It seems that current code only realize the reward loss, how about the pretrain loss mentioned in the paper? Do the reward loss and pretrain loss are trained sequentially?

[lkh329]

No description provided.

[WuJie1010]

Very nice work! I have the same problem. I only confuse that why the grad scale is 1e-3 and only train 100 steps? thanks a lot.

[xujz18]

Because it is simpler to use ReFL alone directly and to achieve decent results, the code provided here is simply ReFL without the inclusion of pre-training data in order to present the core code of the ReFL algorithm in a simple way. If you wish, you can add the pre-training data yourself. The aim of grad scale 1e-3 and train steps 100 is also to simplify.

[lkh329]

Thanks for your replay，very nice work