File size: 8,555 Bytes
778d47d | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 | """
Selector v4 — PAIRWISE selector SFT data builder (Chase-SQL style).
Chase-SQL (Pourreza et al.) frames the selector as a head-to-head judge:
given (question, schema, candidate_A, candidate_B, exec_a, exec_b), the
model outputs which one is more likely correct. At inference, K=8 candidates
are compared in a round-robin tournament (28 calls) or single-elimination
bracket (7 calls); the candidate with the most pairwise wins is picked.
Pros vs pointwise YES/NO:
- Direct preference signal (no calibration of independent probabilities).
- Captures fine-grained discrimination between near-duplicate SQLs.
Data construction:
For each BIRD-train question with at least one YES and one NO trajectory:
- For each (yes_sql, no_sql) pair, emit TWO records:
A = yes, B = no, label = "A"
A = no, B = yes, label = "B"
→ 50/50 label balance, twice the data.
Hard negatives: prefer NO SQLs with high lexical overlap to a YES SQL
(Jaccard on word tokens). Cap at HARDNEG_PER_POS per YES per question.
Output:
data/sft_selector_v4_pairwise/{train,test}
Each row: {"prompt", "completion", "messages", "question", "db_id"}
"""
import json, os, re, sys, random
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
from scripts.rich_schema import render_rich_schema
PAIRWISE_PROMPT = (
"You are a SQL correctness judge. Compare two candidate SQL queries that "
"attempt to answer the same question. Pick the one MORE LIKELY to be correct.\n\n"
"Database schema (with column descriptions, value descriptions, and example values):\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 with a single letter: 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_v4_pairwise"
HARDNEG_PER_POS = 3 # hardest NO partners per YES SQL
MAX_PAIRS_PER_Q = 6 # cap raw (YES, NO) pairs per question (→ 12 records after 2× swap)
MAX_SCHEMA_CHARS = 3000 # smaller than v3 since two SQLs share prompt
EXEC_TIMEOUT = 5 # reduced from 8 to avoid login-node OOM
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):
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
return len(a & b) / max(len(a | b), 1)
def exec_str(db_path, sql):
try:
r, err = _execute_sql("./" + db_path, sql, timeout=EXEC_TIMEOUT)
except Exception as e:
return f"Error: {str(e)[:140]}"
if err:
return f"Error: {str(r)[:140]}"
rows = str(r)[:220]
if rows.strip() and rows.strip() != "[]":
return f"OK. Rows preview: {rows}"
return "OK. (no rows returned)"
def collect_question_groups():
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)
correct = bool(t.get("is_fixed_correct") if t.get("fixed_sql") else t.get("is_planner_correct"))
by_q[key]["cands"].append((sql, correct))
print(f"unique questions: {len(by_q)}", flush=True)
out = []
for k, v in by_q.items():
yes = [c[0] for c in v["cands"] if c[1]]
no = [c[0] 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):
"""For each question, emit at most MAX_PAIRS_PER_Q (yes, no) pairs with hard-neg ranking.
Each pair becomes 2 records (A=YES,B=NO; A=NO,B=YES)."""
raw = []
for sample, yes_list, no_list in qgroups:
# Score every NO by best Jaccard against any YES
no_scored = []
yes_toks = [tokens(y) for y in yes_list]
for ns in no_list:
t_no = tokens(ns)
best = max((jaccard(t_no, ty) for ty in yes_toks), default=0.0)
no_scored.append((best, ns))
no_scored.sort(reverse=True)
pairs = []
for ys in yes_list:
for _, ns in no_scored[:HARDNEG_PER_POS]:
pairs.append((ys, ns))
if len(pairs) >= MAX_PAIRS_PER_Q:
break
if len(pairs) >= MAX_PAIRS_PER_Q:
break
for ys, ns in pairs:
raw.append((sample, ys, ns))
return raw
def render_pair(rng_seed, item):
"""Produce TWO records (swapped A/B) so labels are balanced."""
sample, sql_yes, sql_no = item
rng = random.Random(rng_seed)
db_path = sample["db_path"]
schema = safe_truncate(render_rich_schema(sample, split="train"), MAX_SCHEMA_CHARS)
question = sample.get("question", "")
evidence = sample.get("evidence", "") or "None"
exec_yes = safe_truncate(exec_str(db_path, sql_yes), 220)
exec_no = safe_truncate(exec_str(db_path, sql_no), 220)
out = []
for swap in (False, True):
if not swap:
a, b, ea, eb, label = sql_yes, sql_no, exec_yes, exec_no, "A"
else:
a, b, ea, eb, 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, 600), exec_a=ea,
sql_b=safe_truncate(b, 600), exec_b=eb,
)
out.append({
"prompt": prompt,
"completion": label,
"messages": [
{"role": "user", "content": prompt},
{"role": "assistant", "content": label},
],
"question": question,
"db_id": sample.get("db_id", ""),
})
return out
def main():
rng = random.Random(42)
qg = collect_question_groups()
raw = build_pair_records(rng, qg)
print(f"raw (yes, no) pairs: {len(raw)} → records: {2*len(raw)}", flush=True)
out = []
with ThreadPoolExecutor(max_workers=8) as exe:
futs = [exe.submit(render_pair, i, it) for i, it in enumerate(raw)]
n_done = 0
for fut in as_completed(futs):
try:
out.extend(fut.result())
except Exception as e:
print(f"render err: {e}", flush=True)
n_done += 1
if n_done % 1000 == 0:
print(f" rendered {n_done}/{len(raw)} 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"\n=== v4 PAIRWISE selector data ===")
print(f" train: {len(train)} ({100*n_a/max(len(train),1):.1f}% A-label)")
print(f" test: {len(test)}")
avg = sum(len(r["prompt"]) for r in train) / max(len(train),1)
print(f" avg prompt chars: {avg:.0f}")
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()
|