Datasetpublic

scaleinvariant/cgos-9x9-parquet

CGOS 9x9 Computer Go Games Cleaned 9x9 Go game records from the public monthly archives of the 9x9 Computer Go Server (CGOS), covering 2015-11-10 to 2026-07-31. Players are computer Go programs, many of them KataGo-family bots; per-bot ratings (CGOS BayesElo) are included for strength filtering. Schema column type description game_id string CGOS game number date string date played (from SGF DT) white / black string bot names white_rating /… See the full description on the dataset page: https://huggingface.co/datasets/scaleinvariant/cgos-9x9-parquet.

sourceHugging Faceotherupdated 2mo agoView on Hugging Face

0likes27downloads

Dataset Card

CGOS 9x9 Computer Go Games

Cleaned 9x9 Go game records from the public monthly archives of the 9x9 Computer Go Server (CGOS), covering 2015-11-10 to 2026-07-31.

Players are computer Go programs, many of them KataGo-family bots; per-bot ratings (CGOS BayesElo) are included for strength filtering.

Schema

column	type	description
`game_id`	string	CGOS game number
`date`	string	date played (from SGF `DT`)
`white` / `black`	string	bot names
`white_rating` / `black_rating`	float32	CGOS rating (null if provisional/unknown)
`result`	string	e.g. `W+Resign`, `B+3.5`
`winner`	string	`B` or `W`
`num_moves`	int16	move count
`moves`	list[int16]	move tokens: `0..80` = `row*9+col` (row 0 = top), `81` = pass
`source_archive`	string	source monthly archive, e.g. `9x9_2026_06`

Filtering

—board size verified SZ[9];
—completed games only (RE matches B+.../W+...);
—between 10 and 200 moves (longer 9x9 records are degenerate ko loops);
—exact duplicates removed (CGOS game number, then move-sequence hash).

Per-move engine-analysis comments present in the raw SGFs were dropped; this dataset keeps moves and game metadata only.

Summary

json

{
  "num_games": 1508733,
  "unique_bots": 3553,
  "top_bots": [
    [
      "Gnugo-3.7.10-a1",
      312108
    ],
    [
      "Stop-0.9-005",
      284728
    ],
    [
      "CrazyStone-0006-99",
      213578
    ],
    [
      "Aya783a_50",
      179467
    ],
    [
      "fuego1604_100k",
      155251
    ],
    [
      "DCNN_AyaF128a523x1",
      97923
    ],
    [
      "Aya800e_0721_10k",
      90407
    ],
    [
      "LZ20_128_65_p3000",
      84292
    ],
    [
      "sai_S1_9461_p1000",
      69377
    ],
    [
      "fuego1604_10k",
      57689
    ]
  ],
  "moves_median": 77,
  "moves_max": 200,
  "rated_fraction": 1.0
}

Provenance and license

Game records originate from the public CGOS archives (http://www.yss-aya.com/cgos/). Game records themselves are not original creative works; this redistribution is for research use. Please credit CGOS.

Intended use

Training and evaluation of sequence models over game trajectories. Move streams from distinct games can be merged (e.g. interleaved on one board) to create synthetic mixtures with ground-truth assignment labels.