mkd-chanwoo/keural-QA-en

keural-synthetic-SFT-en-general 是一个从英文维基百科合成的通用监督微调（SFT）数据集，通过两阶段管道生成。该数据集包含约160万条记录，涵盖三类任务：通用问答（general_qa，约33.5%）、推理（reasoning，约33.5%）和结构化输出（structured_output，约33.0%）。生成过程中，模型以维基百科文章为基础锚点，但问题和答案均不引用源文档（如“根据文档”等短语），旨在减少幻觉并模拟自然问答，适用于通用指令调优而非检索增强生成（RAG）微调。数据集使用google/gemma-4-26B-A4B-it模型生成，经过基于规则的质量过滤，记录格式为JSONL，包含问题、答案和原始维基百科文本等字段。

keural-synthetic-SFT-en-general is a synthetic general-purpose supervised fine-tuning (SFT) dataset generated from English Wikipedia using a two-stage pipeline. The dataset contains approximately 1.6 million records across three task types: general_qa (~33.5%), reasoning (~33.5%), and structured_output (~33.0%). During generation, the model uses Wikipedia articles as a grounding anchor, but both questions and answers contain no references to source documents (e.g., no phrases like based on the text), aiming to reduce hallucination and mimic natural Q&A, making it suitable for general instruction tuning rather than retrieval-augmented generation (RAG) fine-tuning. The dataset is generated using the google/gemma-4-26B-A4B-it model, filtered through rule-based quality checks, and formatted in JSONL with fields such as question, answer, and original Wikipedia text.