nbeerbower

/

llama-3-gutenberg-8B

like 7

I think this is one of the best finetunes for writing since a lot of the best writing models have this merged in at some point, and it also improved the model this was trained on in openllm average score if I remember right. The Gemma 9b sppo (iter 3) finetune I tested is surprisingly very good, and performs great in benchmarks (which is why I was skeptical and thought it would be bad or too overfit for real world use). I highly suggest giving dpo training it on the same dataset a try, preferably without any special tokens (the original Gemma 2 tokenizer should be preferred).