Here’s a step-by-step on doing a LoRA fine tune on Flan-UL2 (20B) in 13h for about $15 (40GB A100):
- https://www.reddit.com/r/MachineLearning/comments/12tg2u8/p_finetuning_a_commercially_viable_open_source/
- https://medium.com/@krohling/finetuning-a-commercially-viable-open-source-llm-flan-ul2-3b84e568c458
Training a LLaMA 7B in 8h on 24GB 3090Ti for UE5 docs: https://github.com/bublint/ue5-llama-lora