File size: 4,189 Bytes
1b2d52d | 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 | """
Generate NLI (premise, hypothesis, label) triples for fine-tuning
typeform/distilbert-base-uncased-mnli on legal clause classification.
Sources:
- scripts/train_mlp._DATA (160 synthetic full-text clause sentences)
- data/clause_training_data.csv (keyword-level examples)
Output: data/nli_training_data.jsonl
Usage:
cd ldv-backend && python3 scripts/generate_nli_training_data.py
"""
from __future__ import annotations
import csv
import json
import sys
from pathlib import Path
sys.path.insert(0, str(Path(__file__).parent.parent))
from scripts.train_mlp import _DATA as SYNTH_DATA # noqa: E402
# Hypothesis → expected NLI label for each clause class
_HYPOTHESES: dict[str, list[tuple[str, str]]] = {
"abusive_clause": [
("This clause is abusive and heavily favors one party.", "entailment"),
("This is a balanced and fair contractual provision.", "contradiction"),
],
"payment_risk": [
("This clause imposes excessive payment penalties or interest rates.", "entailment"),
("This clause describes standard and reasonable payment terms.", "contradiction"),
],
"missing_mandatory": [
("This clause is incomplete, a placeholder, or missing key terms.", "entailment"),
("This clause is clearly defined and complete.", "contradiction"),
],
"normal": [
("This clause is abusive and heavily favors one party.", "contradiction"),
("This clause imposes excessive payment penalties or interest rates.", "contradiction"),
("This clause is incomplete, a placeholder, or missing key terms.", "contradiction"),
("This is a standard and complete contractual provision.", "entailment"),
],
}
def _from_master_reasons(datasets_dir: Path) -> list[dict]:
"""Pull Reason text from dangerous_clauses_MASTER.csv as abusive_clause premises."""
rows: list[dict] = []
path = datasets_dir / "dangerous_clauses_MASTERv2.csv"
if not path.exists():
path = datasets_dir / "dangerous_clauses_MASTER.csv" # ponytail: fallback to v1
if not path.exists():
return rows
with open(path, newline="", encoding="utf-8-sig") as f:
for row in csv.DictReader(f):
reason = (row.get("Reason") or "").strip()
if not reason or len(reason) < 20:
continue
for hyp, nli_label in _HYPOTHESES["abusive_clause"]:
rows.append({"premise": reason, "hypothesis": hyp, "label": nli_label})
return rows
def generate(out_path: Path, csv_path: Path, datasets_dir: Path | None = None) -> None:
rows: list[dict] = []
for text, label in SYNTH_DATA:
for hyp, nli_label in _HYPOTHESES.get(label, []):
rows.append({"premise": text, "hypothesis": hyp, "label": nli_label})
if csv_path.exists():
with open(csv_path, newline="", encoding="utf-8") as f:
for row in csv.DictReader(f):
text = row.get("text", "").strip()
label = row.get("label", "").strip()
if not text or label not in _HYPOTHESES:
continue
for hyp, nli_label in _HYPOTHESES[label]:
rows.append({"premise": text, "hypothesis": hyp, "label": nli_label})
if datasets_dir:
master_rows = _from_master_reasons(datasets_dir)
rows.extend(master_rows)
if master_rows:
print(f" +{len(master_rows)} rows from dangerous_clauses_MASTER.csv (Reason field)")
out_path.parent.mkdir(parents=True, exist_ok=True)
with open(out_path, "w", encoding="utf-8") as f:
for row in rows:
f.write(json.dumps(row, ensure_ascii=False) + "\n")
counts: dict[str, int] = {}
for r in rows:
counts[r["label"]] = counts.get(r["label"], 0) + 1
print(f"Wrote {len(rows)} NLI triples → {out_path}")
for k in sorted(counts):
print(f" {k}: {counts[k]}")
if __name__ == "__main__":
base = Path(__file__).parent.parent
generate(
out_path=base / "data" / "nli_training_data.jsonl",
csv_path=base / "data" / "clause_training_data.csv",
datasets_dir=base.parent / "datasets",
)
|