TabArena Leaderboard Code

This repository contains the frontend code to display TabArena leaderboard. The leaderboard is hosted on a HuggingFace space.

Reference:

• Website: https://tabarena.ai
• Paper: https://tabarena.ai/paper-tabular-ml-iid-study
• TabArena Codebase: https://tabarena.ai/code

Install LB Code for Development

pip install -e ".[dev]"
# Or 
uv pip install -r pyproject.toml

Reading the leaderboard programmatically

The published numbers are plain CSVs in this repo, so the cheapest way to read them is a direct fetch (no queue, no token, -L because the raw path redirects to the resolve cache):

curl -sL https://huggingface.co/spaces/TabArena/leaderboard/resolve/main/data/entrants_models/imputation_yes/splits_all/tasks_all/datasets_all/website_leaderboard.csv

The path is data/entrants_{...}/imputation_{yes,no}/splits_{all,lite}/tasks_{...}/datasets_{...}/ for TabArena and data_beyondarena/subsets/{subset}/ for BeyondArena.

entrants is who competed. There are three system categories that can each be admitted or not (open for open-source local systems, llm for systems with an LLM in the loop, api for systems behind a closed-source API), so the eight keys are models (individual models only, the default), open, llm, api, open_llm, open_api, llm_api and open_llm_api. It is not a row filter: Elo is a pairwise rating over whoever competed and Improvability is the gap to the best of them, so each pool is its own evaluation with its own numbers.

The Space also serves four JSON endpoints for agents, defined in api.py: list_leaderboards, get_tabarena_leaderboard, get_beyondarena_leaderboard and get_pareto_frontier. The same four are served as MCP tools at /gradio_api/mcp/. Hugging Face advertises the JSON route through the Agents button on the Space page, which points at a generated agents.md. See AGENTS.md for the contract and the two constraints on how the descriptions must be written.

Current Steps to get results:

1. Run https://github.com/autogluon/tabarena/blob/main/scripts/run_generate_website_artifacts.py
2. Delete the current data folder contents in this repo to remove old results.
3. Unzip the generated clean_website_artifacts.zip zip file into the data folder in this repo.
4. Test the LB locally, otherwise you are done :)