Issue with using model
dishantsingla opened this issue · comments
Hi there! Thank you for reaching out to us.
The above prompt template was used for the versions of Meditron finetuned for specific tasks. The models we uploaded to Huggingface are pretrained Meditron models without any task-specific finetuning.
For prompting the pretrained models, we recommend following the "one-shot" conversation template of FastChat. Here is an example template we used to collect the generations in our qualitative analysis (Appendix E.1):
You can adjust the system message, the one-shot example, and any other decoding parameters to improve the generation.
We hope that this help answer your questions. Looking forward to hearing more feedback!