artemlepin/chess-fraud

ChessFraud

ChessFraud is a tabular benchmark for cheating detection in online chess. It accompanies the KDD 2026 Datasets and Benchmarks Track paper ChessFraud: Exploring the Capabilities of Human-Aligned Models for Cheating Detection in Online Chess.

The two configurations serve complementary roles. ChessFraud supports evaluation against cheating observed under a controlled assistance protocol. ChessFraud-Synth supports training and analysis using alternatives produced by classical engines and human-aligned chess models for public Lichess games.

Dataset at a glance

Configuration	Source	Released rows	Key analysis counts
ChessFraud	505 controlled games, represented as 1,010 player-side games	38,510	407 cheating player-games (40.3%); 9,405 cheating half-moves (24.4%)
ChessFraud-Synth	12,000 public Lichess player-games	1,074,287	417,207 decision points marked by is_used

Loading the data

from datasets import load_dataset

chess_fraud = load_dataset(
    "artemlepin/chess-fraud",
    "chess_fraud",
)

chess_fraud_synth = load_dataset(
    "artemlepin/chess-fraud",
    "chess_fraud_synth",
)

Allie move embeddings

The additive artifacts/allie_embeddings/ subtree contains frozen 1024-dimensional float16 move embeddings produced with Allie-2500 for observed moves and for moves proposed by Stockfish, Lc0, Maia-2, and Allie. It does not change either tabular configuration or its Parquet schema. Allie uses elo_oppo as the opponent's rating. Downstream experiments select rows with is_used = true.

artifacts/allie_embeddings/
├── manifest.json
├── synth/
│   ├── row_index.parquet
│   ├── move_uci.npy
│   ├── move_stockfish_1.npy
│   ├── move_stockfish_5.npy
│   ├── move_stockfish_9.npy
│   ├── move_stockfish_11.npy
│   ├── move_stockfish_15.npy
│   ├── move_lc0_1.npy
│   ├── move_lc0_10.npy
│   ├── move_lc0_100.npy
│   ├── move_maia2_2050.npy
│   └── move_allie_2500.npy
└── tournament/
    ├── row_index.parquet
    └── embs_allie_2500.npz

The eleven Synth arrays preserve their source .npy bytes, row order, shape, and dtype. The tournament .npz likewise preserves its source bytes and contains the named move_uci array. These embeddings are used by the move-level and game-level experiments reported in Tables 4 and 5. The matching GitHub notebooks provide the reproduction protocol and loading details.

Alignment indexes

Join each row_index.parquet to the matching tabular configuration using the documented keys, then use embedding_row to index the arrays. Never assume Parquet row order. Each index provides a zero-based, contiguous, unique embedding_row and a unique key for a one-to-one join.

Index	Columns in schema order	Join key
synth/row_index.parquet	embedding_row (int64), player_id (string), game_id (string), half_move (int64), move_player (string), move_thinking_time (float64)	player_id, game_id, half_move, move_player
tournament/row_index.parquet	embedding_row (int64), game_id (string), player_id (int64), half_move (int64), move_player (string)	game_id, player_id, half_move, move_player

The Synth index retains the finite source-manifest move_thinking_time needed by the embedding experiments. This value is authoritative when an impossible source clock interval is intentionally represented as missing in the public core table.

Provenance and integrity

artifacts/allie_embeddings/manifest.json defines the allie-embeddings-v1 contract. It records the complete file inventory, byte sizes, SHA-256 checksums, NumPy shapes and dtypes, the tournament archive member metadata, index schemas and row counts, normalized row digests, source provenance, and the path-neutral preparation command. Validate the manifest and the one-to-one joins before training; a checksum match alone does not establish row alignment.

Download, cache, and local override

The embedding subtree is approximately 9.56 GB, excluding Hugging Face cache and temporary-file overhead. Ensure that both download storage and working space are available before running a matching tutorial. The tutorials pin a full immutable 40-character dataset revision; use exactly the revision recorded by the tutorial you are reproducing, rather than a mutable branch or tag.

The following helper downloads only the embedding subtree. Its revision argument must be the full immutable revision copied from the matching GitHub tutorial.

from huggingface_hub import snapshot_download


def download_allie_embeddings(*, revision: str, cache_dir: str | None = None) -> str:
    if len(revision) != 40 or any(
        character not in "0123456789abcdef" for character in revision.lower()
    ):
        raise ValueError("revision must be a full 40-character hexadecimal commit SHA")
    return snapshot_download(
        repo_id="artemlepin/chess-fraud",
        repo_type="dataset",
        revision=revision,
        allow_patterns=["artifacts/allie_embeddings/**"],
        cache_dir=cache_dir,
    )

The matching repository tutorials provide a validated resolver with two environment overrides:

• CHESSFRAUD_ALLIE_CACHE_DIR selects the snapshot cache directory;
• CHESSFRAUD_ALLIE_EMBEDDING_DIR bypasses downloading and points to an existing snapshot root that already contains a validated artifacts/allie_embeddings/ subtree.

The local override takes precedence over downloading, but it does not bypass manifest, checksum, schema, or row-index validation.

Scope, privacy, licensing, and limitations

This subtree contains derived embeddings and alignment metadata for rows that are already represented in the public benchmark. The indexes introduce no new identifier domain: tournament identifiers remain pseudonymous, whereas Synth identifiers remain the public Lichess usernames and game IDs described below. The bundle excludes personal information, raw server logs, private engine-hint traces, the tournament plugin, model weights, and detector outputs.

The dataset-level licensing and source terms in this card also apply to the embedding bundle: the ChessFraud release is CC BY 4.0, ChessFraud-Synth derives from the public Lichess database under CC0, and model weights and engine binaries remain under their original licenses. The embeddings cover only the frozen representations and engine configurations used by the embedding experiments; they are not a general model release, do not expand the annotation scope, and should not be used as evidence for sanctions against individual players.

Data representation

Both configurations are organized around a focal player, identified by player_id.

In ChessFraud, each physical half-move is stored once, in the player-side game of the player who made the move. game_id retains the source game hash and has the suffix _w or _b for the White or Black player-side view.

In ChessFraud-Synth, player_id remains fixed within a player-game and the full sequence is retained. Rows for both players provide preceding game context.

All positions are encoded in FEN. All moves are encoded in UCI.

Shared fields

Unless stated otherwise, every field in this section is present in both configurations. All evaluations and outcome estimates use the perspective of player_id: a positive evaluation favors the focal player, irrespective of color.

Identifiers

Field	Type	Values	Meaning
game_id	string	non-empty	Player-side game identifier. ChessFraud uses a pseudonymous game hash with a color suffix; ChessFraud-Synth retains the public Lichess game ID.
player_id	int64 in ChessFraud; string in ChessFraud-Synth	ChessFraud: 1–49; ChessFraud-Synth: non-empty	Focal player. ChessFraud uses a pseudonymous integer; ChessFraud-Synth retains the public Lichess username.
opponent_id	int64 in ChessFraud; string in ChessFraud-Synth	same identifier convention as player_id	Opponent of player_id.

Player context

Field	Type	Values	Meaning
player_color	string	white, black	Color of player_id.
player_elo	int64	non-negative	Rating of player_id for the game.
opponent_elo	int64	non-negative	Rating of opponent_id for the game.
time_control	string	base+increment	Base time and increment in seconds. ChessFraud uses 300+0.
game_result	string	1-0, 0-1, 0.5-0.5	Game result from White's perspective.

ChessFraud ratings are standardized to the Lichess scale. A reported Lichess rating is used when available. For four participants who reported only a Chess.com rating, 300 points are added to approximate the corresponding Lichess rating. This procedure is described in Appendix C.1 of the paper.

Move context

Field	Type	Values	Meaning
half_move	int64	>= 1	One-based half-move index within the source game.
move_player	string	legal UCI move	Observed move at the current half-move. On ChessFraud-Synth analysis rows, it is the focal player's move.
position_before	string	valid FEN	Position immediately before move_player.
position_after	string	valid FEN	Position immediately after move_player.

Timing

Timing values are measured in seconds. Their units are stated here rather than repeated in the field names.

In ChessFraud-Synth, impossible negative intervals in the public Lichess clock telemetry are stored as missing move_thinking_time values.

Field	Type	Range	Meaning
move_thinking_time	float64	[0, +inf)	Time spent on the observed move.
clock_remaining_time	float64	[0, +inf)	Clock time remaining for the player who made the observed move.

Analysis selection

Field	Type	Values	Meaning
is_used	bool	true, false	Whether the row is a paper-aligned decision point.

The opening cutoff is 20 half-moves. In ChessFraud, is_used is true for a retained player move with half_move >= 21. In ChessFraud-Synth, it is true when half_move >= 21 and the side to move matches player_color: odd half-moves for White and even half-moves for Black.

Evaluation features

eval_before, eval_after, and every eval_<source> field are Stockfish 17 depth-15 evaluations. They are expressed in centipawns from the perspective of player_id and clipped to [-1000, 1000], corresponding to [-10, 10] pawns.

Field	Type	Range	Meaning
eval_before	int64	[-1000, 1000]	Clipped evaluation of position_before.
eval_after	int64	[-1000, 1000]	Clipped evaluation after move_player, reoriented to the same player.
centipawn_loss	int64	[0, 2000]	Loss in clipped evaluation after the observed move: max(0, eval_before - eval_after).
normalized_centipawn_loss	float64	[0, 1]	Decrease in the current player's estimated winning chance after the observed move, computed using the Lichess Win% model.

Normalized centipawn loss converts the published clipped evaluations using

p(e) = 1 / (1 + exp(-k * e)), where k = 0.00368208
normalized_centipawn_loss = max(0, p(eval_before) - p(eval_after))

Engine outputs

The move_<source> fields in this and the following section are post hoc features in ChessFraud and candidate synthetic assistance moves in ChessFraud-Synth. All published Stockfish feature outputs were computed with Stockfish 17 and are distinct from the Stockfish 15 assistance recorded during the controlled tournaments.

Field	Type	Values	Meaning
move_stockfish_1	string	UCI move	Top-1 move at depth 1.
move_stockfish_9	string	UCI move	Top-1 move at depth 9.
move_stockfish_15	string	UCI move	Top-1 move at depth 15.

Human-aligned model outputs

Maia-2 uses a focal-player rating of 2050, while Allie uses 2500. Both models receive the opponent's recorded rating for the game. Allie additionally uses the time control and preceding move history available at the current position.

Field	Type	Range or values	Meaning
move_maia2_2050	string	UCI move	Maia-2 top-1 move.
maia2_win_prob_2050	float64	[0, 1]	Player-perspective game-outcome estimate: loss = 0, draw = 0.5, win = 1.
move_allie_2500	string	UCI move	Allie top-1 move.
allie_win_prob_2500	float64	[0, 1]	Player-perspective game-outcome estimate: loss = 0, draw = 0.5, win = 1.

Engine and human-aligned model outputs are filled for every row where is_used is true. Other rows are retained only as sequence context, so these outputs can be empty there.

ChessFraud fields

The fields in this section appear only in the chess_fraud configuration. ChessFraud was collected in two controlled tournaments, numbered 4 and 5. The tournament assistance plugin used Stockfish 15, and each participant was assigned a search depth of 12, 14, 16, or 18.

Field	Type	Range or values	Meaning
tournament_id	int64	4, 5	Controlled tournament.
player_hint_shown	bool	true, false	Whether a Stockfish hint was shown to the current player on this half-move.
assistance_search_depth	int64	12, 14, 16, 18	Search depth assigned to the player in the tournament plugin.
assistance_line_rank	int64	1–5	One-based rank of the displayed engine line followed by the player.
is_cheating_move	bool	true, false	Whether the move belongs to a followed Stockfish line under the tournament annotation protocol.
is_cheating_player_game	bool	true, false	Player-game cheating label used for game-level evaluation.
is_accused_by_opponent	bool	true, false	Whether the opponent accused player_id of cheating after the game.

A cheating series starts when a player follows a displayed Stockfish line and continues while the player follows the same continuation. The series ends when the player diverges from that line. A hint could be requested during the opponent's turn; in that case, the next matching player move is labeled even though player_hint_shown is false on that move's row.

assistance_line_rank is present only when the player followed one of the displayed engine lines.

ChessFraud-Synth fields

The fields in this section appear only in the chess_fraud_synth configuration. They extend the shared fields with the remaining top-1 outputs used in the paper. Stockfish and Lc0 are classical engines; Maia-1, Maia-2, and Allie are human-aligned models.

Sampling metadata

Field	Type	Values	Meaning
rating_bin	string	six 200-point Elo strata up to 2200	Rating stratum used to sample the focal player.

Additional engine outputs

Field	Type	Range or values	Meaning
move_stockfish_5	string	UCI move	Top-1 move at depth 5.
move_stockfish_11	string	UCI move	Top-1 move at depth 11.
eval_stockfish_1	int64	[-1000, 1000]	Evaluation after move_stockfish_1.
eval_stockfish_5	int64	[-1000, 1000]	Evaluation after move_stockfish_5.
eval_stockfish_9	int64	[-1000, 1000]	Evaluation after move_stockfish_9.
eval_stockfish_11	int64	[-1000, 1000]	Evaluation after move_stockfish_11.
eval_stockfish_15	int64	[-1000, 1000]	Evaluation after move_stockfish_15.
move_lc0_1	string	UCI move	Lc0 top-1 move with a 1-node budget.
move_lc0_10	string	UCI move	Lc0 top-1 move with a 10-node budget.
move_lc0_100	string	UCI move	Lc0 top-1 move with a 100-node budget.
eval_lc0_1	int64	[-1000, 1000]	Evaluation after move_lc0_1.
eval_lc0_10	int64	[-1000, 1000]	Evaluation after move_lc0_10.
eval_lc0_100	int64	[-1000, 1000]	Evaluation after move_lc0_100.

Additional human-aligned model outputs

Field	Type	Range or values	Meaning
move_maia1_1900	string	UCI move	Maia-1 1900 top-1 move.
eval_maia1_1900	int64	[-1000, 1000]	Evaluation after move_maia1_1900.
eval_maia2_2050	int64	[-1000, 1000]	Evaluation after move_maia2_2050.
eval_allie_2500	int64	[-1000, 1000]	Evaluation after move_allie_2500.

The eval_<source> fields allow source strength to be computed under one evaluation protocol. For a proposed move, its centipawn loss is max(0, eval_before - eval_<source>). No duplicate centipawn-loss columns are stored.

Splits

ChessFraud has one full split. ChessFraud-Synth uses the paper's 80/20 player-disjoint train and test split, built from the source split_by_player field. The source split_by_player and split_by_games columns are not included in the released tables. ChessFraud-Synth and ChessFraud do not share players.

Reconstructing synthetic examples

Each row stores the observed move once. A binary fair/assisted example for a chosen source can be constructed without generating new chess features:

• use move_player as the fair move;
• use one move_<source> field as the assisted move;
• assign labels 0 and 1, respectively;
• restrict paper-aligned experiments to rows where is_used is true.

The release does not store a generic move_assistance or assistance_model column because all published source moves are explicit columns.

Dataset construction

ChessFraud

The controlled collection and annotation protocol is described in Section 3 and Appendix C of the paper. Appendix C.1 documents rating standardization, and Appendix C.4 gives the engine usage instructions provided to participants.

ChessFraud-Synth

ChessFraud-Synth was constructed from public Lichess Blitz games played in March 2025 using rating-stratified sampling. Matches between two selected players were excluded from the sampled data.

The sampling design, temporal separation from the human-aligned models' training data, split construction, and synthetic cheating procedure are described in Section 4.3 of the paper.

Limitations

1. ChessFraud covers a small controlled cohort playing five-minute games and does not represent every player population or time control.
2. ChessFraud-Synth contains model-generated move substitutions rather than observed cheating behavior.
3. ChessFraud labels cover the logged Stockfish assistance protocol; other forms of assistance are outside the annotation scope.
4. The tournament cheating plugin is not released.

Privacy, consent, and responsible use

All ChessFraud participants gave informed consent to share their gameplay data under the controlled protocol. Compensation was fixed and independent of tournament performance. The release uses pseudonymous participant identifiers; personal information and raw server logs are excluded.

ChessFraud-Synth retains public Lichess usernames and game IDs so that source games remain auditable. These identifiers come from the public Lichess game database.

The datasets are intended for research on cheating detection. They are not a basis for sanctions against individual players in unrelated games.

Licensing and source terms

The ChessFraud release is licensed under CC BY 4.0. ChessFraud-Synth is derived from games distributed through the public Lichess database under CC0. Model weights and engine binaries remain subject to their original licenses.

Citation

If you use ChessFraud or ChessFraud-Synth, cite the accepted paper:

@inproceedings{linich2026chess_fraud,
  title     = {ChessFraud: Exploring the Capabilities of Human-Aligned Models for Cheating Detection in Online Chess},
  author    = {Linich, Anastasiia and Lepin, Artem and Sakhovskiy, Andrey and Toleutaeva, Anita and Lepa, Georgii and Neznamov, Andrei and Budennyy, Semen},
  booktitle = {Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining},
  year      = {2026},
  doi       = {10.1145/3770855.3817587}
}

Paper: https://doi.org/10.1145/3770855.3817587