RL-ckpt is not post-trained from Imitation-ckpt

[AlanLiangC]

When I load the the RL-ckpt and Imitation-ckpt, I found parameters of modules which should be frozon in post-training stage are not same as those in imitation stage. This makes me confused of the advantages of your RL mathod.

[zc-zhao]

Thanks for pointing this out. The previously uploaded IL checkpoint was from an early experiment, which caused the mismatch. I have now updated it with the correct checkpoint.

I rechecked the results and confirmed that they are consistent with those reported in the paper. The RL stage still brings the expected performance improvement.