Guidance on getting started with fine tuned uncensored model

Hi, im completely new with the topic of fine tuning models and such so any guidance would be great. :smiling_face_with_halo:

I want a very specific text generation model that can generate things like dares, truths, challenges, who is questions etc for a drinking game.
So the base model should have alignment removed to be uncensored (nsfw stuff). Saw some models like dolphin and wizardLM. Can i fine tune those? And are those good for what i want to do? I will create my own datasets with like 2000 samples. Will that have enough impact on the model?

And i dont own great hardware to do it locally so i want to run it in the cloud, should i use something like Runpod or Vertex AI? (Or something else?)
I will only use it a few times to generate data and the model needs to be fine tuned, i have no clue how long that takes. Its just a hobby project so i dont wanna spend too much basically, how can I calculate a rough estimation for the costs?

I will start reading more and checking the guides but any help to steer me in the right direction would be really appreciated!

1 Like

If you can find the Transformers format (the format before quantization to GGUF), you can basically train from there again.

There is no definitive method for LLM training (everyone is still searching for it).
So I think it will be a bit of a trial and error process, but if you search, there are probably some reference materials. The training procedure is the same as for other models, so it should be fine.

LLMs

For training

1 Like

Amazing, thanks a lot for the info and all the links! :heart:

1 Like