jescy525/aether-sft-v2-reasoning

aether-sft-v2-reasoning是AETHER家族的监督微调数据集，专注于推理任务。数据格式为JSONL ChatML消息，包含任务类型标签（如函数调用、代码、推理链式思维等），并应用了MinHash去重技术（相似度阈值0.85）。每条记录包括消息序列（角色为系统、用户或助手，内容为文本）、任务类型、来源数据集ID、语言（如英语、法语）和系统来源（如默认archon或从源覆盖）。数据集由prepare_sft.py管道于2026-05-25生成，支持英语和法语，规模在10万到100万样本之间，适用于文本生成和对话任务。

aether-sft-v2-reasoning is an AETHER family supervised fine-tuning dataset for the reasoning group. The format is JSONL ChatML messages with task_type tagging and MinHash deduplication applied (threshold 0.85). The schema includes messages (with roles system, user, or assistant and content text), task_type (e.g., function_calling, code, reasoning_cot), source_ds (HuggingFace dataset ID), lang (e.g., en, fr), and system_source (e.g., archon_default, overridden_from_source). Generated by the prepare_sft.py pipeline on 2026-05-25, it supports English and French, with a size between 100K and 1M samples, and is intended for text-generation and conversational tasks.