>>158539
>8gb of vram
That's barely enough to run a heavily quantized (retarded) LLM locally, much less to train one. It's enough to generate very high quality images with NoobAI/Illustrious models and even train loras for them, so maybe you should consider training a lora to generate heyuri-tan pics or something instead of a chatbot. I don't know how much resources are required to train the LLM equivalent of a lora, so that might be an option if you don't mind the chatbot being a little retarded.