DPO fine-tuned Mistral 7B beats Llama 70B on MT Bench

Posted by clmnt on Tuesday October 10, 2023 @03:26PM
from the huggingface.co dept.

https://huggingface.co/HuggingFaceH4/zephyr-7b-alpha

( 3 comments | on HN )

Re:DPO fine-tuned Mistral 7B beats Llama 70B on MT Bench by amilios on Tuesday October 10, 2023 @03:59PM (#37836720)

Can anyone corroborate this anecdotally? I.e. has anyone actually looked at the output of the two models side-by-side for common tasks? There's lots of talks these days about academic benchmarks being pretty "broken" for modern LMs, and not really properly showcasing the differences between models. I wonder if that's the case here or if the model is genuinely better.

[ Parent | on HN ]

Re:DPO fine-tuned Mistral 7B beats Llama 70B on MT Bench by brucethemoose2 on Wednesday October 11, 2023 @04:03PM (#37849556)

I dunno. I totally missed this and will check it out.

- Huggingface is less likely to "cheat" by training on tests than other orgs, I think.

- Some finetunes are really good at a particular test (like XWin). This isnt necessarily a bad thing, if they are good at a specific niche.

[ Parent | on HN ]

Re:DPO fine-tuned Mistral 7B beats Llama 70B on MT Bench by brucethemoose2 on Wednesday October 11, 2023 @04:05PM (#37849589)

> <|system|>, <|user|> and <|model|>.

Oh hey, thats almost Metharme's format.

It must originate from an older model, as most new models dont use that syntax.

[ Parent | on HN ]