README.md

metadata

license: cc0-1.0

GambitFlow Opening Database

A curated, high-quality chess opening theory database in SQLite format, designed for chess engines and analysis tools. This dataset contains aggregated move statistics derived from thousands of recent, high-rated (2600+ ELO) grandmaster-level games.

The goal of this database is to provide a powerful, statistically-backed foundation for understanding modern opening theory, without relying on traditional, human-annotated opening books.

How to Use

The database is a single SQLite file (opening_theory.db). You can download it directly from the "Files" tab or use the huggingface_hub library to access it programmatically.

Here is a Python example to query the database for the starting position:

import sqlite3
import json
from huggingface_hub import hf_hub_download

# 1. Download the database file
db_path = hf_hub_download(
    repo_id="GambitFlow/Opening-Database",
    filename="opening_theory.db",
    repo_type="dataset"
)

# 2. Connect to the database
conn = sqlite3.connect(db_path)
cursor = conn.cursor()

# 3. Query a position (FEN)
# The FEN should be "canonical" (position, turn, castling, en passant)
start_fen = "rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq -"

cursor.execute("SELECT move_data FROM openings WHERE fen = ?", (start_fen,))
row = cursor.fetchone()

if row:
    # 4. Parse the JSON data
    moves_data = json.loads(row)
    
    # Sort moves by frequency
    sorted_moves = sorted(
        moves_data.items(), 
        key=lambda item: item['frequency'], 
        reverse=True
    )
    
    print(f"Top moves for FEN: {start_fen}\n")
    for move, data in sorted_moves[:5]:
        print(f"- Move: {move}")
        print(f"  Frequency: {data['frequency']:,}")
        print(f"  Avg Score: {data.get('avg_score', 0.0):.3f}\n")

conn.close()

Data Collection and Processing

This dataset was meticulously created through a multi-stage pipeline to ensure the highest quality and statistical relevance.

1. Data Source: The foundation of this dataset is the Lichess Elite Database, a filtered collection of games from high-rated players, originally curated by Nikonoel. We used multiple monthly PGN files from the 2023-2024 period to capture modern theory.

2. Filtering Criteria: Only games meeting the following strict criteria were processed:

• Minimum ELO: Both players had a rating of 2600 or higher.
• Maximum Depth: Only the first 25 moves of each game were analyzed to focus on opening theory.

3. Processing Pipeline:

• Distributed Processing: The PGN files were processed in parallel across multiple sessions to handle the large volume of games efficiently.
• Perspective-Aware Scoring: Each move's quality was scored from the perspective of the player whose turn it was. A win scored 1.0, a loss 0.0, and a draw 0.5.
• Aggregation: For each unique board position (FEN), statistics for every move played were aggregated, including frequency and the sum of scores.
• Merging: The databases from the distributed sessions were merged into a single master file. This process recalculated weighted averages for ELO and move scores, ensuring statistical accuracy.

Dataset Structure

The database contains a single table named openings.

File: opening_theory.db Table: openings

Column	Type	Description
fen	TEXT	The canonical board position (FEN string, Primary Key).
move_data	TEXT	A JSON object containing statistics for each move played.
total_games	INTEGER	The total number of times this position was seen.
avg_elo	INTEGER	The average ELO of players who reached this position.

move_data JSON Structure

The move_data column contains a JSON string with the following structure:

{
  "e4": {
    "frequency": 153944,
    "score_sum": 76972.0,
    "avg_score": 0.5
  },
  "d4": {
    "frequency": 65604,
    "score_sum": 32802.0,
    "avg_score": 0.5
  }
}

• frequency: The number of times this move was played from the given position.
• score_sum: The sum of all perspective-aware scores for this move.
• avg_score: The average score, calculated as score_sum / frequency.

Citation

If you use this dataset in your work, please consider citing it:

@dataset{gambitflow_opening_database_2024,
  author       = {GambitFlow Project},
  title        = {GambitFlow Opening Database},
  year         = {2024},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/GambitFlow/Opening-Database}
}