mystic/firefly-openmath

firefly-openmath数据集是基于nvidia/OpenMathReasoning（具体是cot子集）为Firefly SFT准备的子集。该数据集经过严格过滤，只保留带有提取答案的思维链行，要求包含oxed{...}格式，并拒绝包含长重复字符垃圾的行。数据使用Qwen/Qwen3-4B-Thinking-2507聊天模板渲染，丢弃渲染令牌超过16,384的行，最终包含100,000个训练示例。数据模式包括原始问题陈述、原始接受的助手解决方案、用于SFT的渲染聊天记录以及在指定模型下的渲染聊天令牌长度。firefly-openmath is a prepared subset of nvidia/OpenMathReasoning (cot) for Firefly SFT. The dataset is filtered using Fireflys canonical OpenMath quality gate, keeping only chain-of-thought rows with extracted answers that require oxed{...} format and rejecting pathological repeated-character garbage. Messages are rendered with the Qwen/Qwen3-4B-Thinking-2507 chat template, and rows longer than 16,384 rendered tokens are dropped, resulting in 100,000 training examples. The schema includes original problem statement, original accepted assistant solution, rendered chat transcript for SFT, and token length under the specified model.