OMLCheT/chess-san-base

♟️ OMLCheT Chess SAN Base Dataset

Welcome to the official dataset for OMLCheT (Open Machine Learning Chess Tournament)! This dataset provides curated chess game sequences formatted explicitly for training lightweight Language Models (LMs) and Reinforcement Learning (RL) agents using pure Standard Algebraic Notation (SAN).

The source games are derived from high-quality matchups streamed directly from the Lichess/tournament-chess-games repository, with metadata, timestamps, clock evaluation markers, and move numbers stripped away.

---

🗂️ Dataset Configurations

To accommodate different training methodologies, this repository is split into two distinct subsets. Both configurations feature an independent Train / Test split.

1. clean

Contains raw, strictly legal chess move strings. Best for basic text-generation models or reinforcement learning setups that require error-free syntactic move histories.

• Example: e4 e5 Nf3 Nc6 Bc4 Nf6 d3 Bc5 O-O d6 c3 a6

2. annotated

Preserves human evaluation glyphs (?!, ?, !!) and Numeric Annotation Glyphs (NAGs like $14, $19). This allows advanced models to try and learn contextual tactical evaluation (e.g., recognizing when a move is structurally flagged as a mistake or an advantage).

• Example: Nf1?! d5 c4 e6 Nc3 Nf6 Nf3 Be7 Bf4 O-O e3 c5 Bg5 #5 $14

---

🚀 Quick Start / How to Use

You can load either configuration instantly via the Hugging Face datasets library:

from datasets import load_dataset

# Load the strictly pure move data
dataset_pure = load_dataset("OMLCheT/chess-san-base", "clean")

# Load the text with evaluation annotations
dataset_eval = load_dataset("OMLCheT/chess-san-base", "annotated")

# Inspect a training example
print(dataset_pure["train"][0]["text"])

📊 Dataset Splits (Exact Counts)

Configuration	Train Split Rows	Test Split Rows	Total Rows
🧠 annotated	89,976	4,736	94,712
🤖 clean	82,270	4,330	86,600

(Note: The 'clean' config has slightly fewer rows due to a small number of short games or heavily annotated draw scripts filtering out completely during strict parsing).