File size: 15,152 Bytes
56a0f31 | 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 | #!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""qa_sample.py —— 考释问答 v1 · 采样(PLAN v3 §2.2「考释问答」行)。CPU only。
池:KB addon_slips_v1 的 lib A 出土简块(meta.split=="train", era==EXCAVATED),汉字 15–220。
排除(按作品,先于一切):四篇(王家嘴孔子曰、枣纸简吴王夫差/齐桓公、安大仲尼曰、QYCF14)、演示 76 简(核心包 slips.jsonl)、
A_dev_clean 441 族(split=dev 诸族)、越公其事(holdout)、B_dev_eval 出土同族(睡虎地/马王堆老子+卷前后古佚书/孙膑/孙子/尉缭子/郭店老子)、
9 月 k27 已训练的 3,119 支(按 sample_id;另以 12-gram/包含 挡其它版本的同文块)。
泄漏门:W/scripts/leak_gate.py(gatesets_v1 全部 9 个门集 + 演示 76 简;12-gram / 6-gram 覆盖≥0.3 / ≥6 字包含,双向)。
问题形式:每支简按 (seed 20261023, id) 独立随机;F2(「X」字)只在简内有通假/待考括注字时可选,否则在 F1/F3 中取。
输出 W/data/qa/{slips_v1.jsonl, queries_v1.jsonl, SAMPLE_STATS.json}
"""
import collections, glob, hashlib, json, lzma, math, os, random, re, sqlite3, sys
W = "/root/autodl-tmp/llm_rag_20261023"
sys.path.insert(0, os.path.join(W, "scripts"))
from leak_gate import Gate, norm1, norm2 # noqa: E402
I = "/root/autodl-tmp/llm_rag_20260919/index"
SEED = 20261023
TARGET = int(os.environ.get("TARGET", "8000"))
MIN_HAN, MAX_HAN = 15, 220
CAP_FRAC = 0.04
OUT = os.path.join(W, "data", "qa")
DEMO_SLIPS = os.path.join(W, "vendor", "release_demo_core_v1_2", "data", "slips.jsonl")
K27_DIR = "/root/autodl-tmp/_from_bjb1b/llm_sft_20260918/kaoshi27b/data"
GATESETS = os.path.join(W, "data", "gatesets_v1.jsonl.xz")
TITLE_PAT = ("孔子曰", "仲尼曰", "吴王夫差", "吳王夫差", "齐桓公", "齊桓公", "QYCF", "群玉", "越公其事", "yuegongqishi")
FAMILY_EXCL = {"GD_六位", "SB_柬大王泊旱", "SB_陳公治兵", "SB_融師有成氏", "SB_相邦之道", "QH_摄命", "QH_攝命",
"QH_保训", "QH_保訓", "GD_魯穆公問子思", "QH06_yuegongqishi"}
BDEV_PAT = ("睡虎地", "古佚書", "古佚书", "孫臏", "孙膑", "老子", "孫子兵法", "孙子兵法", "尉繚子", "尉缭子")
EXTRA_STORES = [ # 其余 KB lib A 出土简分片(无 split 字段:训练划分 = 按作品排除 dev/holdout 族);layer 优先序:带通假括注的层在前
("addon_slips_qh15_v1", ("loose", "strict")),
("addon_slips_qhv3_gdll_v1", ("strict", "loose")),
("addon_slips_sbv4_v1", ("strict", "loose")),
("addon_kaoshi_shiwen6_v1", ("strict", "原简释文", "loose")),
("addon_slips_cd14_v1", ("釋文",)),
]
PIAN_EXCL = ("六位", "六德", "柬大王泊旱", "陈公治兵", "融师有成氏", "相邦之道", "摄命", "保训", "鲁穆公问子思", "越公其事")
def norm_text(t, store=""):
t = t or ""
t = re.sub(r"⟦(?:glyph|region):[^⟧]*⟧", "[?]", t)
t = re.sub(r"⟦sym:(?:chongwen|hewen)[^⟧]*⟧", "=", t)
t = re.sub(r"⟦[^⟧]*⟧", "", t)
t = re.sub(r"〔[一二三四五六七八九十百〇零]+〕", "", t)
t = re.sub(r"\s*\$\^\{\[\d+\]\}\$", "", t)
t = re.sub(r"\s*\[\d+\]", "", t)
t = re.sub(r"\(([^()]{1,6})\)", "(\\1)", t)
t = re.sub(r"[ \t\u3000]+", "", t).strip()
if store == "addon_slips_cd14_v1":
t = re.sub(r"(\d+)$", "【\\1】", t)
return t
HAN = re.compile(r"[㐀-鿿\U00020000-\U0002ffff]")
# 通假/待考括注:单字 X(Y);Y 1–4 字,可带 ?
ANN = re.compile(r"([㐀-鿿\U00020000-\U0002ffff])(([^()]{1,6}))")
COMMON = set("亓其又有才在女汝如母毋隹惟唯曰之而於于也者") # 极常见括注:F2 尽量不选
RE_SLIPNO = re.compile(r"【[^】]*】")
RE_BRACE = re.compile(r"\{[^}]*\}")
PUNCT = set(",。、;:?!「」『』“”‘’《》〈〉…—-·,.;:?!\"' \t\n")
def hkey(*p):
return hashlib.sha256(("%d|" % SEED + "|".join(str(x) for x in p)).encode()).hexdigest()
def nhan(s):
return len(HAN.findall(s or ""))
def strip_punct(s):
return "".join(c for c in (s or "") if c not in PUNCT)
def make_query(text_nopunct):
q = RE_SLIPNO.sub("", text_nopunct)
q = RE_BRACE.sub("", q)
q = re.sub(r"[((]([^()()]{1,4})[))]", "", q) # 宽式里偶见的括注:保留本字
q = q.replace("[?]", "").replace("○", "").replace("□", "").replace("〼", "").replace("■", "")
q = re.sub(r"[〔〕\[\]]", "", q)
return re.sub(r"\s+", "", q)[:120]
def ann_chars(*texts):
out = []
for t in texts:
for m in ANN.finditer(t or ""):
x, y = m.group(1), m.group(2)
if not HAN.search(y) and "?" not in y and "?" not in y:
continue
if x == y:
continue
kind = "daikao" if ("?" in y or "?" in y) else "tongjia"
out.append((x, y, kind))
seen, uniq = set(), []
for a in out:
if a[0] not in seen:
seen.add(a[0]); uniq.append(a)
return uniq
def main():
os.makedirs(OUT, exist_ok=True)
drops = collections.Counter()
for k in ("total", "not_train_split", "not_excavated", "han_out_of_range", "title_four_texts",
"family_dev_holdout", "family_bdev_excavated", "k27_sample_id", "k27_text", "gate_eval",
"dup_text"):
drops[k] = 0
# ── 门集
gate = Gate()
gs_n = collections.Counter()
with lzma.open(GATESETS, "rt", encoding="utf-8") as f:
for l in f:
r = json.loads(l)
gate.add(r["set"], r.get("id"), r.get("text") or "")
gs_n[r["set"]] += 1
for l in open(DEMO_SLIPS, encoding="utf-8"):
r = json.loads(l)
for k in ("text_loose_punct", "text_strict_punct"):
if r.get(k):
gate.add("demo76", r["slip_key"], r[k])
gs_n["demo76"] += 1
# ── k27(9 月已训练 3,119 支)
k27_ids, k27_gate = set(), Gate()
for l in open(os.path.join(K27_DIR, "prompts.jsonl"), encoding="utf-8"):
k27_ids.add(json.loads(l)["id"])
for l in open(os.path.join(K27_DIR, "slips_sample.dedup.jsonl"), encoding="utf-8"):
u = json.loads(l)
if u["id"] in k27_ids:
k27_gate.add("k27", u["id"], u.get("text_punct") or u.get("text_nopunct") or "")
# ── 池
con = sqlite3.connect("file:%s/addon_slips_v1/meta.sqlite?mode=ro" % I, uri=True)
cands, seen = [], {}
gate_hits = collections.Counter()
for cid, text, cite, tier, quotable, meta in con.execute(
"select cid,text,cite,tier,quotable,meta from chunks where lib='A' order by rowid"):
drops["total"] += 1
m = json.loads(meta)
if m.get("split") != "train":
drops["not_train_split"] += 1; continue
if m.get("era") != "EXCAVATED":
drops["not_excavated"] += 1; continue
n = nhan(text)
if not (MIN_HAN <= n <= MAX_HAN):
drops["han_out_of_range"] += 1; continue
fam = m.get("family_id") or ""
title_blob = " ".join(str(m.get(k) or "") for k in ("pian", "pian_raw", "book", "corpus_name", "doc_key")) + " " + fam + " " + (cite or "")
if any(p in title_blob for p in TITLE_PAT):
drops["title_four_texts"] += 1; continue
if fam in FAMILY_EXCL:
drops["family_dev_holdout"] += 1; continue
if any(p in title_blob for p in BDEV_PAT):
drops["family_bdev_excavated"] += 1; continue
if m.get("sample_id") and m["sample_id"] in k27_ids:
drops["k27_sample_id"] += 1; continue
strict = m.get("text_strict") or ""
rej, _ = k27_gate.check(text)
if rej:
drops["k27_text"] += 1; continue
rej, info = gate.check(text + "\n" + strict) if False else gate.check(text)
if not rej and strict:
rej, info = gate.check(strict)
if rej:
drops["gate_eval"] += 1
for s in info["sets"]:
gate_hits[s] += 1
continue
key = norm2(text)
if key in seen:
drops["dup_text"] += 1; continue
seen[key] = cid
cands.append((cid, text, cite, tier, quotable, m, strict, n))
# ── 其余 lib A 分片
from leak_gate import t2s as _t2s
extra_ct = collections.Counter()
for store, layers in EXTRA_STORES:
con2 = sqlite3.connect("file:%s/%s/meta.sqlite?mode=ro" % (I, store), uri=True)
rows2 = []
for cid, text, cite, tier, quotable, meta in con2.execute(
"select cid,text,cite,tier,quotable,meta from chunks where lib='A' order by rowid"):
m = json.loads(meta)
layer = m.get("layer") or m.get("section_kind") or m.get("part") or ""
if layer not in layers:
continue
rows2.append((layers.index(layer), cid, text, cite, tier, quotable, m, layer))
rows2.sort(key=lambda x: x[0])
seen_slip = set()
for _, cid, text, cite, tier, quotable, m, layer in rows2:
drops["total"] += 1; extra_ct[(store, "in")] += 1
t = norm_text(text, store)
n = nhan(t)
if not (MIN_HAN <= n <= MAX_HAN) or t.count("[?]") > 0.25 * max(n, 1):
drops["han_out_of_range"] += 1; continue
fam = m.get("doc_key") or ""
blob = _t2s(" ".join(str(m.get(k) or "") for k in ("pian", "book", "doc_key", "corpus_name", "site")) + " " + (cite or ""))
if any(p in blob for p in TITLE_PAT):
drops["title_four_texts"] += 1; continue
if any(p in blob for p in PIAN_EXCL):
drops["family_dev_holdout"] += 1; continue
if any(p in blob for p in BDEV_PAT) or any(_t2s(p) in blob for p in BDEV_PAT):
drops["family_bdev_excavated"] += 1; continue
sk = (fam, str(m.get("slip_no") or m.get("page") or cid))
if sk in seen_slip:
drops["dup_text"] += 1; continue
rej, _ = k27_gate.check(t)
if rej:
drops["k27_text"] += 1; continue
rej, info = gate.check(t)
if rej:
drops["gate_eval"] += 1
for s_ in info["sets"]:
gate_hits[s_] += 1
continue
key = norm2(t)
if key in seen:
drops["dup_text"] += 1; continue
seen[key] = cid; seen_slip.add(sk)
m2 = dict(m, family_id=fam, corpus_name=m.get("corpus_name") or m.get("corpus"), split="(family-level)", layer=layer, store=store)
cands.append((cid, t, cite, tier, quotable, m2, t, n))
extra_ct[(store, "kept")] += 1
print("EXTRA", {"%s:%s" % k: v for k, v in sorted(extra_ct.items())}, file=sys.stderr)
# ── 按作品轮转
by_fam = collections.defaultdict(list)
for c in cands:
by_fam[c[5].get("family_id") or c[0]].append(c)
for f in by_fam:
by_fam[f].sort(key=lambda c: hkey("order", c[0]))
fams = sorted(by_fam, key=lambda f: (hkey("fam", f), f))
cap = max(1, int(math.ceil(CAP_FRAC * TARGET)))
picked, idx = [], {f: 0 for f in fams}
while len(picked) < TARGET:
prog = False
for f in fams:
if len(picked) >= TARGET:
break
i = idx[f]
if i >= len(by_fam[f]) or i >= cap:
continue
picked.append(by_fam[f][i]); idx[f] = i + 1; prog = True
if not prog:
break
# ── 落盘
units, queries = [], []
form_ct, form_elig = collections.Counter(), collections.Counter()
for cid, text, cite, tier, quotable, m, strict, n in picked:
if norm2(strict) == norm2(text) or strict == text:
strict = "" # 其余分片单层:不重复给严式块
nopunct = strip_punct(text)
strict_np = strip_punct(strict)
use_strict_cv = bool(strict) and int(hkey("cv", cid)[:8], 16) % 1000 < 300
anns = ann_chars(strict, text)
rng = random.Random("%d|%s" % (SEED, cid))
forms = ["F1_shidu", "F2_zi", "F3_duidu"] if anns else ["F1_shidu", "F3_duidu"]
if anns:
form_elig["F2_eligible"] += 1
form = rng.choice(forms)
x = None
if form == "F2_zi":
pool = [a for a in anns if a[2] == "daikao"] or [a for a in anns if a[0] not in COMMON] or anns
x = rng.choice(pool)
form_ct[form] += 1
corpus_cn = m.get("corpus_name") or m.get("corpus") or ""
u = {"id": cid, "store": m.get("store") or "addon_slips_v1", "layer": m.get("layer"), "corpus": m.get("corpus"), "corpus_cn": corpus_cn, "book": m.get("book"),
"pian": m.get("pian"), "family_id": m.get("family_id"), "kb_tier": tier, "kb_quotable": quotable,
"sample_id": m.get("sample_id"), "n_han": n,
"text_punct": text, "text_nopunct": nopunct,
"text_strict_punct": strict, "text_strict": strict_np,
"cite": cite, "cv_text": strict_np if use_strict_cv else nopunct,
"cv_source": "strict_nopunct" if use_strict_cv else "loose_nopunct",
"verdict": "confirmed", "form": form,
"x_char": x[0] if x else None, "x_read": x[1] if x else None, "x_kind": x[2] if x else None,
"n_ann": len(anns)}
units.append(u)
queries.append({"id": cid, "query": make_query(nopunct) or nopunct[:120], "verdict": "confirmed", "batch": ""})
with open(os.path.join(OUT, "slips_v1.jsonl"), "w", encoding="utf-8") as f:
for u in units:
f.write(json.dumps(u, ensure_ascii=False) + "\n")
with open(os.path.join(OUT, "queries_v1.jsonl"), "w", encoding="utf-8") as f:
for q in queries:
f.write(json.dumps(q, ensure_ascii=False) + "\n")
per_fam = collections.Counter(u["family_id"] for u in units)
stats = {"seed": SEED, "target": TARGET, "n_units": len(units), "n_candidates_after_guards": len(cands),
"pool": "addon_slips_v1 lib A split=train era=EXCAVATED", "han_range": [MIN_HAN, MAX_HAN],
"family_cap": cap, "dropped": dict(drops), "gate_hits_by_set": dict(gate_hits),
"gatesets_rows": dict(gs_n), "k27_ids": len(k27_ids),
"form_counts": dict(form_ct), "form_eligibility": dict(form_elig),
"by_corpus": dict(collections.Counter(u["corpus_cn"] for u in units)),
"by_book": dict(collections.Counter(u["book"] for u in units)),
"by_tier": dict(collections.Counter(u["kb_tier"] for u in units)),
"by_store": dict(collections.Counter(u["store"] for u in units)),
"n_families": len(per_fam), "top_families": per_fam.most_common(15),
"n_families_capped": sum(1 for f in by_fam if len(by_fam[f]) > cap)}
json.dump(stats, open(os.path.join(OUT, "SAMPLE_STATS.json"), "w", encoding="utf-8"), ensure_ascii=False, indent=1)
print(json.dumps(stats, ensure_ascii=False, indent=1))
if __name__ == "__main__":
main()
|