yyDing1
commented
1 week ago QFT was trained on a very small dataset (15K samples) for only one epoch, so it is unlikely to cause overfitting to the training data. In fact, it does not generate identical samples. In the final 1M samples, less than 0.1% are duplicates.
In the question generation phase, this self-synthesis approach for open-ended generation has been demonstrated to effectively produce large-scale and diverse datasets, as also highlighted in Magpie [1].
作者您好，看完文章我有一个小问题想请教一下。文章中将 Qwen2-Math-7BInstruct 和 DeepSeekMath7B-RL 进行QFT微调激活其问题生成能力。但是，在问题生成的时候只是简单将 temperature 设为 1 ，然后使用一个 “ User ：” 形式的 prompt 来引导生成，感觉这样模型会倾向于生成重复的内容，请问是如何确保生成问题的多样性的。