What strategies could we implement to ensure the model follows negative constraints with higher probability?

HAI-DEF
, ,
1

During our experimentation, we observed that MedGemma struggled to adhere to certain instructions. Notably, the model continued to include ‘References’ in its outputs, even though the system prompt explicitly instructed it to omit them.

3

Hi Anika, would you be able to share sample prompts and responses with us? If you don’t feel comfortable sharing on the public forum, you could use our feedback form.

Thank you,

Fereshteh on behalf of the HAI-DEF team

5

Hi Anika, if you are using the 27B model, ensure that you have thinking enabled which will improve the instruction-following capabilities of the model. See our getting started notebook for an example.