agi-noobs/chess-sft-20k

Chess SFT Training Dataset

A curated dataset of chess positions with deep Stockfish analysis, designed for supervised fine-tuning (SFT) of language models to play and understand chess.

Dataset Description

This dataset contains chess positions extracted from multiple high-quality sources, each analyzed with Stockfish at depth 20 with MultiPV 3 (top 3 candidate moves). The positions are carefully filtered for quality and diversity across game phases, player skill levels, and position types.

Sources

Source	Description	Proportion
lichess-puzzles	Tactical puzzles from Lichess with clear best moves	30%
lichess-elite	Elite games (2400+ vs 2300+ rated) from nikonoel.fr	25%
lichess-tournament	Competitive tournament games	20%
lichess-standard	Standard rated games across all ELO ranges	25%

Key Features

• Deep analysis: Stockfish depth 20, MultiPV 3
• Rich annotations: 100+ fields including evaluation, move quality, tactics, and positional factors
• Balanced phases: ~10% opening, ~65% middlegame, ~25% endgame (for game sources)
• Quality filtered: Positions selected for interesting decisions
• Diverse skill levels: From novice (<1600) to expert (2400+) play

---

Complete Schema

Identification & Source

Field	Type	Description
id	string	Unique identifier: {source}_{game_id}_{ply}
game_id	string	Original game or puzzle ID
source	string	Data source (lichess-puzzles, lichess-elite, lichess-tournament, lichess-standard)
source_type	string	"puzzle" or "game"

Position Representation

Field	Type	Description
fen	string	FEN notation of the position
board_utf8	string	Unicode board visualization
board_ascii	string	ASCII board visualization
side_to_move	string	"white" or "black"
ply	int	Half-move number in the game
fullmove_number	int	Full move number
game_phase	string	"opening", "middlegame", or "endgame"

Legal Moves

Field	Type	Description
legal_moves_uci	string	Space-separated legal moves in UCI notation
legal_moves_san	string	Space-separated legal moves in SAN notation
num_legal_moves	int	Count of legal moves

Current Position Evaluation

Field	Type	Description
current_cp	int	Centipawn evaluation (+ favors side to move)
current_mate	int/null	Mate distance if forced mate exists
current_eval_desc	string	Human-readable eval ("White is winning +3.5")
current_eval_category	string	"winning", "slightly_better", "equal", "slightly_worse", "losing"

Best Move Analysis

Field	Type	Description
best_move_uci	string	Best move in UCI notation
best_move_san	string	Best move in SAN notation
best_move_cp	int	Evaluation after best move
best_move_mate	int/null	Mate distance after best move
best_move_is_capture	bool	Move captures a piece
best_move_captured_piece	string/null	Name of captured piece
best_move_captured_value	int	Material value of capture
best_move_is_check	bool	Move gives check
best_move_is_checkmate	bool	Move is checkmate
best_move_is_castling	bool	Move is castling
best_move_is_promotion	bool	Move is pawn promotion
best_move_promotion_piece	string/null	Promotion piece type
best_move_piece_moved	string	Type of piece moved
best_move_from	string	Origin square (e.g., "e2")
best_move_to	string	Destination square (e.g., "e4")

Move Purpose Analysis (Why the move is good)

Field	Type	Description
best_move_category	string	Primary purpose: "tactic", "attack", "defense", "positional", "developing", "escape"
best_move_is_sacrifice	bool	Move sacrifices material
best_move_sacrifice_type	string/null	"queen_sac", "rook_sac", "piece_sac", "exchange_sac", "pawn_sac"
best_move_material_sacrificed	int	Material value sacrificed

Threats Created

Field	Type	Description
best_move_creates_threats	bool	Move creates new threats
best_move_num_new_attacks	int	Number of new piece attacks
best_move_total_threat_value	int	Total value of threatened pieces
best_move_attacks_queen	bool	Threatens enemy queen
best_move_attacks_rook	bool	Threatens enemy rook
best_move_attacks_minor	bool	Threatens knight or bishop
best_move_new_attacks_json	string/null	JSON array of attack details
best_move_creates_fork	bool	Creates a fork
best_move_creates_pin	bool	Creates a pin
best_move_pinned_piece	string/null	Pinned piece type
best_move_creates_skewer	bool	Creates a skewer
best_move_is_discovered_attack	bool	Is a discovered attack

Defensive Aspects

Field	Type	Description
position_had_hanging_piece	bool	Position had undefended piece
position_hanging_piece_value	int	Value of hanging piece
best_move_saves_hanging	bool	Move saves the hanging piece
best_move_saved_piece	string/null	Type of saved piece
best_move_saved_piece_value	int	Value of saved piece

Positional Improvements

Field	Type	Description
best_move_is_developing	bool	Develops a piece
best_move_improves_center	bool	Improves center control
best_move_center_control_change	int	Change in center control
best_move_increases_king_pressure	bool	Increases attack on enemy king
best_move_king_attack_change	int	Change in king zone attacks
best_move_creates_passed_pawn	bool	Creates a passed pawn
best_move_advances_passed_pawn	bool	Advances existing passed pawn

Mobility

Field	Type	Description
mobility_before	int	Our legal moves before
mobility_after_opponent	int	Opponent's moves after our move
our_mobility_after	int	Our moves after opponent replies

Evaluation After Best Move

Field	Type	Description
after_best_cp	int	Eval after best move (from opponent's view)
after_best_mate	int/null	Mate distance after best move
eval_change	int	Change in evaluation

Principal Variation

Field	Type	Description
pv_uci	string	Best line in UCI (up to 10 moves)
pv_san	string	Best line in SAN
pv_length	int	Length of principal variation

Alternative Moves (2nd and 3rd Best)

Field	Type	Description
second_move_uci	string/null	2nd best move UCI
second_move_san	string/null	2nd best move SAN
second_move_cp	int/null	2nd best move evaluation
second_move_cp_loss	int/null	Centipawn loss vs best
third_move_uci	string/null	3rd best move UCI
third_move_san	string/null	3rd best move SAN
third_move_cp	int/null	3rd best move evaluation
third_move_cp_loss	int/null	Centipawn loss vs best

All Candidates (JSON)

Field	Type	Description
candidates_json	string	Full JSON array of all analyzed moves with complete data
num_candidates	int	Number of candidates analyzed (typically 3)

Move Quality Distribution

Field	Type	Description
num_good_moves	int	Moves within 30cp of best
num_ok_moves	int	Moves 30-100cp worse
num_mistake_moves	int	Moves 100-200cp worse
num_blunder_moves	int	Moves 200+cp worse

Evaluation Gaps

Field	Type	Description
gap_1_2	int	Centipawn gap: best vs 2nd best
gap_1_3	int	Centipawn gap: best vs 3rd best
gap_1_worst	int	Centipawn gap: best vs worst candidate

Position Type Flags

Field	Type	Description
is_clear_best	bool	Best move is clearly superior (gap > 50cp)
is_only_move	bool	Only one reasonable move (gap > 150cp)
is_critical	bool	Large eval swing based on move choice
is_tactical	bool	Position involves tactics
is_forcing	bool	Few good options, forcing sequence
is_quiet	bool	Calm positional position
is_winning	bool	Side to move is winning
is_losing	bool	Side to move is losing
is_equal	bool	Position is roughly equal
is_sharp	bool	High tension, big eval swings possible

Quality Score

Field	Type	Description
interesting_score	int	Composite score for position quality (higher = more interesting)
filter_reason	string	Why position was kept: "puzzle", "critical", "very_interesting", "interesting", "standard", "boring"

Material Count

Field	Type	Description
material_white	int	White's material value
material_black	int	Black's material value
material_diff	int	Material difference (+ favors white)
material_total	int	Total material on board
white_pawns	int	White pawn count
white_knights	int	White knight count
white_bishops	int	White bishop count
white_rooks	int	White rook count
white_queens	int	White queen count
black_pawns	int	Black pawn count
black_knights	int	Black knight count
black_bishops	int	Black bishop count
black_rooks	int	Black rook count
black_queens	int	Black queen count
has_bishop_pair_white	bool	White has bishop pair
has_bishop_pair_black	bool	Black has bishop pair
is_queenless	bool	Neither side has a queen

King Safety

Field	Type	Description
white_king_castled_short	bool	White castled kingside
white_king_castled_long	bool	White castled queenside
white_king_in_center	bool	White king still in center
white_king_zone_attacks	int	Attacks on white king zone
black_king_castled_short	bool	Black castled kingside
black_king_castled_long	bool	Black castled queenside
black_king_in_center	bool	Black king still in center
black_king_zone_attacks	int	Attacks on black king zone

Checks and Threats

Field	Type	Description
is_check	bool	Side to move is in check
num_checkers	int	Number of checking pieces
num_checking_moves	int	Legal moves that give check
num_capture_moves	int	Legal capturing moves
max_capture_value	int	Highest value capture available

Pawn Structure

Field	Type	Description
white_doubled_pawns	int	White's doubled pawns
black_doubled_pawns	int	Black's doubled pawns
white_passed_pawns	int	White's passed pawns
black_passed_pawns	int	Black's passed pawns

Center Control

Field	Type	Description
white_center_control	int	White's attacks on d4,d5,e4,e5
black_center_control	int	Black's attacks on d4,d5,e4,e5

Board State Flags

Field	Type	Description
is_stalemate	bool	Position is stalemate
is_insufficient_material	bool	Draw by insufficient material
can_claim_draw	bool	Draw can be claimed
halfmove_clock	int	Moves since pawn move/capture
is_fifty_move_possible	bool	50-move rule applicable
board_tension	int	Sum of attacked piece values

Player Ratings

Field	Type	Description
white_elo	int	White player's rating
black_elo	int	Black player's rating
avg_elo	int	Average rating
elo_bucket	string	"novice" (<1600), "club" (1600-2000), "strong" (2000-2400), "expert" (2400+)

Game Metadata

Field	Type	Description
time_control	string	Time control of the game
result	string	Game result (1-0, 0-1, 1/2-1/2)

Analysis Metadata

Field	Type	Description
analysis_depth	int	Stockfish search depth used
analysis_multipv	int	Number of lines analyzed

Puzzle-Specific Fields (only for lichess-puzzles)

Field	Type	Description
puzzle_rating	int	Puzzle difficulty rating
puzzle_rating_deviation	int	Rating uncertainty
puzzle_themes	string	Space-separated puzzle themes
puzzle_solution_uci	string	Puzzle solution moves in UCI
puzzle_popularity	int	Puzzle popularity score
puzzle_nb_plays	int	Number of times puzzle was played

Subset Flags

Field	Type	Description
in_100	bool	Included in 100-sample subset
in_500	bool	Included in 500-sample subset
in_1000	bool	Included in 1000-sample subset
in_2000	bool	Included in 2000-sample subset
...	...	Additional subset flags as available

---

Position Quality Distribution

Positions are categorized by why they were kept:

• puzzle: From the puzzles dataset (tactical training)
• critical: Large evaluation swing based on move choice
• very_interesting: High composite interest score (score >= 30)
• interesting: Good interest score (score >= 15)
• standard: Baseline interesting positions (score >= 0)
• boring: Lower interest but included for diversity

Intended Uses

• SFT for chess LLMs: Train language models to suggest good chess moves
• Move quality prediction: Predict evaluation changes from moves
• Position understanding: Train models to describe chess positions
• Chess tutoring: Generate explanations for why moves are good/bad
• Move purpose classification: Classify moves as tactical/positional/defensive

Example Usage

from datasets import load_dataset

dataset = load_dataset("agi-noobs/chess-sft-20k")

# Get a position
pos = dataset["train"][0]

# Basic info
print(f"FEN: {pos['fen']}")
print(f"Best move: {pos['best_move_san']}")
print(f"Evaluation: {pos['current_cp']} cp")

# Move characteristics
print(f"Is tactical: {pos['is_tactical']}")
print(f"Is sacrifice: {pos['best_move_is_sacrifice']}")
print(f"Creates fork: {pos['best_move_creates_fork']}")

# Position insights
print(f"Material diff: {pos['material_diff']}")
print(f"Board tension: {pos['board_tension']}")

# Alternative moves
print(f"2nd best: {pos['second_move_san']} ({pos['second_move_cp_loss']}cp worse)")

Generation Details

• Stockfish version: Latest (via Homebrew)
• Analysis depth: 20 ply
• MultiPV: 3 (top 3 moves analyzed)
• Hash: 128 MB per instance
• Workers: 8 parallel Stockfish instances

Limitations

• Evaluations are Stockfish's assessment, not ground truth
• Some positions may have multiple equally good moves not captured by MultiPV 3
• Puzzle positions are pre-solution (the position before the winning tactic)
• ELO ratings are from Lichess, which may differ from FIDE ratings

License

MIT License

Citation

@dataset{chess_sft_2024,
  title={Chess SFT Training Dataset},
  author={AGI Noobs},
  year={2024},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/agi-noobs/chess-sft-20k}
}

Acknowledgments

• Lichess for the open game and puzzle databases
• nikonoel.fr for the curated elite games database
• Stockfish for the chess engine