maithili-corpus
收藏资源简介:
MaithiliCorpus是一个用于迈蒂利语(Maithili)的大规模文本语料库,迈蒂利语是一种使用天城文字母的低资源印度语言,约有3400万使用者。该语料库是当前最大的开放迈蒂利语文本资源,包含经过语言识别(LID)过滤和去重后的39.8百万单词,对应于47,882个文档。数据构成存在领域不平衡,其中文学作品占比55.9%,儿童文学仅占0.7%。该语料库主要用于训练首个迈蒂利语预训练语言模型MaithiliBERT(基于RoBERTa-base架构),支持掩码语言建模(MLM)等自然语言处理任务。需要注意的是,由于数据来源的领域限制,模型可能无法很好地泛化到方言或不同的语域变体,且BPE分词器在处理罕见迈蒂利语单词时可能效率不高。
MaithiliCorpus is a large-scale text corpus for Maithili, a low-resource Indic language written in Devanagari script with approximately 34 million speakers. It is currently the largest open Maithili text resource, containing 39.8 million words filtered and deduplicated via language identification (LID), corresponding to 47,882 documents. The corpus exhibits domain imbalance: literary works account for 55.9% of the total data, while children's literature only makes up 0.7%. It is primarily utilized to train the first Maithili pre-trained language model MaithiliBERT, which is built on the RoBERTa-base architecture, and supports natural language processing tasks such as Masked Language Modeling (MLM). It should be noted that due to domain limitations of the data sources, the model may fail to generalize well to dialects or various register variants, and the BPE tokenizer may exhibit low efficiency when handling rare Maithili words.
数据集概述
数据集名称:MaithiliCorpus
发布机构:AiventraLab
数据集地址:https://huggingface.co/datasets/AiventraLab/maithili-corpus
语言:迈蒂利语(mai)
许可证:CC-BY-4.0
数据集规模与构成
- 总词数:39.8M 单词
- 文档数量:47,882 篇
- 数据版本:gold_deduped(经过语言识别过滤和去重处理)
- 语域分布:以文学类文本为主(55.9%),儿童文学仅占 0.7%,存在语域不均衡问题
数据集用途
该数据集被用于预训练 MaithiliBERT,一个基于 RoBERTa-base 架构的迈蒂利语首个预训练语言模型,可用于掩码填充(fill-mask)等任务。
模型属性:
- 参数量:124M
- 词汇表:BPE 分词,50K tokens
- 上下文长度:512 tokens
- 训练轮次:10 epochs
- 有效批次大小:32
已知限制
- 预训练数据语域不均衡,可能影响模型对非文学类文本的泛化能力
- 对迈蒂利语方言或语体变化的适应能力有限
- BPE 分词器在处理罕见迈蒂利语词汇时可能效率不高
相关资源
- 模型仓库:AiventraLab/maithili-roberta
- 数据集详细文档:CORPUS_DOCUMENTATION.md(位于数据集仓库内)
- 计算资源致谢:AnK Komputing(https://ank-world.com)提供了训练所需的算力与基础设施




