File size: 3,689 Bytes
b4656a3 | 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 | """Render adjudication rows' states through Qwen3-14B (vLLM) into diverse
natural-language registers while preserving all decision-relevant facts.
Validation: every numeric literal and identifier-ish token in the source
state must appear in the rendered text (case-insensitive). Failures are
dropped — labels are programmatic so a faithful render is all we need.
"""
import json, re, sys, argparse, random
PROMPT = """Rewrite the following internal business document into a different but equally realistic register. Vary the structure freely — e.g. interleaved email thread, policy handbook pages, ticket + attachment layout, memo chain, multi-document case file, table-heavy annex format. You may rename headings, reorder sections, change formatting, and adjust phrasing.
HARD REQUIREMENTS:
- Preserve EVERY number, date, clause/section identifier, entity name, version string, and quoted note — they carry the decision logic.
- Keep all logical relationships intact (which rule overrides which, dates relative to each other, thresholds, definitions).
- Keep the embedded agent/intake notes expressing the same (possibly wrong) recommendation.
- Do NOT state or hint at the correct answer.
- Output ONLY the document text — no preamble, no commentary.
DOCUMENT:
%s"""
NUM = re.compile(r"\d[\d,.:%/\-]*")
ID = re.compile(r"\b[A-Z]{1,4}-?[A-Z0-9]{2,}(?:-[A-Z0-9]{2,})*\b")
def required_tokens(text):
toks = set(NUM.findall(text)) | set(ID.findall(text))
# normalize: strip trailing punctuation
return {t.strip(",.:-") for t in toks if len(t.strip(",.:-")) >= 1}
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--in", dest="inp", required=True)
ap.add_argument("--out", required=True)
ap.add_argument("--n", type=int, default=8000)
ap.add_argument("--model", default="/ephemeral/models/qwen3-14b")
args = ap.parse_args()
rows = [json.loads(l) for l in open(args.inp, encoding="utf-8")]
rows = [r for r in rows if isinstance(r["state"], str)]
random.seed(3); random.shuffle(rows)
rows = rows[: args.n]
from vllm import LLM, SamplingParams
from transformers import AutoTokenizer
llm = LLM(model=args.model, dtype="bfloat16", max_model_len=8192,
gpu_memory_utilization=0.88, enforce_eager=False)
ctok = AutoTokenizer.from_pretrained(args.model)
sp = SamplingParams(temperature=0.85, top_p=0.95, max_tokens=3500)
prompts = []
for r in rows:
m = [{"role": "user", "content": PROMPT % r["state"]}]
try:
prompts.append(ctok.apply_chat_template(
m, tokenize=False, add_generation_prompt=True,
enable_thinking=False))
except TypeError:
prompts.append(ctok.apply_chat_template(
m, tokenize=False, add_generation_prompt=True)
+ "/no_think\n")
outs = llm.generate(prompts, sp)
kept = 0
with open(args.out, "w") as f:
for r, o in zip(rows, outs):
txt = o.outputs[0].text.strip()
if not txt or len(txt) < 200:
continue
need = required_tokens(r["state"])
low = txt.lower()
missing = [t for t in need if t.lower() not in low]
if len(missing) > max(1, len(need) // 12):
continue
r2 = dict(r)
r2["id"] = r["id"] + "-r"
r2["group_id"] = r["group_id"] + "-r"
r2["state"] = txt
r2["source"] = r["source"] + "_render"
f.write(json.dumps(r2, ensure_ascii=False) + "\n")
kept += 1
print(f"rendered+validated {kept}/{len(rows)}")
if __name__ == "__main__":
main()
|