bashqort-raw
收藏资源简介:
Bashqort Raw Corpus 是一个用于大语言模型(LLM)持续训练的原始巴什基尔语文本集合。该数据集是项目“为巴什基尔语适配开源大语言模型”的一部分,旨在评估LlamaTurk和波斯语适配提出的适配方法。语料库汇集了多个来源,为语言建模提供多样化的语言基础。数据来源于四个主要渠道:每日社会政治报纸bashgazet.ru(约1,400篇文章)、地区新闻网站neftcity.ru(约548篇文章)、新闻与分析网站bash.news(约68,708篇文章)以及混合来源(公共领域+乱序文本)的bashkir-corpus(约2090万词元)。数据集总规模在1000万到1亿词元之间。数据经过了文档和句子级别的去重、移除少于5个词的句子、清理HTML残留、广告和元数据等预处理步骤。整个语料库用于自监督学习,未划分训练/验证集。每个数据样本是一个JSON对象,包含以下字段:text(巴什基尔语句子或段落)、source(数据来源标识)和is_shuffled(布尔值,指示文本是否经过乱序处理)。数据集也提供包含原始句子的纯文本列。该数据集适用于LLM(如Llama-2、Llama-3)的持续预训练/进一步预训练、因果语言建模(下一个词元预测)以及任何旨在改进巴什基尔语在NLP中表示的研究。
Bashqort Raw Corpus is a raw Bashkir language text collection for continuous training of large language models (LLMs). It is part of the project Adapting Open-Source Large Language Models for Bashkir, aimed at evaluating adaptation methods proposed by LlamaTurk (Toraman, 2024) and Persian adaptation (Mahdizadeh Sani et al., 2024). The corpus aggregates multiple sources to provide a diverse linguistic foundation for language modeling. Data is sourced from four main channels: the daily socio-political newspaper bashgazet.ru (approximately 1,400 articles), the regional news website neftcity.ru (approximately 548 articles), the news and analysis website bash.news (approximately 68,708 articles), and the mixed-source (public domain + shuffled text) bashkir-corpus (approximately 20.9 million tokens). The total dataset size ranges between 10 million and 100 million tokens. Data has undergone preprocessing steps such as document- and sentence-level deduplication, removal of sentences with fewer than 5 words, and cleaning of HTML remnants, advertisements, and metadata. The entire corpus is used for self-supervised learning without train/validation splits. Each data sample is a JSON object containing the following fields: text (Bashkir sentences or paragraphs), source (data source identifier), and is_shuffled (boolean indicating whether the text has been shuffled). The dataset also provides a plain text column with original sentences. It is suitable for continuous pre-training/further pre-training of LLMs (e.g., Llama-2, Llama-3), causal language modeling (next token prediction), and any research aimed at improving the representation of Bashkir in NLP.
数据集名称
Bashqort Raw Corpus
概述
该数据集包含用于大型语言模型(LLM)持续训练的原始巴什基尔语文本。它是“为巴什基尔语言适配开源LLM”项目的一部分,旨在评估LlamaTurk和波斯语适配提出的方法。
数据规模
- 总计:约10M至100M个token(具体统计请参见Hugging Face页面)
数据来源
| 来源 | 类型 | 数量 |
|---|---|---|
| bashgazet.ru | 日常社会政治报纸 | 约1,400篇文章 |
| neftcity.ru | 区域新闻 | 约548篇文章 |
| bash.news | 新闻与分析 | 约68,708篇文章 |
| bashkir-corpus | 混合(公共领域+打乱) | 约2,090万token |
预处理
- 去重(文档级和句子级)
- 移除少于5个词的句子
- 移除HTML伪影、广告和元数据
- 无训练/验证集划分(整个语料库用于自监督学习)
数据格式
每个样本为JSON对象,包含以下字段:
text:巴什基尔语句子或段落source:数据来源(如 bashgazet.ru、neftcity.ru、bash.news、bash-corpus 等)is_shuffled:是否被打乱(true/false)
同时提供纯文本格式的text列,包含原始句子。
预期用途
- LLM的持续预训练/进一步预训练(例如 Llama-2、Llama-3)
- 下一个词预测(因果语言建模)
- 任何旨在提升巴什基尔语在NLP中表现的研究
许可证
MIT License
引用
若使用该数据集,请引用项目仓库: bibtex @misc{khudiakova2025bashqortraw, author = {Khudiakova, Kseniia}, title = {Bashqort Raw Corpus: Bashkir Text Collection for Continual Training}, year = {2025}, howpublished = {Hugging Face Datasets}, url = {https://huggingface.co/datasets/metuKKhud/bashqort-raw} }




