license: cc-by-2.0
Finetune of miqu-70b-sf dequant of miqudev's leak of Mistral-70B (allegedly an early mistral medium). My diffs are available under CC-0 (That is the Senku-70B repo, full includes the merge), this is a merge with the leaked model, you can use the other repository to save bandwidth.
EQ-Bench: 84.89 GSM8k: 77.18 (71.04 when using ChatML) Hellaswag: 87.67
Edit: Upon further testing a score of 85.09 was achieved using ChatML instead of Mistral's prompt.
I recommend using the ChatML format instead, I will run more benchmarks. This also fixes the bug with Miqu dequant failing to provide a stop. <|im_start|>system Provide some context and/or instructions to the model. <|im_end|> <|im_start|>user The user’s message goes here <|im_end|> <|im_start|>assistant <|im_end|>
Credit to https://twitter.com/hu_yifei for providing GSM & Hellaswag. It is the first open weight model to dethrone GPT-4 on EQ bench,