FEAT: add tagging support to axolotl for DPOTrainer

[filippo82]

Description

Add axolotl tag to HuggingFace model cards created when fine tuning with RL (using the DPOTrainer).

Motivation and Context

This PR is required to extend #1004 to include the DPOTrainer. The relevant issue is #979

How has this been tested?

I have tested this by fine tuning a model and pushing it to the HuggingFace hub.

Types of changes

• create a new AxolotlDPOTrainer class which subclasses the trl DPOTrainer class
• move _sanitize_kwargs_for_tagging outside the AxolotTrainer class

Social Handles (Optional)

X/Twitter: @olgias

[filippo82]

@younesbelkada @winglian This is a follow up for #1004.

I am not entirely convinced by my solution though:

• probably _sanitize_kwargs_for_tagging can be move to utility module
• there is currently code duplication (the two def push_to_hub(self, *args, **kwargs) -> str:) inside AxolotlTrainer and AxolotlDPOTrainer)

[younesbelkada]

Thanks ! Note you can also call model.add_model_tags() now with the newest transformers release but this solution looks great

https://huggingface.co/docs/transformers/v4.37.1/en/main_classes/model#transformers.PreTrainedModel.add_model_tags

[filippo82]

Thank you @younesbelkada 🙏🏻

I rebased the branch and force-pushed a last change to fix a linting issue which was causing a test to fail.

[winglian]

@filippo82 I've fixed the issue and rebased again for you