ShinojiResearch/Senku-70B-Full

UPDATE: 85.09 EQ-Bench with ChatML template

• EQ-Bench: (Mistral) 84.89 -> 85.09 (ChatML)
• GSM8k: (Mistral) 77.18 -> 71.04 (ChatML)
• Hellaswag: (Mistral) 87.67 -> ??

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.

Update: Upon further testing a score of 85.09 was achieved using ChatML instead of Mistral’s prompt.

Prompt Template

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|>