Various errors in orpo_finetuning_example.ipynb notebook

[fairydreaming]

In orpo_finetuning_example.ipynb the currently the used dataset is "trl-lib/ultrafeedback_binarized" and it's loaded with split="train". Previously it was "mlabonne/orpo-dpo-mix-40k" with split="all". The current dataset causes failure during trainer creation:

Shouldn't it be split=None when loading the dataset?

Also the title in the Exercise section of the notebook is:

Exercise: Aligning SmolLM2 with DPOTrainer

Did you mean ORPOTrainer?

Finally, I noticed that you changed the optimizer to optim="paged_adamw_8bit". This optimizer requires bitsandbytes library, but this library is not listed in requirements.txt.

[fairydreaming]

After fixing these problems and finishing the training run I examined the metrics:

Accuracy is 0.5 the whole time, margins barely budged. Does it indicate failure of the training run (at least regarding the preference alignment part)?

[fairydreaming]

@mmeendez8 Did you manage to fix orpo_finetuning_example.ipynb notebook so that the preference alignment training process actually works based on the reported metrics? I see that you fixed most of issues that I reported here, did you get the same metrics values as me?

[mmeendez8]

No sorry, I just went trough the notebook quickly to check and found out a couple of issues. Need to spend some time checking that