mats-sql-bundle / code /scripts /build_selector_v3_pairwise.py
thanhdath's picture
Push code: scripts, slurm sbatch, recipes, utils (v3 + selector series)
778d47d verified
Raw
History Blame Contribute Delete
7.42 kB
"""
Selector v3 SFT data builder: PAIRWISE framing with HARD NEGATIVES.
For each BIRD-train question with at least one YES and one NO trajectory:
- Build (A, B) pairs where the gold answer is A or B (balanced 50/50).
- Hard negatives: prefer NO SQL with highest lexical overlap to YES SQL (Jaccard
on lowercased token n-grams). Falls back to random NO if overlap is uniform.
- Both SQLs include row-preview exec result in the prompt (matching v2 style).
Output: HF dataset at data/sft_selector_v3_pairwise/{train,test}
Format: {"messages": [...chat...], "prompt": str, "completion": "A" or "B", ...}
"""
import json, os, re, sys, random
from collections import defaultdict
from concurrent.futures import ThreadPoolExecutor, as_completed
ROOT = "/weka/s225250685/mats-tist"
os.chdir(ROOT); sys.path.insert(0, ROOT)
os.environ.setdefault("DB_EXEC_API_DISABLE", "1")
os.environ.setdefault("PYTHONNOUSERSITE", "1")
from validator_data.validator import _execute_sql
from datasets import Dataset, DatasetDict
PAIRWISE_PROMPT = (
"You are a SQL correctness judge.\n"
"Schema:\n{schema}\n\n"
"Question: {question}\n"
"External knowledge: {evidence}\n\n"
"Candidate A:\n{sql_a}\n\n"
"Execution result of A:\n{exec_a}\n\n"
"Candidate B:\n{sql_b}\n\n"
"Execution result of B:\n{exec_b}\n\n"
"Which candidate is MORE LIKELY to correctly answer the question? Answer A or B."
)
SRC_PATHS = [
"data/rollouts/bird_train_3stage_K4.jsonl",
"data/rollouts/scaleup_bird_train_2stage_K4.jsonl",
"data/rollouts/scaleup_bird_train_3stage_K4.jsonl",
"data/rollouts/iter2_bird_train_3stage_K8.jsonl",
]
OUT_DIR = "data/sft_selector_v3_pairwise"
HARDNEG_PER_POS = 3 # up to 3 hardest-NO partners per YES SQL per question
MAX_PAIRS_PER_QUESTION = 8 # cap to avoid one easy question dominating
def safe_truncate(s, n=400):
s = str(s) if s is not None else ""
return s if len(s) <= n else s[:n] + "..."
def tokens(sql):
# lowercase token-1 bag (alphanumerics + sql ops)
return set(re.findall(r"[a-zA-Z_][a-zA-Z0-9_]*|[<>=!]+", (sql or "").lower()))
def jaccard(a, b):
if not a or not b: return 0.0
ai, ui = a & b, a | b
return len(ai) / max(len(ui), 1)
def exec_str(db_path, sql):
try:
r, err = _execute_sql("./" + db_path, sql, timeout=10)
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 load_question_groups(rng):
"""Walk all rollouts, return list of (sample, [(sql, label_is_correct), ...]) per question."""
by_q = {}
for src in SRC_PATHS:
if not os.path.exists(src):
print(f"skip missing: {src}", flush=True)
continue
print(f"loading {src}...", flush=True)
with open(src) as f:
for line in f:
line = line.strip()
if not line: continue
s = json.loads(line)
key = (s.get("question",""), s.get("db_id",""))
if key not in by_q:
by_q[key] = {"sample": s, "cands": [], "seen": set()}
for t in s.get("trajectories", []):
sql = (t.get("fixed_sql") or t.get("planner_sql") or "").strip()
if not sql: continue
norm = re.sub(r"\s+", " ", sql.lower())
if norm in by_q[key]["seen"]: continue
by_q[key]["seen"].add(norm)
is_correct = bool(t.get("is_fixed_correct") if t.get("fixed_sql") else t.get("is_planner_correct"))
by_q[key]["cands"].append((sql, is_correct))
print(f"questions: {len(by_q)}", flush=True)
out = []
for k, v in by_q.items():
yes = [c for c in v["cands"] if c[1]]
no = [c for c in v["cands"] if not c[1]]
if not yes or not no:
continue
out.append((v["sample"], yes, no))
print(f"questions with both YES and NO: {len(out)}", flush=True)
return out
def build_pair_records(rng, qgroups):
"""Build pairwise records: for each question, pair each YES with hardest NOs."""
work = [] # (sample, sql_yes, sql_no)
for sample, yes, no in qgroups:
# Score every NO by jaccard against best matching YES
no_scored = []
for ns, _ in no:
best = max(jaccard(tokens(ns), tokens(ys)) for ys, _ in yes)
no_scored.append((best, ns))
no_scored.sort(reverse=True) # hardest first
pairs = []
for ys, _ in yes:
# for each YES, take top HARDNEG_PER_POS NOs not yet paired
chosen = no_scored[:HARDNEG_PER_POS]
for _, ns in chosen:
pairs.append((ys, ns))
if len(pairs) >= MAX_PAIRS_PER_QUESTION:
break
rng.shuffle(pairs)
pairs = pairs[:MAX_PAIRS_PER_QUESTION]
for ys, ns in pairs:
work.append((sample, ys, ns))
return work
def render_one(rng, item):
sample, sql_yes, sql_no = item
db_path = sample["db_path"]
schema = safe_truncate(sample.get("schema", ""), 2000)
question = sample.get("question", "")
evidence = sample.get("evidence", "") or "None"
exec_yes = safe_truncate(exec_str(db_path, sql_yes), 240)
exec_no = safe_truncate(exec_str(db_path, sql_no), 240)
# 50/50 swap: YES at position A or B
if rng.random() < 0.5:
a_sql, b_sql, a_exec, b_exec, label = sql_yes, sql_no, exec_yes, exec_no, "A"
else:
a_sql, b_sql, a_exec, b_exec, label = sql_no, sql_yes, exec_no, exec_yes, "B"
prompt = PAIRWISE_PROMPT.format(
schema=schema, question=question, evidence=evidence,
sql_a=safe_truncate(a_sql, 700), exec_a=a_exec,
sql_b=safe_truncate(b_sql, 700), exec_b=b_exec,
)
return {
"prompt": prompt,
"completion": label,
"messages": [
{"role": "user", "content": prompt},
{"role": "assistant", "content": label},
],
"question": question,
"db_id": sample.get("db_id", ""),
}
def main():
rng = random.Random(42)
qgroups = load_question_groups(rng)
pairs = build_pair_records(rng, qgroups)
print(f"pair records to render: {len(pairs)}", flush=True)
out = []
with ThreadPoolExecutor(max_workers=32) as exe:
futs = [exe.submit(render_one, rng, p) for p in pairs]
n_done = 0
for fut in as_completed(futs):
try:
out.append(fut.result())
except Exception as e:
print(f"render err: {e}", flush=True)
n_done += 1
if n_done % 1000 == 0:
print(f" {n_done}/{len(pairs)}", flush=True)
rng.shuffle(out)
n_test = max(500, len(out) // 25)
test = out[:n_test]; train = out[n_test:]
n_a = sum(1 for r in train if r["completion"] == "A")
print(f"=== v3 pairwise selector data ===")
print(f" train: {len(train)} ({100*n_a/max(len(train),1):.1f}% A-label)")
print(f" test: {len(test)}")
DatasetDict({
"train": Dataset.from_list(train),
"test": Dataset.from_list(test),
}).save_to_disk(OUT_DIR)
print(f" saved {OUT_DIR}", flush=True)
if __name__ == "__main__":
main()