armaanp/gen9-random-doubles-replays

Gen 9 Random Doubles replays

Every gen9randomdoublesbattle replay Pokemon Showdown still lists, plus the ones it has since delisted, deduplicated on replay id.

229,171 battles, 2023-06-26 to 2026-08-26. 73.2% carry a rating: 61,673 at 1400 or better, 43,756 at 1500, 27,728 at 1600, 14,506 at 1700, 5,555 at 1800.

Why this exists

Neither available source is complete on its own.

HolidayOugi/pokemon-showdown-replays holds 203,654 distinct battles for this format and stops at 2026-06-19. It also thins out before that, with May 2026 at about half of April's count, and it is missing a few hundred to 1,837 a month across 2023-12 to 2025-07. Its file for this format reports 395,901 rows, but that is two overlapping snapshots of the same data, 192,247 rows and 203,654 rows, the smaller a strict subset of the larger. Counting rows overstates it by nearly two to one.

Showdown's own index runs to the day you walk it but reaches back only as far as the site still lists. Walking it needs before=<unix seconds>, not page=, which caps at page 100 and reaches about a month.

Against each other: 203,547 battles in both, 25,517 only in the index, and 107 only in the dump. This dataset is the union.

Columns

column	type	notes
id	string	replay id, unique
format	string	display name
formatid	string	always gen9randomdoublesbattle
players	list[string]	both usernames
log	string	the full battle log
uploadtime	int64	unix seconds
rating	double	null when unrated

What a log does and does not contain

Public replays carry no |request| lines, so a player's available moves are not recoverable and no index based action encoding can be labelled from them. Any action target has to be semantic, a move by identity and its target.

A replay is also a poorer witness than the player was. At the moment of decision a log has shown 4.15 of a side's six Pokemon and 1.35 of the four moves on an active, where the player could see all six and all four. Reconstructing a side's team from the whole battle recovers what they knew; doing the same for their opponent would hand a model information the player never had.

Roughly 41.5 decisions per battle are recoverable, split 77.6% moves, 7.9% voluntary switches and 14.5% forced replacements after a faint. In 26.6% of turns at least one slot chose something that never reached the log, its Pokemon having been flinched or put to sleep before it acted, so those slots are absent rather than empty.

Collection

Fetched at no more than 5 requests a second with at most 4 in flight, over roughly 4,500 index requests and 22,767 log requests, with no 429 and no 5xx. Showdown is volunteer run and serving this is a favour. Please do not refetch what is already here.