code/scripts/build_fixer_critique_aware_v6.py

"""
Build critique-aware fixer SFT data.

The OLD fixer SFT (data/hf_fixer_griffith_v5) trains on a fixed critique template, so the fixer
ignores critique content at inference. This breaks the collab signal (HANDOFF_COLLAB_TASK.md §3).

This script rebuilds the fixer SFT data with DIVERSE critiques sampled per question from the
paper-format SFT validators (val-sel + val-cond). The fixer prompt format matches inference
(build_fixer_prompt from run_pipeline_rollouts.py), and the completion is the gold SQL.

Output: HF DatasetDict with (prompt, completion) split 95/5 train/test.
Approach C from the plan: per-question diverse critiques + gold completion. Critique tokens enter
the prompt and the model has to attend to them to know what to output — the critique becomes part
of the conditioning context.
"""
import argparse
import json
import os
import re
import random
import sqlite3
import threading

os.environ.setdefault("PYTHONNOUSERSITE", "1")
os.environ["NO_PROXY"] = "localhost,127.0.0.1"

import requests
from datasets import load_dataset, Dataset, DatasetDict


def safe_exec(db_path, sql, timeout=5):
    r = [None]; e = [None]
    def _run():
        try:
            c = sqlite3.connect(db_path); c.text_factory = lambda b: b.decode(errors="ignore")
            r[0] = c.execute(sql).fetchmany(100); c.close()
        except Exception as ex:
            e[0] = str(ex)
    t = threading.Thread(target=_run, daemon=True); t.start(); t.join(timeout)
    return (None, "TIMEOUT") if t.is_alive() else (r[0], e[0])


def results_match(g, p):
    if g is None or p is None: return False
    def n(rs): return sorted(tuple(str(v).strip().lower() if v is not None else "" for v in r) for r in rs)
    return n(g) == n(p)


def extract_sql(text):
    m = re.search(r"```(?:sql)?\s*(.*?)\s*```", text, re.DOTALL)
    if m:
        s = m.group(1).strip()
        return s[3:].strip() if s.upper().startswith("SQL") else s
    return ""


def qwen_chat(p):
    return f"<|im_start|>user\n{p}<|im_end|>\n<|im_start|>assistant\n"


def llama3_chat(p):
    return (f"<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\n"
            f"{p}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n")


def vllm_complete(host, model, prompt, n, temperature, top_p, max_tokens, seed, stop=None):
    try:
        r = requests.post(f"{host}/v1/completions", json={
            "model": model, "prompt": prompt,
            "n": n, "temperature": temperature, "top_p": top_p,
            "max_tokens": max_tokens, "seed": seed,
            "stop": stop or ["<|eot_id|>", "<|im_end|>"],
        }, timeout=180)
        r.raise_for_status()
        return [c["text"].strip() for c in r.json()["choices"]]
    except Exception as e:
        return []


# Fixer prompt MUST match run_pipeline_rollouts.py:build_fixer_prompt and FIXER_PROMPT_HEADER exactly.
FIXER_PROMPT_HEADER = (
    "You are a SQL fixer. Given the question, schema, original SQL query, "
    "execution response, and the validator's critique below, output ONLY the corrected "
    "final SQL inside ```sql ... ``` markers.\n\n"
)


def build_fixer_prompt(schema_str, question, evidence, planner_sql, exec_response, critique):
    body = (
        f"database schema:\n{schema_str}\n\n"
        f"Question: {question}\n"
        f"External knowledge: {evidence or 'None'}\n\n"
        f"Generated SQL query: {planner_sql}\n\n"
        f"Execution response:\n{exec_response}\n\n"
    )
    return FIXER_PROMPT_HEADER + body + "\n\nValidator critique:\n" + critique + "\n\nFinal SQL:"


def build_validator_body(schema_str, question, evidence, planner_sql, exec_response):
    """Paper-format validator prompt body (val-sel + val-cond share it)."""
    return (
        f"Generate feedbacks to fix the following SQL query:\n"
        f"Database Schema:\n{schema_str}\n\n"
        f"Question: {question}\n"
        f"External knowledge: {evidence or 'None'}\n\n"
        f"SQL query: {planner_sql}\n\n"
        f"Execution response:\n{exec_response}\n\n"
        f"Feedback:"
    )


def main():
    p = argparse.ArgumentParser()
    p.add_argument("--planner_host",   default="http://localhost:8100")
    p.add_argument("--val_sel_host",   default="http://localhost:8101")
    p.add_argument("--val_cond_host",  default="http://localhost:8104")
    p.add_argument("--K", type=int, default=8, help="critiques per question")
    p.add_argument("--temperature", type=float, default=1.0)
    p.add_argument("--max_questions", type=int, default=-1, help="-1 = use full dataset (default)")
    p.add_argument("--seed", type=int, default=42)
    p.add_argument("--out", required=True)
    args = p.parse_args()

    print("Loading BIRD-train + griffith prompts...", flush=True)
    with open("data/sft_bird_with_evidence_train_text2sql.json") as f:
        bird_train = json.load(f)
    ds_g = load_dataset("griffith-bigdata/sft_text2sql", split="train_sft",
                        cache_dir="/weka/s225250685/Huggingface/hub"
                       ).filter(lambda x: x["model_name"] == "deepseek-reasoner")
    griffith = {}
    for row in ds_g:
        sid = int(row["sample_id"])
        if not (0 <= sid < len(bird_train)): continue
        user_msg = row["messages"][1]["content"]
        q_m = re.search(r"Question:\s*(.+?)(?:\n|$)", user_msg)
        if not q_m: continue
        q = q_m.group(1).strip()
        if q.lower() == bird_train[sid]["question"].strip().lower():
            griffith[q.lower()] = {"user_msg": user_msg, "sid": sid}
    print(f"  griffith: {len(griffith)} questions", flush=True)

    DEFAULT_SEL  = "SELECT.\nNo SELECT critique generated.\nConclude: correct."
    DEFAULT_COND = "CONDITION.\nNo CONDITION critique generated.\nConclude: correct."

    rows = []
    n_planner_correct = 0
    n_planner_wrong = 0
    n_no_planner = 0
    random.seed(args.seed)
    items = list(griffith.items()); random.shuffle(items)

    limit = args.max_questions if args.max_questions > 0 else len(items)
    for i, (q_lower, info) in enumerate(items[:limit]):
        bt = bird_train[info["sid"]]
        db_path = bt.get("db_path") or f"data/train_databases/{bt['db_id']}/{bt['db_id']}.sqlite"
        if not os.path.exists(db_path):
            continue
        question = bt["question"]
        evidence = bt.get("evidence", "") or ""
        gold_sql = bt["sql"]

        # Extract rich schema substring from griffith user_msg
        user_msg = info["user_msg"]
        if "Database Schema:" in user_msg:
            schema_str = user_msg.split("Database Schema:", 1)[1].split("Question:", 1)[0].rstrip()
        else:
            schema_str = user_msg

        # 1) Get planner SQL (greedy, T=0.0). Used as the "wrong/right" candidate for fixer.
        planning_prompt = user_msg.rstrip() + "\n\nPlanning:"
        plans = vllm_complete(
            args.planner_host, "planner", qwen_chat(planning_prompt),
            n=1, temperature=0.0, top_p=1.0, max_tokens=1024, seed=args.seed + i,
        )
        if not plans:
            n_no_planner += 1
            continue
        # Extract SQL from planner output (Planning: ... Final SQL query: ```...```)
        planner_text = plans[0]
        m = re.search(r"Final SQL query:\s*```(?:sql)?\s*(.+?)```", planner_text, re.DOTALL | re.IGNORECASE)
        if m:
            planner_sql = m.group(1).strip()
        else:
            planner_sql = extract_sql(planner_text)
        if not planner_sql:
            n_no_planner += 1
            continue

        # 2) Execute planner SQL
        gold_res, gold_err = safe_exec(db_path, gold_sql)
        if gold_res is None:
            continue
        pred_res, perr = safe_exec(db_path, planner_sql)
        planner_correct = (not perr) and results_match(gold_res, pred_res)
        if planner_correct:
            n_planner_correct += 1
        else:
            n_planner_wrong += 1
        exec_response = (f"Error: {perr[:200]}" if perr
                         else f"OK. Result rows (preview): {str(pred_res)[:300]}")

        # 3) Generate K val-sel critiques and K val-cond critiques (paper format)
        val_body = build_validator_body(schema_str, question, evidence, planner_sql, exec_response)
        # Seed with the clause token so the val-sel/val-cond model continues directly.
        sel_seeded  = val_body + "\nSELECT.\n"
        cond_seeded = val_body + "\nCONDITION.\n"

        sel_outs = vllm_complete(
            args.val_sel_host, "validator", llama3_chat(sel_seeded),
            n=args.K, temperature=args.temperature, top_p=0.9,
            max_tokens=384, seed=args.seed + i,
        )
        cond_outs = vllm_complete(
            args.val_cond_host, "validator", llama3_chat(cond_seeded),
            n=args.K, temperature=args.temperature, top_p=0.9,
            max_tokens=384, seed=args.seed + i + 1,
        )
        if not sel_outs and not cond_outs:
            continue
        # Re-prepend the clause token (vLLM returns only the continuation)
        sel_outs  = [f"SELECT.\n{c.lstrip()}"   if c else DEFAULT_SEL  for c in sel_outs]
        cond_outs = [f"CONDITION.\n{c.lstrip()}" if c else DEFAULT_COND for c in cond_outs]
        # Pad to K with defaults
        while len(sel_outs)  < args.K: sel_outs.append(DEFAULT_SEL)
        while len(cond_outs) < args.K: cond_outs.append(DEFAULT_COND)

        # 4) Combine each (sel, cond) pair into the inference critique format
        gold_completion = f"```sql\n{gold_sql}\n```"
        for j in range(args.K):
            s_out, c_out = sel_outs[j], cond_outs[j]
            combined = (
                f"<select>\n{s_out}\n</select>\n\n"
                f"<condition>\n{c_out}\n</condition>\n\n"
                "<join>\nJOIN.\nNone\n</join>\n\n"
                "<order>\nORDER BY.\nNone\n</order>"
            )
            prompt = build_fixer_prompt(schema_str, question, evidence, planner_sql, exec_response, combined)
            rows.append({"prompt": prompt, "completion": gold_completion})

        if (i + 1) % 50 == 0:
            print(f"  [{i+1}/{limit}] rows={len(rows)} planner_ok={n_planner_correct} "
                  f"planner_wrong={n_planner_wrong} no_planner={n_no_planner}", flush=True)

    print(f"\nGenerated {len(rows)} fixer SFT rows", flush=True)
    print(f"  Planner correct: {n_planner_correct}  Planner wrong: {n_planner_wrong}  No planner: {n_no_planner}",
          flush=True)
    if not rows:
        print("ERROR: no rows generated"); return

    random.seed(42); random.shuffle(rows)
    n_train = int(0.95 * len(rows))
    DatasetDict({
        "train": Dataset.from_list(rows[:n_train]),
        "test":  Dataset.from_list(rows[n_train:]),
    }).save_to_disk(args.out)
    print(f"Saved → {args.out}  train={n_train}  test={len(rows) - n_train}", flush=True)


if __name__ == "__main__":
    main()