JingerAI
JingerAI
@guijuzhejiang The author's suggestion is that both models belong to the same model family, as confirmed in the paper "instructgpt."
@xiangrongzeng Will the project support GLM？ It is a benefit for the Chinese.
Otherwise, is it necessary to load weights from SFT in step 2? In this project, the RW model only starts from the pre-trained model, which is different from instructGPT.
@yaozhewei You are right, many thanks. Additionally, in this project, it appears that the KL penalty is added starting from the last token. Does this approach make sense in light...
> @JingerAI please create another issue so we can separate different problems. @panganqi and others, please try the latest branch to see if the accuracy problem is resolved. Also, please...