I started with unsloth's tutorials and documentations, but nowadays finetunes are mostly for non-creative writing purposes (business, support chats etc) or simply abliterating them. That said, you're one of the few fine-tuning llms for creative writing purposes. Can you drop a few guides/resources for someone looking to start fine-tuning llms, or generally how you fine-tuned this model (dataset, etc)?
Honestly, I'd love to find resources or guides too. Everyone who does this stuff is being very secretive about it. For one, maybe join Drummer's discord, since that's a pretty active community that will have a lot of details.
As for this particular finetune, it was devastatingly simple: pure completion dataset, i.e. no instructions or anything, just pure text, chopped up into chunks equal your target context size. Train against the base model, then merge your LoRA with the instruction model. Unless you used an insanely high learning rate or something else, this should give you a model that is more "aligned" with your writing inputs, but otherwise preserves the instruction tune's IQ and performance.
For more advanced stuff, it's all very much one gigantic research project. Nobody knows, some people strike gold, and of those that do, few share the specifics that matter, so it's mostly guesswork. Wouldn't recommend unless you have gobs of patience and time and don't mind exorbitant electricity bills.
Much appreciated. Although rather than being secretive in the sense that there are gatekept communities, it feels more like people don't have an incentive to share resources/guides these days. I was able to find some using RAGs and deep research setups, but most of them were in Chinese or some other language. Not really an option to translate highly technical terms using LLMs.
