llm-data-eval-20261030 / scripts /check_gate6.py
Raphael2099's picture
scripts
b2a67ba verified
Raw History Blame Contribute Delete
4.81 kB
# check_gate6.py —— score the 13-probe WW-6 gate panel per arm, per criterion.
# C1 fail : 楚王 appears (越王 must not be rewritten)
# C2 pass : the 三 of slip 63 is carried into slip 64's 日不食 clause (三步/三天 window)
# C3 pass : 赏谷 read as awarded grain ready/handed out, and NOT as grain ripening
# 用法: python check_gate6.py [OUT.json]
import json, os, re, sys
D = "/root/autodl-tmp/llm_rag_20261023/runs/G_baselines/_ww6_gate_panel_20261030"
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",
]
NEG = ("不是", "非", "未", "没有", "不")
# applicability by probe id (the generator's "crit" key is not forwarded by vllm_gen_multi_lora.py)
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",),
}
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):
san = False
for m in re.finditer(r"不食|绝食|不吃|没吃|没有吃", h):
seg = h[max(0, m.start() - 30): m.end() + 8]
if "三" in seg or "叁" in seg:
san = True
return san, ("三∈window" if san else "no 三")
def c3(h):
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))
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|已经成|结成)", h):
pre = h[max(0, m.start() - 12): m.start()]
if not any(n in pre for n in NEG):
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 = {}
mismatch = []
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", p)
continue
rows = [json.loads(l) for l in open(p, encoding="utf-8")]
per = {}
clean = 0
crit_fail = {"C1": 0, "C2": 0, "C3": 0}
crit_n = {"C1": 0, "C2": 0, "C3": 0}
detail = []
for r in rows:
pid = r["id"]
crit = APPL.get(pid, ("C1", "C2", "C3"))
hyp = r["hyp"]
st = {}
ok_all = True
for c in crit:
ok, ev = CHK[c](hyp)
st[c] = {"ok": ok, "ev": ev}
crit_n[c] += 1
if not ok:
crit_fail[c] += 1
ok_all = False
# a probe that answers something else entirely (empty / cut) is not a clean pass
if r.get("finish") not in (None, "stop") or len(hyp.strip()) < 8:
ok_all = False
st["bad_finish"] = r.get("finish")
if ok_all:
clean += 1
detail.append({"id": pid, "ok": ok_all, "st": st, "finish": r.get("finish"),
"n_tok": r.get("n_tok"), "hyp": hyp})
if not ok_all:
mismatch.append({"arm": tag, "probe": pid,
"why": {k: v for k, v in st.items() if k == "bad_finish" or (isinstance(v, dict) and not v["ok"])},
"hyp": hyp[:600]})
per = {"rows": len(rows), "clean": clean, "of": len(rows),
"crit_fail": crit_fail, "crit_n": crit_n,
"three_of_three": (crit_fail["C1"] == 0 and crit_fail["C2"] == 0 and crit_fail["C3"] == 0),
"detail": detail}
summary[tag] = per
print("\n%s clean=%d/%d C1fail=%d/%d C2fail=%d/%d C3fail=%d/%d 3of3=%s" % (
tag, clean, len(rows), crit_fail["C1"], crit_n["C1"], crit_fail["C2"], crit_n["C2"],
crit_fail["C3"], crit_n["C3"], per["three_of_three"]))
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) > 1:
with open(sys.argv[1], "w", encoding="utf-8") as f:
json.dump({"summary": summary, "mismatches": mismatch}, f, ensure_ascii=False, indent=1)
print("\nWROTE", sys.argv[1])