mats-sql-bundle / code /scripts /build_selector_v6_pointwise.py
thanhdath's picture
Push code: scripts, slurm sbatch, recipes, utils (v3 + selector series)
778d47d verified
Raw
History Blame Contribute Delete
6.43 kB
"""
Phase 1 (v6) — Build POINTWISE selector training data.
Each record = (question, rich_schema, evidence, single SQL, exec_result) → YES/NO.
Source: data/qwen72b_candidates_bird_train.jsonl + gold injection.
Output: HF DatasetDict at data/sft_selector_v6_pointwise_rich/{train,test}
"""
import argparse
import json
import os
import re
import sys
import random
os.environ.setdefault("PYTHONNOUSERSITE", "1")
os.environ.setdefault("DB_EXEC_API_DISABLE", "1")
ROOT = "/weka/s225250685/mats-tist"
os.chdir(ROOT)
sys.path.insert(0, ROOT)
from validator_data.validator import _execute_sql
from datasets import Dataset, DatasetDict
from scripts.rich_schema import render_rich_schema
POINTWISE_PROMPT = (
"You are a SQL correctness judge for the BIRD benchmark.\n"
"Database Schema (with column meanings, value descriptions, and example values):\n"
"{schema}\n\n"
"Question: {question}\n"
"External knowledge: {evidence}\n\n"
"Candidate SQL:\n{sql}\n\n"
"Execution result of the candidate:\n{exec_result}\n\n"
"Does this SQL correctly answer the question, given the schema, the column "
"descriptions, the external knowledge, and the execution result? Answer YES or NO."
)
MAX_SCHEMA_CHARS = 3000
def safe_truncate(s, n):
if s is None:
return ""
s = str(s)
return s if len(s) <= n else s[:n] + "..."
def gold_exec_str(db_path, sql, timeout=10):
if not sql or not sql.strip():
return "Error: empty SQL"
try:
r, err = _execute_sql("./" + db_path if not db_path.startswith("./") else db_path, sql, timeout=timeout)
except Exception as e:
return f"Error: {str(e)[:160]}"
if err:
return f"Error: {str(r)[:160]}"
rows = str(r)[:260]
if rows.strip() and rows.strip() != "[]":
return f"OK. Rows preview: {rows}"
return "OK. (no rows returned)"
def render(sample, sql, exec_result, label):
schema = safe_truncate(render_rich_schema(sample, split="train"), MAX_SCHEMA_CHARS)
prompt = POINTWISE_PROMPT.format(
schema=schema,
question=sample.get("question", ""),
evidence=sample.get("evidence", "") or "None",
sql=safe_truncate(sql, 800),
exec_result=safe_truncate(exec_result, 300),
)
return {
"prompt": prompt,
"completion": label,
"messages": [
{"role": "user", "content": prompt},
{"role": "assistant", "content": label},
],
"question": sample.get("question", ""),
"db_id": sample.get("db_id", ""),
"is_yes": int(label == "YES"),
}
def gold_record_for(rec):
"""Returns the gold-injected record for one question, or None if gold errors."""
if not rec.get("sql"):
return None
seen = set()
for c in rec.get("candidates", []):
norm = re.sub(r"\s+", " ", (c.get("sql") or "").strip().lower())
if norm:
seen.add(norm)
gold_norm = re.sub(r"\s+", " ", rec["sql"].strip().lower())
if gold_norm in seen:
return None
ge = gold_exec_str(rec["db_path"], rec["sql"])
if ge.startswith("Error"):
return None
return render(rec, rec["sql"], ge, "YES")
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--input", default="data/qwen72b_candidates_bird_train.jsonl")
ap.add_argument("--out", default="data/sft_selector_v6_pointwise_rich")
ap.add_argument("--inject_gold", action="store_true", default=True)
args = ap.parse_args()
rng = random.Random(42)
records = []
n_gold = 0
n_yes = 0
n_no = 0
raw_rows = []
with open(args.input) as f:
for line in f:
line = line.strip()
if not line: continue
raw_rows.append(json.loads(line))
print(f"input rows: {len(raw_rows)}", flush=True)
# Phase 1: render all candidate records (CPU-bound, fast — no exec needed since exec_str already in JSONL).
for r in raw_rows:
seen = set()
for c in r.get("candidates", []):
norm = re.sub(r"\s+", " ", (c.get("sql") or "").strip().lower())
if not norm or norm in seen:
continue
seen.add(norm)
label = "YES" if c.get("is_correct") else "NO"
records.append(render(r, c["sql"], c["exec_str"], label))
if label == "YES":
n_yes += 1
else:
n_no += 1
print(f"after cand render: YES={n_yes} NO={n_no}", flush=True)
# Phase 2: parallel gold injection
if args.inject_gold:
from concurrent.futures import ThreadPoolExecutor, as_completed
with ThreadPoolExecutor(max_workers=32) as exe:
futs = {exe.submit(gold_record_for, r): r for r in raw_rows}
n_proc = 0
for fut in as_completed(futs):
n_proc += 1
try:
gr = fut.result()
except Exception:
gr = None
if gr is not None:
records.append(gr)
n_gold += 1
n_yes += 1
if n_proc % 500 == 0:
print(f" gold-injected {n_proc}/{len(raw_rows)} total_gold={n_gold}", flush=True)
print(f"records: {len(records)} YES={n_yes} NO={n_no} gold_added={n_gold}")
# Downsample NO to ~equal YES (balance) — currently NO probably >> YES
yes_rec = [r for r in records if r["is_yes"]]
no_rec = [r for r in records if not r["is_yes"]]
rng.shuffle(no_rec)
keep_no = no_rec[: min(len(no_rec), int(1.2 * len(yes_rec)))]
final = yes_rec + keep_no
rng.shuffle(final)
print(f"after balance: {len(final)} YES={len(yes_rec)} NO={len(keep_no)}")
# 96/4 split by question (so identical Q never split).
by_q = {}
for r in final:
by_q.setdefault(r["question"], []).append(r)
qs = list(by_q.keys())
rng.shuffle(qs)
n_test_q = max(40, len(qs) // 25)
test_qs = set(qs[:n_test_q])
train, test = [], []
for q, recs in by_q.items():
(test if q in test_qs else train).extend(recs)
rng.shuffle(train)
rng.shuffle(test)
print(f"train: {len(train)} test: {len(test)}")
DatasetDict({
"train": Dataset.from_list(train),
"test": Dataset.from_list(test),
}).save_to_disk(args.out)
print(f"SAVED: {args.out}")
if __name__ == "__main__":
main()