Spaces:

open-llm-leaderboard

/

open_llm_leaderboard

like 11.8k

Running on CPU Upgrade

I posted about this before (after noticing that the leaderboard uses almost as much ram as Skyrim when using Firefox) and they closed the ticket. The tech they’re using for the leaderboard is very, very inefficient, and not suited for creating a website like this. Every action that takes 4-10 seconds to process would be instantaneous if they changed what it’s built with.

They should be using this: https://datatables.net/

Which would make the HF leaderboard as fast as this other LLM leaderboard:
https://llm.extractum.io/

My current issue is with viewing model sizes. To view all 7b models I have to uncheck 70b, wait 4 seconds, uncheck 30b, wait 4 seconds, etc, until 7b is the last one active. A data re-sort like this would take 0 seconds with DataTables.

It really doesn’t reflect well on HuggingFace since they’re trying to position themselves as tech industry leaders here. Sites like this worked seamlessly decades ago; it’s just a plain HTML table.

It takes a full 30 seconds for my phone to load the leaderboard, and 17 seconds every time I sort a column. Something has to change.

Hi!
Thank you for your issue @Ostixe360 ! Yes, we have, after @spacecowgoesmoo opened their issue here.
The Gradio team is actually working on an upgrade for this, which will be released very soon.
We are also working on adding a lightweight version of the leaderboard as a very basic HTML page for people with connectivity problems, but it's not at the top of our priority list atm, we are doing our best to fix this soon!

Okay thank you!

Tagging @pngwn who is working on this on the gradio side :)

The worst of this should be addressed in the next week.

Hi @Ostixe360 @spacecowgoesmoo @TNTOutburst !
An upgrade of the leaderboard has been pushed thanks to the amazing efforts of @pngwn in the Gradio team, and the leaderboard is now blazing fast!
We hope you like it!