♟️ GambitFlow Synapse-Edge-Data (Massive Corpus)

Associated Model • Source: Lichess

📖 Dataset Overview

Synapse-Edge-Data is a high-performance, large-scale dataset specifically engineered for training Gen-3 chess engines like Synapse-Edge v1. It represents a revolutionary fusion of high-level human positional play and rigorous tactical patterns.

Unlike previous datasets that focus solely on game outcomes, this corpus includes 3 million+ tactical puzzles, enabling models to learn "the art of the kill" alongside "the art of strategy."

---

🏗️ Dataset Structure

The dataset is delivered as a sharded SQLite system, optimized for Distributed Training across multiple machines or accounts.

SQLite Schema (training_data table)

Column	Type	Description
fen	TEXT (PK)	The Forsyth-Edwards Notation of the board state (Position).
position_stats	TEXT (JSON)	Detailed win/loss/draw statistics from master games (if available).
best_move	TEXT	The calculated optimal UCI move for the position.
is_tactical	INTEGER	Flag: 1 for tactical puzzles, 0 for positional master games.
difficulty	TEXT	Skill classification: beginner, intermediate, or advanced.

Data Distribution (Shards)

To facilitate training on hardware like Google Colab, the dataset is split into 4 shards:

• shards/synapse_shard_1.db (~1.38M positions)
• shards/synapse_shard_2.db (~1.38M positions)
• shards/synapse_shard_3.db (~1.38M positions)
• shards/synapse_shard_4.db (~1.38M positions)

---

🛠️ Dataset Creation

Curation Rationale

Most open-source chess datasets are either purely positional (games) or purely tactical (puzzles). Synapse-Edge-Data bridges this gap. By combining 5.5M elite positions with 3M tactical patterns, we ensure that trained models don't just "play well" but can also find clinical tactical solutions under pressure.

Source Data

1. Elite Positional Data: Extracted from Lichess rated games where both players had an ELO > 2000.
2. Tactical Database: Processed from the official Lichess Puzzle Database, filtering for themes like Forks, Pins, Skewers, Discovered Attacks, and Mate-in-X.

Data Processing Pipeline

1. Decompression: Raw .pgn.zst and .csv.zst files were streamed and decoded.
2. Cleaning: Duplicate FENs were removed to prevent overfitting.
3. Normalization: FENs were truncated to 4 fields (removing half-move clock) for higher uniqueness.
4. Best Move Calculation: Positional data was enriched by selecting the move with the highest success rate in master games.

---

🚀 Use Cases

Direct Use

• Training Chess Engines: Specifically designed for Multi-Head Neural Networks (Policy/Value/Tactical heads).
• Game Analysis: Can be used to train models that detect "blunders" or "sharp" positions.

Out-of-Scope Use

• Training for variants like Chess960 or Antichess (this dataset is for Standard Chess only).
• Non-chess board games.

---

🛡️ Bias, Risks, and Limitations

• ELO Bias: Since the data is sourced from players > 2000 ELO, the model may struggle to exploit beginner-level "random" blunders effectively without further fine-tuning.
• Engine Bias: Some "best moves" in the positional set reflect human theory, which might differ from pure engine evaluations at extreme depths.

📜 License

This dataset is published under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license. It is intended for research and educational purposes. Commercial use requires prior authorization.

---

Maintained by Rafsan @ GambitFlow Labs