Skip to content

Conversation

KarelKenens
Copy link

What does this PR do?

In DataCollatorForVisionLanguageModeling, attempt to dynamically determine additional processor outputs, on top of input_ids and attention_mask. For example in the Gemma 3 model family also token_type_ids are part of the output (and required futher downstream).

Fixes #4189

@qgallouedec , similar to #4190

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

sft_trainer.DataCollatorForVisionLanguageModelling does not account for "non-standard" processor outputs (e.g. Gemma 3)
1 participant