search

karpathy / nanoGPT

The simplest, fastest repository for training/finetuning medium-sized GPTs.
MIT License
37.65k stars 6.01k forks source link

finetuning the model with cherry picked answers #318

Open frroossst opened 1 year ago

frroossst commented 1 year ago

Is it possible to fine tune the model, let's say it produces 10 samples or outputs, can I then rank the outputs or pick the most appropriate one and fine tune it to answer as such? I am working on a code completion bot, and am very impressed at how well the model is doing. Was wanting to deploy it internally and get feedback from developers on the most appropriate autocomplete. Thanks!

VatsaDev commented 1 year ago

You can finetune the model, just give it a giant text file full of code, the process is in the readme. Ranking the outputs is a totally different task though, more like classification. Thats similar to how chatbot prompts work. you might be able to take a classification model and use that with a dataset you make similar to the imdb dataset but instead of movie reviews you have Good/Bad code instead. Using this would give you a workable model, but might not be production ready.

for a full from scratch code complete bot look at --> this

AnioL13 commented 1 year ago

thank you