TRL - Fine tuned small model (facebook350m) yields many empty inferences

Intermediate
2

The options to be given to the trainer may be quite different from other models?
https://stackoverflow.com/questions/76857722/huggingface-sft-for-completion-only-not-working