Currently I’m using davinci (latest) to generate long-form completions for an app I have created but it’s eating up funds quite quickly. I’ve been looking into alternative models like curie and babbage but the output is clearly of much lower quality compared to davinci. I’ve also looked into fine-tuning, and I was wondering if I were to fine-tune curie or babbage for my specific use, could I get davinci quality output for my specific need? Obviously, I assume the more training data you feed it, the better, but I was wondering if anyone has any experience with this and can tell me the quality you can achieve with fine-tuned models?
Maybe… but probably not. Too many variables to say for sure, but likely not. Fine-tuning is really hit and miss (depending on a lot of variables…)
I’d recommend a quick search of the forums for similar issues and discussions about fine-tuning.
Hope this helps.
I am fine-tuning curie for my project and the results are great! (chatgpt uses the same number of parameters as curie, so don’t underestimate it) - you just need good training data (quality+quantity)
I am now making tests with babbage as well, but I’m guessing that will be no-go…let’s see