hotdogs/claude46_filtered
收藏资源简介:
claude46_filtered数据集是从angrygiraffe/claude-opus-4.6-4.7-reasoning-8.7k过滤而来的子集,移除了26%的拒绝回答样本,保留了28,576个干净的推理链样本。这些样本主要来自Claude Opus 4.6和4.7的推理能力蒸馏,涵盖了编码、数学、科学、哲学、创意写作等多个类别。数据集格式为包含角色和内容的对话消息,用于训练基于Qwen3.6-27B变体的无审查推理模型。过滤过程通过识别特定的拒绝模式(如I cannot、as an AI等)来移除不符合要求的样本,确保模型训练时不重新引入拒绝行为。
The claude46_filtered dataset is a filtered subset of angrygiraffe/claude-opus-4.6-4.7-reasoning-8.7k with 9,928 refusals removed, leaving 28,576 clean reasoning examples. These examples are primarily distilled from the reasoning capabilities of Claude Opus 4.6 and 4.7, covering categories such as coding, math, science, philosophy, creative writing, and more. The dataset format consists of messages with roles and content, intended for training uncensored reasoning models based on a variant of Qwen3.6-27B. The filtering process identified and removed refusal patterns (e.g., I cannot, as an AI, etc.) to prevent reintroducing refusal behavior during model training.




