File size: 19,495 Bytes
5c45a62 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 | """
ORPO data generation for MATS pipeline (paper §4 / Alg. 1, Alg. 2).
Modes:
--agent planner — Alg. 1: K rollouts on BIRD-TRAIN, chosen=correct SQL, rejected=wrong
--agent validator_sel — Alg. 2 collaborative: validator critique is chosen if FIXER (using it)
produces correct SQL, rejected otherwise. Uses previous-iter fixer.
--agent validator_cond — same as validator_sel but for condition critique
--agent fixer — fixer chosen=correct corrected SQL, rejected=wrong
--mode collab — use the trained fixer to judge validator outputs (paper §4.3)
--mode collab_v2 — inference-aligned: critique-says-None ⇒ keep planner SQL; else run fixer.
Chosen/rejected by FINAL pipeline SQL correctness. Filters pairs where
critique-text actually influenced final outcome.
--mode independent — use a heuristic (e.g., string "INCORRECT" in critique when SQL is wrong)
to mark chosen/rejected, no fixer involvement. For baseline comparison.
Output: HF dataset with {prompt, chosen, rejected} for ORPO training.
"""
import argparse, os, re, json, random, sqlite3, threading
os.environ.setdefault("PYTHONNOUSERSITE", "1")
os.environ["NO_PROXY"] = "localhost,127.0.0.1"
import requests
from datasets import load_dataset, Dataset, DatasetDict
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=180)
r.raise_for_status()
return [c["text"].strip() for c in r.json()["choices"]]
except Exception as e:
return []
def build_planner_data(args, griffith, bird_train):
"""Alg. 1 — planner ORPO data."""
rows = []
random.seed(args.seed)
items = list(griffith.items()); random.shuffle(items)
n_correct_only = 0; n_wrong_only = 0; n_pairs = 0
for i, (q_lower, info) in enumerate(items[:args.max_questions if args.max_questions > 0 else len(items)]):
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): continue
planning_prompt = info["user_msg"].rstrip() + "\n\nPlanning:"
chat = qwen_chat(planning_prompt)
outs = vllm_complete(args.planner_host, "planner", chat,
n=args.K, temperature=args.temperature, top_p=0.9,
max_tokens=1024, seed=args.seed + i)
if not outs: continue
gold_res, _ = safe_exec(db_path, bt["sql"])
if gold_res is None: continue
correct, wrong = [], []
for cot in outs:
sql = extract_sql(cot)
if not sql: continue
pred_res, err = safe_exec(db_path, sql)
if err or not results_match(gold_res, pred_res):
wrong.append(cot)
else:
correct.append(cot)
if correct and wrong:
for c in correct[:2]:
for w in wrong[:2]:
rows.append({"prompt": planning_prompt, "chosen": c, "rejected": w})
n_pairs += 1
elif correct: n_correct_only += 1
elif wrong: n_wrong_only += 1
if (i+1) % 200 == 0:
print(f" [{i+1}] pairs={n_pairs}, only_c={n_correct_only}, only_w={n_wrong_only}", flush=True)
return rows
def build_validator_data(args, griffith, bird_train, side):
"""Alg. 2 — collaborative validator ORPO data.
For each (planner_sql, planner_exec_response):
generate K validator critiques (sel or cond)
For each critique: feed to FIXER, check if fixer output is correct.
Chosen = critique that led to correct fix
Rejected = critique that led to wrong fix (or no improvement)
Mode 'independent': mark chosen/rejected by heuristic on SQL correctness alone (no fixer).
"""
# Paper format: validator prompt uses "Generate feedbacks ... Feedback:" (data_processing/
# generate_sft_data_for_validator.py) and completion ends with "Conclude: correct/incorrect."
# The val-sel and val-cond models share this prompt; they differ only by their training
# completion (SELECT. vs CONDITION. block).
FIXER_INSTR = ("You are a SQL fixer. Given the question, schema, original SQL query, "
"execution response, and the validator's critique below, output ONLY the corrected "
"final SQL inside ```sql ... ``` markers.")
clause_token = "SELECT." if side == "sel" else "CONDITION."
rows = []
random.seed(args.seed)
items = list(griffith.items()); random.shuffle(items)
n_pairs = 0
for i, (q_lower, info) in enumerate(items[:args.max_questions if args.max_questions > 0 else len(items)]):
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): continue
# Step 1: get a planner SQL (greedy)
planning_prompt = info["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)
if not plans: continue
planner_sql = extract_sql(plans[0])
if not planner_sql: continue
# Step 2: execute planner SQL
gold_res, _ = safe_exec(db_path, bt["sql"])
pred_res, err = safe_exec(db_path, planner_sql)
if gold_res is None: continue
planner_correct = (not err) and results_match(gold_res, pred_res)
exec_response = (f"Error: {err[:200]}" if err
else f"OK. Result rows (preview): {str(pred_res)[:300]}")
# Step 3: generate K validator critiques (paper format)
schema = info["user_msg"].split("Database Schema:", 1)[1].split("Question:", 1)[0] \
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.rstrip()}\n\n"
f"Question: {bt['question']}\n"
f"External knowledge: {bt.get('evidence','')}\n\n"
f"SQL query: {planner_sql}\n\n"
f"Execution response:\n{exec_response}\n\n"
f"Feedback:")
# Seed each critique with the clause token so the val-sel/val-cond model continues directly
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 + i)
if not critiques: continue
# Re-prepend the clause token (vLLM returns only the continuation)
critiques = [f"{clause_token}\n{c.lstrip()}" for c in critiques]
chosen, rejected = [], []
if args.mode == "collab":
# Use fixer to judge each critique
for crit in critiques:
# Build fixer prompt with this critique
# Wrap paper-format critique in legacy <select>/<condition> tags so the
# existing wrapper-tag-trained fixer SFT model sees the format it expects.
wrapped_crit = f"<{'select' if side == 'sel' else 'condition'}>\n{crit}\n</{'select' if side == 'sel' else 'condition'}>"
fix_prompt = (FIXER_INSTR + "\n\nDatabase Schema:\n" +
info["user_msg"].split("Database Schema:")[1].split("Question:")[0].rstrip() +
f"\n\nQuestion: {bt['question']}\nExternal knowledge: {bt.get('evidence','None')}\n\n"
f"Generated SQL query: {planner_sql}\n\n"
f"Execution response:\n{exec_response}\n\n"
f"Validator critique:\n{wrapped_crit}\n\n"
f"Final SQL:")
fix_outs = vllm_complete(args.fixer_host, "fixer", llama3_chat(fix_prompt),
n=1, temperature=0.0, top_p=1.0, max_tokens=512,
seed=args.seed + i)
if not fix_outs: continue
fix_sql = extract_sql(fix_outs[0])
if not fix_sql: continue
fix_res, fix_err = safe_exec(db_path, fix_sql)
fix_correct = (not fix_err) and results_match(gold_res, fix_res)
if fix_correct:
chosen.append(crit)
else:
rejected.append(crit)
elif args.mode == "collab_v2":
# Inference-aligned: Conclude:correct ⇒ keep planner SQL; else run fixer.
# Chosen/rejected by FINAL pipeline SQL correctness.
def critique_says_no_fix(crit):
# Paper format: "Conclude: correct." means no fix needed
return "Conclude: correct" in crit
outcomes = [] # (crit, final_correct, says_no_fix)
for crit in critiques:
says_no_fix = critique_says_no_fix(crit)
if says_no_fix:
final_sql = planner_sql
else:
fix_prompt = (FIXER_INSTR + "\n\nDatabase Schema:\n" +
info["user_msg"].split("Database Schema:")[1].split("Question:")[0].rstrip() +
f"\n\nQuestion: {bt['question']}\nExternal knowledge: {bt.get('evidence','None')}\n\n"
f"Generated SQL query: {planner_sql}\n\n"
f"Execution response:\n{exec_response}\n\n"
f"Validator critique:\n{crit}\n\n"
f"Final SQL:")
fix_outs = vllm_complete(args.fixer_host, "fixer", llama3_chat(fix_prompt),
n=1, temperature=0.0, top_p=1.0, max_tokens=512,
seed=args.seed + i)
fix_sql = extract_sql(fix_outs[0]) if fix_outs else ""
final_sql = fix_sql if fix_sql else planner_sql
fres, ferr = safe_exec(db_path, final_sql)
fcorrect = (not ferr) and results_match(gold_res, fres)
outcomes.append((crit, fcorrect, says_no_fix))
# Filter: only keep pairs where critique-text actually influenced outcome.
# Skip questions where all critiques landed in same bucket OR all say the same thing.
distinct_says = len(set(o[2] for o in outcomes))
if distinct_says >= 2: # at least one None-critique and one fix-critique
for crit, fcorrect, _ in outcomes:
(chosen if fcorrect else rejected).append(crit)
# Fallback (single-bucket-says, but outcomes differ): still use end-to-end signal
elif len({o[1] for o in outcomes}) >= 2:
for crit, fcorrect, _ in outcomes:
(chosen if fcorrect else rejected).append(crit)
else: # independent mode
# Paper format: critique should "Conclude: correct" if planner SQL is correct,
# "Conclude: incorrect" if wrong.
for crit in critiques:
says_correct = "Conclude: correct" in crit
says_incorrect = "Conclude: incorrect" in crit
if planner_correct and says_correct:
chosen.append(crit)
elif not planner_correct and says_incorrect:
chosen.append(crit)
elif says_correct or says_incorrect:
rejected.append(crit)
# critiques with no conclusion are skipped
if chosen and rejected:
for c in chosen[:2]:
for r in rejected[:2]:
rows.append({"prompt": val_prompt, "chosen": c, "rejected": r})
n_pairs += 1
if (i+1) % 100 == 0:
print(f" [{i+1}] pairs={n_pairs}", flush=True)
return rows
def build_fixer_data(args, griffith, bird_train):
"""Fixer ORPO: K fixer outputs, chosen=correct, rejected=wrong."""
FIXER_INSTR = ("You are a SQL fixer. Given the question, schema, original SQL query, "
"execution response, and the validator's critique below, output ONLY the corrected "
"final SQL inside ```sql ... ``` markers.")
rows = []
random.seed(args.seed)
items = list(griffith.items()); random.shuffle(items)
n_pairs = 0
for i, (q_lower, info) in enumerate(items[:args.max_questions if args.max_questions > 0 else len(items)]):
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): continue
# Get planner SQL
planning_prompt = info["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)
if not plans: continue
planner_sql = extract_sql(plans[0])
if not planner_sql: continue
gold_res, _ = safe_exec(db_path, bt["sql"])
pred_res, err = safe_exec(db_path, planner_sql)
if gold_res is None: continue
if (not err) and results_match(gold_res, pred_res): continue # planner already correct, skip
exec_response = (f"Error: {err[:200]}" if err
else f"OK. Result rows (preview): {str(pred_res)[:300]}")
# Get validator critiques
val_critique = "<select>\nSELECT.\nINCORRECT\n</select>\n\n<condition>\nCONDITION.\nINCORRECT\n</condition>"
# Build fixer prompt
fix_prompt = (FIXER_INSTR + "\n\nDatabase Schema:\n" +
info["user_msg"].split("Database Schema:")[1].split("Question:")[0].rstrip() +
f"\n\nQuestion: {bt['question']}\nExternal knowledge: {bt.get('evidence','None')}\n\n"
f"Generated SQL query: {planner_sql}\n\n"
f"Execution response:\n{exec_response}\n\n"
f"Validator critique:\n{val_critique}\n\n"
f"Final SQL:")
outs = vllm_complete(args.fixer_host, "fixer", llama3_chat(fix_prompt),
n=args.K, temperature=args.temperature, top_p=0.9,
max_tokens=512, seed=args.seed + i)
if not outs: continue
correct, wrong = [], []
for fix_text in outs:
fix_sql = extract_sql(fix_text)
if not fix_sql: continue
fix_res, fix_err = safe_exec(db_path, fix_sql)
if (not fix_err) and results_match(gold_res, fix_res):
correct.append(fix_text)
else:
wrong.append(fix_text)
if correct and wrong:
for c in correct[:2]:
for w in wrong[:2]:
rows.append({"prompt": fix_prompt, "chosen": c, "rejected": w})
n_pairs += 1
if (i+1) % 200 == 0:
print(f" [{i+1}] fixer pairs={n_pairs}", flush=True)
return rows
def main():
p = argparse.ArgumentParser()
p.add_argument("--agent", required=True, choices=["planner", "validator_sel", "validator_cond", "fixer"])
p.add_argument("--mode", default="collab", choices=["collab", "collab_v2", "independent"])
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("--K", type=int, default=8)
p.add_argument("--temperature", type=float, default=1.0)
p.add_argument("--max_questions", type=int, default=-1)
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)
if args.agent == "planner":
rows = build_planner_data(args, griffith, bird_train)
elif args.agent == "validator_sel":
rows = build_validator_data(args, griffith, bird_train, "sel")
elif args.agent == "validator_cond":
rows = build_validator_data(args, griffith, bird_train, "cond")
elif args.agent == "fixer":
rows = build_fixer_data(args, griffith, bird_train)
print(f"\nGenerated {len(rows)} (chosen, rejected) pairs", flush=True)
if not rows:
print("ERROR: no pairs generated"); return
random.seed(42); random.shuffle(rows)
n_train = int(0.95 * len(rows))
DatasetDict({
"train_dpo": Dataset.from_list(rows[:n_train]),
"test_dpo": Dataset.from_list(rows[n_train:]),
}).save_to_disk(args.out)
print(f"Saved → {args.out} train={n_train} test={len(rows)-n_train}", flush=True)
if __name__ == "__main__":
main()
|