How to limit the loss computation to the answer ?
#44
by
schwarzwalder
- opened
In the idefics2 paper, there is a mention of computing the loss only for the answer part of the VQA task. I could not find such in the fine-tune colab.
Could you please provide a short snippet for that ?
Thanks in advance.
Yesn it's true that it's not present in the google colab.
In our codebase, it is done in a hacky way in the packing, by tokenizing the input, getting the positions between Assistant: and the next <end_of_utterance>, and not computing the loss on those ids.
Thanks!
schwarzwalder
changed discussion status to
closed