Davinci Fine Tune base model is sweet

She knows her name and answers questions correctly . But her style is more like personal than business oriented. I have not figured out how to control her style. Temp or top doesn’t seem to matter☺️. Give her more tokens and she just uses them all talking jibberish. I really don’t know what else to do with her except keep building the training file larger hoping for the best.

You could try adjusting your system prompt or adding a couple of example responses say the head of the chat to focus the time and behavior of the model.

Beyond that I suspect much more fine-tuning would be required to align the expressed personality of your model.


I tried classification for ada model and it worked fine.
Process followed is like train with a data set, validate with another data set, train again with other data set if we see any failures in validation, and repeat until we achieve required responses. We can use gpt itself to generate data sets.
It took 8 sets for me to get what I expected for ada model.

I am new to this and this is just my experience.

Thank you,

1 Like

Between using fine tuning and providing sampled to the GPT model in the format of the output you want it to be, the model should be able to mimic the format of the output you want it to do so.

Also, ensuring a low value of the temperature will also force the model to align more to your samples and fine tuning.

1 Like

The power of “completion”

1 Like

Thank you all for your suggestions and help. I will keep working on my training file every job seems better than the previous one.