Fine-tuning doesn't seem to work for me so far

Fine-tuning on Curie, and the results were not much different than few-shot learning, if not worse.

But I only used ~150 samples, all human-verified.

Should I continue to work on writing more samples? Or could there be something I’m missing?

1 Like

What’s your use case? Generally 150 is on the low side. Find tuned curie is also a lot faster and cheaper.

1 Like

Just a simple paraphrase, as I can’t get a good result with either few-shot or instruct series.

I thought it’s a quite straightforward so I only prepared 150. Will generate more and attempt again.

Any advice meanwhile will be much appreciated, thanks!