"conf", "logits", and "probs" are all NaN

[chrisole]

Hi! Thanks for your work.
I try to reproduce 'run_gpt.py' on several datasets.
However, the 'eval_res.json' shows: {"index": 0, "true": 0, "pred": 0, "conf": NaN, "logits": [NaN, NaN], "probs": [NaN, NaN]} from the first time save the checkpointss. I try different learning rates but the results are the same.
Foundation model is llama2-7b-chat.
Do you know the reason? Thanks!

[adamxyang]

Hi, thanks for your interests! It looks like a NaN error occurs during training, could you check if the model's outputs and loss go NaN during training? I have previously experienced similar issues when training reward models with PEFT, there seems to be some issue with PEFT version and weights' precision that can cause NaN during training, you could also try adding torch_dtype=torch.bfloat16 when loading the model.

[chrisole]

Many thanks for your reply!
By the way, may I ask if this method also adaptive to auto-regressive languge generation? I noticed that these reasoning datasets are all like classification. I plan to extent your approach to longer sequence generation (e.g., Q&A task), is it feasible？I would appreciate if you give some advice : )

[adamxyang]

The current linearised Laplace framework only looks at uncertainty on a single prediction token, this may not translate easily to sentence-level uncertainty, and computing Jacobians could be quite costly. Currently, we're only working on extending the single token uncertainty to other settings, e.g. reward modelling. Also we have released a new more efficient and self-contained library here, that we plan to maintain in the future.

[chrisole]

Thanks a lot!