TobiasLogic/chess-arena-results

Chess Arena Results

Every finished game played in Chess Arena, kept here because a Space has no filesystem that survives a restart. The Space reads this on startup and writes back a few seconds after each game, so the leaderboard carries across restarts, sleeps and rebuilds rather than resetting every time the container is recycled.

It is append only. Nothing in the arena can delete a result.

Format

results.json is a JSON array of games, one array each:

[
  ["chessmamba2", "chess-llama", 1.0],
  ["engine-3", "random", 1.0],
  ["chessaggro", "chessgpt", 0.5]
]

Position	Meaning
0	player id with the white pieces
1	player id with the black pieces
2	result from white's point of view: 1.0 white won, 0.5 drawn, 0.0 black won

Player ids are the arena's internal keys. Built in players use short names such as chessmamba2, chess-llama, chessaggro, chessslm, chessgpt, chessformer, bluesunflower, chesspythia, engine-1 through engine-4, and random. Anything added at runtime is stored as hub:owner/name.

How the ratings are worked out

The Space does not accumulate ratings with a K factor. That would depend on the order games happened in and would never settle. Instead every game is fitted at once by maximum likelihood under the Bradley Terry model, which is what BayesElo and Ordo do, so the ratings are a property of the results rather than of their sequence. A draw counts as half a win. The scale is anchored so the field averages 1500, which makes these ratings meaningful within this pool only and not comparable to a FIDE or Lichess number.

Two rules shape what goes in:

The random mover is not rated. It plays no chess, so beating it says nothing about strength and would only inflate whoever met it most. Its games are stored, and its record shows in the cross table as a floor to check against, but they are excluded from the fit.

Games stopped by the ply cap are adjudicated on material rather than written down as draws. A cap out never finished, and recording it as a draw would put a result into the ratings that neither side earned. Four pawns of material decides it, otherwise it stands as drawn.

Reading it

import json
from huggingface_hub import hf_hub_download

path = hf_hub_download("TobiasLogic/chess-arena-results", "results.json", repo_type="dataset")
games = json.load(open(path))

wins = {}
for white, black, score in games:
    wins[white] = wins.get(white, 0) + score
    wins[black] = wins.get(black, 0) + (1 - score)

The fitting code the Space uses is elo.py in the Space repo, about eighty lines with no dependencies beyond the standard library.

Caveats worth knowing

Games are played at 500 ms a move on shared hardware, so a model that wants a lot of search is measured a little short of its best.

The earliest handful of rows were recorded before two bugs were fixed: the search engine picked moves from partially searched positions when it ran out of time, and unfinished games were written down as draws. Those rows are still here, since the store is append only. They are few and will be swamped as games accumulate, but they are not clean data.