File size: 8,043 Bytes
72805b8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ac49ad8
 
 
 
 
72805b8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""

Inference Script - SQL Arena OpenEnv Environment

Baseline agent that uses an LLM to solve SQL challenges.

"""

import os
import sys
import textwrap
from typing import List, Optional

from openai import OpenAI

sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
from src.sql_arena.environment import SQLArenaEnvironment
from src.sql_arena.models import SQLArenaAction

# =====================================================
# Configuration
# =====================================================

API_KEY = os.getenv("HF_TOKEN") or os.getenv("API_KEY")
API_BASE_URL = os.getenv("API_BASE_URL") or "https://router.huggingface.co/v1"
MODEL_NAME = os.getenv("MODEL_NAME") or "Qwen/Qwen2.5-72B-Instruct"

BENCHMARK = "sql_arena"
TEMPERATURE = 0.3
MAX_TOKENS = 500

TASKS = [
    {"difficulty": "basic_select", "task_id": "easy_001", "name": "basic_select", "max_steps": 5},
    {"difficulty": "join_aggregate", "task_id": "medium_001", "name": "join_aggregate", "max_steps": 7},
    {"difficulty": "complex_analysis", "task_id": "hard_001", "name": "complex_analysis", "max_steps": 10},
]


# =====================================================
# Logging (MANDATORY format)
# =====================================================

def log_start(task: str, env: str, model: str) -> None:
    print(f"[START] task={task} env={env} model={model}", flush=True)


def log_step(step: int, action: str, reward: float, done: bool, error: Optional[str]) -> None:
    error_val = error if error else "null"
    done_val = str(done).lower()
    action_short = action.replace('\n', ' ').strip()[:100]
    print(
        f"[STEP] step={step} action={action_short} reward={reward:.2f} done={done_val} error={error_val}",
        flush=True,
    )


def log_end(success: bool, steps: int, score: float, rewards: List[float]) -> None:
    rewards_str = ",".join(f"{r:.2f}" for r in rewards)
    print(
        f"[END] success={str(success).lower()} steps={steps} score={score:.2f} rewards={rewards_str}",
        flush=True,
    )


# =====================================================
# LLM Agent
# =====================================================

SYSTEM_PROMPT = textwrap.dedent("""

    You are an expert SQL query writer. You are interacting with a SQL challenge environment.



    Each turn you receive: database schema, a question, previous query results, and feedback.

    Your goal: Write a SQL query that correctly answers the question.



    Rules:

    - Output ONLY the SQL query, nothing else

    - No explanations, no markdown, no code fences

    - Use standard SQLite syntax

    - Be precise with column names and table names

    - If your previous query had errors, fix them based on the feedback

""").strip()


def build_user_prompt(observation: dict, step: int, history: List[str]) -> str:
    parts = []
    parts.append(f"=== SQL Challenge (Step {step}) ===")
    parts.append(f"\nDifficulty: {observation.get('difficulty', 'unknown')}")
    parts.append(f"\n--- Database Schema ---\n{observation.get('schema_description', '')}")
    parts.append(f"\n--- Question ---\n{observation.get('question', '')}")

    if observation.get('expected_columns'):
        parts.append(f"\n--- Expected Columns ---\n{observation['expected_columns']}")
    if observation.get('query_result'):
        parts.append(f"\n--- Previous Query Result ---\n{observation['query_result']}")
    if observation.get('error_message'):
        parts.append(f"\n--- Error ---\n{observation['error_message']}")
    if observation.get('feedback'):
        parts.append(f"\n--- Feedback ---\n{observation['feedback']}")

    parts.append(f"\nAttempts remaining: {observation.get('attempts_remaining', 0)}")

    if history:
        parts.append("\n--- Previous Attempts ---")
        for h in history[-3:]:
            parts.append(h)

    parts.append("\nWrite your SQL query now:")
    return "\n".join(parts)


def get_sql_from_llm(client: OpenAI, observation: dict, step: int, history: List[str]) -> str:
    user_prompt = build_user_prompt(observation, step, history)
    try:
        completion = client.chat.completions.create(
            model=MODEL_NAME,
            messages=[
                {"role": "system", "content": SYSTEM_PROMPT},
                {"role": "user", "content": user_prompt},
            ],
            temperature=TEMPERATURE,
            max_tokens=MAX_TOKENS,
            stream=False,
        )
        raw = (completion.choices[0].message.content or "").strip()
        sql = raw
        if sql.startswith("```sql"):
            sql = sql[6:]
        if sql.startswith("```"):
            sql = sql[3:]
        if sql.endswith("```"):
            sql = sql[:-3]
        sql = sql.strip()
        return sql if sql else "SELECT 1"
    except Exception as exc:
        print(f"[DEBUG] LLM request failed: {exc}", flush=True)
        return "SELECT 1"


# =====================================================
# Main Inference Loop
# =====================================================

def run_task(client: OpenAI, env: SQLArenaEnvironment, task_config: dict) -> float:
    difficulty = task_config["difficulty"]
    task_id = task_config["task_id"]
    task_name = task_config["name"]
    max_steps = task_config["max_steps"]

    history: List[str] = []
    rewards: List[float] = []
    steps_taken = 0
    best_score = 0.0

    log_start(task=task_name, env=BENCHMARK, model=MODEL_NAME)

    try:
        result = env.reset(difficulty=difficulty, task_id=task_id)
        obs_dict = result.observation.model_dump()

        for step in range(1, max_steps + 1):
            if result.done:
                break

            sql_query = get_sql_from_llm(client, obs_dict, step, history)

            action = SQLArenaAction(sql_query=sql_query)
            result = env.step(action)

            obs_dict = result.observation.model_dump()
            reward = result.reward
            done = result.done
            error = obs_dict.get("error_message")

            rewards.append(reward)
            steps_taken = step
            best_score = max(best_score, result.info.get("score", 0.0))

            log_step(step=step, action=sql_query, reward=reward, done=done, error=error)

            history.append(
                f"Step {step}: {sql_query[:80]}... -> reward={reward:.2f}"
            )

            if done:
                break

        final_score = min(max(best_score, 0.0), 1.0)
        # Clamp to strictly between 0 and 1
        if final_score <= 0.0:
            final_score = 0.01
        if final_score >= 1.0:
            final_score = 0.99
        success = final_score >= 0.5

    except Exception as e:
        print(f"[DEBUG] Task {task_name} error: {e}", flush=True)
        final_score = 0.0
        success = False

    finally:
        log_end(success=success, steps=steps_taken, score=final_score, rewards=rewards)

    return final_score


def main() -> None:
    client = OpenAI(base_url=API_BASE_URL, api_key=API_KEY)
    env = SQLArenaEnvironment()

    all_scores = []

    for task_config in TASKS:
        print(f"\n{'='*60}", flush=True)
        print(f"Running task: {task_config['name']} ({task_config['difficulty']})", flush=True)
        print(f"{'='*60}", flush=True)

        score = run_task(client, env, task_config)
        all_scores.append(score)
        print(f"\nTask {task_config['name']} final score: {score:.2f}\n", flush=True)

    avg_score = sum(all_scores) / len(all_scores) if all_scores else 0.0
    print(f"\n{'='*60}", flush=True)
    print("SUMMARY", flush=True)
    print(f"{'='*60}", flush=True)
    for tc, sc in zip(TASKS, all_scores):
        print(f"  {tc['name']:20s}: {sc:.2f}", flush=True)
    print(f"  {'Average':20s}: {avg_score:.2f}", flush=True)

    env.close()


if __name__ == "__main__":
    main()