Works for T5/BART?

[danyaljj]

Very cool work!

Does this work for T5/BART models as well?

[danyaljj]

Side note: it'd be good to update the transformers dependency to the latest (v4.0.0).

[lvwerra]

You are right, when I have time I'll upgrade it to v4.0.0. I haven't tested it but I suspect if you take a model with a text generation head it should work. Note that you need add a value head to your model architecture (see here).

[danyaljj]

I can try it. Other than running with no errors, what other ways I can test that the code is working fine? Is there a benchmark or a quantitative way of verifying the code?

[lvwerra]

Monitoring the rewards on the IMDb dataset would be a good start. For GPT-2 it takes only 1-2h to train.