code/scripts/build_fixer_orpo_iter2_conservative.py

"""Build fixer ORPO iter-2 data biased toward CONSERVATIVE behavior.

Source: BIRD-TRAIN 3-stage rollouts (multiple files merged).
- Bad flips (P correct, F wrong): chosen=planner_sql, rejected=fixed_sql  → teaches "don't mangle correct SQL"
- Good flips (P wrong, F correct): chosen=fixed_sql, rejected=planner_sql → teaches "do fix when needed"
- Same-correct synthetic pairs: when both P and F end up correct but F differs from P, use planner_sql as chosen
  (slight preference for the simpler / closer-to-input SQL).

Total: targeting ~500-1000 pairs.

Output: data/llm_alignment/scaleup_iter2_v2/hf_fixer_conservative
"""
import json
import os
from datasets import Dataset, DatasetDict

OUT_DIR = "/home/datht/mats-sql-tist/data/llm_alignment/scaleup_iter2_v2/hf_fixer_conservative"

SRC_PATHS = [
    "/home/datht/mats-sql-tist/data/rollouts/bird_train_3stage_K4.jsonl",
    "/home/datht/mats-sql-tist/data/rollouts/scaleup_bird_train_3stage_K4.jsonl",
    "/home/datht/mats-sql-tist/data/rollouts/iter2_bird_train_3stage_K8.jsonl",
]


def normalize_sql(sql):
    return " ".join(sql.split()).lower().strip()


def main():
    bad_flip_pairs = []
    good_flip_pairs = []
    same_correct_pairs = []
    seen_prompts = set()  # dedup across files

    for p in SRC_PATHS:
        if not os.path.exists(p):
            continue
        with open(p) as f:
            for line in f:
                s = json.loads(line)
                for t in s.get("trajectories", []):
                    pc = t.get("is_planner_correct", False)
                    fc = t.get("is_fixed_correct", False)
                    fixer_prompt = (t.get("fixer_prompt") or "").strip()
                    planner_sql = (t.get("planner_sql") or "").strip()
                    fixed_sql = (t.get("fixed_sql") or "").strip()
                    if not fixer_prompt or not planner_sql or not fixed_sql:
                        continue
                    if normalize_sql(planner_sql) == normalize_sql(fixed_sql):
                        continue
                    # Dedup by prompt
                    key = fixer_prompt[:1000] + "|" + planner_sql[:200]
                    if key in seen_prompts:
                        continue
                    seen_prompts.add(key)

                    chosen_planner = f"```sql\n{planner_sql}\n```"
                    chosen_fix = f"```sql\n{fixed_sql}\n```"
                    base = {
                        "prompt": fixer_prompt,
                        "db_path": s.get("db_path", ""),
                        "question": s.get("question", ""),
                        "db_id": s.get("db_id", ""),
                    }
                    if pc and (not fc):
                        # bad flip: prefer planner
                        bad_flip_pairs.append({**base, "chosen": chosen_planner, "rejected": chosen_fix})
                    elif (not pc) and fc:
                        # good flip: prefer fix
                        good_flip_pairs.append({**base, "chosen": chosen_fix, "rejected": chosen_planner})
                    elif pc and fc:
                        # both correct but different SQL — prefer planner (don't change unnecessarily)
                        same_correct_pairs.append({**base, "chosen": chosen_planner, "rejected": chosen_fix})

    # Cap same_correct_pairs to balance
    target_same = max(0, 600 - len(bad_flip_pairs) - 3 * len(good_flip_pairs))
    same_correct_pairs = same_correct_pairs[:target_same]
    # Upsample good_flip_pairs (rare but informative)
    good_flip_aug = good_flip_pairs * 3
    # Upsample bad_flip too (main signal — fixer should NOT mangle)
    bad_flip_aug = bad_flip_pairs * 3

    new_pairs = bad_flip_aug + good_flip_aug + same_correct_pairs

    # Merge with existing scaleup_iter2/hf_fixer_shared
    from datasets import load_from_disk
    try:
        existing = load_from_disk("/home/datht/mats-sql-tist/data/llm_alignment/scaleup_iter2/hf_fixer_shared")
        for split in ("train_dpo", "test_dpo"):
            for r in existing[split]:
                new_pairs.append({
                    "prompt": r["prompt"],
                    "chosen": r["chosen"],
                    "rejected": r["rejected"],
                    "db_path": r.get("db_path", ""),
                    "question": r.get("question", ""),
                    "db_id": r.get("db_id", ""),
                })
        print(f"  Merged {len(existing['train_dpo']) + len(existing['test_dpo'])} pairs from scaleup_iter2/hf_fixer_shared")
    except Exception as e:
        print(f"  WARN: could not merge existing data: {e}")

    import random
    rng = random.Random(42)
    rng.shuffle(new_pairs)

    n_test = max(20, len(new_pairs) // 30)
    test = new_pairs[:n_test]
    train = new_pairs[n_test:]
    all_pairs = new_pairs

    dd = DatasetDict({
        "train_dpo": Dataset.from_list(train),
        "test_dpo": Dataset.from_list(test),
    })
    dd.save_to_disk(OUT_DIR)

    print(f"=== Fixer ORPO iter-2 conservative dataset built ===")
    print(f"  bad_flip_pairs:      {len(bad_flip_pairs)}")
    print(f"  good_flip_pairs:     {len(good_flip_pairs)} (x3 → {len(good_flip_aug)})")
    print(f"  same_correct pairs:  {len(same_correct_pairs)}")
    print(f"  Total train: {len(train)}, test: {len(test)}")
    print(f"  Saved to {OUT_DIR}")


if __name__ == "__main__":
    main()