qwen3.6-plus中文思维链数据集

这是一个极小规模的思维链数据集。从魔搭Qwen3的数据集中选择了一些了问题作为输入。使用 Qwen3.6-Plus 模型输出回复。使用 Qwen3.5-Max 模型对思维链部分进行英中翻译，生成对应的中文思维链。测试dpo的效果不是很好，也可能是我设置问题。sft的话100-200条在5e-5的学习率下3轮训练即可很好的生成中文思维链，当然因为数据规模较少，仍会存在英文的输出。

This is an extremely small-scale chain-of-thought (CoT) dataset. Some questions were selected from the ModelScope Qwen3 dataset as inputs. Responses were generated using the Qwen3.6-Plus model. The chain-of-thought segments were translated from English to Chinese by the Qwen3.5-Max model to produce corresponding Chinese chain-of-thought content. The tested DPO (Direct Preference Optimization) performance was unsatisfactory, which may be attributed to improper setup on my side. For SFT (Supervised Fine-Tuning), training with 100–200 samples, a learning rate of 5e-5, and 3 epochs can yield good results in generating Chinese chain-of-thought content. Nevertheless, due to the limited dataset scale, English outputs may still occasionally occur.

搜集汇总

背景与挑战

背景概述

该数据集是一个名为'qwen3.6-plus中文思维链数据集'的中文思维链数据集，由x963082438在ModelScope平台发布，采用Apache License 2.0许可证，更新于2026年4月18日，大小为1.53MB，下载需登录。数据集未提供详细描述，但可通过SDK或GIT命令获取。