Trilogix1/reasoning-distill-opus-4-7-max-sft

该数据集包含7,823个单轮推理对话，这些对话来自Claude Opus 4.7模型，并经过重新格式化以适用于监督微调（SFT）。每个对话都包含一个完整的Qwen风格聊天模板对话，其中包含模型生成的思考过程和最终答案。数据集适用于使用`trl.SFTTrainer`和`train_on_responses_only`进行训练，特别关注模型生成的思考部分。数据集的平均每行约4k tokens，部分长尾推理链可达32k tokens。此外，该数据集已用于训练一个名为`lordx64/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled`的模型。

This dataset contains 7,823 single-turn reasoning conversations from Claude Opus 4.7, reformatted for supervised fine-tuning (SFT). Each row is a complete Qwen-style chat-template conversation, including the models generated thinking process (`<think>...</think>` block) and final answer. The dataset is designed for use with `trl.SFTTrainer` and `train_on_responses_only`, focusing on the models generated reasoning. On average, each row contains ~4k tokens (Qwen3 tokenizer), with some long-tail reasoning chains extending up to 32k tokens. Additionally, this dataset has been used to train a model named `lordx64/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled`.