bug fix

Author: patricebechard

browsergym/experiments/src/browsergym/experiments/benchmark/configs.py

@@ -244,7 +244,7 @@
         ),
         task_metadata=task_metadata("workarena"),
     ),
-    "workarena_dynamic_guidance": lambda n_repeats=1: Benchmark(
+    "workarena_dynamic_guidance": lambda n_repeats=3: Benchmark(
         name="workarena_dynamic_guidance",
         high_level_action_set_args=DEFAULT_HIGHLEVEL_ACTION_SET_ARGS["workarena++"],
         is_multi_tab=True,
@@ -256,6 +256,7 @@
             meta_seed=42,  # meta seed for evaluation curriculum
             max_steps=50,
             curriculum_type="agent",
+            seeds_l1=n_repeats,
         ),
         task_metadata=task_metadata("workarena"),
     ),

browsergym/experiments/src/browsergym/experiments/benchmark/utils.py

@@ -12,7 +12,7 @@
 
 
 def make_env_args_list_from_workarena_curriculum(
-    level: Literal["l1", "l2", "l3"],
+    level: Literal["l1", "l2", "l3", "dg"],
     task_category_filter: str,
     meta_seed: int,
     max_steps: int,
@@ -22,7 +22,7 @@ def make_env_args_list_from_workarena_curriculum(
     """
     Returns a WorkArena predefined task curriculum (e.g., task and seed combination).
     """
-    assert level in ("l1", "l2", "l3")
+    assert level in ("l1", "l2", "l3", "dg")
     assert curriculum_type in ("human", "agent")
 
     env_args_list = []