Gimmenoto/Warinza008

---
language:
- en
license: other
task_categories:
- reinforcement-learning
- other
tags:
- chess
- board-game
- deep-learning
- neural-network
- lichess
- tensor-dataset
pretty_name: Lichess Elite Chess Tensor Dataset (2013–May 2020)
size_categories:
- 10M<n<700M
---

# ♟️ Lichess Elite Chess Tensor Dataset (2013 – May 2020)

## Overview

This dataset contains high-dimensional chess board representations extracted exclusively from the **Lichess Elite Database**, covering games from **January 2013 to May 2020**.

Each chess position is encoded as a structured **125 × 8 × 8 float tensor**, designed for deep learning, policy learning, and reinforcement learning research.

This dataset was prepared strictly for **educational and academic research purposes only**, under the Department of Computer Science,
King Mongkut’s University of Technology Thonburi (KMUTT), Thailand.

---

# Data Source

Original dataset:

- Lichess Elite Database
https://database.nikonoel.fr/

Time range used:

- January 2013 – May 2020

No data outside this range is included.

This dataset is not affiliated with Lichess. All rights remain with the original data provider.

---

# Dataset Structure

## Split

| Split | Description |
|-------|------------|
| train | All extracted board positions |

---

# Features

| Column | Type | Description |
|--------|------|-------------|
| board | list<list<list<float32>>> | 125-channel 8×8 tensor |
| win | bool | True if game result is 1-0 |
| draw | bool | True if game result is 1/2-1/2 |
| lose | bool | True if game result is 0-1 |
| piece_to_move | int64 | Encoded move target (from_square × 64 + to_square) |

---

# Board Tensor Specification (125 Channels)

The tensor is composed of structured feature groups described below.

---

## 1. Piece Planes (14 channels)

Binary bitboards (1.0 where piece exists, else 0.0).

White pieces:
- Pawn
- Knight
- Bishop
- Rook
- Queen
- King

Black pieces:
- Pawn
- Knight
- Bishop
- Rook
- Queen
- King

Additional duplicated bishop planes:
- White bishops
- Black bishops

Total: 14 channels

---

## 2. Castling Rights (4 channels)

Each channel is a full 8×8 plane filled with 1.0 if the right exists.

- White kingside
- White queenside
- Black kingside
- Black queenside

---

## 3. En Passant (2 channels)

- En passant target square mask
- En passant capture pawn square

---

## 4. Game State Scalars (4 channels)

Each is broadcasted over the entire 8×8 plane.

- Side to move (1.0 = White, 0.0 = Black)
- Halfmove clock normalized (halfmove_clock / 100)
- Repetition ≥ 2
- Repetition ≥ 3

---

## 5. Tactical Information (6 channels)

- Checkers mask
- Squares attacked by White
- Squares attacked by Black
- White pinned pieces
- Black pinned pieces
- Threatened pieces (current side pieces under attack)

---

## 6. King Mobility (2 channels)

- White king legal moves
- Black king legal moves

---

## 7. Pawn Structure (5 channels)

- White passed pawns
- Black passed pawns
- Isolated pawns
- Doubled pawns
- Structural pressure proxy

---

## 8. History Planes (84 channels)

Last 6 board positions × 14 piece channels each.

Order:
- Most recent position first

Each historical board encodes:
- 6 white piece planes
- 6 black piece planes
- 2 bishop duplicate planes

---

## 9. Move Masks and Evaluation (4 channels)

- Legal move destination squares
- Capture squares
- Promotion squares
- Mobility scalar (legal_moves / 218)

---

## 10. Material Balance (1 channel)

Normalized material difference:

Piece values:
- Pawn = 1
- Knight = 3
- Bishop = 3
- Rook = 5
- Queen = 9

---

# Channel Summary

| Category | Channels |
|----------|----------|
| Pieces | 14 |
| Castling | 4 |
| En passant | 2 |
| Game state | 4 |
| Tactical info | 6 |
| King mobility | 2 |
| Pawn structure | 5 |
| History | 84 |
| Move masks + mobility | 4 |
| Material balance | 1 |
| **TOTAL** | **125** |

---

# Intended Use

This dataset is suitable for:

- Policy prediction (move selection)
- WDL classification (win/draw/lose)
- Value networks
- AlphaZero-style supervised pretraining
- Position evaluation modeling
- Tactical learning

---

# Example Usage (PyTorch)

```python
from datasets import load_dataset
import torch

dataset = load_dataset("your-username/lichess-elite-125ch")

sample = dataset["train"][0]

board_tensor = torch.tensor(sample["board"]) # shape [125, 8, 8]
move_target = sample["piece_to_move"]