h2o-llmstudio VS ue5-llama-lora

Vector databases are probably a good place to start, though you've already tried LlamaIndex. You might want to try https://github.com/h2oai/h2o-llmstudio and https://github.com/h2oai/h2ogpt.

Training framework: https://github.com/h2oai/h2o-llmstudio

Check out LLM Studio for fine tuning LLMs. Open source: https://github.com/h2oai/h2o-llmstudio

https://github.com/bublint/ue5-llama-lora https://www.reddit.com/r/LocalLLaMA/comments/157vzq6/unleashing_the_power_of_language_learning_models/

You can follow this guy's steps. https://github.com/bublint/ue5-llama-lora He basic just throw a txt file let AI read it. Pygmalion 7B or 13B are both llama base, so you can just follow his step to fine-tuning Pygmalion you want. You will need to prepare the character's script,like chat,reaction,background to a file,then just hope AI will learn it correctly,it might need to fine-tuning multiple times if your feed is not enough or AI learn it wrong somehow.

ue5 lora might be a good starting point. It doesn't use any advanced features at all:

I got the idea from a reddit user that started a project that uses LLMs (forgot his name but here's a link to his repo, and I based my implementation from this TDS article

Someone did a LoRA finetuning example using the UE5 documentation, which I replicated to make sure, and you do end up producing word patterns from the document, but it doesn't get incorporated as concepts very well

GitHub - bublint/ue5-llama-lora: A proof-of-concept project that showcases the potential for using small, locally trainable LLMs to create next-generation documentation tools.

This would be an interesting experiment. People are already doing similar stuff, such as expanding the model's knowledge domain into something specific. Here's an example of how someone created a LoRA for UE5 documentation: https://github.com/bublint/ue5-llama-lora

I basically copied the steps from this Unreal Engine 5 LoRA repo and adjusted as needed.