JohnTdi/polish-llm-sft-pl

Polish LLM SFT Dataset是一个用于改进和训练波兰语大型语言模型（LLM）的数据集，旨在帮助各种LLM学习和提升波兰语能力。数据集包含38,781个样本，覆盖多种任务类型，包括文本生成、问答、多项选择等。数据格式为JSON，每个样本包含用户和助手角色的对话消息。数据集结构分为多个类别：code（Python编码任务）、empathy（情感支持对话）、handmade（手写混合数据）、history（波兰和世界历史问答）、idioms（波兰成语，包含277个成语的3种表达方式）、math（数学应用题）、mcq_8tags（新闻文章分类）、mcq_belebele（长段落阅读理解多项选择）、mcq_dyk（冷知识多项选择）、mcq_general（多领域多项选择）、mcq_polemo（情感分析多项选择）、mcq_ppc（释义分类多项选择）、mcq_psc（相同/不同文章分类）、mcq_rc（阅读理解多项选择）、rag（基于上下文的问答）、reasoning（链式思维推理）、translation（波兰语-英语翻译对）、trivia（波兰文化和地理常识）、wiki_qa（波兰维基百科问答）。数据集采用质量评分系统：q10表示手写并验证的数据，q9表示合成生成且约95%准确率，q8表示良好数据但有轻微问题。注意事项：合成数据可能包含事实错误，q8文件比q10有更高幻觉风险，部分文件来源混合且无精确追踪。许可证为Apache 2.0，但source_corpus/部分使用CC-BY-SA 4.0（基于波兰维基百科）。

Polish LLM SFT Dataset is a dataset prepared to help various LLMs learn and improve their Polish language capabilities. It contains 38,781 samples covering multiple task categories such as text generation, question-answering, and multiple-choice. Each sample is in JSON format with messages containing user and assistant roles. The dataset is structured into categories: code (Python coding tasks), empathy (emotional support dialogues), handmade (handwritten ground truth mix), history (Polish & world history QA), idioms (Polish idioms with 277 idioms in 3 phrasings), math (math word problems), mcq_8tags (news article category classification), mcq_belebele (long-passage reading comprehension MCQ), mcq_dyk (Did you know trivia MCQ), mcq_general (multi-domain MCQ), mcq_polemo (sentiment analysis MCQ), mcq_ppc (paraphrase classification MCQ), mcq_psc (same/different article classification), mcq_rc (reading comprehension MCQ), rag (context-grounded QA), reasoning (chain-of-thought reasoning), translation (PL↔EN translation pairs), trivia (Polish cultural & geo trivia), and wiki_qa (QA from Polish Wikipedia). A quality scale is used: q10 indicates handwritten and verified data, q9 indicates synthetically generated data with ~95% accuracy, and q8 indicates good data with minor issues. Caveats: Synthetic data may contain factual errors, q8 files carry higher hallucination risk than q10, and some files originate from mixed sources without per-sample provenance tracking. The license is Apache 2.0, except for the source_corpus/ part which is CC-BY-SA 4.0 (based on Polish Wikipedia).