Datasets:
File size: 21,395 Bytes
6477e62 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 | #!/usr/bin/env python3
"""validate_engine.py — correctness validator for Chinese Standard Mahjong (MCR) engines
against the IJCAI-2026 Final Stage-2 golden game set (Dannibal/mcr-final2026-testset).
Two modes:
1. Self-test (validates the dataset itself against the reference replay semantics):
python3 validate_engine.py --self-test mcr_final2026_full.jsonl.gz [--jobs 64]
Replays every game with a pure-python reference checker (the corrected
"replay_harness2" semantics: PENG/CHI-embedded discards, live-discard GANG
discrimination, BUGANG handling, ron/zimo/qianggang terminal labeling) and
asserts, for every game:
- wall integrity (136 tiles, 4 of each of the 34 kinds)
- deal & draw order: seat s is dealt/draws from wall[34*s .. 34*s+33],
consumed from the BACK (index 34*s+33 first); tileCnt bookkeeping
- per-step hand/meld legality (every discard from hand, PENG needs 2 in
hand, CHI only from the left neighbour with both fillers in hand,
melded GANG needs 3, AnGang needs 4 after own draw, BUGANG upgrades an
existing PENG)
- claim resolution priority (HU > PENG/GANG > CHI) and actor rotation
- response<->event consistency (the granted claim's response string must
match the following judge event, including the embedded discard)
- terminal: ending classification (zimo/ron/draw + qianggang), fan_total
== sum(value*cnt) of the fan list, exact score arithmetic
(zimo: winner +3*(8+fan), others -(8+fan);
ron: winner +(3*8+fan), discarder -(8+fan), bystanders -8),
scores sum to 0, and agreement with the record's `expected` block.
2. Engine mode (validates YOUR engine — the product):
python3 validate_engine.py --engine mymodule:MyEngine mcr_final2026_golden.jsonl
Your engine must implement the interface below. The validator drives it
with the logged bot responses turn by turn and requires the engine to
reproduce the judge's protocol stream exactly: every per-seat request
string, every display event (claim resolutions, gang discrimination, win
detection) and the terminal fan/score result.
class MyEngine:
def reset(self, wall: list[str], quan: int, srand: int) -> dict:
'''Start a game. wall = 136 tiles; seat s draws from
wall[34*s : 34*s+34], back to front. Returns the first turn:
{"requests": {"0": "0 0 0", ...}, "display": {"action": "INIT", ...}}'''
def step(self, responses: dict[str, str]) -> dict:
'''Feed the four seats' response strings to the pending
requests; return the next turn dict:
{"requests": {...} or {}, "display": {...}}.
The terminal turn's display has action HU (with fan/fanCnt/
score) or HUANG.'''
By default `canHu` and `tileCnt` inside displays are compared too (they
are exactly reproducible); use --loose to compare only
action/player/tile/tileCHI/hand/fan/fanCnt/score.
`--demo` runs engine mode with a built-in engine that replays the stored
turns (interface smoke test).
Dataset: https://huggingface.co/datasets/Dannibal/mcr-final2026-testset
"""
import argparse
import gzip
import json
import sys
from collections import Counter
TILE_KINDS = ([c + str(n) for c in "WBT" for n in range(1, 10)]
+ ["F" + str(n) for n in range(1, 5)]
+ ["J" + str(n) for n in range(1, 4)])
CLAIM_PRIO = {"HU": 3, "GANG": 2, "PENG": 2, "CHI": 1}
class CheckFail(Exception):
pass
def _ck(cond, msg):
if not cond:
raise CheckFail(msg)
def classify_terminal(displays):
"""Harness2 terminal semantics from the ordered display stream.
Returns dict(ending, winner, discarder, qianggang) — ending in
zimo|ron|draw. Ron off a PENG/CHI-embedded discard counts the claimer
as discarder; HU immediately after BUGANG is qianggang (robbed kong),
discarder = the BUGANG player."""
live = None # (tile, seat) of the unclaimed discard
prev = None
for d in displays:
a = d["action"]
if a == "DRAW":
live = None
elif a in ("PLAY", "PENG", "CHI"):
live = (d["tile"], d["player"])
elif a == "GANG":
live = None
elif a == "BUGANG":
live = None
elif a == "HUANG":
return dict(ending="draw", winner=None, discarder=None, qianggang=False)
elif a == "HU":
w = d["player"]
if prev is not None and prev["action"] == "BUGANG":
return dict(ending="ron", winner=w,
discarder=prev["player"], qianggang=True)
if live is not None:
return dict(ending="ron", winner=w, discarder=live[1], qianggang=False)
return dict(ending="zimo", winner=w, discarder=None, qianggang=False)
prev = d
return dict(ending="unknown", winner=None, discarder=None, qianggang=False)
def expected_scores(ending, winner, discarder, fan_total):
if ending == "draw":
return [0, 0, 0, 0]
sc = [0, 0, 0, 0]
if ending == "zimo":
for s in range(4):
sc[s] = 3 * (8 + fan_total) if s == winner else -(8 + fan_total)
else: # ron (incl. qianggang)
for s in range(4):
if s == winner:
sc[s] = 3 * 8 + fan_total
elif s == discarder:
sc[s] = -(8 + fan_total)
else:
sc[s] = -8
return sc
def check_record(rec):
"""Full structural + semantic check of one dataset record.
Raises CheckFail on the first violation. Returns a dict of facts
(event counts etc.) usable for tagging."""
wall = rec["wall"]
_ck(len(wall) == 136, "wall length %d != 136" % len(wall))
_ck(Counter(wall) == Counter({t: 4 for t in TILE_KINDS}),
"wall is not 4x each of the 34 tile kinds")
segs = [wall[34 * s:34 * s + 34] for s in range(4)]
remain = [34, 34, 34, 34] # tiles left in each seat's segment
hands = [None] * 4
packs = [[] for _ in range(4)] # (type, tile, offer_seat)
live = None # (tile, seat) live unclaimed discard
prev_a = prev_p = None
turns = rec["turns"]
facts = Counter()
displays = [t["display"] for t in turns]
def hand_ok(s):
_ck(len(hands[s]) == 13 - 3 * len(packs[s]),
"seat %d hand size %d with %d melds" % (s, len(hands[s]), len(packs[s])))
for k, turn in enumerate(turns):
d = turn["display"]
a = d["action"]
nxt = turns[k + 1]["display"] if k + 1 < len(turns) else None
resp = turn.get("responses") or {}
if a == "INIT":
_ck(d.get("quan") == rec["quan"], "INIT quan mismatch")
_ck(d.get("srand") == rec["srand"], "INIT srand mismatch")
elif a == "DEAL":
for s in range(4):
expect = [segs[s][33 - i] for i in range(13)]
_ck(d["hand"][s] == expect,
"seat %d dealt hand != last-13-reversed of wall segment" % s)
hands[s] = list(d["hand"][s])
remain[s] = 21
elif a == "DRAW":
s, t = d["player"], d["tile"]
_ck(remain[s] > 0, "seat %d draws from empty segment" % s)
_ck(segs[s][remain[s] - 1] == t,
"draw %s != wall order tile %s" % (t, segs[s][remain[s] - 1]))
remain[s] -= 1
_ck(d["tileCnt"][s] == remain[s], "tileCnt mismatch after draw")
# rotation: replacement draw after own kong, else next seat
if prev_a in ("GANG", "BUGANG"):
_ck(s == prev_p, "replacement draw by wrong seat")
elif prev_a in ("PLAY", "PENG", "CHI"):
_ck(s == (prev_p + 1) % 4, "draw out of rotation")
elif prev_a == "DEAL":
_ck(s == 0, "first draw not by dealer (seat 0)")
live = None
hands[s].append(t)
# drawer's own response drives the next event
r = (resp.get(str(s)) or "").split()
if r:
if r[0] == "PLAY":
_ck(nxt and nxt["action"] == "PLAY" and nxt["player"] == s
and nxt["tile"] == r[1], "PLAY response not honored")
elif r[0] == "GANG":
_ck(nxt and nxt["action"] == "GANG" and nxt["player"] == s
and nxt.get("tile") == r[1], "AnGang response not honored")
elif r[0] == "BUGANG":
_ck(nxt and nxt["action"] == "BUGANG" and nxt["player"] == s
and nxt["tile"] == r[1], "BuGang response not honored")
elif r[0] == "HU":
_ck(nxt and nxt["action"] == "HU" and nxt["player"] == s,
"zimo HU response not honored")
elif a in ("PLAY", "PENG", "CHI"):
s = d["player"]
out = d["tile"]
if a == "PLAY":
_ck(prev_a == "DRAW" and prev_p == s, "PLAY not after own draw")
elif a == "PENG":
_ck(live is not None and live[1] != s, "PENG with no live discard")
ct = live[0]
_ck(hands[s].count(ct) >= 2, "PENG without 2 matching tiles in hand")
hands[s].remove(ct); hands[s].remove(ct)
packs[s].append(("PENG", ct, live[1]))
facts["minggang_or_peng"] += 1
facts["peng"] += 1
else: # CHI
_ck(live is not None, "CHI with no live discard")
ct, ds = live
_ck((s - ds) % 4 == 1, "CHI from non-left neighbour")
mid = d["tileCHI"]
_ck(ct[0] == mid[0] and mid[0] in "WBT", "CHI suit mismatch")
m, c = int(mid[1]), int(ct[1])
_ck(c in (m - 1, m, m + 1), "claimed tile not part of chi run")
need = [mid[0] + str(x) for x in (m - 1, m, m + 1) if x != c]
for tt in need:
_ck(tt in hands[s], "CHI filler %s not in hand" % tt)
hands[s].remove(tt)
packs[s].append(("CHI", mid, ds))
facts["chi"] += 1
# the embedded (PENG/CHI) or plain discard
_ck(out in hands[s], "discard %s not in hand" % out)
hands[s].remove(out)
hand_ok(s)
live = (out, s)
# claim arbitration on this discard
claims = {}
for ss in range(4):
if ss == s:
continue
rr = (resp.get(str(ss)) or "PASS").split()
if rr and rr[0] != "PASS":
claims[ss] = rr
if claims:
best = max(CLAIM_PRIO.get(rr[0], 0) for rr in claims.values())
winners = [ss for ss, rr in claims.items()
if CLAIM_PRIO.get(rr[0], 0) == best]
if len(winners) > 1: # multiple HU on one tile: nearest downstream
facts["multi_hu"] += 1
winners = sorted(winners, key=lambda ss: (ss - s) % 4)[:1]
if len(claims) > 1:
facts["multi_claim"] += 1
ws = winners[0]
rr = claims[ws]
_ck(nxt is not None, "claim pending but log ended")
_ck(nxt["player"] == ws,
"claim priority: expected seat %d to act next" % ws)
_ck({"HU": "HU", "PENG": "PENG", "CHI": "CHI",
"GANG": "GANG"}[rr[0]] == nxt["action"],
"granted claim %s != next event %s" % (rr[0], nxt["action"]))
if rr[0] == "PENG":
_ck(nxt["tile"] == rr[1], "PENG follow-up discard mismatch")
elif rr[0] == "CHI":
_ck(nxt["tileCHI"] == rr[1] and nxt["tile"] == rr[2],
"CHI mid/follow-up mismatch")
else:
_ck(nxt is not None and nxt["action"] in ("DRAW", "HUANG"),
"no claim: expected DRAW/HUANG next, got %s"
% (nxt and nxt["action"]))
elif a == "GANG":
s = d["player"]
t = d.get("tile")
if live is not None: # melded gang of the live discard
ct, ds = live
_ck(s != ds, "GANG own discard")
_ck(t == ct, "melded GANG tile %s != live discard %s" % (t, ct))
_ck(hands[s].count(ct) == 3, "melded GANG without 3 in hand")
for _ in range(3):
hands[s].remove(ct)
packs[s].append(("GANG", ct, ds))
live = None
facts["minggang"] += 1
else: # concealed kong after own draw
_ck(prev_a == "DRAW" and prev_p == s, "AnGang not after own draw")
_ck(hands[s].count(t) == 4, "AnGang without 4 in hand")
for _ in range(4):
hands[s].remove(t)
packs[s].append(("ANGANG", t, s))
facts["angang"] += 1
_ck(nxt is not None and ((nxt["action"] == "DRAW" and nxt["player"] == s)
or nxt["action"] == "HUANG"),
"no replacement draw after GANG")
elif a == "BUGANG":
s, t = d["player"], d["tile"]
up = [p for p in packs[s] if p[0] == "PENG" and p[1] == t]
_ck(up, "BUGANG without matching PENG meld")
_ck(t in hands[s], "BUGANG tile not in hand")
hands[s].remove(t)
packs[s][packs[s].index(up[0])] = ("GANG", t, up[0][2])
live = None
facts["bugang"] += 1
# qianggang arbitration
hus = [ss for ss in range(4) if ss != s
and (resp.get(str(ss)) or "PASS").split()[:1] == ["HU"]]
if hus:
_ck(nxt is not None and nxt["action"] == "HU"
and nxt["player"] in hus, "qianggang HU not honored")
else:
_ck(nxt is not None and ((nxt["action"] == "DRAW"
and nxt["player"] == s)
or nxt["action"] == "HUANG"),
"no replacement draw after BUGANG")
elif a == "HUANG":
nxt_drawer = (prev_p + 1) % 4 if prev_a in ("PLAY", "PENG", "CHI") \
else prev_p
_ck(remain[nxt_drawer] == 0,
"HUANG but next drawer still has %d wall tiles" % remain[nxt_drawer])
elif a == "HU":
pass # terminal block checked below
else:
raise CheckFail("unknown display action %r" % a)
prev_a, prev_p = a, d.get("player", prev_p)
# ---- terminal / expected block ----
exp = rec["expected"]
cls = classify_terminal(displays)
_ck(cls["ending"] == exp["ending"], "ending %s != expected %s"
% (cls["ending"], exp["ending"]))
_ck(cls["winner"] == exp["winner"], "winner mismatch")
_ck(cls["discarder"] == exp["discarder"], "discarder mismatch")
_ck(cls["qianggang"] == exp["qianggang"], "qianggang flag mismatch")
last = displays[-1]
if exp["ending"] == "draw":
_ck(last["action"] == "HUANG", "draw game does not end with HUANG")
_ck(exp["scores"] == [0, 0, 0, 0], "draw scores not all 0")
_ck(exp["fan_total"] == 0 and exp["fan"] == [], "draw with fan")
else:
_ck(last["action"] == "HU", "hu game does not end with HU")
fan_sum = sum(v * c for _, v, c in exp["fan"])
_ck(fan_sum == exp["fan_total"],
"fan list sums to %d != fan_total %d" % (fan_sum, exp["fan_total"]))
_ck(last["fanCnt"] == exp["fan_total"], "display fanCnt != fan_total")
_ck(list(last["score"]) == list(exp["scores"]), "display score != expected")
want = expected_scores(exp["ending"], exp["winner"], exp["discarder"],
exp["fan_total"])
_ck(want == list(exp["scores"]),
"score arithmetic: computed %s != logged %s" % (want, exp["scores"]))
_ck(exp["fan_total"] >= 8, "winning hand below the 8-fan minimum")
_ck(sum(exp["scores"]) == 0, "scores do not sum to 0")
facts[exp["ending"]] += 1
return dict(facts)
# ---------------------------------------------------------------- engine mode
LOOSE_KEYS = ("action", "player", "tile", "tileCHI", "hand", "fan", "fanCnt", "score")
def _disp_eq(a, b, loose):
if loose:
return all(a.get(k) == b.get(k) for k in LOOSE_KEYS)
ka = {k: v for k, v in a.items() if v is not None}
kb = {k: v for k, v in b.items() if v is not None}
return ka == kb
class DemoReplayEngine:
"""Interface demo: 'engine' that replays the record's own stored turns."""
def bind(self, rec):
self._turns = rec["turns"]
self._i = 0
def reset(self, wall, quan, srand):
self._i = 0
return self._emit()
def step(self, responses):
return self._emit()
def _emit(self):
t = self._turns[self._i]
self._i += 1
return {"requests": t.get("request") or {}, "display": t["display"]}
def run_engine(engine, rec, loose=False):
"""Drive `engine` through one game; raise CheckFail on first divergence."""
turns = rec["turns"]
out = engine.reset(list(rec["wall"]), rec["quan"], rec["srand"])
for k, turn in enumerate(turns):
_ck(_disp_eq(out["display"], turn["display"], loose),
"turn %d display diverged:\n engine: %s\n oracle: %s"
% (k, json.dumps(out["display"], ensure_ascii=False),
json.dumps(turn["display"], ensure_ascii=False)))
if turn.get("request"):
_ck(out.get("requests") == turn["request"],
"turn %d request strings diverged" % k)
if k + 1 < len(turns):
out = engine.step(dict(turn.get("responses") or {}))
# --------------------------------------------------------------------- driver
def iter_records(path):
op = gzip.open if path.endswith(".gz") else open
with op(path, "rt", encoding="utf-8") as f:
for line in f:
line = line.strip()
if line:
yield json.loads(line)
def _self_one(line):
rec = json.loads(line)
try:
check_record(rec)
return (rec["game_id"], None, rec.get("tags", []))
except CheckFail as e:
return (rec["game_id"], str(e), rec.get("tags", []))
except Exception as e: # noqa: BLE001
return (rec["game_id"], "internal: %r" % e, rec.get("tags", []))
def main():
ap = argparse.ArgumentParser(description=__doc__.split("\n")[0])
ap.add_argument("dataset", help=".jsonl or .jsonl.gz test file")
ap.add_argument("--self-test", action="store_true",
help="validate the dataset against the reference replay")
ap.add_argument("--engine", metavar="module:Class",
help="validate an engine implementation")
ap.add_argument("--demo", action="store_true",
help="engine mode with the built-in replay engine")
ap.add_argument("--loose", action="store_true",
help="engine mode: compare only core display keys")
ap.add_argument("--jobs", type=int, default=1)
ap.add_argument("--limit", type=int, default=0)
args = ap.parse_args()
if args.self_test:
op = gzip.open if args.dataset.endswith(".gz") else open
with op(args.dataset, "rt", encoding="utf-8") as f:
lines = [ln for ln in f if ln.strip()]
if args.limit:
lines = lines[:args.limit]
if args.jobs > 1:
import multiprocessing as mp
with mp.Pool(args.jobs) as pool:
results = pool.map(_self_one, lines, chunksize=16)
else:
results = [_self_one(ln) for ln in lines]
fails = [(g, e) for g, e, _ in results if e]
n_disc = sum(1 for _, _, tags in results if "gb_discrepancy" in tags)
print("self-test: %d/%d games PASS (%d known-discrepancy games included"
" — they pass: the judge is the ground truth)"
% (len(results) - len(fails), len(results), n_disc))
for g, e in fails[:20]:
print("FAIL %s: %s" % (g, e))
sys.exit(1 if fails else 0)
if args.demo or args.engine:
if args.demo:
eng = DemoReplayEngine()
else:
modname, clsname = args.engine.split(":")
mod = __import__(modname)
eng = getattr(mod, clsname)()
n = ok = 0
for rec in iter_records(args.dataset):
n += 1
if args.limit and n > args.limit:
n -= 1
break
if args.demo:
eng.bind(rec)
try:
run_engine(eng, rec, loose=args.loose)
ok += 1
except CheckFail as e:
print("FAIL %s: %s" % (rec["game_id"], e))
print("engine mode: %d/%d games reproduced exactly" % (ok, n))
sys.exit(0 if ok == n else 1)
ap.error("choose --self-test, --engine or --demo")
if __name__ == "__main__":
main()
|