Updated grader values

inference.py

@@ -251,7 +251,7 @@ async def run_task(task_name: str) -> None:
                     break
 
             if rewards:
-                score   = min(max(sum(rewards) / len(rewards), 0.0), 1.0)
+                score   = min(max(sum(rewards) / len(rewards), 0.01), 0.99)
             success = score >= SUCCESS_SCORE_THRESHOLD
 
     except Exception as exc:

sql_env/env.py

@@ -88,12 +88,16 @@ class SQLCorrectionEnv:
 
         done = reward >= 0.95 or self._step_count >= self._task.max_steps
         self._done = done
+        
+        safe_reward = float(f"{reward:.4f}")
+        safe_reward = max(0.02, min(0.98, safe_reward))
+
 
         obs = self._make_observation()
 
         return StepResult(
             observation=obs,
-            reward=round(reward, 4),
+            reward=safe_reward,
             done=done,
             info={
                 "grader_reason": reward_model.reason,

sql_env/grader.py

@@ -101,7 +101,7 @@ def grade(action: SQLAction, task: SQLTask) -> SQLReward:
         )
 
     # ── Level 0: No recognizable SQL ─────────────────────────────────────────
-    return SQLReward(value=_clamp(0.01), reason="Response is not valid SQL.")
+    return SQLReward(value=_clamp(0.02), reason="Response is not valid SQL.")
 
 
 def generate_feedback(action: SQLAction, task: SQLTask, reward: SQLReward) -> str: