的创建源于目前缺乏公开的巴什基尔语主题分类基准,其主要目的是为适应巴什基尔语的大语言模型(LLMs)提供任务特定的微调与评估数据。数据来源于对 bash.news 网站新闻标题的随机抽样,并由巴什基尔语母语者 Ilyas Khatipov 进行手动标注。为了提升类别平衡性,在数据清洗过程中移除了样本数量少于17个的主题类别。数据集包含总计267个新闻标题样本,均匀分布于10个主题类别,具体分布为:文化(21.7%)、医疗(13.1%)、教育(12.0%)、社会领域(9.4%)、政治(9.4%)、体育(7.9%)、兵役(6.7%)、经济(6.7%)、事件(6.7%)和安全(6.4%)。数据以CSV或JSON格式提供,包含两个核心字段:`title`(巴什基尔语新闻标题,字符串类型)和 `topic`(主题标签,字符串类型,为上述10个类别之一)。数据集未预设固定的训练集和测试集划分,鼓励使用者根据实验需要(例如80/20比例)自行划分以确保可复现性。该数据集适用于以下场景:对大语言模型进行巴什基尔语主题分类任务的任务特定微调;评估模型在巴什基尔语理解上的零样本和少样本性能;作为未来巴什基尔语自然语言处理研究工作的基准数据集。数据集遵循MIT许可证。
Bashqort Topic Classification is a topic classification dataset for Bashkir news headlines. The creation of this dataset stems from the current lack of publicly available benchmarks for Bashkir topic classification, with the main purpose of providing task-specific fine-tuning and evaluation data for large language models (LLMs) adapted to Bashkir. The data is sourced from random sampling of news headlines on the bash.news website and manually annotated by native Bashkir speaker Ilyas Khatipov. To improve category balance, topic categories with fewer than 17 samples were removed during data cleaning. The dataset contains a total of 267 news headline samples, evenly distributed across 10 topic categories, with specific distributions as follows: culture (21.7%), healthcare (13.1%), education (12.0%), social sphere (9.4%), politics (9.4%), sports (7.9%), military service (6.7%), economy (6.7%), events (6.7%), and security (6.4%). The data is provided in CSV or JSON format, with two core fields: `title` (Bashkir news headline, string type) and `topic` (topic label, string type, one of the 10 categories mentioned above). The dataset does not have a predefined split for training and test sets; users are encouraged to split it according to experimental needs (e.g., 80/20 ratio) to ensure reproducibility. This dataset is suitable for the following scenarios: task-specific fine-tuning of large language models for Bashkir topic classification tasks; evaluating zero-shot and few-shot performance of models in Bashkir understanding; serving as a benchmark dataset for future Bashkir natural language processing research. The dataset is licensed under the MIT License.
数据集名称
Bashqort Topic Classification
数据集描述
这是一个面向巴什基尔语新闻标题的主题分类数据集。由于巴什基尔语缺乏公开的主题分类基准,该数据集旨在用于针对该语言的任务特定微调和评估。
数据来源与构建
- 来源:随机采样自 bash.news 的新闻标题。
- 标注:由巴什基尔语母语者 Ilyas Khatipov 进行人工标注。
- 清洗:移除了样本数少于17个的类别,以改善类别平衡。
类别分布(清洗后)
| 主题(Topic) | 数量(Count) | 占比(Proportion) |
|---|---|---|
| 文化(Мәҙәниәт) | 58 | 21.7% |
| 医疗(Һаулыҡ һаҡлау) | 35 | 13.1% |
| 教育(Мәғариф) | 32 | 12.0% |
| 社会领域(Социаль өлкә) | 25 | 9.4% |
| 政治(Сәйәсәт) | 25 | 9.4% |
| 体育(Спорт) | 21 | 7.9% |
| 兵役(Хәрби хеҙмәт) | 18 | 6.7% |
| 经济(Иҡтисад) | 18 | 6.7% |
| 事件(Ваҡиғалар) | 18 | 6.7% |
| 安全(Хәүефһеҙлек) | 17 | 6.4% |
总计:267 条标题,覆盖 10 个类别。
数据格式
数据集以 CSV/JSON 格式提供,包含以下列:
title:巴什基尔语新闻标题(字符串)topic:主题标签(字符串,来自上述10个类别)
示例: json { "title": "Өфөлә мәктәптәрҙә яңы уҡыу йылы башланды", "topic": "Education" }
数据集划分
- 总样本数:267
- 划分方式:无固定的训练/测试集划分,建议用户自行创建(如80/20)以保证可重复性。
数据集统计
- 下载大小:20,514 字节
- 数据集大小:40,302 字节
- 样本规模:100 < n < 1000
预期用途
- 对大型语言模型进行任务特定的微调,用于主题分类
- 巴什基尔语理解的零样本和少样本评估
- 为未来的巴什基尔语自然语言处理工作提供基准
许可协议
MIT License
引用格式
bibtex @misc{khudiakova2025bashqorttask, author = {Khudiakova, Kseniia and Khatipov, Ilyas}, title = {Bashqort Topic Classification: News Headlines with 10 Topics}, year = {2025}, howpublished = {Hugging Face Datasets}, url = {https://huggingface.co/datasets/metuKKhud/bashqort-task} }




