mkd-chanwoo/normalized-datasets-for-koreanLLM
收藏资源简介:
这是一个用于韩语大语言模型预训练的规范化数据集,属于Keural韩语LLM预训练管道的阶段0.5输出。该数据集将来自38个源数据集的原始文本转换为统一的JSONL模式,未应用任何过滤,是干净、结构化的原始形式。数据集规模约为8.36亿文档,涵盖英语、韩语、代码和科学领域,每个文档都遵循统一的JSONL模式(包括doc_id、source_name、domain、language、text等字段)。规范化过程解决了不同源数据集格式和字段名不一致的问题,将所有文本提取并写入统一模式,并添加元数据(如领域、语言、许可证)。数据集主要用于下游处理阶段(如过滤、去重),是预训练数据准备的基础步骤。
This is a normalized dataset for pre-training Korean Large Language Models (LLMs), which is the stage 0.5 output of the Keural Korean LLM pre-training pipeline. This dataset converts raw texts from 38 source datasets into a unified JSONL schema without applying any filtering, retaining a clean and structured raw form. The dataset contains approximately 836 million documents, covering English, Korean, code, and scientific domains. Each document adheres to the unified JSONL schema, including fields such as doc_id, source_name, domain, language, and text. The normalization process resolves the inconsistencies in formats and field names across different source datasets, extracts all texts and writes them into the unified schema, while adding metadata such as domain, language, and license. The dataset is primarily used for downstream processing stages (e.g., filtering, deduplication) and serves as a foundational step in pre-training data preparation.




