How is the risk of malicious up/down voting in the side-by-side tab handled?

#3
by MoritzLaurer HF staff - opened

In the tab "Arena (side-by-side)" I can select specific models to compare and then I can cast a vote, knowing which model I am voting for, enabling people/organisations to intentionally upvote/downvote their own or competitor models. Are these votes somehow handled differently to votes from the tab "Arena (battle)", where the user truly doesn't know which model they are evaluating?

Very cool arena/leaderboard btw :)

Massive Text Embedding Benchmark org

Please use the associated github for issues (see also https://huggingface.co/spaces/mteb/arena/discussions/2).

Though quick answer: We do believe we keep track of how the votes were given so it is possible to filter based on "side-by-side"

Massive Text Embedding Benchmark org

they dont count for the lb :)
Screenshot 2024-08-15 at 7.09.40 AM.png

Sign up or log in to comment