Goader/kobza

Kobza是一个目前为止最大的公开可用乌克兰语语料库，包含近600亿个token，跨越9700万份文档。它旨在支持乌克兰语大型语言模型（LLM）的预训练和微调，以及乌克兰语在多语种设置中的低代表性。该语料库从广泛的网络来源中汇集高质量乌克兰语文本，并应用严格的去重步骤以确保语言建模任务的高效性。每个文档都包含源、子源、时间戳和URL等元数据，以便进行灵活过滤。

Kobza is the largest publicly available Ukrainian corpus to date, comprising nearly 60 billion tokens across 97 million documents. It is designed to support pretraining and fine-tuning of large language models (LLMs) in Ukrainian, as well as multilingual settings where Ukrainian is underrepresented. The corpus aggregates high-quality Ukrainian text from a wide range of web sources and applies rigorous deduplication steps to ensure high utility for language modeling tasks.