Upload 16 files

inference.py

@@ -285,6 +285,10 @@ Generate your next action (SQL or action name):"""
                     log_step(step=step, action=current_action, reward=0.0, done=False, error=error_msg)
                     break
             
+            # Fix: Ensure score is not 0.0 when loop ends without done=True
+            if not success and steps_taken == MAX_STEPS:
+                score = max(0.001, score)
+            
         except Exception as e:
             print(f"[DEBUG] Task execution failed: {e}", flush=True, file=sys.stderr)
             
@@ -323,7 +327,8 @@ def main():
     success, steps, score, rewards = agent.run_task(TASK_NAME, seed=42)
     
     # Log end
-    log_end(success=success, steps=steps, score=score, rewards=rewards)
+    safe_score = round(max(0.001, min(0.999, score)), 4)
+    log_end(success=success, steps=steps, score=safe_score, rewards=rewards)
 
 
 if __name__ == "__main__":

server/graders.py

@@ -204,7 +204,7 @@ def _evaluate_fix_outcome(
         return False, 0.0, False
 
     if task_type == "type_repair":
-        target_col = ground_truth.get("column", ground_truth.get("target_column", "amount"))
+        target_col = ground_truth.get("column", ground_truth.get("target_column", ""))
         if uploaded_mode:
             try:
                 rows = test_conn.execute(
@@ -343,7 +343,7 @@ def _evaluate_fix_outcome(
         return issue_resolved, preservation, True
 
     if task_type == "null_audit":
-        target_col = ground_truth.get("column", ground_truth.get("null_column", "city"))
+        target_col = ground_truth.get("column", ground_truth.get("null_column", ""))
         if uploaded_mode:
             try:
                 rows = test_conn.execute(
@@ -432,7 +432,7 @@ def _evaluate_fix_outcome(
             return False, 0.0, True
 
     if task_type == "dedup_bug":
-        key_cols = ground_truth.get("key_columns") or [ground_truth.get("column", ground_truth.get("offending_key", "order_id"))]
+        key_cols = ground_truth.get("key_columns") or [ground_truth.get("column", ground_truth.get("offending_key", ""))]
         expected_clean = ground_truth.get("expected_clean_count", 0)
         if uploaded_mode:
             try:
@@ -488,7 +488,7 @@ def _evaluate_fix_outcome(
     if task_type == "fk_integrity":
         parent_table = ground_truth.get("parent_table", "customers")
         parent_key = ground_truth.get("parent_key", "id")
-        fk_col = ground_truth.get("column", "customer_id")
+        fk_col = ground_truth.get("column", "")
         if uploaded_mode:
             try:
                 orphan_count = test_conn.execute(
@@ -555,7 +555,7 @@ def _evaluate_fix_outcome(
     if task_type == "schema_drift":
         expected_columns = ground_truth.get(
             "expected_columns",
-            ["event_id", "user_id", "event_type", "amount", "event_time"],
+            [],
         )
         expected_rows = ground_truth.get("expected_row_count", 0)
         canonical_table = ground_truth.get("table", "canonical_events")
@@ -1027,7 +1027,7 @@ def compute_episode_score(
         d5 = 0.0
         if task_type == "dedup_bug":
             correct_table = ground_truth.get("table", "order_items").lower()
-            correct_key = ground_truth.get("column", "order_id").lower()
+            correct_key = ground_truth.get("column", "").lower()
             table_ok = (submit_action.root_cause_table or "").lower() == correct_table
             key_ok = (submit_action.offending_key or "").lower() == correct_key
             if table_ok and key_ok:
@@ -1082,8 +1082,8 @@ def compute_episode_score(
         dimensions.update(null_meta)
 
     return SQLReward(
-        score=final_score,
-        step_reward=final_score,
+        score=_clamp(final_score),
+        step_reward=_clamp(final_score),
         feedback=" | ".join(feedback_parts),
         dimensions=dimensions,
         partial_scores={
@@ -1193,7 +1193,7 @@ def grade_dedup_bug(
     """Single-step dedup_bug scorer for exploit checks and compatibility."""
     fix_sql = (action.fix_query or action.sql or "").strip()
     target_table = ground_truth.get("table", ground_truth.get("root_cause_table", "order_items"))
-    target_key = ground_truth.get("column", ground_truth.get("offending_key", "order_id"))
+    target_key = ground_truth.get("column", ground_truth.get("offending_key", ""))
     table_ok = (action.root_cause_table or "").lower() == target_table.lower()
     key_ok = (action.offending_key or "").lower() == target_key.lower()
     partial_scores = {