Hiya, bit of a slump and some conflicting information.
We are building a math platform with an LLM-tutor. The tutor should 1) act pedagogically and 2) according to the exercises, i. e. it should rather bring the pupil to find the solution than spoil it outright and should also be able to provide either instructional help (“in this exercise you have to do this by doing that”) or help according to content (“you can add fractions by…”). If possible in the same model, also do function calling, e. g. highlight a specific part of the exercise.
Now, one agency created something similar with RAG for medical documents. Another swears on fine tuning. As a rule of thumb I thought that knowledge is RAG and behaviour is fine tuning.
But, there’s also fine tuning and merging, which allegedly provides both knowledge and behaviour. I think Rombodawg is a proponent: Fine tune base, then merge the adapter, the fine tuned model (?), and the instruct model.
Which is most advisable?