jescy525/nous-sft-v1-sources

nous-sft-v1-sources 是一个属于AETHER家族的监督微调（SFT）数据集，属于nous_core组。数据以JSONL ChatML消息格式存储，包含消息（角色为system、user或assistant，内容为文本）、任务类型（如function_calling、code、reasoning_cot等）、来源数据集ID、语言（如英语、法语）和系统来源（如archon_default或overridden_from_source）。应用了任务类型标记和MinHash去重技术，阈值为0.85，旨在提高数据质量。数据集由prepare_sft.py管道于2026年5月25日生成，适用于文本生成和对话任务，规模在10万到100万条数据之间。The nous-sft-v1-sources is a Supervised Fine-tuning (SFT) dataset belonging to the AETHER family and the nous_core group. Stored in JSONL ChatML message format, the dataset contains entries with messages (with roles of system, user, or assistant and text content), task types (e.g., function_calling, code, reasoning_cot, etc.), source dataset IDs, languages (e.g., English, French), and system sources (e.g., archon_default or overridden_from_source). Task type tagging and MinHash deduplication with a threshold of 0.85 have been applied to improve data quality. Generated by the prepare_sft.py pipeline on May 25, 2026, this dataset is suitable for text generation and conversational tasks, with a scale ranging from 100,000 to 1,000,000 data entries.