Spaces:
Sleeping
Sleeping
savvy
#1
by shambhalala - opened
- .gitignore +1 -2
- README.md +1 -1
- inference.py +166 -365
.gitignore
CHANGED
|
@@ -30,5 +30,4 @@ Thumbs.db
|
|
| 30 |
# Pytest / coverage
|
| 31 |
.pytest_cache/
|
| 32 |
.coverage
|
| 33 |
-
htmlcov/
|
| 34 |
-
ai_env/
|
|
|
|
| 30 |
# Pytest / coverage
|
| 31 |
.pytest_cache/
|
| 32 |
.coverage
|
| 33 |
+
htmlcov/
|
|
|
README.md
CHANGED
|
@@ -254,4 +254,4 @@ debugging where production data isn't available.
|
|
| 254 |
`run_partial` + `ask_clarification` before `submit_query` models real
|
| 255 |
developer behaviour: check syntax, check schema, then commit. The turn
|
| 256 |
penalty in the reward function ensures the agent learns to use these
|
| 257 |
-
efficiently rather than exhausting the budget.
|
|
|
|
| 254 |
`run_partial` + `ask_clarification` before `submit_query` models real
|
| 255 |
developer behaviour: check syntax, check schema, then commit. The turn
|
| 256 |
penalty in the reward function ensures the agent learns to use these
|
| 257 |
+
efficiently rather than exhausting the budget.
|
inference.py
CHANGED
|
@@ -1,425 +1,226 @@
|
|
| 1 |
"""
|
| 2 |
-
|
| 3 |
-
============
|
|
|
|
| 4 |
|
| 5 |
-
|
|
|
|
| 6 |
[START] task=<task_id> env=sql-repair model=<model_name>
|
| 7 |
[STEP] step=<n> action=<action_str> reward=<0.00> done=<true|false> error=<msg|null>
|
| 8 |
-
[END] success=<true|false> steps=<n> score=<score> rewards=<r1,r2,...
|
| 9 |
-
|
| 10 |
-
ENV VARS
|
| 11 |
-
HF_TOKEN / API_KEY — LLM API key
|
| 12 |
-
API_BASE_URL — LLM endpoint (default: HF router)
|
| 13 |
-
MODEL_NAME — Model ID (default: Qwen2.5-72B-Instruct)
|
| 14 |
-
ENV_SERVER_URL — Running server (default: http://localhost:8000)
|
| 15 |
-
SQL_TASK_ID — Task to run (default: easy_001)
|
| 16 |
-
SQL_DIFFICULTY — Difficulty (default: easy)
|
| 17 |
"""
|
| 18 |
|
| 19 |
-
import
|
|
|
|
|
|
|
| 20 |
import os
|
| 21 |
-
import sys
|
| 22 |
import textwrap
|
| 23 |
from typing import List, Optional
|
| 24 |
|
| 25 |
-
import httpx
|
| 26 |
from openai import OpenAI
|
| 27 |
|
| 28 |
-
#
|
| 29 |
-
|
| 30 |
-
|
| 31 |
-
|
| 32 |
-
|
| 33 |
-
|
| 34 |
-
|
| 35 |
-
|
| 36 |
-
BENCHMARK
|
| 37 |
-
|
| 38 |
-
|
| 39 |
-
|
| 40 |
-
|
| 41 |
-
|
| 42 |
-
|
| 43 |
-
|
| 44 |
-
|
| 45 |
-
|
| 46 |
-
|
| 47 |
-
|
| 48 |
-
|
| 49 |
-
|
| 50 |
-
|
| 51 |
-
|
| 52 |
-
|
| 53 |
-
|
| 54 |
-
tags (id, product_id, tag)
|
| 55 |
-
|
| 56 |
-
RULES:
|
| 57 |
-
- You have a limited number of turns. Use them wisely.
|
| 58 |
-
- You have THREE action types:
|
| 59 |
-
1. ask_clarification — ask ONE focused question about schema or intent.
|
| 60 |
-
Use ONLY when the intent is genuinely ambiguous.
|
| 61 |
-
2. run_partial — test a candidate query to see if it runs without error.
|
| 62 |
-
Use to validate your fix before submitting.
|
| 63 |
-
3. submit_query — your final corrected query. Ends the episode.
|
| 64 |
-
|
| 65 |
-
STRATEGY:
|
| 66 |
-
- Read the broken query and the intent carefully.
|
| 67 |
-
- Identify the exact bug (wrong table name, wrong operator, bad join key, etc.)
|
| 68 |
-
- For easy tasks: fix directly and submit.
|
| 69 |
-
- For medium/hard tasks: run_partial once to verify, then submit.
|
| 70 |
-
- Only ask_clarification if you truly cannot determine the schema/intent.
|
| 71 |
-
- Never submit a query you have not at least mentally validated.
|
| 72 |
-
|
| 73 |
-
RESPONSE FORMAT — respond with a JSON object and nothing else.
|
| 74 |
-
Do NOT wrap in markdown fences.
|
| 75 |
-
|
| 76 |
-
To ask a clarification:
|
| 77 |
-
{"action": "ask_clarification", "question": "<your question>"}
|
| 78 |
-
|
| 79 |
-
To test a query:
|
| 80 |
-
{"action": "run_partial", "query": "<SELECT ...>"}
|
| 81 |
-
|
| 82 |
-
To submit your final fix:
|
| 83 |
-
{"action": "submit_query", "query": "<SELECT ...>"}
|
| 84 |
-
""").strip()
|
| 85 |
-
|
| 86 |
-
|
| 87 |
-
def build_user_prompt(obs: dict, history: List[str]) -> str:
|
| 88 |
-
history_block = "\n".join(history) if history else "None"
|
| 89 |
-
clarification = obs.get("clarification_response") or "None"
|
| 90 |
-
error_msg = obs.get("error_message") or "None"
|
| 91 |
-
return textwrap.dedent(f"""
|
| 92 |
-
TASK ID: {obs.get('task_id', TASK_ID)}
|
| 93 |
-
DIFFICULTY: {obs.get('difficulty', DIFFICULTY)}
|
| 94 |
-
INTENT: {obs.get('intent', '')}
|
| 95 |
-
SCHEMA HINT: {obs.get('schema_hint', '')}
|
| 96 |
-
BROKEN QUERY: {obs.get('broken_query', '')}
|
| 97 |
-
TURN: {obs.get('turn', '?')} / {obs.get('max_turns', '?')}
|
| 98 |
-
LAST ERROR: {error_msg}
|
| 99 |
-
LAST CLARIFICATION RESPONSE: {clarification}
|
| 100 |
-
LAST ACTION TYPE: {obs.get('last_action_type') or 'None'}
|
| 101 |
-
|
| 102 |
-
HISTORY:
|
| 103 |
-
{history_block}
|
| 104 |
-
|
| 105 |
-
Respond with a JSON object for your next action.
|
| 106 |
-
""").strip()
|
| 107 |
-
|
| 108 |
|
| 109 |
-
# ── Logging ───────────────────────────────────────────────────────────────────
|
| 110 |
def log_start(task: str, env: str, model: str) -> None:
|
| 111 |
print(f"[START] task={task} env={env} model={model}", flush=True)
|
| 112 |
|
| 113 |
|
| 114 |
def log_step(step: int, action: str, reward: float, done: bool, error: Optional[str]) -> None:
|
| 115 |
-
|
| 116 |
-
action_clean = action.replace("\n", " ").replace("\r", "")[:200]
|
| 117 |
print(
|
| 118 |
-
f"[STEP] step={step} action={action_clean}
|
| 119 |
-
f"done={str(done).lower()} error={
|
| 120 |
flush=True,
|
| 121 |
)
|
| 122 |
|
| 123 |
|
| 124 |
def log_end(success: bool, steps: int, score: float, rewards: List[float]) -> None:
|
| 125 |
-
rewards_str = ",".join(f"{r:.2f}" for r in rewards)
|
| 126 |
print(
|
| 127 |
f"[END] success={str(success).lower()} steps={steps} "
|
| 128 |
-
f"score={score:.3f} rewards={
|
| 129 |
flush=True,
|
| 130 |
)
|
| 131 |
|
|
|
|
|
|
|
|
|
|
| 132 |
|
| 133 |
-
|
| 134 |
-
|
| 135 |
-
|
| 136 |
-
|
| 137 |
-
|
| 138 |
-
|
| 139 |
-
|
| 140 |
-
|
| 141 |
-
)
|
| 142 |
-
resp.raise_for_status()
|
| 143 |
-
print(f"[DEBUG] Reset successful", flush=True)
|
| 144 |
-
return resp.json()
|
| 145 |
-
except httpx.TimeoutException as exc:
|
| 146 |
-
print(f"[DEBUG] /reset timed out: {exc}", flush=True)
|
| 147 |
-
raise
|
| 148 |
-
except httpx.HTTPStatusError as exc:
|
| 149 |
-
print(f"[DEBUG] /reset HTTP {exc.response.status_code}: {exc.response.text[:300]}", flush=True)
|
| 150 |
-
raise
|
| 151 |
-
except httpx.RequestError as exc:
|
| 152 |
-
print(f"[DEBUG] /reset connection error: {exc}", flush=True)
|
| 153 |
-
raise
|
| 154 |
-
except Exception as exc:
|
| 155 |
-
print(f"[DEBUG] /reset unexpected error: {exc}", flush=True)
|
| 156 |
-
raise
|
| 157 |
-
|
| 158 |
-
|
| 159 |
-
def api_step(http: httpx.Client, payload: dict) -> dict:
|
| 160 |
-
try:
|
| 161 |
-
resp = http.post(
|
| 162 |
-
f"{ENV_SERVER_URL}/step",
|
| 163 |
-
json=payload,
|
| 164 |
-
timeout=30,
|
| 165 |
-
)
|
| 166 |
-
resp.raise_for_status()
|
| 167 |
-
return resp.json()
|
| 168 |
-
except httpx.TimeoutException as exc:
|
| 169 |
-
print(f"[DEBUG] /step timed out: {exc}", flush=True)
|
| 170 |
-
raise
|
| 171 |
-
except httpx.HTTPStatusError as exc:
|
| 172 |
-
print(f"[DEBUG] /step HTTP {exc.response.status_code}: {exc.response.text[:300]}", flush=True)
|
| 173 |
-
raise
|
| 174 |
-
except httpx.RequestError as exc:
|
| 175 |
-
print(f"[DEBUG] /step connection error: {exc}", flush=True)
|
| 176 |
-
raise
|
| 177 |
-
except Exception as exc:
|
| 178 |
-
print(f"[DEBUG] /step unexpected error: {exc}", flush=True)
|
| 179 |
-
raise
|
| 180 |
|
| 181 |
|
| 182 |
-
def
|
| 183 |
-
|
| 184 |
-
|
| 185 |
-
|
| 186 |
-
return resp.json()
|
| 187 |
-
except httpx.TimeoutException as exc:
|
| 188 |
-
print(f"[DEBUG] /grader timed out: {exc}", flush=True)
|
| 189 |
-
raise
|
| 190 |
-
except httpx.HTTPStatusError as exc:
|
| 191 |
-
print(f"[DEBUG] /grader HTTP {exc.response.status_code}: {exc.response.text[:300]}", flush=True)
|
| 192 |
-
raise
|
| 193 |
-
except httpx.RequestError as exc:
|
| 194 |
-
print(f"[DEBUG] /grader connection error: {exc}", flush=True)
|
| 195 |
-
raise
|
| 196 |
-
except Exception as exc:
|
| 197 |
-
print(f"[DEBUG] /grader unexpected error: {exc}", flush=True)
|
| 198 |
-
raise
|
| 199 |
|
|
|
|
|
|
|
|
|
|
| 200 |
|
| 201 |
-
|
| 202 |
-
|
| 203 |
-
|
| 204 |
-
|
| 205 |
-
|
| 206 |
-
|
| 207 |
-
|
| 208 |
-
|
|
|
|
| 209 |
|
| 210 |
-
# Build prompt
|
| 211 |
-
try:
|
| 212 |
-
user_prompt = build_user_prompt(obs, history)
|
| 213 |
-
except Exception as exc:
|
| 214 |
-
print(f"[DEBUG] prompt build error: {exc}", flush=True)
|
| 215 |
-
return fallback
|
| 216 |
|
| 217 |
-
|
| 218 |
-
|
| 219 |
-
# Call LLM
|
| 220 |
try:
|
| 221 |
-
|
| 222 |
model=MODEL_NAME,
|
| 223 |
messages=[
|
| 224 |
{"role": "system", "content": SYSTEM_PROMPT},
|
| 225 |
-
{"role": "user",
|
| 226 |
],
|
| 227 |
temperature=TEMPERATURE,
|
| 228 |
max_tokens=MAX_TOKENS,
|
| 229 |
-
stream=False,
|
| 230 |
)
|
| 231 |
-
|
| 232 |
except Exception as exc:
|
| 233 |
-
print(f"[DEBUG] LLM
|
| 234 |
-
return
|
| 235 |
|
| 236 |
-
# Extract text
|
| 237 |
-
try:
|
| 238 |
-
raw = (completion.choices[0].message.content or "").strip()
|
| 239 |
-
except Exception as exc:
|
| 240 |
-
print(f"[DEBUG] LLM response read failed: {exc}", flush=True)
|
| 241 |
-
return fallback
|
| 242 |
|
| 243 |
-
|
| 244 |
-
|
| 245 |
-
|
| 246 |
-
|
| 247 |
-
|
| 248 |
-
|
| 249 |
-
|
| 250 |
-
|
| 251 |
-
|
| 252 |
-
|
| 253 |
|
| 254 |
-
# Parse JSON
|
| 255 |
-
try:
|
| 256 |
-
parsed = json.loads(raw)
|
| 257 |
-
except json.JSONDecodeError as exc:
|
| 258 |
-
print(f"[DEBUG] JSON parse failed ({exc}): {raw[:200]}", flush=True)
|
| 259 |
-
return fallback
|
| 260 |
|
| 261 |
-
|
| 262 |
-
|
| 263 |
-
|
| 264 |
-
|
| 265 |
-
|
| 266 |
-
|
| 267 |
-
|
| 268 |
-
|
| 269 |
-
return fallback
|
| 270 |
-
|
| 271 |
-
return parsed
|
| 272 |
-
|
| 273 |
-
|
| 274 |
-
def action_to_payload(parsed: dict) -> dict:
|
| 275 |
-
"""Convert LLM response dict to the JSON shape /step expects."""
|
| 276 |
-
action = parsed.get("action", "")
|
| 277 |
-
if action == "submit_query":
|
| 278 |
-
return {"type": "submit_query", "query": parsed.get("query", "")}
|
| 279 |
-
elif action == "run_partial":
|
| 280 |
-
return {"type": "run_partial", "query": parsed.get("query", "")}
|
| 281 |
-
elif action == "ask_clarification":
|
| 282 |
-
return {"type": "ask_clarification", "question": parsed.get("question", "")}
|
| 283 |
-
# Safe fallback (should not be reached after validation)
|
| 284 |
-
return {"type": "submit_query", "query": parsed.get("query", "SELECT 1")}
|
| 285 |
-
|
| 286 |
-
|
| 287 |
-
def action_repr(parsed: dict) -> str:
|
| 288 |
-
"""One-line string for [STEP] logging."""
|
| 289 |
-
action = parsed.get("action", "unknown")
|
| 290 |
-
if action in ("submit_query", "run_partial"):
|
| 291 |
-
q = parsed.get("query", "").replace("\n", " ")
|
| 292 |
-
return f'{action}("{q}")'
|
| 293 |
-
elif action == "ask_clarification":
|
| 294 |
-
return f'ask_clarification("{parsed.get("question", "")}")'
|
| 295 |
-
return action
|
| 296 |
-
|
| 297 |
-
|
| 298 |
-
# ── Main ──────────────────────────────────────────────────────────────────────
|
| 299 |
-
def main() -> None:
|
| 300 |
-
# Validate environment variables first
|
| 301 |
-
print(f"[DEBUG] Checking environment variables...", flush=True)
|
| 302 |
-
print(f"[DEBUG] API_KEY present: {bool(API_KEY)}", flush=True)
|
| 303 |
-
print(f"[DEBUG] API_BASE_URL: {API_BASE_URL[:50] if API_BASE_URL else 'MISSING'}...", flush=True)
|
| 304 |
-
print(f"[DEBUG] ENV_SERVER_URL: {ENV_SERVER_URL}", flush=True)
|
| 305 |
-
print(f"[DEBUG] MODEL_NAME: {MODEL_NAME}", flush=True)
|
| 306 |
-
|
| 307 |
-
if not API_KEY or not API_BASE_URL:
|
| 308 |
-
print("[ERROR] Missing required environment variables: API_KEY and API_BASE_URL must be set", flush=True)
|
| 309 |
-
# Still emit start/end for validator
|
| 310 |
-
log_start(task=TASK_ID, env=BENCHMARK, model=MODEL_NAME)
|
| 311 |
-
log_end(success=False, steps=0, score=0.0, rewards=[])
|
| 312 |
-
sys.exit(1)
|
| 313 |
-
|
| 314 |
-
rewards: List[float] = []
|
| 315 |
-
steps_taken = 0
|
| 316 |
-
score = 0.0
|
| 317 |
-
success = False
|
| 318 |
-
|
| 319 |
-
# Emit [START] before anything that can fail so the validator always sees it
|
| 320 |
-
log_start(task=TASK_ID, env=BENCHMARK, model=MODEL_NAME)
|
| 321 |
|
| 322 |
-
|
| 323 |
-
|
| 324 |
-
|
| 325 |
-
llm = OpenAI(base_url=API_BASE_URL, api_key=API_KEY)
|
| 326 |
-
print(f"[DEBUG] OpenAI client initialized successfully", flush=True)
|
| 327 |
-
except Exception as exc:
|
| 328 |
-
print(f"[DEBUG] OpenAI client init failed: {exc}", flush=True)
|
| 329 |
-
log_end(success=False, steps=0, score=0.0, rewards=[])
|
| 330 |
-
sys.exit(1)
|
| 331 |
|
| 332 |
-
|
|
|
|
|
|
|
| 333 |
history: List[str] = []
|
| 334 |
|
|
|
|
| 335 |
try:
|
| 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 |
-
reward, done, info = 0.0, True, {}
|
| 383 |
-
|
| 384 |
-
error_msg: Optional[str] = obs.get("error_message") if isinstance(obs, dict) else None
|
| 385 |
-
feedback: str = info.get("feedback", "") if isinstance(info, dict) else ""
|
| 386 |
-
|
| 387 |
-
rewards.append(reward)
|
| 388 |
-
steps_taken = step
|
| 389 |
-
|
| 390 |
-
log_step(step=step, action=action_str, reward=reward, done=done, error=error_msg)
|
| 391 |
-
|
| 392 |
-
history.append(
|
| 393 |
-
f"Step {step} [{parsed.get('action', '?')}]: {action_str} "
|
| 394 |
-
f"-> reward={reward:+.2f} | feedback: {feedback}"
|
| 395 |
-
)
|
| 396 |
-
|
| 397 |
-
if done:
|
| 398 |
-
print(f"[DEBUG] Episode done at step {step}", flush=True)
|
| 399 |
-
break
|
| 400 |
-
|
| 401 |
-
# ── Grader ────────────────────────────────────────────────────
|
| 402 |
-
try:
|
| 403 |
-
print(f"[DEBUG] Getting final score from grader", flush=True)
|
| 404 |
-
grader_result = api_grader(http)
|
| 405 |
-
score = float(grader_result.get("score", 0.0))
|
| 406 |
-
score = min(max(score, 0.0), 1.0)
|
| 407 |
-
print(f"[DEBUG] Grader score: {score}", flush=True)
|
| 408 |
-
except Exception as exc:
|
| 409 |
-
print(f"[DEBUG] /grader failed, falling back to reward sum: {exc}", flush=True)
|
| 410 |
-
score = min(max(sum(rewards), 0.0), 1.0)
|
| 411 |
-
|
| 412 |
-
success = score >= SUCCESS_SCORE_THRESHOLD
|
| 413 |
|
| 414 |
except Exception as exc:
|
| 415 |
-
print(f"[DEBUG]
|
| 416 |
-
|
| 417 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 418 |
|
| 419 |
-
|
| 420 |
-
|
| 421 |
-
log_end(success=success, steps=steps_taken, score=score, rewards=rewards)
|
| 422 |
|
| 423 |
|
| 424 |
if __name__ == "__main__":
|
| 425 |
-
main()
|
|
|
|
| 1 |
"""
|
| 2 |
+
inference.py
|
| 3 |
+
============
|
| 4 |
+
Baseline inference script for the Interactive SQL Query Repair environment.
|
| 5 |
|
| 6 |
+
Mandatory stdout format
|
| 7 |
+
-----------------------
|
| 8 |
[START] task=<task_id> env=sql-repair model=<model_name>
|
| 9 |
[STEP] step=<n> action=<action_str> reward=<0.00> done=<true|false> error=<msg|null>
|
| 10 |
+
[END] success=<true|false> steps=<n> score=<score> rewards=<r1,r2,...>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 11 |
"""
|
| 12 |
|
| 13 |
+
from __future__ import annotations
|
| 14 |
+
|
| 15 |
+
import asyncio
|
| 16 |
import os
|
|
|
|
| 17 |
import textwrap
|
| 18 |
from typing import List, Optional
|
| 19 |
|
|
|
|
| 20 |
from openai import OpenAI
|
| 21 |
|
| 22 |
+
# ---------------------------------------------------------------------------
|
| 23 |
+
# Config — exactly matching competition sample pattern
|
| 24 |
+
# ---------------------------------------------------------------------------
|
| 25 |
+
|
| 26 |
+
API_KEY = os.getenv("HF_TOKEN") or os.getenv("API_KEY")
|
| 27 |
+
API_BASE_URL = os.getenv("API_BASE_URL") or "https://router.huggingface.co/v1"
|
| 28 |
+
MODEL_NAME = os.getenv("MODEL_NAME") or "Qwen/Qwen2.5-72B-Instruct"
|
| 29 |
+
SERVER_URL = os.getenv("ENV_SERVER_URL", "http://localhost:7860")
|
| 30 |
+
BENCHMARK = "sql-repair"
|
| 31 |
+
|
| 32 |
+
EVAL_TASKS = [
|
| 33 |
+
("easy_001", "easy"),
|
| 34 |
+
("easy_007", "easy"),
|
| 35 |
+
("medium_001", "medium"),
|
| 36 |
+
("medium_005", "medium"),
|
| 37 |
+
("hard_001", "hard"),
|
| 38 |
+
("hard_004", "hard"),
|
| 39 |
+
]
|
| 40 |
+
|
| 41 |
+
MAX_STEPS = 8
|
| 42 |
+
TEMPERATURE = 0.2
|
| 43 |
+
MAX_TOKENS = 400
|
| 44 |
+
|
| 45 |
+
# ---------------------------------------------------------------------------
|
| 46 |
+
# Logging helpers (mandatory stdout format)
|
| 47 |
+
# ---------------------------------------------------------------------------
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 48 |
|
|
|
|
| 49 |
def log_start(task: str, env: str, model: str) -> None:
|
| 50 |
print(f"[START] task={task} env={env} model={model}", flush=True)
|
| 51 |
|
| 52 |
|
| 53 |
def log_step(step: int, action: str, reward: float, done: bool, error: Optional[str]) -> None:
|
| 54 |
+
action_clean = action.replace("\n", " ").replace("\r", "")
|
|
|
|
| 55 |
print(
|
| 56 |
+
f"[STEP] step={step} action={action_clean!r} "
|
| 57 |
+
f"reward={reward:.2f} done={str(done).lower()} error={error or 'null'}",
|
| 58 |
flush=True,
|
| 59 |
)
|
| 60 |
|
| 61 |
|
| 62 |
def log_end(success: bool, steps: int, score: float, rewards: List[float]) -> None:
|
|
|
|
| 63 |
print(
|
| 64 |
f"[END] success={str(success).lower()} steps={steps} "
|
| 65 |
+
f"score={score:.3f} rewards={','.join(f'{r:.2f}' for r in rewards)}",
|
| 66 |
flush=True,
|
| 67 |
)
|
| 68 |
|
| 69 |
+
# ---------------------------------------------------------------------------
|
| 70 |
+
# HTTP helpers
|
| 71 |
+
# ---------------------------------------------------------------------------
|
| 72 |
|
| 73 |
+
def _post(path: str, body: dict) -> dict:
|
| 74 |
+
import urllib.request, json
|
| 75 |
+
url = SERVER_URL.rstrip("/") + path
|
| 76 |
+
req = urllib.request.Request(
|
| 77 |
+
url, json.dumps(body).encode(),
|
| 78 |
+
headers={"Content-Type": "application/json"}
|
| 79 |
+
)
|
| 80 |
+
with urllib.request.urlopen(req, timeout=30) as r:
|
| 81 |
+
return json.loads(r.read())
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 82 |
|
| 83 |
|
| 84 |
+
def _get(path: str) -> dict:
|
| 85 |
+
import urllib.request, json
|
| 86 |
+
with urllib.request.urlopen(SERVER_URL.rstrip("/") + path, timeout=30) as r:
|
| 87 |
+
return json.loads(r.read())
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 88 |
|
| 89 |
+
# ---------------------------------------------------------------------------
|
| 90 |
+
# LLM helpers
|
| 91 |
+
# ---------------------------------------------------------------------------
|
| 92 |
|
| 93 |
+
SYSTEM_PROMPT = textwrap.dedent("""
|
| 94 |
+
You are an expert SQL repair agent.
|
| 95 |
+
Fix the broken SQL query so it satisfies the given intent.
|
| 96 |
+
Rules:
|
| 97 |
+
- SQLite-compatible SQL only.
|
| 98 |
+
- No window functions — use subqueries or CTEs instead.
|
| 99 |
+
- Output ONLY the fixed SQL, no explanation, no markdown, no backticks.
|
| 100 |
+
- For clarification requests, output ONLY the question text.
|
| 101 |
+
""").strip()
|
| 102 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 103 |
|
| 104 |
+
def _ask_llm(client: OpenAI, user_prompt: str) -> str:
|
|
|
|
|
|
|
| 105 |
try:
|
| 106 |
+
resp = client.chat.completions.create(
|
| 107 |
model=MODEL_NAME,
|
| 108 |
messages=[
|
| 109 |
{"role": "system", "content": SYSTEM_PROMPT},
|
| 110 |
+
{"role": "user", "content": user_prompt},
|
| 111 |
],
|
| 112 |
temperature=TEMPERATURE,
|
| 113 |
max_tokens=MAX_TOKENS,
|
|
|
|
| 114 |
)
|
| 115 |
+
return (resp.choices[0].message.content or "").strip()
|
| 116 |
except Exception as exc:
|
| 117 |
+
print(f"[DEBUG] LLM error: {exc}", flush=True)
|
| 118 |
+
return ""
|
| 119 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 120 |
|
| 121 |
+
def repair_prompt(intent, schema, broken, clarif, error, history) -> str:
|
| 122 |
+
return textwrap.dedent(f"""
|
| 123 |
+
Intent : {intent}
|
| 124 |
+
Schema : {schema}
|
| 125 |
+
Broken query : {broken}
|
| 126 |
+
Clarification: {clarif or 'None'}
|
| 127 |
+
Last error : {error or 'None'}
|
| 128 |
+
History : {chr(10).join(history[-4:]) or 'None'}
|
| 129 |
+
Output the corrected SQL query only.
|
| 130 |
+
""").strip()
|
| 131 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 132 |
|
| 133 |
+
def clarify_prompt(intent, schema, broken) -> str:
|
| 134 |
+
return textwrap.dedent(f"""
|
| 135 |
+
Intent : {intent}
|
| 136 |
+
Schema : {schema}
|
| 137 |
+
Broken query : {broken}
|
| 138 |
+
Ask ONE clarification question to help fix this query.
|
| 139 |
+
Output ONLY the question text.
|
| 140 |
+
""").strip()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 141 |
|
| 142 |
+
# ---------------------------------------------------------------------------
|
| 143 |
+
# Episode runner
|
| 144 |
+
# ---------------------------------------------------------------------------
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 145 |
|
| 146 |
+
def run_episode(client: OpenAI, task_id: str, difficulty: str) -> float:
|
| 147 |
+
rewards: List[float] = []
|
| 148 |
+
steps_taken, score, success = 0, 0.0, False
|
| 149 |
history: List[str] = []
|
| 150 |
|
| 151 |
+
log_start(task_id, BENCHMARK, MODEL_NAME)
|
| 152 |
try:
|
| 153 |
+
obs = _post("/reset", {"task_id": task_id, "difficulty": difficulty})
|
| 154 |
+
intent = obs["intent"]
|
| 155 |
+
schema = obs["schema_hint"]
|
| 156 |
+
broken_q = obs["broken_query"]
|
| 157 |
+
clarif_resp: Optional[str] = None
|
| 158 |
+
last_error: Optional[str] = None
|
| 159 |
+
|
| 160 |
+
for step in range(1, MAX_STEPS + 1):
|
| 161 |
+
steps_taken = step
|
| 162 |
+
|
| 163 |
+
if step == 1:
|
| 164 |
+
q = _ask_llm(client, clarify_prompt(intent, schema, broken_q)) \
|
| 165 |
+
or "What are the correct table and column names?"
|
| 166 |
+
action_body = {"type": "ask_clarification", "question": q}
|
| 167 |
+
action_str = f"ask_clarification({q!r})"
|
| 168 |
+
elif step == 2:
|
| 169 |
+
fq = _ask_llm(client, repair_prompt(intent, schema, broken_q, clarif_resp, last_error, history)) or broken_q
|
| 170 |
+
action_body = {"type": "run_partial", "query": fq}
|
| 171 |
+
action_str = f"run_partial({fq!r})"
|
| 172 |
+
else:
|
| 173 |
+
fq = _ask_llm(client, repair_prompt(intent, schema, broken_q, clarif_resp, last_error, history)) or broken_q
|
| 174 |
+
action_body = {"type": "submit_query", "query": fq}
|
| 175 |
+
action_str = f"submit_query({fq!r})"
|
| 176 |
+
|
| 177 |
+
result = _post("/step", action_body)
|
| 178 |
+
step_obs = result["observation"]
|
| 179 |
+
reward = float(result["reward"])
|
| 180 |
+
done = bool(result["done"])
|
| 181 |
+
info = result["info"]
|
| 182 |
+
err = step_obs.get("error_message")
|
| 183 |
+
|
| 184 |
+
rewards.append(reward)
|
| 185 |
+
clarif_resp = step_obs.get("clarification_response")
|
| 186 |
+
last_error = err
|
| 187 |
+
history.append(
|
| 188 |
+
f"step={step} type={action_body['type']} "
|
| 189 |
+
f"reward={reward:+.2f} feedback={info['feedback']!r}"
|
| 190 |
+
)
|
| 191 |
+
|
| 192 |
+
log_step(step, action_str, reward, done, err)
|
| 193 |
+
if done:
|
| 194 |
+
break
|
| 195 |
+
|
| 196 |
+
g = _get("/grader")
|
| 197 |
+
score = float(g["score"])
|
| 198 |
+
success = bool(g["solved"])
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 199 |
|
| 200 |
except Exception as exc:
|
| 201 |
+
print(f"[DEBUG] Episode error: {exc}", flush=True)
|
| 202 |
+
|
| 203 |
+
log_end(success, steps_taken, score, rewards)
|
| 204 |
+
return score
|
| 205 |
+
|
| 206 |
+
# ---------------------------------------------------------------------------
|
| 207 |
+
# Main — async to match competition sample pattern
|
| 208 |
+
# ---------------------------------------------------------------------------
|
| 209 |
+
|
| 210 |
+
async def main() -> None:
|
| 211 |
+
client = OpenAI(base_url=API_BASE_URL, api_key=API_KEY)
|
| 212 |
+
|
| 213 |
+
scores: List[float] = []
|
| 214 |
+
for task_id, difficulty in EVAL_TASKS:
|
| 215 |
+
try:
|
| 216 |
+
scores.append(run_episode(client, task_id, difficulty))
|
| 217 |
+
except Exception as exc:
|
| 218 |
+
print(f"[DEBUG] Task {task_id} failed: {exc}", flush=True)
|
| 219 |
+
scores.append(0.0)
|
| 220 |
|
| 221 |
+
avg = sum(scores) / len(scores) if scores else 0.0
|
| 222 |
+
print(f"\n[SUMMARY] tasks={len(scores)} avg_score={avg:.3f}", flush=True)
|
|
|
|
| 223 |
|
| 224 |
|
| 225 |
if __name__ == "__main__":
|
| 226 |
+
asyncio.run(main())
|