the nice thing too is that because you are freezing almost all the parameters, and generally in lower precision (eg QLoRA loads the full model in 4-bit), it's s

the nice thing too is that because you are freezing almost all the parameters, and generally in lower precision (eg QLoRA loads the full model in 4-bit), it's super low gpu memory usage. a free Colab will suffice for finetuning a 7b definitely, renting a 3090 is less than 50 cents an hour, pretty low barrier to entry to try something!