We used two models, one trained on files with traceability back to the original source (llama) and one trained on Q&A generated from files (curie). One way to combine these two might be:
collect all relevant files
run a preprocessing step that creates Q&A for each file.
create a new file containing the original data + the Q&A
train the llama model on this
This could give us the "here's where I pulled the answer from" structure of llama plus the nice answers coming from the curie model. Thoughts?
We used two models, one trained on files with traceability back to the original source (llama) and one trained on Q&A generated from files (curie). One way to combine these two might be:
This could give us the "here's where I pulled the answer from" structure of llama plus the nice answers coming from the curie model. Thoughts?