Chess Cheating Dataset

Introduction

In recent times, there's been a surge of interest from armchair data scientists and programmers eager to develop algorithms for detecting cheaters in online chess games. While the enthusiasm is commendable, effective cheating detection requires access to high-quality, diverse datasets. This dataset aims to be a starting point for gathering comprehensive data about cheating in chess. Currently, there is limited public data available for evaluating cheating detection methods, making this initiative even more crucial.

This dataset consists of 48,000 games played between bots. Each move made by these bots has a chance of being generated by either Stockfish or Maia. To save computing time, most of the game, one side has a probability of zero to use stockfish (Only one side is "cheating").

In these games, the first 10 moves are sourced from the Lichess database. The Maia bot, which participates in these games, has a fixed Elo rating for each game. This rating is determined based on the Elo ratings of the players from the original game from which the first 10 moves were sourced.

For example, if the first 10 moves are taken from a game where one player has an Elo rating of 1864 and the other has 1765, the Maia bot's Elo rating will be set to 1900 and 1800, respectively.

Row information

• Liste cheat white : 0 for not cheating and 1 for cheating for the white side
• Liste cheat black : 0 for not cheating and 1 for cheating for the black side
• Game : game PGN with the result of the game at the end
• Elo White : Elo of the player who did the first 10 white moves from the lichess database
• Elo Black : Elo of the player who did the first 10 black moves from the lichess database
• score : result of the game '1-0', '0-1' or '1/2-1/2'

Additional information

• stockfish version : 16
• stockfish depth : 20
• maia weights : directly from the [github] (https://github.com/CSSLab/maia-chess/tree/master/maia_weights)
• If the elo is equal to -1 it means that the original elo is either bellow 1000 or above 2000

Some numbers

• 48933 games total
• 12278 games has no side cheating
• 11028 games has both side cheating
• 25627 games has only one side cheating