Is there a boost in performance for full fine-tuning versus LoRA?

gururise / AlpacaDataCleaned

Alpaca dataset from Stanford, cleaned and curated

Apache License 2.0

1.46k stars 146 forks source link

Is there a boost in performance for full fine-tuning versus LoRA? #56

Closed timothylimyl closed 12 months ago

timothylimyl commented 1 year ago

It seems that the evaluation comparison made was all under the training scheme of LoRA. Any ideas on full fine-tuning versus the LoRA approach?

gururise commented 1 year ago

No idea how well the full fine-tuned models work. They could perform better. You can test on academic benchmarks using lm-eval-harness to get some idea; however, what we really need is a way to test conversational ability in a chatbot.

Wheeledparasite commented 1 year ago

Training on the latest version of this dataset results in a very capable alpaca model. I'm seeing much better results than with the original alpaca dataset.