Dataset Overview

Dataset Name

Dataset Name

Source

The dataset’s game sequences are sourced from https://database.lichess.org/.

Content

The dataset contains chess board states and their associated move sequences. PGN files were downloaded from the Lichess database; each game is parsed into multiple records, each beginning with a FEN string followed by 1‑10 SAN moves.

Intended Use

The data will be used to train the ChessRoberta model, but because they are unfiltered they may not be suitable for building a high‑performance chess engine.

Structure

Each record includes the following fields:

• text (Str): String containing the FEN and multiple SAN moves.
• pgn_start (Int): Index of the first SAN within the text string.
• num_sans (Int): Number of half‑moves present in the text string.
• num_prior_moves (Int): Number of half‑moves that occurred before the FEN (one move for each side counts as two moves).
• game_id (Str): Lichess identifier for the source game.

Special markers are used as delimiters in the text field:

• PGN_START: "~"
• MOVE_SEP: ">"

Example Record

{
  text: "rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq - 0 1~e4>e6>d3>d5>Nd2",
  pgn_start: 57,
  num_sans: 5,
  num_prior_moves: 0,
  game_id: "https://lichess.org/PwE2cWn3"
}