kilicai/turkish-sft-from-scratch-120k

Turkish SFT From Scratch 120K是一个从零开始生成的土耳其语监督微调（SFT）数据集，包含12个类别，每个类别10,000个样本，总计120,000个样本。类别包括：指令跟随、问答、摘要、思维链、多轮对话、重写、文本分类、错误纠正、正式写作、翻译、代码解释和创意写作。数据集经过验证，包括规范的消息格式（system/user/assistant）、重复项哈希检查、NSFW/有毒词汇过滤，思维链数学答案使用#### answer进行断言验证，多轮示例至少包含3个用户和3个助手轮次。审计数据显示无不良生成、重复或安全问题。

Turkish SFT From Scratch 120K is a Turkish SFT dataset generated from scratch with category control, containing 12 categories x 10,000 examples = 120,000 examples. Categories include: instruction-following, qa, summarization, cot, multi-turn-dialogue, rewriting, text-classification, error-correction, formal-writing, translation, code-explanation, and creative-writing. Validations include canonical messages format (system/user/assistant), exact duplicate hash check, NSFW/toxic word filter, CoT math answers with `#### answer` assertion verification, and multi-turn examples with at least 3 user + 3 assistant turns. Audit shows no bad generated data, duplicates, or safety hits.