Dannibal/ijcai-mahjong-round2

IJCAI-2026 Chinese Standard Mahjong — Finalist & Strong-Bot Game Corpus

Botzone game logs of the top-16 finalists of the IJCAI-2026 Mahjong AI Competition (Chinese Standard Mahjong / MCR, duplicate format), plus the top-3 players of every public simulation round, collected for opponent modeling and on-distribution evaluation. Each game is seat-tagged (we know which seat each tracked bot played), which makes the logs directly usable for per-player imitation / behavior cloning.

• Game: Chinese Standard Mahjong (国标麻将, MCR rules, 8-fan floor), Botzone game id 5e37dcf74019f43051e53201.
• Group / competition: Botzone group 69d9eca875ce41301557e099 (IJCAI-2026 Mahjong AI Competition).
• Format: 4-player duplicate tournaments + simulation rounds + global ladder.

---

Files

File	Size	Contents
finalist_games.tar.gz	~165 MB	~44,400 raw match JSONs (<match_id>.json), seat-tagged. The primary corpus.
finalist_decisions.tar.gz	~41 MB	Per-player .npz of extracted discard decisions {obs, mask, act, score} — ready for training.
imit_v3.tar.gz	~1.4 GB	27 imitation policy networks (one per tracked player), fused torch state-dicts — usable as evaluation opponents.
*.log	small	Result logs from our experiments (on-distribution gauntlet, candidate verification).

---

What was collected, and from whom

We tracked two groups of players (by Botzone user-id and bot-id, so all of a player's games are captured regardless of which bot version they ran):

The 16 finalists (Round-1 final standings)

Rank	Player	Tournament bot	Total score
1	Rouxqdd	dd_mahjong	1692.66
2	player152	player2	1681.11
3	Legendx	Greed	1668.41
4	cspsept	SeptMahjongVer0	1635.15
5	pyccc	PHOD	1630.26
6	蓝月_歪歪	Bot	1623.32
7	kong	shiro	1581.11
8	whatcanisayman	final_bot	1579.55
9	狗子酱pia	rts	1551.65
10	Lilangi	跟傻子似的	1548.38
11	moyu	distill	1526.56
12	Amy_xue	Lilith	1507.65
13	flbb	flbb2	1496.35
14	我什么都不会	小试强化	1495.61
15	Chinese_zjc_	春日影	1457.00
16	QiuQiuR	丘丘人	1443.63

Strong non-finalist bots (top-3 of each simulation round)

Captured because they are useful imitation teachers / evaluation opponents. Notable ones: pycc / chunjiandu (a very strong reference bot), mythos, 但闻海棠 / lee, 逐日小精灵, Toriel / Supervised_v1, 天胡豪七 / 小寻歌, MUFC / Bound. A few top sim slots were LLM-API players (e.g. gpt_5_mini) — included as beat-targets, not imitation teachers (their policy is not recoverable from observable discards).

Sources

• Global ladder — GET /globalmatchlist?game=5e37dcf74019f43051e53201&startid=<cursor>, paginated backward in time, filtered to the tracked ids.
• Simulation contests — GET /contest/detail/<id> (the table field lists /match/<id> links). Sim-7 (69f0440183ee0a54c18b5709) and Sim-8 (6a1d3d1258ebe27b197a8cca) contain finalists; the per-sim top-3 span Sim-1…8.
• Tournament (Round-1) matches are NOT included — those contests are marked secret on Botzone (table empty, excluded from the global list), so their logs are not retrievable.

---

How it was collected (method)

1. Crawl match pages, extract var _rawLogJSON (the per-turn game record) → the log field.
2. Seat tagging (key step): the match page also carries a playerNames JS array in seat order, where each entry's user-id is embedded in its avatar URL (/avatar/<uid>.png). We parse this into a players roster [{seat, name, uid, target}], so every decision can be attributed to the correct player/seat. (Logs without this roster are useless for per-player imitation — this is why we re-fetched.)
3. Version hygiene: Botzone match ids are MongoDB ObjectIds, so int(match_id[:8], 16) is the unix timestamp. We date-filter to keep recent bot versions and avoid mixing a player's old/weak submissions with their current one.
4. Decision extraction: each game is replayed with a feature encoder; the tracked seats' discard (Play) decisions are recorded as (obs, mask, act) plus that seat's final duplicate score.

---

Data formats

Raw match JSON (finalist_games/<match_id>.json)

{
  "match_id": "6a33c8c0b38f704f97b5f885",
  "our_players": { "<user_id>": {"player": "Legendx", "bot_name": "Greed"} },
  "players": [
    {"seat": 0, "name": "morrow",        "uid": "6a2117...", "target": false},
    {"seat": 1, "name": "[Legendx]Greed","uid": "691418...", "target": true },
    ...
  ],
  "log": [ /* per-turn records; each has output.display.action in
             {INIT,DEAL,DRAW,PLAY,CHI,PENG,GANG,BUGANG,HU} and a final
             record with display.score = [s0,s1,s2,s3] (duplicate scores) */ ]
}

Decision npz (finalist_decisions/<player>.npz)

key	shape / dtype	meaning
obs	(N, 38, 4, 9) int8	board/hand feature tensor at each discard decision
mask	(N, 235) bool	legal-action mask (235-action space: Pass / Hu / Play×34 / Chi / Peng / Gang …)
act	(N,) int16	the action taken (these npz contain discard/Play decisions only)
score	(N,) float32	that seat's final duplicate game score (for advantage-weighted BC)

---

How to use

Load the raw games:

import json, tarfile
tf = tarfile.open("finalist_games.tar.gz")
g = json.load(tf.extractfile("finalist_games/6a33c8c0b38f704f97b5f885.json"))
seat = next(p["seat"] for p in g["players"] if p["name"].startswith("[Rouxqdd]"))
# replay g["log"], collecting that seat's decisions

Train a behavior-cloning / imitation policy from a player's npz:

import numpy as np
z = np.load("finalist_decisions/Rouxqdd.npz")
obs, mask, act, score = z["obs"], z["mask"], z["act"], z["score"]
# masked cross-entropy on (obs -> act); optionally weight by score (advantage-weighted BC)

Use the imitation nets (imit_v3.tar.gz) as evaluation opponents in a 4-player gauntlet (each is a fused ResNet policy over the 235-action space; build a net sized from the checkpoint's own keys and load_state_dict).

Suggested uses

• Opponent modeling / best-response against specific finalists.
• On-distribution evaluation — gauntlet your bot against imitations of the actual field instead of synthetic opponents.
• Behavior cloning / offline RL research on a real, strong, imperfect-information card game.
• Style/meta analysis — discard tendencies, deal-in rates, fan preferences per player.

---

Caveats & honest notes

• Discard-only decisions: the npz capture Play decisions, not claim decisions (Chi/Peng/Gang/Hu). The raw logs contain everything if you want to extract more.
• Imitation ceiling: in our experiments, behavior-cloning / KL-distillation toward this data produced policies that tie-or-lose a strong reference bot once verified at scale — small-sample gauntlets repeatedly threw false positives that collapsed under 200-game bias-corrected re-tests. Treat any "beats baseline" claim skeptically and verify with large N + a matched null.
• Sim ≠ tournament version: a player's simulation-round bot may differ from their tournament submission. Sim-8 is closest in time to the final.
• No tournament logs: the actual Round-1 tournament games are not public/retrievable.

Provenance

Collected June 2026 from public Botzone pages for research on the IJCAI-2026 Mahjong AI Competition. Game logs are public match records. Player handles are as displayed on Botzone. If you are a listed player and want your handle anonymized, open an issue.