Dev question, does anyone have experience fine-tuning an openai model?

[karmic dagger]

https://platform.openai.com/docs/guides/fine-tuning. Did it have good results? How big was your training data set? How much did it cost? I'm doing a AI VTuber project, but I'm despairge, cuz it's using openai's gpt-3.5-turbo and the customization/personality is just...not there, no matter how I tweak the prompts or model settings. I have everything else working - the VTuber/TTS/STT, etc - but I feel like my project might have come to a grinding halt, cuz I don't know if I have the hardware/money to implement a good customized AI model. Sorry if this is not the right place to ask a question like this but I'm a burned out mentally ill loner dev and I had to get this off my chest. DM's appreciated if you have knowledge on this. :c.

[hollow hull]

Assuming its fine to answer here, I'd recommend using an offline conversational or instruct model like those found on https://huggingface.co/models?pipeline_tag=conversational&sort=trending . You can fine tune it yourself. Idk your hardware, but even with 8gbs of VRAM you can run 6B 4bit models just fine. It takes some work to get started, but even pre-trained models ca have a lot of personality just from the prompt.

[karmic dagger]

Hmm, that seems like a great starting point! Thanks Pladis, I’ll take a look. Do you know if it’s expensive to fine tune your own model (like, electricity costs?) or, if it’s GPU intensive to run it? Like, could you run the model while gaming? (I have a 4070ti on the way.)

[hollow hull]

There are a few methods of fine tuning like Lora and soft prompts. I think yeah it'll cost some money with electricity (whatever wattage your GPU has at near constant usage, id recommend looking at underclocking). I'm not sure how long each method takes.

While running your model, it depends on the game. But yes you can. There are also ways to run the model on the cloud for less than a dollar an hour.