Some small requests for learning

[TomLucidor]

It would be great to see this updated to include LLaMA (and possibly be able to hot-swap other open source models)
Side note: wanted to re-tweak this such that it can mash-up Inspirobot quotes with regular quotes to have more sass.

Question: how would instruction prompting work? Currently I can only think of 4 prompts for authorship styles

1. Generate a quote in the writing style of X
2. Generate a quote in the style of X
3. Generate a quote by X
4. Generate a X quote

Side note: what is the performance difference between LoRA and direct finetuning?

Dataset:
inspirobot_quotes.txt
inspirobot_stories.txt

[basedrhys]

Hey Tom,

I think all of those instruction prompts would work well for a model like LLaMA this model was trained back before LLMs really took off (it uses GPT-2) so is fine-tuned with a specific input/output format (_TOPIC_ ... _QUOTE_).

As for the performance difference, I've never tried LoRA so can't speak to it in this use-case. From what I know its comparable, although for something like quote generation (and most open-ended text generation tasks) its non-trivial to create a strict notion of "performance"