format_dataset.py

#!/usr/bin/env python3
"""
Format BothBosu scam-dialogue (and optionally Scammer-Conversation)
into standardized chat-template JSONL for SFT.

REQUIREMENTS:
  pip install datasets transformers

USAGE:
  python format_dataset.py --out_dir ./formatted_scam_data

OUTPUT:
  formatted_scam_data/
    train.jsonl   (chat-format messages per line)
    test.jsonl
    README.md     (dataset card fragment)

Each JSONL line:
  {"messages": [
    {"role": "system",    "content": "You are a phone scam detection expert."},
    {"role": "user",      "content": "Read this transcript...\n\n{transcript}"},
    {"role": "assistant", "content": "SCAM"}
  ]}
"""

import argparse
import json
from pathlib import Path

from datasets import load_dataset, concatenate_datasets


PROMPT_TEMPLATE = (
    "Read this phone call transcript and classify it:\n\n"
    "{transcript}\n\n"
    "Answer with exactly ONE word: SCAM or LEGITIMATE."
)
SYSTEM = "You are a phone scam detection expert."


def parse_args():
    p = argparse.ArgumentParser()
    p.add_argument("--primary", default="BothBosu/scam-dialogue")
    p.add_argument("--secondary", default="BothBosu/Scammer-Conversation",
                   help="Optional extra dataset to merge into train")
    p.add_argument("--out_dir", default="./formatted_scam_data")
    return p.parse_args()


def row_to_chat(row):
    """Convert a raw dataset row → ChatML dict."""
    answer = "SCAM" if row["label"] == 1 else "LEGITIMATE"
    # Handle different column names across datasets
    transcript = row.get("dialogue") or row.get("conversation")
    return {
        "messages": [
            {"role": "system",    "content": SYSTEM},
            {"role": "user",      "content": PROMPT_TEMPLATE.format(transcript=transcript)},
            {"role": "assistant", "content": answer},
        ]
    }


def save_jsonl(rows, path: Path):
    path.parent.mkdir(parents=True, exist_ok=True)
    with open(path, "w", encoding="utf-8") as f:
        for r in rows:
            f.write(json.dumps(r, ensure_ascii=False) + "\n")
    print(f"Saved {len(rows)} rows → {path}")


def main():
    args = parse_args()
    out_dir = Path(args.out_dir)

    # Load primary
    print(f"Loading primary dataset: {args.primary}")
    ds_train = load_dataset(args.primary, split="train")
    ds_test  = load_dataset(args.primary, split="test")

    # Optional secondary merge
    if args.secondary:
        try:
            ds_extra = load_dataset(args.secondary, split="train")
            n_before = len(ds_train)
            ds_train = concatenate_datasets([ds_train, ds_extra])
            print(f"Merged {args.secondary}: {n_before} → {len(ds_train)} train rows")
        except Exception as e:
            print(f"Skipped secondary dataset: {e}")

    # Convert
    train_rows = [row_to_chat(r) for r in ds_train]
    test_rows  = [row_to_chat(r) for r in ds_test]

    # Save
    save_jsonl(train_rows, out_dir / "train.jsonl")
    save_jsonl(test_rows,  out_dir / "test.jsonl")

    # Stats
    n_scam_train = sum(1 for r in train_rows if r["messages"][2]["content"] == "SCAM")
    n_scam_test  = sum(1 for r in test_rows  if r["messages"][2]["content"] == "SCAM")

    stats = {
        "train": {"total": len(train_rows), "scam": n_scam_train, "legit": len(train_rows) - n_scam_train},
        "test":  {"total": len(test_rows),  "scam": n_scam_test,  "legit": len(test_rows)  - n_scam_test},
    }

    (out_dir / "stats.json").write_text(json.dumps(stats, indent=2))
    print(f"\nStats:\n{json.dumps(stats, indent=2)}")

    # README fragment
    readme = f"""# Formatted Scam-Call Dataset (ChatML)

Generated by `format_dataset.py`.

## Sources
- Primary: {args.primary}
- Secondary: {args.secondary or "None"}

## Statistics
```json
{json.dumps(stats, indent=2)}
```

## Schema
Each `.jsonl` line is a ChatML message list compatible with TRL / Unsloth SFTTrainer.
"""
    (out_dir / "README.md").write_text(readme)
    print(f"\nDone. Output directory: {out_dir.absolute()}")


if __name__ == "__main__":
    main()