I am waiting on a final proof of my upcoming book to arrive, and in the meantime I dove back into coding and research. Benevolent By Design will be released next week! In the meantime, here’s a video about my finetuning experiments and synthetic data.
I’ve been able to finetune davinci with data from field service technicians. The model produces content that resembles what the technicians write in the field, so it worked well.
Now: My original expectation was to have a model capable of being an assistant to the techs, but it turns out the way I fine-tuned it generates content in their style but not useful at all.
I will start playing with the ‘Customer-Support’ prompt format proposed by OpenAI in its documentation and see what happens there.
Now to the credits:
I was able to finally train and use a fine-tuned model thanks to David Shapiro, so THANK YOU AGAIN @daveshapautomator !
If anyone has a question about the HOW-TO, I’ll be happy to assist.