TannerGladson/chess-roberta-pretraining

Dataset Card for Dataset Name

A collection of move sequences from chess games found at https://database.lichess.org/

Dataset Details

A collection of chess board states and the associated move sequences. PGN files have been downloaded from https://database.lichess.org/. Each game has been parsed into multiple records, where each record begins with an FEN and is followed by 1 to 10 SANs.

Uses

Will be used to train ChessRoberta. This should likely not be used to create a high performing chess model. These games have not been filtered.

Dataset Structure

Each line is a record, and each record has the following fields:

• text: (Str) the aforementioned string where an FEN and several SANs have been concatentated
• pgn_start: (Int) the index of the first character of the first SAN (in the text field's string)
• num_sans: (Int) the number of half-moves contained in the text field's string
• num_prior_moves (Int) the number of half-moves which elapsed before this record's FEN (remark: black and white each moving once is two elapsed moves)
• game_id (Str) the lichess identifier where the game was taken from

Within the text field, special tokens are used as seperators

SPECIAL_TOKENS = {
  "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"
}