Bauxitiego/mtg-deck-matchup

MTG Deck Matchup

13,036 real Magic: The Gathering Standard decklists from 408 real MTGO Standard Challenge events (September 2025 through August 2026), each joined to its own real tournament result: rank, match wins and losses, game win percentage, opponent match win percentage. Every card resolved against Scryfall where possible. Built for Bauxitiego/mtg-deck-matchup; see that repo for the collection pipeline, the card-resolution investigation, and the model built on top of this data.

Most public MTG datasets give you an archetype label and its aggregate win rate. This one gives you the actual 60-card list behind a specific real result, so the deck itself is available as a feature, not just its label.

What's in a row

One row is one player's decklist in one event.

column	type	notes
event_id, event_slug, event_description, event_date	str	e.g. "Standard Challenge 32", 2026-04-09
player_count	int	field size for that event
loginid, player	str	MTGO account ID and handle, as published on mtgo.com
rank	int	final standing
wins, losses	int	real match record for the event
game_win_pct	float	the target used in the linked project
opponent_match_win_pct	float	strength-of-schedule tiebreaker, as MTGO reports it
eliminated	bool	whether the player was cut before the event ended
archetype_cluster	int	inferred, not an official archetype name - see below
main_deck, sideboard	list of struct	{name, qty, cmc, colors, supertype, rarity, resolved_source}
main_deck_text, sideboard_text	str	the same decklist as plain "qty name" lines
n_lands, land_ratio, avg_cmc_nonland, n_colors, curve_0..6, color_W/U/B/R/G, type_*	float/int	deck-level summary features aggregated from main_deck

archetype_cluster comes from TF-IDF-weighted K-means (k=20) over each deck's nonland card-inclusion vector, run against this same corpus. MTGO's raw data has no archetype field at all, so this is inferred, not a metagame-tool label - useful as a rough shell grouping, not as ground truth.

Card resolution

Card names are matched against Scryfall's Oracle Cards export by exact name, then by split/adventure/double-faced front face. About 3% of card slots (tagged with MTGO's internal OM1 cardset code) don't resolve to any Scryfall card under any name - a real, unexplained MTGO-internal set-code anomaly, not a formatting bug, written up in notes/01-data-sources-verified.md in the linked repo. For those cards, cmc, colors, supertype, and rarity come directly from MTGO's own card data instead of Scryfall, and resolved_source is set to mtgo_attributes so you can filter on it if exact Scryfall provenance matters for your use case.

Collection

Decklists and results come from mtgo.com's own official event pages: real, individual, per-player outcomes, not an archetype-level aggregate. Two third-party aggregators (aetherhub.com, mtgdecks.net) were considered and not used, because both declare Content-Signal: ai-train=no in their robots.txt. Card data is Scryfall's Oracle Cards bulk export, downloaded the same week as collection.

Example

from datasets import load_dataset

ds = load_dataset("Bauxitiego/mtg-deck-matchup", split="train")
row = ds[0]
print(row["player"], row["game_win_pct"], row["main_deck_text"][:100])

License and attribution

The compiled dataset (this joined, resolved table) is released under CC BY 4.0. Magic: The Gathering is a trademark of Wizards of the Coast; this is unofficial Fan Content under Wizards' Fan Content Policy, not endorsed by Wizards. Card data is from Scryfall, used under its stated terms for exactly this kind of bulk use. Decklist and tournament result data is republished from Wizards' own public mtgo.com event pages.