Spaces:
Sleeping
Sleeping
File size: 5,604 Bytes
1a1713a e965a47 1a1713a e965a47 f6bd747 e965a47 6c6f994 e965a47 f52aed3 e965a47 82a5b1b e965a47 2194233 e965a47 f52aed3 2194233 f52aed3 7a6f18c 82a5b1b 7a6f18c 2194233 fd851a9 2194233 fd851a9 2194233 2be7532 2194233 82a5b1b fd851a9 82a5b1b fd851a9 2194233 82a5b1b 2194233 fd851a9 82a5b1b 2194233 fd851a9 f52aed3 82a5b1b f52aed3 2194233 f52aed3 2194233 f52aed3 2194233 f52aed3 e965a47 f6bd747 e965a47 f6bd747 c4a66be 2194233 1a1713a 66cc86e | 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 | """
FastAPI server using openenv.core base classes — required for validator.
"""
import random
from openenv.core.env_server.http_server import create_app
from openenv.core.env_server.interfaces import Environment
try:
from sql_env.models import SQLAction, SQLObservation, SQLState
from sql_env.tasks import TASK_SETS
from sql_env.grader import grade, generate_feedback
except ImportError:
from models import SQLAction, SQLObservation, SQLState
from tasks import TASK_SETS
from grader import grade, generate_feedback
class SQLCorrectionEnvironment(Environment):
SUPPORTS_CONCURRENT_SESSIONS = True
def __init__(self):
self._difficulty = "easy"
self._current_task = None
self._step_count = 0
self._done = False
self._last_reward = 0.01
self._rewards_history = []
self._stagnation_count = 0
def reset(self, seed=None, episode_id=None, **kwargs) -> SQLObservation:
actual_difficulty = (
kwargs.get("task_id")
or kwargs.get("difficulty")
or "easy"
)
self._difficulty = actual_difficulty
tasks = TASK_SETS.get(actual_difficulty, TASK_SETS["easy"])
self._current_task = random.choice(tasks)
self._step_count = 0
self._done = False
self._last_reward = 0.01
self._rewards_history = []
self._stagnation_count = 0
return self._make_observation(previous_attempt=None, feedback=None)
def step(self, action: SQLAction) -> SQLObservation:
if self._current_task is None:
self.reset()
self._step_count += 1
reward_obj = grade(action, self._current_task)
reward = reward_obj.value
# Stagnation penalty: penalize repeating the same score
if abs(reward - self._last_reward) < 0.01 and self._step_count > 1:
self._stagnation_count += 1
if self._stagnation_count >= 2:
reward = max(0.01, reward - 0.1)
else:
self._stagnation_count = 0
# Final Clamp
reward = max(0.01, min(0.98, reward))
self._last_reward = reward
self._rewards_history.append(reward)
done = (reward >= 0.95) or (
self._step_count >= self._current_task.max_steps
)
self._done = done
feedback = generate_feedback(action, self._current_task, reward)
return self._make_observation(
previous_attempt=action.corrected_query,
feedback=feedback,
)
@property
def state(self) -> SQLState:
if self._current_task is None:
return SQLState(
task_id="none",
difficulty="none",
step_count=0,
max_steps=0,
done=False,
last_reward=0.01,
rewards_history=[],
)
return SQLState(
task_id=self._current_task.task_id,
difficulty=self._difficulty,
step_count=self._step_count,
max_steps=self._current_task.max_steps,
done=self._done,
last_reward=self._last_reward,
rewards_history=self._rewards_history,
)
# ── Internal helpers ──────────────────────────────────────────────────────
def _make_observation(
self,
previous_attempt: str | None,
feedback: str | None,
) -> SQLObservation:
assert self._current_task is not None
max_steps = self._current_task.max_steps
steps_remaining = max(0, max_steps - self._step_count)
return SQLObservation(
task_id=self._current_task.task_id,
broken_query=self._current_task.broken_query,
schema_context=self._current_task.schema_context,
# Only surface the hint on easy tasks
error_hint=(
self._current_task.error_hint
if self._difficulty == "easy"
else None
),
step_number=self._step_count,
steps_remaining=steps_remaining,
previous_attempt=previous_attempt,
feedback=feedback,
)
app = create_app(
SQLCorrectionEnvironment,
SQLAction,
SQLObservation,
env_name="sql-correction-env",
)
@app.get("/tasks")
async def list_tasks():
"""List available task difficulties with metadata."""
return {
"tasks": [
{
"id": "easy",
"difficulty": "easy",
"description": "Fix a single SQL keyword typo. Error hint provided.",
"max_steps": 5,
"count": 15,
},
{
"id": "medium",
"difficulty": "medium",
"description": (
"Fix multiple errors across keywords and clauses. No hint."
),
"max_steps": 5,
"count": 15,
},
{
"id": "hard",
"difficulty": "hard",
"description": (
"Fix complex multi-join queries including column name errors. "
"Schema provided, no hint."
),
"max_steps": 4,
"count": 10,
},
]
}
def main():
import uvicorn
uvicorn.run(app, host="0.0.0.0", port=7860)
if __name__ == "__main__":
main()
|