File size: 8,030 Bytes
82ae30c | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 | """
Compute Best-of-N metrics with a TRAINED selector (binary YES/NO classifier).
For each question, run the selector on each of N candidates and pick the one
with the highest YES probability. Also compute greedy / pass@N for comparison.
Usage:
python scripts/compute_bestofn_with_selector.py <rollout.jsonl> <selector_ckpt> <label> [--selector_host URL]
If --selector_host given (a vLLM endpoint), use it. Otherwise load model in-process.
"""
import argparse
import json
import os
import re
import sys
from collections import Counter
from concurrent.futures import ThreadPoolExecutor
# Bypass HTTP proxy for local vLLM endpoints
os.environ["NO_PROXY"] = "localhost,127.0.0.1"
os.environ["no_proxy"] = "localhost,127.0.0.1"
ROOT = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
os.chdir(ROOT)
sys.path.insert(0, ROOT)
from validator_data.validator import _execute_sql
from data_processing.planner import is_execution_correct
import requests
PROMPT_TEMPLATE = (
"You are a SQL correctness judge.\n"
"Schema:\n{schema}\n\n"
"Question: {question}\n"
"External knowledge: {evidence}\n\n"
"Candidate SQL:\n{sql}\n\n"
"Execution result:\n{exec_result}\n\n"
"Is this SQL correct for the question? Answer YES or NO."
)
def qwen_chat(prompt: str) -> str:
return f"<|im_start|>user\n{prompt}<|im_end|>\n<|im_start|>assistant\n"
def safe_truncate(s, n=400):
if s is None:
return "(empty)"
s = str(s)
return s if len(s) <= n else s[:n] + "..."
def safe_execute(db_path, sql):
if not sql or sql.strip() == "":
return ("", True)
try:
return _execute_sql("./" + db_path, sql)
except Exception:
return ("", True)
def score_via_vllm(host, prompt_chat, model_name="selector"):
"""Get P(YES) − P(NO) via vLLM completions with logprobs."""
payload = {
"model": model_name,
"prompt": prompt_chat,
"max_tokens": 1,
"n": 1,
"temperature": 0.0,
"logprobs": 20,
}
try:
r = requests.post(f"{host}/v1/completions", json=payload, timeout=60)
r.raise_for_status()
choice = r.json()["choices"][0]
# Look at logprobs of top tokens; find YES vs NO
if "logprobs" in choice and choice["logprobs"]:
top = choice["logprobs"]["top_logprobs"][0]
yes_lp = max((top[k] for k in (" YES", "YES", " yes", "yes", " Yes", "Yes") if k in top), default=-100.0)
no_lp = max((top[k] for k in (" NO", "NO", " no", "no", " No", "No") if k in top), default=-100.0)
return yes_lp - no_lp
# Fallback: text match
text = choice.get("text", "").strip().upper()
return 1.0 if text.startswith("YES") else (-1.0 if text.startswith("NO") else -100.0)
except Exception as e:
sys.stderr.write(f"score_via_vllm err: {type(e).__name__}: {e}\n")
return -100.0
def build_prompt_chat(sample, t, exec_result_str=None):
"""Build selector prompt.
`exec_result_str` MUST be the actual SQL execution result preview (or error message).
Do NOT pass the gold-graded label — that leaks the correctness label into the prompt
and makes the selector trivially match the oracle.
"""
schema = sample.get("schema", "")
question = sample.get("question", "")
evidence = sample.get("evidence", "") or "None"
fixed_sql = t.get("fixed_sql") or t.get("planner_sql") or ""
if exec_result_str is None:
# Safe default at inference time: signal unknown (selector must judge from SQL alone)
exec_result_str = "(execution result not available)"
prompt = PROMPT_TEMPLATE.format(
schema=safe_truncate(schema, 3000),
question=question,
evidence=evidence,
sql=safe_truncate(fixed_sql, 800),
exec_result=safe_truncate(exec_result_str, 300),
)
return qwen_chat(prompt)
def main():
parser = argparse.ArgumentParser()
parser.add_argument("rollout_jsonl")
parser.add_argument("label")
parser.add_argument("--selector_host", default="http://localhost:8103")
args = parser.parse_args()
n_q = 0
n_greedy = 0
n_pass_at_N = 0
n_majority = 0 # rule-based majority (for compare)
n_selector = 0 # trained selector pick
K_used = None
samples = []
with open(args.rollout_jsonl) as f:
for line in f:
line = line.strip()
if line:
samples.append(json.loads(line))
print(f"Loaded {len(samples)} samples")
for sample in samples:
traj = sample.get("trajectories", [])
if not traj:
continue
n_q += 1
if K_used is None:
K_used = len(traj)
# Greedy = first traj
if traj[0].get("is_fixed_correct"):
n_greedy += 1
# pass@N = any correct
if any(t.get("is_fixed_correct") for t in traj):
n_pass_at_N += 1
# Rule-based majority: pick most-common non-empty execution result
db_path = sample["db_path"]
gold_sql = sample["sql"]
true_exec = safe_execute(db_path, gold_sql)
if true_exec[1]:
continue # gold has error; skip
# Execute all candidates' fixed SQLs once
with ThreadPoolExecutor(max_workers=8) as exe:
exec_results = list(exe.map(
lambda t: safe_execute(db_path, t.get("fixed_sql") or t.get("planner_sql") or ""),
traj
))
# Rule-based majority
majority_picks = []
for i, (er, t) in enumerate(zip(exec_results, traj)):
if er[1]:
continue
res_str = str(er[0]).strip()
if not res_str or "(no rows)" in res_str or res_str == "[]":
continue
majority_picks.append((i, res_str))
if majority_picks:
counter = Counter(s for _, s in majority_picks)
top_res, _ = counter.most_common(1)[0]
for i, s in majority_picks:
if s == top_res:
if traj[i].get("is_fixed_correct"):
n_majority += 1
break
else:
if traj[0].get("is_fixed_correct"):
n_majority += 1
# Trained selector: score each candidate using REAL execution result (no gold-label leak).
# Re-use the exec_results computed above for rule-based majority.
scores = []
with ThreadPoolExecutor(max_workers=8) as exe:
def _make_prompt(idx, t_):
er = exec_results[idx]
if er[1]:
exec_str = f"Error: {er[0]}"
else:
rows_str = str(er[0])
if not rows_str.strip() or rows_str.strip() == "[]":
exec_str = "OK. Result rows (preview): (no rows)"
else:
exec_str = f"OK. Result rows (preview): {rows_str[:300]}"
return build_prompt_chat(sample, t_, exec_result_str=exec_str)
futs = [exe.submit(score_via_vllm, args.selector_host, _make_prompt(i, t)) for i, t in enumerate(traj)]
for f in futs:
scores.append(f.result())
best_idx = max(range(len(scores)), key=lambda i: scores[i])
if traj[best_idx].get("is_fixed_correct"):
n_selector += 1
if n_q == 0:
print(f"{args.label}: no questions evaluated")
return
print()
print(f"=== {args.label} ===")
print(f" questions evaluated: {n_q}")
print(f" K used per question: {K_used}")
print(f" greedy (1st traj): {n_greedy}/{n_q} = {100*n_greedy/n_q:.2f}%")
print(f" rule-based majority: {n_majority}/{n_q} = {100*n_majority/n_q:.2f}%")
print(f" trained selector: {n_selector}/{n_q} = {100*n_selector/n_q:.2f}%")
print(f" pass@{K_used} (oracle): {n_pass_at_N}/{n_q} = {100*n_pass_at_N/n_q:.2f}%")
print()
if __name__ == "__main__":
main()
|