Hacker News Reader

Comment by GaggiX | Hacker News Reader

I like using the OpenChat model based on Mistral, I think it is by far the best 7B model: https://openchat.team/ (you can also use it online here)

View on HN

Mistral is a big step up from llama2 in my experience, really impressive to see a 7B model that can do so much. I haven't noticed a considerable improvement from the 13b version, and by using 7b I can keep my context size very large.

Holy balls, this is amazing. Thank you for sharing. At that size, I could probably go one up higher and still put it on my 3090. Thank you thank you.

It actually does really well for simple queries like I'm asking about writing systemd files.

thanks for this recommendation. have you been using this one for a while? How do you like it?

It's pretty new but it performs very well when I use it, it is also the third best non-proprietary model on the chatbot arena: https://huggingface.co/spaces/lmsys/chatbot-arena-leaderboar... (the ELO is calculated from blinded comparisons so it really shows the actual performance of the model).

Color me impressed. This is what I typed: “ageh man beh zabooneh khodemoon benevisam, to mifahmi?”

—

So above is a question in Farsi written using Latin alphabet. The q is “if i write in our own language would you understand?”

Answer: نه، من نمی فهمم. این زبانی است که می تواند به زبان فارسی به شما در این بخش کار نماید. اگر سؤالی دارید یا می خواهید به سوالی پاسخ دهید، لطفاً به زبان انگلیسی بگویید.

‘No, I will not understand. ... please ask in the English language.’

Reply on news.ycombinator.com