cc @edbeeching @lewtun @cmpatino

This is a good point, and was a conversation we had while designing the chat template.

Leaving the tool call as user was a design choice we made during the SFT phase of training to mask the loss for the tool role. Our SFT pipeline already masked the loss for the user role, so mapping the tool call to the user role enables masking the loss through the chat template.

This mapping had a limited impact on BFCL scores, so we decided to continue with that approach.