mkd-chanwoo/keural-rag-ko

keural-synthetic-SFT是一个专门用于检索增强生成（RAG）的韩语监督微调（SFT）合成数据集。它基于1,545,735个韩语网络文档（种子），通过google/gemma-4-26B-A4B-it模型生成了1,402,611个问答对，涵盖五种任务类型：general_qa（通用问答）、reasoning（推理）、structured_output（结构化输出）、critique（批判性分析）和summary（摘要）。数据集设计用于文档基础的问答模型训练或RAG微调，其答案基于提供的上下文（种子文档）生成，因此可能包含如根据文档等表述。数据集覆盖广泛领域，基于3,436个唯一关键词，并经过质量过滤（接受率86.9%）和重复数据删除（删除率16.63%）。数据以JSON格式存储，包含ID、种子ID、关键词、任务类型、问题、答案等字段。许可证为CC BY 4.0。

keural-synthetic-SFT is a Korean synthetic dataset for supervised fine-tuning (SFT) specialized in Retrieval-Augmented Generation (RAG). It is based on 1,545,735 Korean web documents (seeds) and generates 1,402,611 question-answer pairs across five task types: general_qa, reasoning, structured_output, critique, and summary, using the google/gemma-4-26B-A4B-it model. The dataset is designed for document-based QA model training or RAG fine-tuning, where answers are generated based on provided context (seed documents) and may include phrases like according to the document. It covers a wide range of domains based on 3,436 unique keywords and has undergone quality filtering (86.9% acceptance rate) and deduplication (16.63% removal rate). The data is stored in JSON format with fields such as ID, seed ID, keyword, task type, question, and answer. The license is CC BY 4.0.