Add ds model card, rebased #2101
Conversation
src/axolotl/core/trainer_builder.py
Outdated
| @@ -410,10 +426,12 @@ def __init__( | |||
| *_args, | |||
| bench_data_collator=None, | |||
| eval_data_collator=None, | |||
| dataset_tag_names=None, | |||
There was a problem hiding this comment.
dataset_tag_names isn't used for the base Trainer is it? Only for DPOTrainer?
There was a problem hiding this comment.
it is used for both AxolotlTrainer and AxolotlDPOTrainer within def push_to_hub via _sanitize_kwargs_for_ds_tagging
There was a problem hiding this comment.
I'm not fond of introducing arguments that don't match the original Trainer signature for this case. The call to _sanitize_kwargs_for_ds_tagging should handle the case where the attribute isn't available.
There was a problem hiding this comment.
Thanks for the advice. While trl trainer has an override for def create_model_card over the base transformers trainer, it uses the same def push_to_hub inherited from the base transformers trainer. I changed dataset_tag_names to dataset_tags to match def push_to_hub
Description
Continuation of PR#1015,#2094 with rebasing
Motivation and Context
How has this been tested?
Screenshots (if appropriate)
Types of changes
Social Handles (Optional)