4 ms·
Anecdotally, I finetuned Mistral 7B for a specific (and slightly unusual) natural language processing task just a few days ago. GPT-4 can do the task, but it ne
by thorum 3y ago
Anecdotally, I finetuned Mistral 7B for a specific (and slightly unusual) natural language processing task just a few days ago. GPT-4 can do the task, but it needs a long complex prompt and only gets it right about 80-90% of the time - the finetuned model performs significantly better with fewer tokens. (In fact it does so well that I suspect I could get good results with an even smaller model.)
- oceanplexian 3y agoI have a fine tuned version of Mistral doing a really simple task and spitting out some JSON. I'm getting equivalent performance to GPT-4 on that specialized task. It's lower latency, it's outputting more tokens/sec., more reliable, private, and completely free. I don't think we will have an Open Source GPT4 for a long time so this is sorta clickbait, but for the small, specialized tasks, tuned on high quality data, we are already in the "Linux" era of OSS models. They can do real, practical work.
- YetAnotherNick 3y ago> completely free Not according to my calculation. For low request rate it is likely more expensive than GPT4.
- tomrod 3y agoBeen my thought for awhile now. Can you recommend where I can learn more about hardware requirements for running Mistral/Mixtral?
- cced 3y agoHow are you guys fine tuning?
- skeletonjelly 3y agoCan you please point me in the direction of the guide you used for fine tuning? Did you use QLoRA?