Logits Shift Problem

[guoyd1999]

Do trado models shift their logits during the SFT/RL process? I checked their SFT forward function, and it looks the same as the SDAR model’s. So they do not shift their logits to predict the next token, right? Also, in the sampling process, it does not shift its prediction. Any information that I missed?

[yinjjiew]

TraDo model is the same as SDAR because we only conducted RL from SDAR.

And yes, the logits are not shifted in the training code, because SDAR does not too. I haven't checked the modeling code of SDAR in detail. But it must have shifted the logits in their modeling part, I think.