code/scripts/build_selector_v8_orpo_pairs.py

"""
Build ORPO preference pairs for selector v8.

Each preference record: same prompt, chosen='YES' (correct candidate), rejected='NO' (wrong candidate).
Within each BIRD-train question, pair up (YES candidate, NO candidate) records from v8 SFT data
that share the SAME question + db.

Reads: data/sft_selector_v8_pointwise_enriched/train
Writes: data/sft_selector_v8_orpo/{train,test}

For ORPO trainer expected format:
  {"prompt": str, "chosen": "YES", "rejected": "NO", ...metadata}

Per Q, with N YES and M NO records, we can make N*M pairs. Cap to max_pairs_per_q for balance.
"""
import argparse, os, sys, random
from collections import defaultdict
os.environ.setdefault("PYTHONNOUSERSITE", "1")
ROOT = "/weka/s225250685/mats-tist"
sys.path.insert(0, ROOT)
from datasets import load_from_disk, Dataset, DatasetDict


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--sft", default=os.path.join(ROOT, "data/sft_selector_v8_pointwise_enriched"))
    ap.add_argument("--out", default=os.path.join(ROOT, "data/sft_selector_v8_orpo"))
    ap.add_argument("--max_pairs_per_q", type=int, default=4)
    args = ap.parse_args()

    rng = random.Random(42)
    dd = load_from_disk(args.sft)

    def make_pairs(rows):
        # Group by (question, db_id)
        groups = defaultdict(lambda: {"yes": [], "no": []})
        for r in rows:
            k = (r["question"], r["db_id"])
            (groups[k]["yes" if r["is_yes"] else "no"]).append(r)
        out = []
        for k, g in groups.items():
            if not g["yes"] or not g["no"]:
                continue
            rng.shuffle(g["yes"]); rng.shuffle(g["no"])
            pairs_emitted = 0
            for y in g["yes"]:
                for n in g["no"]:
                    if pairs_emitted >= args.max_pairs_per_q: break
                    out.append({
                        "prompt": y["prompt"],
                        "chosen": "YES",
                        "rejected": "NO",
                        "messages": [
                            {"role": "user", "content": y["prompt"]},
                            {"role": "assistant", "content": "YES"},
                        ],
                        "rejected_messages": [
                            {"role": "user", "content": n["prompt"]},
                            {"role": "assistant", "content": "NO"},
                        ],
                        "question": y["question"],
                        "db_id": y["db_id"],
                    })
                    pairs_emitted += 1
                if pairs_emitted >= args.max_pairs_per_q: break
        return out

    train_pairs = make_pairs(list(dd["train"]))
    test_pairs = make_pairs(list(dd["test"]))
    rng.shuffle(train_pairs); rng.shuffle(test_pairs)
    print(f"train pairs: {len(train_pairs)}  test pairs: {len(test_pairs)}")

    DatasetDict({
        "train": Dataset.from_list(train_pairs),
        "test": Dataset.from_list(test_pairs),
    }).save_to_disk(args.out)
    print(f"SAVED: {args.out}")


if __name__ == "__main__":
    main()