[ExI] Tiny But Mighty LLM

[Gadersd]

There is a new kid on the block: Mistral 7B. Mistral is a European AI company that seeks to rival the big players such as Meta and OpenAI by releasing powerful fully open source AI models. It just released a 7 billion parameter model and it is dominating the performance charts, exceeding models nearly twice its size. There are quantized versions of it that are just about 6 gigabytes in size, small enough to potentially run on smartphones.

I’ve been testing it and am very impressed. It successfully solved a quadratic equation by factoring it and even performed complicated arithmetic nearly perfectly that even I am unable to easily do in my head. I asked it to count the number of words in a sentence and it succeeded even while GPT 4 failed at the same task, being off by one word. It has some weaknesses however. It failed to reason that drying multiple shirts at once takes the same amount of time as a single shirt and committed various other trivial logical errors.

In summary, unlike tiny language models of the past, Mistral 7B OpenOrca (a fine-tuned version) is actually useful for real tasks although it is still prone to mistakes. One example use-case is filtering and deleting old private emails that you don’t have time to do manually, something I am excited about as I have thousands of emails only some of which are spam. Since the model is small you can run it on your own hardware and keep all your data private from the companies.

You can try it out easily at the webpage https://huggingface.co/spaces/Open-Orca/Mistral-7B-OpenOrca <https://huggingface.co/spaces/Open-Orca/Mistral-7B-OpenOrca>
Have a go and see how far the small models (soon to be on your phones) have come!
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20231008/91d42764/attachment.htm>