GambitFlow/Elite-Data

📚 GambitFlow Elite Training Data (Unified)

View on GitHub • Target Models: Nexus-Core & Synapse-Base

📖 Dataset Overview

This repository hosts the foundational knowledge bases for the GambitFlow chess engines. It consolidates two distinct, powerful datasets:

1. chess_stats_v2.db: The original, large-scale dataset used to train the Nexus-Core engine.
2. match_positions_v2.db: A new, ultra-high-quality dataset specifically curated for the next-generation Synapse-Base engine.

Together, they provide a comprehensive training resource covering different eras of chess theory and rating levels.

---

💎 Dataset 1: Synapse-Base Match Data (match_positions_v2.db)

This is the newly added, highly-focused dataset designed to teach Synapse-Base advanced middlegame strategy and endgame technique. It prioritizes quality over quantity.

Data Engineering & Filtering

• Source: Lichess Elite Database (2024-2025 monthly archives).
• Critical Filters:
  • Player Rating: Both players must have an ELO of 2400 or higher.
  • Game Phase: Skips the first 10 moves of every game to focus on non-theoretical positions.
  • Position Selection: An intelligent filtering algorithm was used to select only "interesting" positions (e.g., positions with material imbalance, tactical complexity, or critical endgame structures).
• Final Volume: A dense collection of approximately 3,000,000 strategically rich positions.

Schema: positions table

Column	Type	Description
fen	TEXT	The board position (FEN).
phase	TEXT	'midgame' or 'endgame'.
value_target	REAL	The game's outcome scored from -1.0 (loss) to 1.0 (win) from the current player's perspective.
move_played	TEXT	The move played by the 2400+ ELO human in that position.
avg_elo	INTEGER	The average rating of the two players.

---

🕰️ Dataset 2: Nexus-Core Legacy Data (chess_stats_v2.db)

This is the original, large-scale dataset that powered the Nexus-Core engine. It provides a broad foundation of solid, club-level chess knowledge.

Data Engineering & Filtering

• Source: Lichess Public Database (January 2017).
• Critical Filter: Only games where both players had an ELO greater than 2000 were accepted.
• Extraction: Positions were extracted up to the first 20 moves (Opening/Early Middlegame).
• Final Volume: Over 5,000,000 total positions processed, resulting in 2,488,753 unique positions.
• File Size: 882 MB.

Schema: positions table

Column	Type	Description
fen	TEXT (PK)	The board position, truncated to 4 parts (Position, Turn, Castling, En Passant).
stats	TEXT (JSON)	A JSON string containing aggregated move counts and game outcomes (Win/Draw/Loss).

---

🚀 Usage Example (Python)

This example shows how to load and sample the new Synapse-Base data.

import sqlite3
from huggingface_hub import hf_hub_download

# Download the new Match Data
db_path = hf_hub_download(
    repo_id="GambitFlow/Elite-Data",
    filename="match_positions_v2.db",
    repo_type="dataset"
)

# Connect and sample data
conn = sqlite3.connect(db_path)
cursor = conn.cursor()

# Get 5 random middlegame positions
cursor.execute("SELECT fen, move_played, value_target FROM positions WHERE phase='midgame' ORDER BY RANDOM() LIMIT 5")

for row in cursor.fetchall():
    print(f"FEN: {row}")
    print(f"Grandmaster Move: {row} | Outcome Score: {row}")
    print("-" * 30)

conn.close()

---