File size: 12,106 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 | """
FAST regen of COLLAB ORPO validator data using Qwen-2.5-72B-Instruct-AWQ as the fixer.
Why: the OLD Llama-1B fixer used in the previous iter1 collab data-gen ignored critique content,
so chosen/rejected of critiques was essentially uncorrelated with critique CONTENT/VERDICT —
the validator had no learnable signal from the conclusion token (chosen and rejected had
identical verdict distributions in the iter1 data). A critique-responsive fixer (72B) makes
each critique produce a genuinely different fixer output, restoring a real collab signal.
Speed: ThreadPoolExecutor over questions, vLLM batches concurrent requests internally. The
72B is the bottleneck; running with --threads 16+ saturates the vLLM batch scheduler.
Output: HF DatasetDict with {prompt, chosen, rejected} for ORPO training, splits
`train_dpo` / `test_dpo`, matching the iter1 collab schema.
"""
import argparse
import json
import os
import re
import random
import sqlite3
import threading
from concurrent.futures import ThreadPoolExecutor, as_completed
os.environ.setdefault("PYTHONNOUSERSITE", "1")
os.environ["NO_PROXY"] = "localhost,127.0.0.1"
import requests
from datasets import load_dataset, Dataset, DatasetDict
_db_lock = threading.Lock()
def safe_exec(db_path, sql, timeout=5):
r = [None]; e = [None]
def _run():
try:
c = sqlite3.connect(db_path); c.text_factory = lambda b: b.decode(errors="ignore")
r[0] = c.execute(sql).fetchmany(100); c.close()
except Exception as ex:
e[0] = str(ex)
t = threading.Thread(target=_run, daemon=True); t.start(); t.join(timeout)
return (None, "TIMEOUT") if t.is_alive() else (r[0], e[0])
def results_match(g, p):
if g is None or p is None: return False
def n(rs): return sorted(tuple(str(v).strip().lower() if v is not None else "" for v in r) for r in rs)
return n(g) == n(p)
def extract_sql(text):
m = re.search(r"```(?:sql)?\s*(.*?)\s*```", text, re.DOTALL)
if m:
s = m.group(1).strip()
return s[3:].strip() if s.upper().startswith("SQL") else s
return ""
def qwen_chat(p):
return f"<|im_start|>user\n{p}<|im_end|>\n<|im_start|>assistant\n"
def llama3_chat(p):
return (f"<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\n"
f"{p}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n")
def vllm_complete(host, model, prompt, n, temperature, top_p, max_tokens, seed, stop=None):
try:
r = requests.post(f"{host}/v1/completions", json={
"model": model, "prompt": prompt,
"n": n, "temperature": temperature, "top_p": top_p,
"max_tokens": max_tokens, "seed": seed,
"stop": stop or ["<|eot_id|>", "<|im_end|>"],
}, timeout=300)
r.raise_for_status()
return [c["text"].strip() for c in r.json()["choices"]]
except Exception:
return []
FIXER_INSTR = (
"You are an expert SQL judge and fixer. You will see a candidate SQL, its execution result, "
"and a validator's critique.\n\n"
"Your task:\n"
"1. Decide if the candidate SQL correctly answers the question. Consider the validator's "
"critique as a hint, but verify with your own SQL expertise.\n"
"2. If the candidate SQL is correct, output it UNCHANGED.\n"
"3. If the candidate SQL has a real issue, output a corrected SQL.\n"
"4. Prefer keeping the candidate unchanged when in doubt.\n\n"
"Output ONLY the final SQL inside ```sql ... ``` markers."
)
def build_fixer_prompt(schema_str, question, evidence, planner_sql, exec_response, wrapped_critique):
body = (
f"\n\nDatabase Schema:\n{schema_str.rstrip()}\n\n"
f"Question: {question}\n"
f"External knowledge: {evidence or 'None'}\n\n"
f"Candidate SQL:\n{planner_sql}\n\n"
f"Execution response:\n{exec_response}\n\n"
f"Validator critique:\n{wrapped_critique}\n\nFinal SQL:"
)
return FIXER_INSTR + body
def process_one(args, q_lower, info, bird_train, side, idx):
"""Process one BIRD-train question. Returns (status, list of pair dicts)."""
bt = bird_train[info["sid"]]
db_path = bt.get("db_path") or f"data/train_databases/{bt['db_id']}/{bt['db_id']}.sqlite"
if not os.path.exists(db_path):
return ("skip_no_db", [])
question = bt["question"]
evidence = bt.get("evidence", "") or ""
user_msg = info["user_msg"]
if "Database Schema:" in user_msg:
schema_str = user_msg.split("Database Schema:", 1)[1].split("Question:", 1)[0].rstrip()
else:
schema_str = user_msg
# 1) Planner SQL (greedy)
planning_prompt = user_msg.rstrip() + "\n\nPlanning:"
plans = vllm_complete(
args.planner_host, "planner", qwen_chat(planning_prompt),
n=1, temperature=0.0, top_p=1.0, max_tokens=1024, seed=args.seed + idx,
)
if not plans:
return ("no_planner", [])
m = re.search(r"Final SQL query:\s*```(?:sql)?\s*(.+?)```", plans[0], re.DOTALL | re.IGNORECASE)
planner_sql = m.group(1).strip() if m else extract_sql(plans[0])
if not planner_sql:
return ("no_planner", [])
# 2) Execute planner SQL
with _db_lock:
gold_res, _ = safe_exec(db_path, bt["sql"])
pred_res, perr = safe_exec(db_path, planner_sql)
if gold_res is None:
return ("no_gold", [])
planner_correct = (not perr) and results_match(gold_res, pred_res)
exec_response = (f"Error: {perr[:200]}" if perr
else f"OK. Result rows (preview): {str(pred_res)[:300]}")
# 3) Generate K validator critiques (paper format, seeded with clause token)
clause_token = "SELECT." if side == "sel" else "CONDITION."
schema_in_val_prompt = (info["user_msg"]
.split("Database Schema:", 1)[1].split("Question:", 1)[0]).rstrip() \
if "Database Schema:" in info["user_msg"] else info["user_msg"]
val_prompt = (
f"Generate feedbacks to fix the following SQL query:\n"
f"Database Schema:{schema_in_val_prompt}\n\n"
f"Question: {question}\n"
f"External knowledge: {evidence}\n\n"
f"SQL query: {planner_sql}\n\n"
f"Execution response:\n{exec_response}\n\n"
f"Feedback:"
)
seeded_prompt = val_prompt + "\n" + clause_token + "\n"
critiques = vllm_complete(
args.validator_host, "validator", llama3_chat(seeded_prompt),
n=args.K, temperature=args.temperature, top_p=0.9,
max_tokens=384, seed=args.seed + idx,
)
if not critiques:
return ("no_val", [])
critiques = [f"{clause_token}\n{c.lstrip()}" for c in critiques]
# 4) For each critique, ask the 72B fixer (qwen_chat format) and check correctness
chosen, rejected = [], []
for crit in critiques:
wrapped_crit = (
f"<{'select' if side == 'sel' else 'condition'}>\n{crit}\n"
f"</{'select' if side == 'sel' else 'condition'}>"
)
fix_prompt = build_fixer_prompt(schema_str, question, evidence, planner_sql, exec_response, wrapped_crit)
fix_outs = vllm_complete(
args.fixer_host, "fixer_72b", qwen_chat(fix_prompt),
n=1, temperature=0.0, top_p=1.0, max_tokens=512,
seed=args.seed + idx,
)
if not fix_outs:
rejected.append(crit)
continue
fix_sql = extract_sql(fix_outs[0])
if not fix_sql:
rejected.append(crit)
continue
with _db_lock:
fix_res, fix_err = safe_exec(db_path, fix_sql)
fix_correct = (not fix_err) and results_match(gold_res, fix_res)
(chosen if fix_correct else rejected).append(crit)
pairs = []
if chosen and rejected:
for c in chosen[:2]:
for r in rejected[:2]:
pairs.append({"prompt": val_prompt, "chosen": c, "rejected": r})
status = "planner_correct" if planner_correct else "planner_wrong"
return (status, pairs, len(chosen), len(rejected))
def main():
p = argparse.ArgumentParser()
p.add_argument("--planner_host", default="http://localhost:8100")
p.add_argument("--validator_host", default="http://localhost:8101")
p.add_argument("--fixer_host", default="http://localhost:8102")
p.add_argument("--side", required=True, choices=["sel", "cond"])
p.add_argument("--K", type=int, default=4)
p.add_argument("--temperature", type=float, default=1.0)
p.add_argument("--max_questions", type=int, default=-1, help="-1 = use full dataset (default)")
p.add_argument("--threads", type=int, default=16)
p.add_argument("--seed", type=int, default=42)
p.add_argument("--out", required=True)
args = p.parse_args()
print("Loading BIRD-train + griffith prompts...", flush=True)
with open("data/sft_bird_with_evidence_train_text2sql.json") as f:
bird_train = json.load(f)
ds_g = load_dataset("griffith-bigdata/sft_text2sql", split="train_sft",
cache_dir="/weka/s225250685/Huggingface/hub"
).filter(lambda x: x["model_name"] == "deepseek-reasoner")
griffith = {}
for row in ds_g:
sid = int(row["sample_id"])
if not (0 <= sid < len(bird_train)): continue
user_msg = row["messages"][1]["content"]
q_m = re.search(r"Question:\s*(.+?)(?:\n|$)", user_msg)
if not q_m: continue
q = q_m.group(1).strip()
if q.lower() == bird_train[sid]["question"].strip().lower():
griffith[q.lower()] = {"user_msg": user_msg, "sid": sid}
print(f" griffith: {len(griffith)} questions", flush=True)
random.seed(args.seed)
items = list(griffith.items()); random.shuffle(items)
chunk = items[:(args.max_questions if args.max_questions > 0 else len(items))]
rows_all = []
counters = {"planner_correct": 0, "planner_wrong": 0,
"no_planner": 0, "skip_no_db": 0, "no_gold": 0, "no_val": 0}
total_chosen = 0
total_rejected = 0
print(f"Processing {len(chunk)} questions with {args.threads} threads, K={args.K}, side={args.side}...",
flush=True)
with ThreadPoolExecutor(max_workers=args.threads) as ex:
futures = []
for idx, (q_lower, info) in enumerate(chunk):
futures.append(ex.submit(process_one, args, q_lower, info, bird_train, args.side, idx))
done = 0
for fut in as_completed(futures):
try:
result = fut.result()
if len(result) == 4:
status, pairs, n_c, n_r = result
total_chosen += n_c
total_rejected += n_r
else:
status, pairs = result
except Exception as e:
print(f" worker exception: {e}", flush=True)
continue
counters[status] = counters.get(status, 0) + 1
rows_all.extend(pairs)
done += 1
if done % 50 == 0:
print(f" [{done}/{len(chunk)}] pairs={len(rows_all)} "
f"chosen_traj={total_chosen} rejected_traj={total_rejected} "
f"ok={counters['planner_correct']} wrong={counters['planner_wrong']} "
f"no_planner={counters['no_planner']} no_gold={counters['no_gold']} no_val={counters['no_val']}",
flush=True)
print(f"\nGenerated {len(rows_all)} (chosen, rejected) pairs", flush=True)
print(f" counters: {counters}", flush=True)
print(f" total critiques labeled chosen={total_chosen}, rejected={total_rejected}", flush=True)
if not rows_all:
print("ERROR: no rows generated"); return
random.seed(42); random.shuffle(rows_all)
n_train = int(0.95 * len(rows_all))
DatasetDict({
"train_dpo": Dataset.from_list(rows_all[:n_train]),
"test_dpo": Dataset.from_list(rows_all[n_train:]),
}).save_to_disk(args.out)
print(f"Saved → {args.out} train={n_train} test={len(rows_all) - n_train}", flush=True)
if __name__ == "__main__":
main()
|