Recommendations on underlying model fine-tuning

[codybum]

This work is very interesting and potentially useful in many domains.

Do you have any recommendations on how we might fine-tune models in specific domains to better support structured extraction? Specifically, we would be interested extracting structured data from medical reports where things such as a fixed set of conditions, location site, and other semantic labels are specific to the input text, but patterns could be learned from fine-tuned training. While we could fine-tune the model with a json response to report input, it is not clear that this would be the best approach.

Thanks for this work and any future response.

[tranlm]

I would also benefit from being able to fine tune models while enforcing a valid json output. Wondering if the developers might look into this.