Finetuned model from scratch in 1 hour flat with synthetic data

Sometimes people don’t believe me when I say how trivial this stuff is. Here’s my proof that it is fast and easy. This ultimately took about 3 hours but most of that was waiting. In terms of coding and testing it was less than an hour.

Code is here: GitHub - daveshap/MovieScriptGenerator: Finetuning project for GPT-3

Commits to see how long it took:
image

10 Likes

I did the finetune with DAVINCI as the base but I probably only needed CURIE, which is 10x cheaper.

All told this cost about $9 or so in DAVINCI calls and then $14.56 for the finetuned model. If I did CURIE instead, it would have been about $12 total.

1 Like

Great video, David. Thank you!

1 Like

I did a quick search for ‘Jenna’ and ‘Beach House’, and it looks like the model might have been obsessing over American Beach House (2015)? American Beach House (2015) - Full Cast & Crew - IMDb

No idea why it’d fixate on that.

1 Like

Very impressive! One can learn a lot just watching you.

1 Like