---
tags:
- chess
- embeddings
- puzzles
- computer-vision
- pytorch
- tabular
license: other # TODO: Specify the license (e.g., apache-2.0, cc-by-sa-4.0)
language:
- en
- zxx # For FEN, Moves, Embeddings
pretty_name: "Lichess Puzzle Embeddings (ChessLM Encoder v4)"
dataset_info:
  features:
    - name: FEN
      dtype: string
    - name: Moves
      dtype: string
    - name: Themes
      dtype: string
    - name: Embedding
      dtype: sequence # Sequence of floats
      # For older datasets versions: dtype: list[float32]
      # Specify embedding dimension if known, e.g., sequence: float32[256]
      # The exact dimension depends on the d_model of the loaded encoder.
      # Assuming d_model=256 based on typical defaults.
      # sequence: float32[256]
---

# Lichess Puzzle Embeddings (ChessLM Encoder v4)

## Dataset Description

This dataset contains pre-computed vector embeddings for chess puzzle positions sourced from the Lichess Open Puzzle Database. The embeddings were generated using the odestorm1/chesslm (https://huggingface.co/datasets/odestorm1/chesslm_puzzles) Encoder Transformer model.

Each row corresponds to a unique chess puzzle, providing its initial FEN (Forsyth–Edwards Notation) string, the sequence of moves in the puzzle solution, associated themes, and the generated embedding vector for the starting position.

The dataset was created by sampling up to 1,000,000 puzzles from the Lichess Database (https://database.lichess.org/) and processing the starting FEN position of each puzzle through the pre-trained encoder model. The embedding represents the state of the board *before* the first puzzle move is made.

More details about the model can be found at https://github.com/bluehood/Encoder-ChessLM and the technical writeup https://bluehood.github.io/research/benh_Beyond_Evaluation__Learning_Contextual_Chess_Position_Representations_2025.pdf. 

## Supported Tasks and Leaderboards

This dataset can be used for various tasks related to chess puzzle analysis and understanding, including:

*   **Puzzle Similarity Search:** Find puzzles with similar starting positions based on embedding distance.
*   **Theme Prediction/Clustering:** Analyze if embeddings cluster according to puzzle themes.
*   **Difficulty Prediction:** Use embeddings as features to predict puzzle ratings or solve rates.
*   **Downstream Model Training:** Use embeddings as input features for models tackling other chess-related tasks.

## Languages

The data primarily uses:
*   **English (`en`)**: For the `Themes` column.
*   **No linguistic content (`zxx`)**: For `FEN`, `Moves`, and `Embedding` columns, which represent chess notation or numerical data.

## Dataset Structure

### Data Instances

A typical row in the dataset looks like this (embedding truncated for brevity):

```json
{
  "FEN": "r1b1k2r/1p1n1ppp/pq2p3/3pP3/1P1N4/P1N5/2PQ1PPP/R3K2R w KQkq - 0 14",
  "Moves": "f2f4 b6d4 d2d4 f8c5",
  "Themes": "crushing deflection middlegame short",
  "Embedding": [-0.0123, 0.4567, -0.7890, ..., 0.1122]
}
```

### Data Fields

*   `FEN` (string): The Forsyth–Edwards Notation string representing the starting position of the puzzle.
*   `Moves` (string): A space-separated string of moves representing the puzzle's solution (player moves and opponent responses).
*   `Themes` (string): A space-separated string of themes associated with the puzzle (e.g., "mateIn2", "fork", "endgame").
*   `Embedding` (sequence/list of float32): The vector embedding generated by the `ChessLM Encoder v4` for the starting FEN position. The dimension corresponds to the `d_model` of the encoder (e.g., 256).


## Dataset Creation

The primary source data is the Lichess Open Puzzle Database. This file contains information about millions of chess puzzles generated from Lichess games.

### Annotations

The `FEN`, `Moves`, and `Themes` columns are taken directly from the source Lichess Database. The `Embedding` column is generated by processing the `FEN` string through the ChessLM model.

### Citation Information

If you use this dataset or the underlying model methodology, please cite:

```bibtex
@misc{hull2025beyond,
      title={Beyond Evaluation: Learning Contextual Chess Position Representations},
      author={Ben Hull},
      year={2025},
      howpublished={Accessed via \url{[https://bluehood.github.io/](https://bluehood.github.io/)}},
      note={Technical report}
}
```

### Contributions

Thanks to the Lichess team for providing the open puzzle database.

## How to Use

You can load the dataset using the Hugging Face `datasets` library or `pandas`.

**Using `datasets`:**

```python
from datasets import load_dataset

# Load from Hugging Face Hub
dataset = load_dataset("odestorm1/chesslm_puzzles")

# Load local parquet file
dataset = load_dataset("parquet", data_files="puzzle_embeddings.parquet")
```

**Using `pandas`:**

```python
import pandas as pd

df = pd.read_parquet("puzzle_embeddings.parquet")
print(df.head())

# Access the embedding (list of floats) for the first puzzle
first_embedding = df.iloc[0]['Embedding']
print(f"Embedding dimension: {len(first_embedding)}")
```