⛔ Add EOS token to processed input in SFT

[qgallouedec]

Learn to generate EOS.

[kashif]

for the gemma3 generation issue?

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[qgallouedec]

No, actually there have been recurrent reports that SFT can't learn to generate EOS. I'm pretty sure #2405 re-introduced the bug reported in #1623

[skandermoalla]

@qgallouedec I've faced this multiple times. I think it's just because of the (not so good) practice in the examples everywhere setting the pad token to the eos token. Then the SFT preprocessing masks everything that's a pad token (=eos token), including the real eos token in the chat template.

[skandermoalla]

Personally, I don't think these forced patches are a good design. I understand that you want the Trainers to work out of the box, but users should still make sure they have a chat template that adds an eos properly. In case someone doesn't want an eos they can't do that anymore now.
(Same for the DPOTrainer btw, I think it adds an extra eos token somewhere.)

[HwangYej1]

i got this TypeError: 'Qwen2TokenizerFast' object is not subscriptabl, after change this code