datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
pawn-stockfish-100m
PAWN Stockfish 100M
100,000,000 self-play chess games generated with Stockfish 18, each
annotated with per-position, per-legal-move evaluations — for chess
policy-learning and NNUE-distillation research.
Dataset Summary
100,000,000 machine-generated self-play chess games. Every position in
every game is annotated with an evaluation of every legal move, not just
the move played. The dataset was built as training data for
PAWN — a testbed for finetuning
and… See the full description on the dataset page: https://huggingface.co/datasets/thomas-schweich/pawn-stockfish-100m.lichess-stockfish-normalized
Lichess Chess Positions: ML-Ready Deduplicated Evaluations
Dataset Description
A curated dataset of 316,072,343 unique chess positions with Stockfish evaluations, optimized for training neural networks. This is a deduplicated, ML-ready version of the Lichess evaluation database.
Why This Dataset?
While Lichess provides deduplicated evaluations in JSONL.zst format, and HuggingFace hosts the full (non-deduplicated) version, this dataset offers:
Unique advantages:… See the full description on the dataset page: https://huggingface.co/datasets/mateuszgrzyb/lichess-stockfish-normalized.stockfish-19-soft-targets
avewright/stockfish-19-soft-targets
Official Stockfish 19 MultiPV soft targets, mined from the Lichess ECO
opening set. In-progress snapshot toward 1M unique positions.
600,000 rows in this upload. Source id 4. Vocab compact (1968).
How positions are chosen
Games start from the Lichess Chess Openings dataset
(lichess-org/chess-openings): 3,810 named
leaves (HF card still lists 3,704) plus
book prefixes, 7,852 unique starts.
ECO volumes: A 817 / B 772 /
C 1,250 / D… See the full description on the dataset page: https://huggingface.co/datasets/avewright/stockfish-19-soft-targets.Stockfish_4PC_HCEchess-sft-10m-stockfishstockfish-evaluation-SAN
Dataset Card for the Stockfish Evaluations
A dataset of chess positions evaluated with various flavours of Stockfish running within user browsers. Produced by, and for, the Lichess analysis board. Evaluations are formatted as JSON; one position per line.
The schema of a position looks like this:
{
"fen": "8/8/2B2k2/p4p2/5P1p/Pb6/1P3KP1/8 w - -",
"depth": 42,
"evaluation": 5.64,
"best_move": "Kg1",
"best_line": "Kg1 Ke6 Kh2 Kd6 Be8 Kc5 Kh3 Kd6 Bb5 Ke7"
}
fen: string, the… See the full description on the dataset page: https://huggingface.co/datasets/bingbangboom/stockfish-evaluation-SAN.stockfish_datasetstockfish-debug
Dataset Card for stockfish-debug
See my blog post for additional details.
Columns
The datase contain the following columns:
fen: The FEN string of the board.
move: The move that was played.
result: The result of the game (with "-" for unfinished games).
Data details
Pre-processing of the Stockfish games provided by BlueSunflower/chess_games_base.
Code used:
import jsonlines
import chess
import tqdm
def preprocess_games(in_path, out_path):
with… See the full description on the dataset page: https://huggingface.co/datasets/yp-edu/stockfish-debug.chess-stockfish-regret
Chess RLVR Stockfish Regret 1400/100 Snapshot
This snapshot dataset stores chess positions for reinforcement learning with verifiable rewards.
Each row contains:
{
"id": "chess_rlvr_000001",
"fen": "rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq - 0 1",
"legal_moves": "{\"Nf3\": -0.015, \"e4\": 0.0}"
}
legal_moves is a JSON object encoded as a string. The object maps each legal SAN move to a Stockfish-derived negative regret score for the player to move.
The RLVR… See the full description on the dataset page: https://huggingface.co/datasets/albertklorer/chess-stockfish-regret.stockfish-policy
Stockfish Policy Teacher (100k)
Root-policy supervision for 100,000 chess positions, produced by a fixed-node
MultiPV Stockfish search. Each row carries a full probability distribution over
every legal move at the root — not just the best move — so it can be used as
a soft-label teacher for policy distillation.
What makes a row
For every legal root move the generator first evaluates the child position with
go depth 0, negating the child score into the root… See the full description on the dataset page: https://huggingface.co/datasets/Pawitt/stockfish-policy.chess-rlvr-stockfish-regret
Chess RLVR Stockfish WDL
This dataset stores chess positions for reinforcement learning with verifiable rewards.
Each row contains:
{
"id": "chess_rlvr_000001",
"fen": "rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq - 0 1",
"legal_moves": {
"Nf3": -0.015,
"e4": 0.0
}
}
legal_moves maps each legal SAN move to a Stockfish-derived negative regret score for the player to move.
The RLVR reward is negative expected-score regret:
reward =… See the full description on the dataset page: https://huggingface.co/datasets/albertklorer/chess-rlvr-stockfish-regret.GAMBIT-stockfish18-selfplay
GAMBIT: Generalization or Memorization? Brittleness Testing for Chess-Trained Language Models
Overview
1050 self-play games between base Stockfish 18 and Stockfish Skill Level 0-20 variants.
Each skill level variant plays 25 White/25 Black games against a base instance of Stockfish.
All games use depth=15 and no opening books were used.
sf18_selfplay_games.pgn — PGN file containing all labeled 1050 SF18 self-play games
sf18_selfplay_match-summary.txt— match… See the full description on the dataset page: https://huggingface.co/datasets/ethanjtang/GAMBIT-stockfish18-selfplay.chess-stockfish-analysisstockfish-nodes1
Stockfish Self-Play (nodes=1)
1M games of Stockfish 17 self-play at 1 node per move. Pre-tokenized in the PAWN training format.
At nodes=1, Stockfish evaluates each position with a single NNUE forward pass (no tree search). Despite the lack of search, the NNUE evaluation head produces surprisingly strong play — far from random. Games exhibit coherent openings, reasonable piece development, and tactical awareness, though blunders are more frequent than at higher node counts. This… See the full description on the dataset page: https://huggingface.co/datasets/thomas-schweich/stockfish-nodes1.Chess-Stockfish-BestMove-MessagesChess_Stockfish_BestMove_ExplainStockfish-BestMove-Messages-jsonChess-Stockfish-BestMove-Messages-no_jsonStockfishDepthSixlichess-stockfish-canonical