ltg
/

How to fine-tune this model?

#2
by alvations - opened

When trying to finetune the model with accelerate, it's throwing an error:

>>> trainer.train()
...
Traceback (most recent call last):
...
File "../models/modules/transformers_modules/ltg/deberta-xxlarge-fixed/.../modeling_deberta.py", line 1312 in forward
    assert labels is None, "only inference is supported for now"
AssertError: only inference is supported for now

Any plans to let users finetune the model? Or is it just inference only for the near future ~1-2 months?

Language Technology Group (University of Oslo) org

Hi, to be honest, I don't know how to finetune DeBERTa on a causal-LM task and I haven't even tried :) One possibility is to simply finetune it with a MLM objective and then use the generation method proposed in our paper; another possibility is to finetune it on a prefix-LM objective and then use standard autoregressive generation. And there are probably many other possibilities, it's an uncharted territory.

davda54 changed discussion status to closed

Sign up or log in to comment