senaro/opus-4.6-reasoning-sft-12k

Opus 4.6 Reasoning SFT 12k是一个统一的、经过预清理的推理数据集，由4个Claude Opus 4.6蒸馏来源构建而成，专为监督微调设计。数据集修复了源数据中的不同模式、空值和非标准键存储的推理问题，将所有模式统一为标准`messages`格式，并合并了推理跟踪到助手内容中（使用`<think>...</think>`标签）。数据集包含12,929个样本，其中97.5%包含推理跟踪，格式为`messages`（`{role, content}`字典列表），角色包括`user`和`assistant`。内容涵盖数学、逻辑、编程、科学、废话检测和一般知识等领域。

Opus 4.6 Reasoning SFT 12k is a unified, pre-cleaned reasoning dataset built from 4 Claude Opus 4.6 distillation sources, ready for supervised fine-tuning. The dataset fixes issues in source data such as different schemas, null values, and reasoning stored in non-standard keys by unifying all schemas to the standard `messages` format and merging reasoning traces into assistant content using `<think>...</think>` tags. It contains 12,929 samples, with 97.5% including reasoning traces, formatted as `messages` (list of `{role, content}` dicts) with roles `user` and `assistant`. Content spans mathematics, logic, programming, science, bullshit detection, and general knowledge.