4 ms·

> When you say "fine tune" here what are you looking to do? As an example of fine tuning, I might take a pretrained model and then continue training it with a

> When you say "fine tune" here what are you looking to do? As an example of fine tuning, I might take a pretrained model and then continue training it with a custom dataset that is tailored to a specific text generation task (not classification). Here is an example of a custom dataset that I might fine tune on: https://github.com/baobabKoodaa/future/blob/8d2ae91e6a6f00c7c47ac7e8dc24fb9271b13a49/public/index.html#L487 https://github.com/baobabKoodaa/future/blob/8d2ae91e6a6f00c7... I would like the LLM to generate fictional text in the same style as the fine tuned dataset.

simonw 4y ago
I've not yet managed to convince myself if fine tuning LLMs works for that kind of example. Have you tried fine tuning GPT3 via the OpenAI APIs for this? It should only cost a few dollars for that smaller set of examples, and it would at least help demonstrate it it's possible to get the results you want with the current best-in-class language model before you try to run that against a smaller model that you can fit on your own hardware.
- baobabKoodaa 4y ago
  > Have you tried fine tuning GPT3 via the OpenAI APIs for this I haven't. That's not a bad idea. > it would at least help demonstrate it it's possible to get the results you want with the current best-in-class language model before you try to run that against a smaller model that you can fit on your own hardware The dataset you saw was (mostly) generated with ChatGPT and davinci-002, by using prompt engineering instead of fine tuning. So it's definitely possible to produce good results like this (though no judgment here on the question of prompt engineering vs fine tuning).
simonw 4y ago
Wow, this is really fun: I pasted your prompt from https://raw.githubusercontent.com/baobabKoodaa/future/8d2ae91e6a6f00c7c47ac7e8dc24fb9271b13a49/prompt-instructions.txt https://raw.githubusercontent.com/baobabKoodaa/future/8d2ae9... into ChatGPT and now it's acting exactly like I'm stuck in a weird future.
- baobabKoodaa 4y ago
  Glad to hear you like it! I should probably note that the way I used it was by appending a few selected examples to the prompt. Having some examples improves the quality a lot. (I'm on mobile now so can't easily link but you can find the prompt examples starting from server.js)

simonw 4y ago

I've not yet managed to convince myself if fine tuning LLMs works for that kind of example. Have you tried fine tuning GPT3 via the OpenAI APIs for this? It should only cost a few dollars for that smaller set of examples, and it would at least help demonstrate it it's possible to get the results you want with the current best-in-class language model before you try to run that against a smaller model that you can fit on your own hardware.

baobabKoodaa 4y ago
> Have you tried fine tuning GPT3 via the OpenAI APIs for this I haven't. That's not a bad idea. > it would at least help demonstrate it it's possible to get the results you want with the current best-in-class language model before you try to run that against a smaller model that you can fit on your own hardware The dataset you saw was (mostly) generated with ChatGPT and davinci-002, by using prompt engineering instead of fine tuning. So it's definitely possible to produce good results like this (though no judgment here on the question of prompt engineering vs fine tuning).

simonw 4y ago

Wow, this is really fun: I pasted your prompt from https://raw.githubusercontent.com/baobabKoodaa/future/8d2ae91e6a6f00c7c47ac7e8dc24fb9271b13a49/prompt-instructions.txt https://raw.githubusercontent.com/baobabKoodaa/future/8d2ae9... into ChatGPT and now it's acting exactly like I'm stuck in a weird future.

baobabKoodaa 4y ago
Glad to hear you like it! I should probably note that the way I used it was by appending a few selected examples to the prompt. Having some examples improves the quality a lot. (I'm on mobile now so can't easily link but you can find the prompt examples starting from server.js)