Final OpenEnv submission fix

evaluation/grader.py

@@ -1,24 +1,12 @@
-def grade_easy(episode_rewards):
-    # maximize engagement
-    return min(1.0, sum(episode_rewards) / max(1, len(episode_rewards)))
+def grade_easy(rewards):
+    return min(1.0, sum(rewards) / max(1, len(rewards)))
 
 
-def grade_medium(episode_rewards):
-    # engagement + diversity (simulated)
-    score = sum(episode_rewards) / max(1, len(episode_rewards))
-    diversity_bonus = 0.1
-    return min(1.0, score + diversity_bonus)
+def grade_medium(rewards):
+    score = sum(rewards) / max(1, len(rewards))
+    return min(1.0, score + 0.1)
 
 
-def grade_hard(episode_rewards):
-    # engagement + alignment + penalties
-    score = sum(episode_rewards) / max(1, len(episode_rewards))
-    penalty = 0.2
-    return max(0.0, min(1.0, score - penalty))
-
-
-TASK_GRADERS = {
-    "easy": grade_easy,
-    "medium": grade_medium,
-    "hard": grade_hard,
-}
+def grade_hard(rewards):
+    score = sum(rewards) / max(1, len(rewards))
+    return max(0.0, min(1.0, score - 0.2))

openenv.yaml

@@ -1,9 +1,14 @@
 name: value-aware-feed-ranking
 
 tasks:
-  - easy
-  - medium
-  - hard
+  - name: easy
+    grader: evaluation.grader:grade_easy
+
+  - name: medium
+    grader: evaluation.grader:grade_medium
+
+  - name: hard
+    grader: evaluation.grader:grade_hard
 
 observation_space:
   type: UserState