Henhen1227/rps-strategy-games

RPS Strategy game archive

Every finished game played on RPS Strategy, updated once a day. Complete records: every move, both clocks after every action, draw offers and time extensions, and how the game ended.

The games are not my invention. Intransitive, Total War and Infiltration were designed by WebGoatGuy, who runs the original implementation at meaf.us/rps2. This site is a fan implementation, and this data set is games played on it. Games from meaf.us's own server are not included here — they are Meaf's to publish.

Loading it

from datasets import load_dataset

games = load_dataset("Henhen1227/rps-strategy-games", split="train")
print(games[0]["pgn"])

The pgn/ directory carries the same games as gzipped PGN text, for tooling that would rather read that than a dataframe.

What a row is

One finished game. The pgn column is self-sufficient — it replays without anything else in the file — and the other columns are there so you can filter without parsing it.

column	type
game_id	string
mode_id, mode_name	string
red_player, blue_player	string
red_key, blue_key	string
players	string
opener	string
pgn_dialect	int64
winner_color, outcome, end_reason, termination	string
event	string
ranked	bool
tournament_id, series_id, opening_seed	string
book_plies	int64
red_elo, red_elo_after, blue_elo, blue_elo_after	int64
ply_count, initial_time_ms, increment_ms	int64
started_at_unix_ms, finished_at_unix_ms	int64
pgn	string

Three columns worth reading before you use this

opener — the rules changed on 2026-09-03. Blue moves first now; Red moved first before. Both sets of games are in here, and mixing them is the mistake this column exists to prevent. Filter on opener == "blue" for games under current rules. pgn_dialect is a different fact — a move-numbering convention, where dialect 1 numbered pairs by colour and dialect 2 numbers by whoever opened — and both dialects replay correctly. Read opener for the rules, not pgn_dialect.

players — bot, mixed or human. The mix is heavily skewed, and so are the game lengths:

players	share	median plies
bot — engine vs engine	77%	76
mixed — engine vs person	19%	23
human — person vs person	3%	51

Any statistic over the whole file is a statistic about engine play. Say which subset you meant. Note that mixed games are the shortest: people resign against engines quickly.

book_plies — moves nobody chose. In bot matches the first few plies are dealt from a seeded random opening rather than selected by either engine. Skip them when training or grading: they are not decisions.

Names and keys

Games finished from 2026-09-09T00:00:00Z onward carry the usernames they were played under. Games finished before that instant read Anonymous player on both seats — they were played before the site's privacy policy said games would be published under names, so publishing them that way would not have been agreed to.

red_key and blue_key are opaque per-player identifiers: an HMAC of the account under a key that is not published. They let you group one person's games without the key leading back to an account. They are stable across renames, and a player who deletes their account keeps their key while their name becomes Deleted player.

What is never in here

No Discord user ID or handle, no account ID, no account key hash, no password hash, no IP address, and no in-game chat. Nor the internal timestamp recording when a row was written. The site's privacy policy is the binding version of this list.

How the files are laid out

data/YYYY/games-YYYY-MM-DD.parquet   # the dataset
pgn/YYYY/games-YYYY-MM-DD.pgn.gz     # the same games as PGN text

File dates are publication dates, not play dates. A file named 2026-09-11 holds the games first published that day, which is almost always the games played the day before — but a game the server recorded late lands in whichever file was next. Sort by finished_at_unix_ms for chronological order.

Files are never rewritten. Once a day is out it stays as it was, which is also why a name removed at the source only changes in files published afterwards: copies already downloaded cannot be recalled.

License

CC0-1.0. Public domain dedication — use it for anything, no permission needed and no attribution required. Game records are records of facts, and this makes that explicit rather than asserting a copyright over them.

Attribution is welcome but not required. If it is useful to you, I would enjoy hearing about it.

Known limitations

• Two rule sets are present; see opener above.
• Bot and human games are mixed; see players above.
• Compiled opening statistics are not published here.
• Games interrupted by a server restart are excluded — only finished games are in the file.
• Modes are lopsided: Intransitive (V6) is almost everything, with a few hundred Infiltration (V3) and Total War (V5) games.
• Every game carries before-and-after ratings, including unranked ones, where the two are simply equal. ranked is the column to filter on, not the ratings.