File size: 8,028 Bytes
4e1037f | 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 | """Vote across provider runs by execution result β the reproducible voting layer.
The archive's headline numbers (85.5%, and the 94.0% built on top of it) came
from voting: run the same questions through several different models, then keep
the answer they agree on. That layer is frozen in `scripts/archive/` because it
scored itself with the unsafe comparator β it called `compare_results` directly,
which blesses a pred that crashed as a match whenever gold also returns no rows.
This is that idea, rebuilt on the safe primitives:
* candidates come from finished eval reports (one per provider, same slice);
* every candidate SQL is re-executed against the live database;
* candidates cluster on `fingerprint_rows` β the execution result, not the SQL
text, so the same answer spelled two ways votes together;
* the largest cluster wins, ties break on report order (the strongest provider
first);
* the winner is scored with `safe_compare_pred`, which cannot bless a crash.
Diversity here comes from using *different models*, not from re-sampling one
model at a higher temperature β that was measured (config F) and it lost.
Usage:
python scripts/ensemble_providers.py \
eval/reports/2026-07-11/E_dense_fewshot_repair-evfirst.json \
eval/reports/2026-07-11/E_dense_fewshot_repair-groq.json \
--out eval/reports/2026-07-11/ensemble.json
"""
from __future__ import annotations
import argparse
import json
import sys
from collections import defaultdict
from dataclasses import dataclass
from pathlib import Path
from typing import Any
from sqlalchemy.exc import SQLAlchemyError
from nl_sql.db.registry import get_default_registry
from nl_sql.eval.metrics.execution_accuracy import safe_compare_pred
from nl_sql.eval.self_consistency import fingerprint_rows
from nl_sql.execution.guards import validate_sql
_ROW_CAP = 10_000
@dataclass(frozen=True, slots=True)
class Candidate:
provider: str
sql: str
rank: int # position of its report on the command line; lower = more trusted
@dataclass(frozen=True, slots=True)
class Executed:
candidate: Candidate
rows: list[tuple[Any, ...]] | None # None = the SQL failed to execute
def _run(engine: Any, sql: str) -> list[tuple[Any, ...]] | None:
"""Execute one candidate. A failure is a vote for nothing, not an empty answer."""
from nl_sql.db.connection import execute_readonly
if not sql.strip():
return None
if not validate_sql(sql, "sqlite").ok: # the guard the product runs behind
return None
try:
with execute_readonly(engine, sql, row_cap=_ROW_CAP) as result:
return list(result.rows)
except (SQLAlchemyError, Exception):
return None
def _vote(executed: list[Executed]) -> Executed | None:
"""Largest cluster of identical execution results; ties go to the better report.
A candidate that failed to execute is dropped: it agrees with nothing. An
empty result is a real answer and clusters normally β "no rows" is often
correct in BIRD.
"""
alive = [e for e in executed if e.rows is not None]
if not alive:
return None
clusters: dict[str, list[Executed]] = defaultdict(list)
for item in alive:
clusters[fingerprint_rows(item.rows or [])].append(item)
best = max(
clusters.values(),
key=lambda group: (len(group), -min(e.candidate.rank for e in group)),
)
return min(best, key=lambda e: e.candidate.rank)
def main(argv: list[str] | None = None) -> int:
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("reports", nargs="+", help="eval report JSONs, strongest first")
parser.add_argument("--out", default="", help="write the merged report here")
args = parser.parse_args(argv)
reports = [json.loads(Path(p).read_text(encoding="utf-8")) for p in args.reports]
names = [Path(p).stem for p in args.reports]
print(f"[ensemble] {len(reports)} providers: {', '.join(names)}")
# question_id -> candidates, in report order
by_qid: dict[Any, list[Candidate]] = defaultdict(list)
gold: dict[Any, dict[str, Any]] = {}
for rank, (report, name) in enumerate(zip(reports, names, strict=True)):
for rec in report["records"]:
qid = rec["question_id"]
by_qid[qid].append(Candidate(provider=name, sql=rec.get("pred_sql") or "", rank=rank))
gold.setdefault(qid, rec)
registry = get_default_registry()
engines: dict[str, Any] = {}
records: list[dict[str, Any]] = []
matched = 0
for i, (qid, candidates) in enumerate(sorted(by_qid.items(), key=lambda kv: str(kv[0])), 1):
base = gold[qid]
db_id = base["db_id"]
# Report records store the bare BIRD schema name (e.g. "california_schools"),
# matching BirdExample.db_id. The registry keys BIRD SQLite specs by
# BirdExample.registry_db_id == f"bird_{db_id}" (dataset.py:74-76). Calling
# registry.get(db_id) directly raised KeyError on the very first question β
# this script had never actually completed a run. Mirror registry_db_id here.
registry_db_id = f"bird_{db_id}"
if registry_db_id not in engines:
engines[registry_db_id] = registry.get(registry_db_id).make_engine()
engine = engines[registry_db_id]
executed = [Executed(candidate=c, rows=_run(engine, c.sql)) for c in candidates]
winner = _vote(executed)
gold_rows = _run(engine, base["gold_sql"])
if winner is None or gold_rows is None:
comparison = safe_compare_pred(
gold_rows or [],
[],
gold_sql=base["gold_sql"],
pred_failed=winner is None,
gold_failed=gold_rows is None,
)
else:
comparison = safe_compare_pred(
gold_rows,
winner.rows or [],
gold_sql=base["gold_sql"],
pred_failed=False,
gold_failed=False,
)
matched += int(comparison.match)
agreement = sum(
1
for e in executed
if e.rows is not None
and winner is not None
and fingerprint_rows(e.rows) == fingerprint_rows(winner.rows or [])
)
records.append(
{
**{k: base[k] for k in ("question_id", "db_id", "difficulty", "question")},
"gold_sql": base["gold_sql"],
"pred_sql": winner.candidate.sql if winner else "",
"winning_provider": winner.candidate.provider if winner else None,
"agreement": agreement,
"candidates": len(candidates),
"match": comparison.match,
"comparison_reason": comparison.reason,
}
)
if i % 25 == 0:
print(f" [{i:>3}/{len(by_qid)}] running EA {matched / i:.1%}")
n = len(records)
ea = matched / n if n else 0.0
print(f"\n[ensemble] EA {ea:.1%} ({matched}/{n})")
by_provider: dict[str, int] = defaultdict(int)
for rec in records:
if rec["winning_provider"]:
by_provider[rec["winning_provider"]] += 1
print("[ensemble] questions won per provider:")
for name, count in sorted(by_provider.items(), key=lambda kv: -kv[1]):
print(f" {name:<50} {count}")
unanimous = sum(1 for r in records if r["agreement"] == r["candidates"])
print(f"[ensemble] all providers agreed on {unanimous}/{n} questions")
if args.out:
out = {
"configuration": "ensemble+" + "+".join(names),
"sql_model": " + ".join(names),
"overall": {"n": n, "ea": round(ea, 4), "matched": matched},
"records": records,
}
Path(args.out).write_text(json.dumps(out, indent=2, ensure_ascii=False), encoding="utf-8")
print(f"[json] {args.out}")
return 0
if __name__ == "__main__":
sys.exit(main())
|