Will scaling the training set influence the conclusions? And why this work "support" the idea of "weak to strong alignment"?

[yucc-leon]

Well done work. My colleagues and I have found similar results to yours in training a domain-specific LLM. But we cannot do such clear and clean experiments like yours.

But after reading it came out to me that there were other works suggesting millions of IFT data can still be used to train domain LLMs, such as Huatuo-I/II (https://arxiv.org/abs/2304.06975, https://arxiv.org/pdf/2311.09774.pdf) and Adapting Large Language Models via Reading Comprehension from MSR. So why is it so different or what are the borders/pre-conditions of two types of works? There may be more work to be done.

And would you explain why this paper supports the "weak to strong alignment" especially if the weak model cannot identify the knowledge it learns badly?