My first run, 8192 ctx qlora, trained on AEZAKMI-3_6 dataset. Base seems to not be too slopped but finetune is not great - lots of slopped GPTisms, "It's important to remember" etc. It does seem uncensored though, so if you're not fine with Llama-3-8B-Instruct, this might be an option until more better finetunes come out. ChatML prompt format.
