Fine-Tuning Gemma 270M (KeyError: input_ids)

[cemalgnlts]

Hi,

I want to fine-tune this dataset using the Gemma 270m notebook: https://huggingface.co/datasets/neurae/dnd_style_intents

However, I'm getting the following error: In [11] KeyError: 'input_ids'

Here's my notebook: https://gist.github.com/cemalgnlts/1701373bab2f4e2365dd628b9d038e31

I'm not sure what I should do — I didn't have any issues with the chess dataset, but when I added a different dataset, this error occurred.

Thanks.

[Parveshiiii]

Actually, the train_on_responses_only function expects pre-tokenized text. After reviewing your notebook, I think you applied the chat template and passed raw text directly to train_on_responses_only, but it requires tokenized input — that’s why it’s throwing the 'input_ids' error.

[cemalgnlts]

This was my first attempt, I hadn't paid attention. Thank you.