Data set created and used for the project on https://github.com/fenilgmehta/Chess-Force

• The PGN file have been taken from the following websites:

  • https://www.kingbase-chess.net/
  • http://utzingerk.com/test_kaufman
  • Kaufman's 25 boards postions: https://www.chessprogramming.org/Kaufman_Test

• One may either download files from 03_csv_score_data or 03_csv_score_data_compressed. The only difference is that 03_csv_score_data contains raw CSV files, whereas 03_csv_score_data_compressed contain the same files in compressed ZIP format.

• File/Folder	Size
  03_csv_score_data	1.68 GB
  03_csv_score_data_compressed	0.6 GB

• PGN source	Game count
  KingBase2019-pgn/KingBase2019-A00-A39.pgn	58968
  KingBase2019-pgn/KingBase2019-A40-A79.pgn	35706
  KingBase2019-pgn/KingBase2019-A80-A99.pgn	7020
  KingBase2019-pgn/KingBase2019-B00-B19.pgn	38419
  KingBase2019-pgn/KingBase2019-B20-B49.pgn	44497
  KingBase2019-pgn/KingBase2019-B50-B99.pgn	29985
  KingBase2019-pgn/KingBase2019-C00-C19.pgn	23555
  KingBase2019-pgn/KingBase2019-C20-C59.pgn	25927
  KingBase2019-pgn/KingBase2019-C60-C99.pgn	20437
  KingBase2019-pgn/KingBase2019-D00-D29.pgn	31801
  KingBase2019-pgn/KingBase2019-D30-D69.pgn	25397
  KingBase2019-pgn/KingBase2019-D70-D99.pgn	19695
  KingBase2019-pgn/KingBase2019-E00-E19.pgn	24082
  KingBase2019-pgn/KingBase2019-E20-E59.pgn	12034
  KingBase2019-pgn/KingBase2019-E60-E99.pgn	22579
  Kaufman games	25

  ∴ Total chess games = 420127

• Stockfish 10 was used for the score generation

• Stockfish 10 configuration:

  • Max CPU cores/threads = 1
  • Max Hash table size = 16 MB
  • Max search depth = 20
  • Max thinking time = 1 second(for KingBase), 2 seconds(for Kaufman)

• The output of Stockfish 10 is normalized using the method described in the paper Predicting Chess Moves with Multilayer Perceptron and Limited Lookahead under the section III Methodology - A. Dataset, centipawn score generation and normalization, and board representation, and then it is divided by 𝑚𝑎𝑥_𝑠𝑐𝑜𝑟𝑒 so that the score is between -1 and 1.

  • Formula used if it is a checkmate 𝑐𝑝 = 𝑚𝑎𝑥_𝑠𝑐𝑜𝑟𝑒 − 𝑚𝑎𝑡𝑒_𝑠𝑡𝑒𝑝𝑠 ∗ 𝑚𝑖𝑛_𝑑𝑖𝑓 ∗ 𝑠𝑖𝑔𝑛
    • 𝑐𝑝 = centi-pawn score
    • 𝑚𝑎𝑥_𝑠𝑐𝑜𝑟𝑒 = 10,000
    • 𝑚𝑎𝑡𝑒_𝑠𝑡𝑒𝑝𝑠 = it’s the absolute value of the number of steps in which one of the sides can be checkmated
    • 𝑚𝑖𝑛_𝑑𝑖𝑓 = 50
    • 𝑠𝑖𝑔𝑛 = +1 if black is getting checkmated, -1, if white is getting checkmated
  • If it is not a checkmate, then the score is directly used and divided by 𝑚𝑎𝑥_𝑠𝑐𝑜𝑟𝑒.

• The CSV files are in the following format:

  • First line is fen_board,cp_score

  • All line after that are: FEN notation of the chess boards,centipawn-score

  • The first few lines of 03_csv_score_data/z__r_A00-A39__aa.csv are:

    fen_board	cp_score
    2r5/p3prb1/6kp/1PNp2p1/P2P4/6PP/1B2RPK1/8 b - - 0 33	-0.0187
    5r1k/pQ4pp/3p2b1/8/2P3q1/1PN3P1/P5BP/4R2K w - - 1 31	0.0824
    rnbq1rk1/ppp1bppp/4pn2/3p4/8/3P1NP1/PPPNPPBP/R1BQ1RK1 b - - 2 6	-0.0059
    rnbq1rk1/ppp1npbp/6p1/4p3/1PPp4/3P1NP1/P2NPPBP/R1BQ1RK1 b - - 2 8	0.0029
    8/8/1B4k1/6p1/5p1p/8/8/6K1 b - - 2 74	-0.0825