the costs of fine-tune

[TheCutestCat]

Your idea is really interesting, but I'm more worried about the fine-tune program spending too much money.

1. Can you share the approximate cost of your previous work?
2. As well as GPT-4 has been released and it shows very strong thinking power, perhaps GPT-4 can further improve the prediction performance of the model. Do you have a plan to use the GPT-4 ?

[asoyemi1]

To piggyback on the above question, is there a way to save and load a fined-tuned for later inference? This could save some cost on having to retrain the model from scratch.

[kjappelbaum]

Sorry for not following up on this issue.

1. It depends. A large chunk was for the davinci runs that we did not even include in the main text, and there are things we have done very often. We report around 4k for the full work.
2. GPT-4 can currently not be fine-tuned. But we have seen some promising results with few-shot inference (e.g. check my Twitter)

Yes, you can load a fine-tuned model for later inference, you can save the model name from the output the Tuner gives you.

[TheCutestCat]

@kjappelbaum Thanks for your response. Large language model can be so powerful. Perhaps we can develop a open-source model that based on LLaMA(Meta's "Chatgpt") and all researchers can participate. This could be very interesting🤣

[kjappelbaum]

I'll soon upload a revised model of our paper; we also have some results on consumer hardware in there :)

[bmiles]

From my experience I did a finetune with 4000 rows of data (943,748 trained tokens) and the training cost with openai was $0.38