Go game archive =============== This archive contains over 600,000 games of Go, collected by Nicol N. Schraudolph from a variety of sources, for machine learning purposes. At the top level you you should see the following files & directories: README - this file tools - software utilities sgf - the raw game data in "Smart-Go" format, filed by source & date olf - the same data, cleaned up & reorganized, in "one-line" format The tools directory ------------------- sgf2olf - (Gnu) awk script to translate from Smart-Go to one-line format olfilt - shell script for filtering pathological games out of olf files bzip2 - (de)compression utility for the sgf files The sgf directory ----------------- pro - transcribed professional games igs - games collected from the Internet Go Server (IGS) nngs - games collected from the No Name Go Server (NNGS) The igs/92 & igs/93 subdirectories contain over 8,200 games between players ranked 4th dan or higher (that is, professionals and strong amateurs) played on the IGS between 10/92 and 10/93. (Note that some of the earliest games use various illegitimate ways to record a pass; the "sgf2olf" script in the tools directory can handle these.) The igs/95 and igs/96 subdirectories contain *all* games played on the IGS between 01/95 and 03/96 - including abandoned/aborted games, games between rank beginners, games with/between computer programs, games on non-standard board sizes, etc. (The following players are known to me to be computer programs: ManyFaces, Explorer, Perception, Golife, and Poka.) From 04/93 onwards, all IGS games contain the copyright notice: Copyright 1993 Tim L. Casey Permission to reproduce this game is given, provided proper credit is given. No warrantee, implied or explicit, is understood. Use of this game is an understanding and agreement of this notice. The earlier games lack such a notice, but since they come from the same source they should probably be understood implicitly as being subject to the same conditions. The copyright holder is given as "Tim L. Casey" until 05/96, and "I.Net Technologies, Inc." thereafter. The olf directory ----------------- This directory contains all non-pathological games from the sgf directory, but in a more compact representation that uses just one line per game, containing the following space-separated fields (in this order): 1) date (format: "yyyy-mm-dd"); 2) name of black player; 3) rank/rating of black player (decimal digit = confirmed, see Appendix for rank/ratings equivalence table); 4) name of white player; 5) rating of white player (format as for black player); 6) board size (format: "width:height"); 7) time limit (in minutes, zero means "none"); 8) result (888.8 = opponent ran out of time, 999.9 = opponent resigned, otherwise win by points, positive = for black, negative = for white); 9) komi (extra points to compensate white for black moving first); 10) handicap (format: "h-0", "h-2", "h-3", "h-4", etc.); 11) moves (two letters from "a" to "s" = coordinates for each move, upper case = black, lower case = white, "TT" resp. "tt" = pass, handicap stones given as black move followed by white pass); 12) (optional) territory at end of game (format as for moves). Pathological games that have been removed from the olf files include: games that could not be reconstructed correctly due to corrupt sgf files, games in which five stones or less were played, games of players against themselves lasting less than 3.5*board_size moves (often "experiments" of some sort), duplicate games (all above fields identical), games with recorded score greater than twice the number of points on the board, games in which recorded score + komi is not an integer, and games containing illegal moves (including suicide moves and ko violations). The remaining games are organized in files/directories according to board size: 9x9 - games on a 9x9 board 13x13 - games on a 13x13 board 19x19 - games on the standard (19x19) board other - games using other board sizes Directories are further subdivided according to how the game ended: score - game was completed & scored, no territory recorded terry - game was completed & scored, territory was recorded time - game ended with one player running out of time resign - game ended with one player resigning There may be further subdivision according to whether it was an even or handicap game, and finally, the rating of the weaker player (black), using the numerical scheme given in 3) above. Note that it was at some point fashionable on the IGS for very strong players to claim extremely low ranks; games in which black is ranked 28-30 kyu have therefore been classified by opponent rank less handicap. ----------------- Appendix: rank/ratings translation ------------------ translation: 30 kyu = 2 ... 1 kyu = 31, 1 dan = 32 ... 9 dan = 40, 1 dan (pro) = 41 ... 9 dan (pro) = 49); see equivalence table olf rating AGA CHINA EUROPE IGS JAPAN KOREA ---------- --- ----- ------ --- ----- ----- 0 - - - NR - - 2 20K 23K 22K 30K 19K 26Gup 3 19K 22K 21K 29K 18K 25Gup 4 18K 21K 20K 28K 17K 24Gup 5 17-18K 20-21K 19-20K 27K 16-17K 23-24Gup 6 17K 20K 19K 26K 16K 23Gup 7 16K 19K 18K 25K 15K 22Gup 8 15-16K 18-19K 17-18K 24K 14-15K 21-22Gup 9 15K 18K 17K 23K 14K 21Gup 10 14K 17K 16K 22K 13K 20Gup 11 13K 16K 15K 21K 12K 19Gup 12 12-13K 15-16K 14-15K 20K 11-12K 18-19Gup 13 12K 15K 14K 19K 11K 18Gup 14 11K 14K 13K 18K 10K 17Gup 15 10K 13K 12K 17K 9K 16Gup 16 9-10K 12-13K 11-12K 16K 8-9K 15-16Gup 17 9K 12K 11K 15K 8K 15Gup 18 8K 11K 10K 14K 7K 14Gup 19 7K 10K 9K 13K 6K 13Gup 20 6-7K 9-10K 8-9K 12K 5-6K 12-13Gup 21 6K 9K 8K 11K 5K 12Gup 22 5K 8K 7K 10K 4K 11Gup 23 4-5K 7-8K 6-7K 9K 3-4K 10-11Gup 24 4K 7K 6K 8K 3K 10Gup 25 3K 6K 5K 7K 2K 9Gup 26 2K 5K 4K 6K 1K 8Gup 27 1-2K 4-5K 3-4K 5K 1K-1D 7-8Gup 28 1K 4K 3K 4K 1D 7Gup 29 1D 3K 2K 3K 2D 6Gup 30 2D 2K 1K 2K 3D 5Gup 31 3D 1K 1D 1K 4D 4Gup 32 4D 1D 2D 1D 4-5D 3Gup 33 5D 2D 3D 2D 5-6D 2Gup 34 6D 3D 4D 3D 6D 1-2Gup 35 6D 3-4D 4-5D 4D 6D 1Gup 36 7D 4-5D 5-6D 5D 6-7D 1Gup