danbrooks/draftzero-selfplay-games

DraftZero self-play games

Games of Magic: The Gathering limited (Foundations, FDN) played by DraftZero's agent against itself, with tree search at every decision. Each searched decision is stored as a training example: the position, the search's answer and the game's result. They are what DraftZero's self-play training learns from (docs/021).

Status: empty for now. Games arrive once the self-play loop starts. The layout and format below may still change before the first upload; this card will say when they do.

What a game is

• The engine: XMage, a full rules implementation of Magic, driven through MageZero.
• The decks: two decks drawn at random from top players' Foundations Premier Draft decks published by 17lands (players with a 60%+ win rate).
• The players: a policy-value network (first trained to imitate top players' decisions from 17lands replays, then by self-play) searching each decision with Monte Carlo tree search over sampled versions of the hidden cards. Neither side sees the other's hand or decklist.
• The budget: the number of search simulations per decision is recorded with each game (e.g. il_bc@100).
• About 169 searched decisions a game; about 100 KB a game, gzipped.

Format

Files are gzipped JSON lines. One line is one seat's view of one game:

Field	Meaning
pair	the deck pair's index in its run
swap	whether the seats were swapped (each deck pair is played twice, once each way)
seat	A or B
bot	the player and its search budget, e.g. il_bc@100
result	the game's result for this seat: 1 win, -1 loss, null no winner (e.g. the 50-turn cap)
records	the seat's searched decisions, in order (below)

Each entry of records is one decision:

Field	Meaning
features	the position as MageZero encodes it, from this seat's point of view (the opponent's hand hidden): sorted ids of discrete features (cards in zones with their status, life totals, turn and step, ...). Ids are XMage feature hashes (xmage_feature_hash, version 1)
type	the decision: PRIORITY (cast, play or pass), CHOOSE_USE (yes or no, e.g. "attack with this creature?"), CHOOSE_TARGET (a target, or which attacker to block)
turn	the game's turn number
legal	the legal options, as indices into DraftZero's action vocabulary (assets/vocab/FDN_SPG.tsv); for CHOOSE_USE, 0 is no and 1 is yes
visits	how many search simulations went to each legal option (same order as legal): the policy target
q	the search's value of the position for this seat, from -1 (lost) to 1 (won)
heuristic	MageZero's hand-written evaluation of the position for this seat

Planned additions, before or with the first games: the network's prior and each option's mean value, the option actually played, the network version, and a replay log (seeds and choices, so any future encoder can rebuild the positions).

Reading it

import gzip, json
from huggingface_hub import hf_hub_download, list_repo_files

repo = "danbrooks/draftzero-selfplay-games"
files = [f for f in list_repo_files(repo, repo_type="dataset") if f.endswith(".jsonl.gz")]
path = hf_hub_download(repo, files[0], repo_type="dataset")
for line in gzip.open(path, "rt"):
    game = json.loads(line)
    for d in game["records"]:
        best = d["legal"][max(range(len(d["visits"])), key=d["visits"].__getitem__)]

DraftZero's tools/imitation_scale/selfplay_tables.py turns these files into training tables.

License and credits

• This dataset: CC BY 4.0.
• 17lands: every deck comes from its public datasets, licensed CC BY 4.0. Thank you to 17lands and the players who share their data.
• XMage, the rules engine every game runs in, and MageZero by Will Wroble, the framework DraftZero builds on.
• The code that produces the games is MIT-licensed: danieljbrooks/draft-zero.
• Magic: The Gathering is a trademark of Wizards of the Coast. This dataset is unofficial and not affiliated with or endorsed by Wizards of the Coast.