LLM Arena: Benchmarking LLM Models in Russian

LLM Arena (https://llmarena.ru/) is an open crowd-sourced platform designed for evaluating large language models (LLMs) in the Russian language. Our goal is to provide an objective, open, and up-to-date benchmark of LLM models tailored for the Russian language.

Features

The platform offers three main features:

1. Anonymous Chat (https://llmarena.ru/?arena): Users can engage in a blind test, posing questions to two anonymous models simultaneously. This allows for unbiased comparisons and helps us gather valuable user feedback for building our leaderboard.

2. Named Chat (https://llmarena.ru/?compare): Users can select specific models to chat with side-by-side, allowing for direct comparisons and a more controlled evaluation experience.

3. Leaderboard (https://llmarena.ru/?leaderboard): The platform displays a leaderboard based on the Elo rating system and the Bradley-Terry model, showcasing the rankings of different LLMs based on the aggregated user evaluations.

   • Arena Hard Benchmark: This offline benchmark utilizes a more potent model to evaluate the performance of other models based on a curated set of 500 prompts.
   • Data Visualization: The leaderboard section also features four plots visualizing the performance of models across different categories.

How It Works

1. User Interaction: Users interact with LLMs in either the anonymous or named chat setting.
2. User Evaluation: Users rate which model provided a better response.
3. Data Collection: The platform collects these user evaluations.
4. Leaderboard Generation: The Elo rating system and Bradley-Terry model use the collected data to rank the LLMs and generate the leaderboard.