smallchess/kaggle-chess-positions

Kaggle Chess Positions (mirror)

Public mirror of the Chess Positions dataset by koryakinp on Kaggle, hosted for easier access in ML pipelines (including 2d-chess-ocr).

Source & attribution

Field	Value
Original dataset	koryakinp/chess-positions
Original author	koryakinp
License	CC0-1.0
Generation tool	koryakinp/chess-generator
Board/piece assets	Chess.com

This repository is a mirror of the Kaggle release. Please cite the original Kaggle dataset and respect the CC0-1.0 license.

Dataset summary

• 100,000 synthetic chess board images, 400×400 pixels
• Positions contain 5–15 pieces (always includes both kings)
• 28 board styles × 32 piece styles = 896 visual combinations
• Training set: 80,000 images (train/)
• Test set: 20,000 images (test/)

Piece sampling probabilities in the original generator:

• Pawn 30%, Bishop 20%, Knight 20%, Rook 20%, Queen 10%, plus both kings

Some positions may be chess-illegal (e.g. both kings in check). Treat labels accordingly.

Label format (FEN in filename)

Each image filename encodes the piece-placement portion of FEN, with dashes between ranks instead of slashes:

1B1B1K2-3p1N2-6k1-R7-5P2-4q3-7R-1B6.jpeg

• Uppercase = White pieces (PNBRQK)
• Lowercase = Black pieces (pnbrqk)
• Digits 1–8 = empty squares on a rank

There is no separate label file; parse the filename to recover the 8×8 board.

Usage

Load with Hugging Face Datasets

from datasets import load_dataset

ds = load_dataset("smallchess/kaggle-chess-positions")
# ds["train"], ds["test"]

Use with 2d-chess-ocr (piece classifier training)

In the 2d-chess-ocr dataset pipeline, Step 2 extracts 64×64 piece tiles from these board images (labels come from the filename):

# after downloading this dataset locally
uv run -m dataset_creation.chess_piece_sets.create_piece_dataset \
    --source-dir /path/to/kaggle-chess-positions \
    --output-dir datasets/pieces-dataset \
    --tile-size 64

Point --source-dir at the folder containing train/ and test/.

Download from the Hub

huggingface-cli download smallchess/kaggle-chess-positions --repo-type dataset --local-dir ./kaggle-chess-positions

Repository layout

train/   # 80,000 JPEG board images
test/    # 20,000 JPEG board images

Citation

If you use this data, please credit the original Kaggle dataset:

@misc{koryakinp_chess_positions,
  author       = {koryakinp},
  title        = {Chess Positions},
  year         = {2019},
  publisher    = {Kaggle},
  howpublished = {\url{https://www.kaggle.com/datasets/koryakinp/chess-positions}}
}