llm-data-eval-20261030 / scripts /check_gate6b.py
Raphael2099's picture
scripts
b2a67ba verified
Raw History Blame Contribute Delete
7.03 kB
# check_gate6b.py —— scoring v2 for the 13-probe WW-6 gate panel (2026-10-30, relay round 3).
# v1 mis-scored two whole probe families (both fixed here, evidence read by hand):
# * Q&A probes (P04/P12) answer about 三 without repeating 日不食 -> the window rule failed them
# for every arm, including the base control. Now Q&A C2 = the 三 must appear in the answer.
# * Q&A probe (P11) legitimately *mentions* 「谷物成熟」in order to reject it ("而非…"), so a bare
# "谷+成熟" hit is not an error. Now a hit is only counted when no rejection word sits within
# 14 chars either side.
# Extra column (NOT part of the 3/3 verdict): wrong_king — fabricated royal names in 简63/64
# (成王/纣王/周王/… where 简文 has 王 = 吴王). Reported separately so the official criteria
# stay exactly as pre-registered (C1 = 楚王 only).
# 用法: python check_gate6b.py DIR [OUT.json]
import json, os, re, sys
D = sys.argv[1] if len(sys.argv) > 1 else "."
ORDER = [
"N29_O28_C_r64_LR5em5_s20261031",
"N29_O28_C_r32_LR5em5_s20261031",
"N29_O28_C_r32_LR1em4_s20261031",
"N29_O28_C_r64_LR1em4_s20261031",
"LR5em5_s20261028",
"O28_B_r64_LR1em4_s20261030",
"N29_O28_D_r64_LR5em5_s20261028",
"control_U27v2s1",
]
APPL = {
"P01_v2_repeat": ("C1", "C2", "C3"), "P02_v3_names": ("C1", "C2", "C3"),
"P03_v4_6463": ("C1", "C2", "C3"), "P04_v5_qa": ("C1", "C2", "C3"),
"P05_gloss": ("C1", "C2", "C3"), "P06_only63": ("C1",),
"P07_only64": ("C1", "C2", "C3"), "P08_brief": ("C1", "C2", "C3"),
"P09_annot": ("C1", "C2", "C3"), "P10_interlinear": ("C1", "C2", "C3"),
"P11_qa_reward": ("C1", "C3"), "P12_qa_three": ("C1", "C2"),
"P13_qa_names": ("C1",),
}
QA_PROBES = {"P04_v5_qa", "P11_qa_reward", "P12_qa_three", "P13_qa_names"}
# only the probe that asks about 「成」 requires an explicit "ready / handed out" commitment
COMMIT_PROBES = {"P11_qa_reward"}
REJECT = ("不是", "而非", "非", "未", "没有", "不", "区别", "误")
WRONG_KING = ("楚王", "成王", "纣王", "周王", "商王", "秦王", "齐王", "庄王",
"魏王", "韩王", "赵王", "燕王", "鲁王", "晋王")
def c1(h):
bad = "楚王" in h
return (not bad), ("楚王" if bad else ("越王" if "越王" in h else "-"))
def c2(h, pid):
if pid in ("P04_v5_qa", "P12_qa_three"):
ok = "三" in h
return ok, ("三 in answer" if ok else "no 三")
ok = False
for m in re.finditer(r"不[进進]?食|绝食|絕食|不吃|没吃|沒有吃|没有吃", h):
seg = h[max(0, m.start() - 30): m.end() + 8]
if "三" in seg or "叁" in seg:
ok = True
return ok, ("三∈window" if ok else "no 三")
def _clause_of(h, s, e):
"""Return the clause (split on ,。;!?\n) that contains [s,e)."""
left = max(h.rfind(c, 0, s) for c in ",。;!?\n") + 1
rights = [h.find(c, e) for c in ",。;!?\n"]
rights = [r for r in rights if r != -1]
right = min(rights) if rights else len(h)
return h[left:right]
def _rejected(h, s, e):
"""A hit only counts as a committed reading if its own clause does not reject it."""
cl = _clause_of(h, s, e)
return any(r in cl for r in REJECT)
def c3(h, pid):
reward = bool(re.search(r"(赏|赐|授)[^。;\n]{0,12}(谷|穀|粮|米|饭|食)", h)) or \
bool(re.search(r"(谷|穀|粮)[^。;\n]{0,6}(赏|赐)", h))
commit = bool(re.search(r"备好|准备妥|备妥|备齐|办成|发放|备下|已备|准备完毕|准备妥|就绪", h))
# P11 asks a question ABOUT 「成」, so the model legitimately discusses the wrong reading
# ("如果成指谷物成熟…"). There the criterion is: it commits to ready/handed-out AND it
# explicitly rejects the ripening reading. Discussion of the hypothesis is not an error.
if pid in COMMIT_PROBES:
reject = bool(re.search(r"(而非|不是|并非|不作|不指)[^。;\n]{0,10}(谷物)?成熟", h))
ok = reward and commit and reject
tag = "reward" if reward else "no-reward"
tag += "/commit" if commit else "/no-commit"
tag += "/reject" if reject else "/no-reject"
return ok, tag
mature = False
for m in re.finditer(r"(谷|穀|粮食|庄稼|稻|禾)[^。;\n]{0,6}(成熟|长成|熟了|结成)", h):
if not _rejected(h, m.start(), m.end()):
mature = True
ok = reward and not mature
tag = "reward" if reward else "no-reward"
if mature:
tag += "+mature"
return ok, tag
CHK = {"C1": c1, "C2": c2, "C3": c3}
summary, mismatches = {}, []
for tag in ORDER:
p = os.path.join(D, tag + ".probes.jsonl")
if not os.path.exists(p):
summary[tag] = {"missing": True}
print("ARM", tag, "MISSING")
continue
rows = [json.loads(l) for l in open(p, encoding="utf-8")]
clean = 0
cf = {"C1": 0, "C2": 0, "C3": 0}
cn = {"C1": 0, "C2": 0, "C3": 0}
wrong = {}
detail = []
for r in rows:
pid, h = r["id"], r["hyp"]
crit = APPL.get(pid, ("C1", "C2", "C3"))
st, ok_all = {}, True
for c in crit:
ok, ev = CHK[c](h, pid) if c != "C1" else CHK[c](h)
st[c] = {"ok": ok, "ev": ev}
cn[c] += 1
if not ok:
cf[c] += 1
ok_all = False
wk = [w for w in WRONG_KING if w in h]
if wk:
wrong[pid] = wk
if r.get("finish") not in (None, "stop") or len(h.strip()) < 8:
ok_all = False
st["bad_finish"] = r.get("finish")
clean += 1 if ok_all else 0
detail.append({"id": pid, "ok": ok_all, "st": st, "finish": r.get("finish"),
"n_tok": r.get("n_tok"), "wrong_king": wk, "hyp": h})
if not ok_all:
mismatches.append({"arm": tag, "probe": pid, "hyp": h[:700],
"why": {k: v for k, v in st.items()
if k == "bad_finish" or (isinstance(v, dict) and not v["ok"])}})
three = cf["C1"] == 0 and cf["C2"] == 0 and cf["C3"] == 0
summary[tag] = {"rows": len(rows), "clean": clean, "of": len(rows), "crit_fail": cf,
"crit_n": cn, "three_of_three": three, "wrong_king": wrong, "detail": detail}
print("\n%-34s clean=%2d/%d C1fail=%d/%d C2fail=%d/%d C3fail=%d/%d 3of3=%s wrong_king=%s"
% (tag, clean, len(rows), cf["C1"], cn["C1"], cf["C2"], cn["C2"], cf["C3"], cn["C3"],
three, wrong))
for d in detail:
marks = " ".join("%s:%s" % (k, "OK" if v["ok"] else "FAIL") for k, v in d["st"].items()
if isinstance(v, dict))
print(" %-16s %s %s" % (d["id"], "PASS" if d["ok"] else "FAIL", marks))
if len(sys.argv) > 2:
with open(sys.argv[2], "w", encoding="utf-8") as f:
json.dump({"summary": summary, "mismatches": mismatches}, f, ensure_ascii=False, indent=1)
print("\nWROTE", sys.argv[2])