jescy525/archon-sft-v1-reasoning
收藏资源简介:
archon-sft-v1-reasoning是一个属于AETHER家族的监督微调(SFT)数据集,专注于推理任务组。该数据集采用JSONL ChatML消息格式,每条记录包含消息序列(角色包括系统、用户或助手,以及内容)、任务类型标签(如函数调用、代码、推理链等)、来源数据集ID、语言(主要为英语)和系统来源字段。应用了MinHash去重技术,阈值设为0.85。数据100%来源于microsoft/orca-math-word-problems-200k数据集,任务类型全部为math_symbolic,涉及数学文字问题推理。尽管元数据标签包含英语和法语,但实际检测显示100%为英语内容,语言字段一致标记为en。该数据集通过prepare_sft.py管道生成,旨在用于数学问题解决的监督微调,代表实际数据而非初始设计计划。
archon-sft-v1-reasoning is an AETHER family supervised fine-tuning (SFT) dataset belonging to the reasoning group. It is formatted as JSONL ChatML messages with task_type tagging and MinHash deduplication applied (threshold 0.85). The schema includes fields for messages (with roles such as system, user, or assistant and content), task_type (e.g., function_calling, code, reasoning_cot), source_ds (indicating the original dataset ID), lang (primarily English), and system_source. The data is 100% sourced from microsoft/orca-math-word-problems-200k, with all task_type entries being math_symbolic, focusing on math word problems reasoning. Despite language tags including both English and French, sampled rows show 100% English content, and the lang field is consistently en. Generated via the prepare_sft.py pipeline, this dataset represents the actual SFT data for mathematical reasoning, distinct from initial planning documents.




