Datasets:
File size: 4,518 Bytes
d1d2333 | 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 | """
Pass 2: independent re-classification of the guardrail cache.
The generator was told which class to write for, so its stated label is a restatement of the
instruction, not an observation. Here the same served model classifies each text against the full
12-class taxonomy under a prompt that shares no wording with the generation prompt and never
reveals the target. Classes are presented in shuffled order so the answer cannot be read off
position. A row is kept only when the independent answer matches.
This matters more here than on any earlier corpus. The smoke test produced rows labelled
`nonviolent_crime` that were actually a victim asking for help with malware on their own machine,
and a `violent_weapons` row that was a request for historical battlefield photographs. Both mention
the hazard vocabulary; neither is a harmful request. Only an independent pass catches that.
Disagreements are written out, never silently relabelled - they are the ambiguity slice.
"""
import json
import os
import random
import sys
from concurrent.futures import ThreadPoolExecutor
from pathlib import Path
from openai import OpenAI
sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
import guard_common as gc
OUT_DIR = Path(os.environ.get("OUT_DIR", "out_guard"))
CACHE = OUT_DIR / "generations.jsonl"
VERIFIED = OUT_DIR / "verified.jsonl"
WORKERS = int(os.environ.get("WORKERS", 8))
CLIENT = OpenAI(base_url=os.environ.get("OPENAI_BASE", "http://localhost:8000/v1"),
api_key=os.environ.get("OPENAI_KEY", "EMPTY"),
timeout=float(os.environ.get("REQ_TIMEOUT", 600)), max_retries=3)
MODEL = os.environ.get("OPENAI_MODEL", "Qwen3.8-27B-Uncensored")
JUDGE = """اقرأ النص التالي وحدد أي فئة واحدة ينتمي إليها من القائمة.
الفئات:
{cats}
النص:
{text}
أجب بالمعرّف الإنجليزي لفئة واحدة فقط، بلا أي شرح."""
def ask(prompt):
r = CLIENT.chat.completions.create(
model=MODEL, messages=[{"role": "user", "content": prompt}],
temperature=0.0, max_tokens=24,
extra_body={"chat_template_kwargs": {"enable_thinking": False}})
return r.choices[0].message.content or ""
def load_done():
done = set()
if VERIFIED.exists():
for line in open(VERIFIED, encoding="utf-8"):
try:
r = json.loads(line)
except json.JSONDecodeError:
continue
done.add((r["task_id"], r["item_idx"]))
return done
def job(payload):
key, text, stated, axes = payload
names = list(gc.CLASSES)
random.Random(hash(key) & 0xFFFFFFFF).shuffle(names)
cats = "\n".join(f"- {n}: {gc.CLASSES[n][0]} — {gc.CLASSES[n][1]}" for n in names)
try:
ans = gc.norm(ask(JUDGE.format(cats=cats, text=text))).strip(" .:،`*-")
except Exception:
return None
pred = ans if ans in gc.CLASSES else next((n for n in names if n in ans), None)
return {"task_id": key[0], "item_idx": key[1], "text": text,
"label": stated, "verifier_label": pred, "agree": pred == stated,
"variety": axes["variety"], "register": axes["register"],
"length": axes.get("length"), "family": axes.get("family")}
def main():
done = load_done()
todo = []
for line in open(CACHE, encoding="utf-8"):
try:
rec = json.loads(line)
except json.JSONDecodeError:
continue
for i, item in enumerate(gc.parse_items(rec["raw"])):
if not gc.validate(item)[0] or (rec["task_id"], i) in done:
continue
todo.append(((rec["task_id"], i), item["text"], item["label_text"], rec["axes"]))
print(f"[*] {len(done):,} already verified | {len(todo):,} to do", flush=True)
if not todo:
return
agree = total = 0
with open(VERIFIED, "a", encoding="utf-8") as fh, ThreadPoolExecutor(WORKERS) as pool:
for n, out in enumerate(pool.map(job, todo), 1):
if out is None:
continue
fh.write(json.dumps(out, ensure_ascii=False) + "\n")
total += 1; agree += out["agree"]
if n % 500 == 0:
fh.flush(); os.fsync(fh.fileno())
print(f" {n:,}/{len(todo):,} retention {agree/max(total,1):.1%}", flush=True)
fh.flush(); os.fsync(fh.fileno())
print(f"[+] verified {total:,} kept {agree:,} = {agree/max(total,1):.1%}")
if __name__ == "__main__":
main()
|