Spaces:

open-llm-leaderboard

/

open_llm_leaderboard

like 10.1k

Running on CPU Upgrade

It also needs to be manually refreshed whenever someone wants to see up to date data instead of defaulting to the latest leader board.

I don't think this is undesirable, it prevents models from appearing as high quality while some evals are missing which is less misleading than them appearing really high because they did one test well.

The missing benchmarks (HellaSwag and MMLU) have random baselines at 25%, so at least it shouldn't be 0, assuming no model will perform worse than random, which should be the case for HellaSwag and MMLU, though not necessarily for the other two benchmarks.

@natolambert was this fixed?