README.md

---
license: cc-by-nc-4.0
task_categories:
- reinforcement-learning
- tabular-classification
language:
- en
tags:
- chess
- gambitflow
- big-data
- elite
- sqlite
size_categories:
- 1M<n<10M
pretty_name: GambitFlow Elite Training Data
---

# 📚 GambitFlow Elite Training Data

<div align="center">

![Dataset Banner](https://capsule-render.vercel.app/api?type=waving&color=0:27ae60,100:2c3e50&height=200&section=header&text=Elite%20Training%20Data&fontSize=50&animation=fadeIn&fontAlignY=35&desc=5%20Million%2B%20Master%20Level%20Positions%20(ELO%202000%2B)&descAlignY=60)

[![License: CC BY-NC 4.0](https://img.shields.io/badge/License-CC%20BY--NC%204.0-lightgrey.svg)](https://creativecommons.org/licenses/by-nc/4.0/)
![Format](https://img.shields.io/badge/Format-SQLite3-green)
![Volume](https://img.shields.io/badge/Size-882MB-blue)
![Quality](https://img.shields.io/badge/Quality-Filtered%202000%2B%20ELO-red)

[**View on GitHub**](https://github.com/GambitFlow/GambitFlow) • [**Source Model: Nexus-core CE**](https://huggingface.co/GambitFlow/gambitflow-nexus-core)

</div>

## 📖 Dataset Description

This dataset is the highly curated input required to train **strong, club-level chess evaluation models** like the **Nexus-core CE**. It is designed to maximize the signal-to-noise ratio in chess data by removing moves made by lower-rated players.

By exclusively training on **Elite-level games**, the resulting AI avoids learning common amateur mistakes and focuses on solid positional principles.

## 🛠️ Data Engineering & Filtering

The database was created through a multi-stage, streaming pipeline to handle the massive volume efficiently without memory overflow.

1.  **Source:** Lichess Public Database (January 2017).
2.  **CRITICAL FILTER:** Only games where **White ELO > 2000 AND Black ELO > 2000** were accepted.
3.  **Extraction:** Positions (FENs) were extracted only up to the first **20 moves** of each filtered game (the Opening/Early Middlegame phase).
4.  **Optimization:** The data was aggregated by unique FEN and stored in a compressed **SQLite** file.

- **Final Volume:** Over **5,000,000 Total Positions** processed, resulting in **2,488,753 Unique Positions**.
- **File Size:** **882 MB**.

## 📂 File Structure & Schema

The main file is `chess_stats_v2.db`.

### Table: `positions`

| Column | Type | Description |
|--------|------|-------------|
| `fen` | **TEXT (Primary Key)** | The board position. **Truncated to 4 parts** (Position, Turn, Castling, En Passant) for maximum data aggregation across transpositions. |
| `stats` | **TEXT (JSON)** | JSON string containing aggregated move counts and game outcomes (W/D/L) for subsequent training. |

## 🚀 Usage (Model Training)

This database is meant to be read by the **`SQLiteIterableDataset`** class in PyTorch, ensuring only small batches of data are streamed at a time, preventing RAM crashes even with large datasets.

## ⚠️ License

This dataset is licensed under **CC BY-NC 4.0**.
It is a derivative work of the Lichess Open Database (CC0).
Commercial use is strictly prohibited.

---
<div align="center">
  <p>Curated by <a href="https://github.com/GambitFlow">GambitFlow</a></p>
</div>