LMSYS
Statistics for Chatbot Arena
We added some additional figures to show more statistics. The code for generating them is also included in this notebook. Please note that you may see different orders from different ranking methods. This is expected for models that perform similarly, as demonstrated by the confidence interval in the bootstrap figure. Going forward, we prefer the classical Elo calculation because of its scalability and interpretability. You can find more discussions in this blog post.
Figure 1: Fraction of Model A Wins for All Non-tied A vs. B Battles
Figure 2: Battle Count for Each Combination of Models (without Ties)
Figure 3: Confidence Intervals on Model Strength (via Bootstrapping)
Figure 4: Average Win Rate Against All Other Models (Assuming Uniform Sampling and No Ties)