fix eval api_key leak in metadata; fix llm config in run infer (#2998)

evaluation/EDA/run_infer.py

@@ -62,7 +62,7 @@ def process_instance(
     reset_logger: bool = True,
 ):
     # Create the agent
-    agent = Agent.get_cls(metadata.agent_class)(llm=LLM(llm_config=metadata.llm_config))
+    agent = Agent.get_cls(metadata.agent_class)(llm=LLM(config=metadata.llm_config))
     # Setup the logger properly, so you can run multi-processing to parallelize the evaluation
     eval_output_dir = metadata.eval_output_dir
     if reset_logger:

evaluation/agent_bench/run_infer.py

@@ -37,7 +37,7 @@ def process_instance(
     reset_logger: bool = True,
 ):
     # Create the agent
-    agent = Agent.get_cls(metadata.agent_class)(llm=LLM(llm_config=metadata.llm_config))
+    agent = Agent.get_cls(metadata.agent_class)(llm=LLM(config=metadata.llm_config))
 
     inst_id = instance.instance_id
     question = instance.description

evaluation/biocoder/run_infer.py

@@ -87,7 +87,7 @@ def process_instance(
     reset_logger: bool = True,
 ):
     # Create the agent
-    agent = Agent.get_cls(metadata.agent_class)(llm=LLM(llm_config=metadata.llm_config))
+    agent = Agent.get_cls(metadata.agent_class)(llm=LLM(config=metadata.llm_config))
     instance = BiocoderData(**instance)
     print(instance)
     workspace_dir_name = (

evaluation/bird/run_infer.py

@@ -126,7 +126,7 @@ def process_instance(
     reset_logger: bool = True,
 ):
     # Create the agent
-    agent = Agent.get_cls(metadata.agent_class)(llm=LLM(llm_config=metadata.llm_config))
+    agent = Agent.get_cls(metadata.agent_class)(llm=LLM(config=metadata.llm_config))
     workspace_mount_path = os.path.join(
         config.workspace_mount_path, 'bird_eval_workspace'
     )

evaluation/browsing_delegation/run_infer.py

@@ -31,7 +31,7 @@ def process_instance(
     reset_logger: bool = True,
 ):
     # Create the agent
-    agent = Agent.get_cls(metadata.agent_class)(llm=LLM(llm_config=metadata.llm_config))
+    agent = Agent.get_cls(metadata.agent_class)(llm=LLM(config=metadata.llm_config))
     env_id = instance.instance_id
     # Setup the logger properly, so you can run multi-processing to parallelize the evaluation
     if reset_logger:

evaluation/gaia/run_infer.py

@@ -48,7 +48,7 @@ def process_instance(
     reset_logger: bool = True,
 ):
     # Create the agent
-    agent = Agent.get_cls(metadata.agent_class)(llm=LLM(llm_config=metadata.llm_config))
+    agent = Agent.get_cls(metadata.agent_class)(llm=LLM(config=metadata.llm_config))
     # create process-specific workspace dir
     # we will create a workspace directory for EACH process
     # so that different agent don't interfere with each other.

evaluation/gpqa/run_infer.py

@@ -120,7 +120,7 @@ def process_instance(
     reset_logger: bool = True,
 ):
     # Create the agent
-    agent = Agent.get_cls(metadata.agent_class)(llm=LLM(llm_config=metadata.llm_config))
+    agent = Agent.get_cls(metadata.agent_class)(llm=LLM(config=metadata.llm_config))
     old_workspace_mount_path = config.workspace_mount_path
     old_workspace_base = config.workspace_base
     try:

evaluation/humanevalfix/run_infer.py

@@ -108,7 +108,7 @@ def process_instance(
     reset_logger: bool = True,
 ):
     # Create the agent
-    agent = Agent.get_cls(metadata.agent_class)(llm=LLM(llm_config=metadata.llm_config))
+    agent = Agent.get_cls(metadata.agent_class)(llm=LLM(config=metadata.llm_config))
     old_workspace_mount_path = config.workspace_mount_path
     old_workspace_base = config.workspace_base
 

evaluation/logic_reasoning/run_infer.py

@@ -103,7 +103,7 @@ def process_instance(
     reset_logger: bool = True,
 ):
     # Create the agent
-    agent = Agent.get_cls(metadata.agent_class)(llm=LLM(llm_config=metadata.llm_config))
+    agent = Agent.get_cls(metadata.agent_class)(llm=LLM(config=metadata.llm_config))
     old_workspace_mount_path = config.workspace_mount_path
     old_workspace_base = config.workspace_base
 

evaluation/miniwob/run_infer.py

@@ -41,7 +41,7 @@ def process_instance(
     reset_logger: bool = True,
 ):
     # Create the agent
-    agent = Agent.get_cls(metadata.agent_class)(llm=LLM(llm_config=metadata.llm_config))
+    agent = Agent.get_cls(metadata.agent_class)(llm=LLM(config=metadata.llm_config))
     env_id = instance.id
     # Setup the logger properly, so you can run multi-processing to parallelize the evaluation
     if reset_logger:

evaluation/ml_bench/run_infer.py

@@ -67,7 +67,7 @@ ID2CONDA = {
 
 
 def process_instance(instance: Any, metadata: EvalMetadata, reset_logger: bool = True):
-    agent = Agent.get_cls(metadata.agent_class)(llm=LLM(llm_config=metadata.llm_config))
+    agent = Agent.get_cls(metadata.agent_class)(llm=LLM(config=metadata.llm_config))
     old_workspace_mount_path = config.workspace_mount_path
     old_workspace_base = config.workspace_base
     try:

evaluation/swe_bench/run_infer.py

@@ -172,7 +172,7 @@ def process_instance(
     reset_logger: bool = True,
 ):
     # Create the agent
-    agent = Agent.get_cls(metadata.agent_class)(llm=LLM(llm_config=metadata.llm_config))
+    agent = Agent.get_cls(metadata.agent_class)(llm=LLM(config=metadata.llm_config))
 
     workspace_mount_path = os.path.join(config.workspace_mount_path, '_eval_workspace')
     # create process-specific workspace dir

evaluation/toolqa/run_infer.py

@@ -35,7 +35,7 @@ AGENT_CLS_TO_INST_SUFFIX = {
 
 
 def process_instance(instance: Any, metadata: EvalMetadata, reset_logger: bool = True):
-    agent = Agent.get_cls(metadata.agent_class)(llm=LLM(llm_config=metadata.llm_config))
+    agent = Agent.get_cls(metadata.agent_class)(llm=LLM(config=metadata.llm_config))
     # create process-specific workspace dir
     # we will create a workspace directory for EACH process
     # so that different agent don't interfere with each other.

evaluation/utils/shared.py

@@ -29,6 +29,14 @@ class EvalMetadata(BaseModel):
     data_split: str | None = None
     details: dict[str, Any] | None = None
 
+    def model_dump_json(self, *args, **kwargs):
+        dumped = super().model_dump_json(*args, **kwargs)
+        dumped_dict = json.loads(dumped)
+        logger.debug(f'Dumped metadata: {dumped_dict}')
+        # avoid leaking sensitive information
+        dumped_dict['llm_config'] = self.llm_config.to_safe_dict()
+        return json.dumps(dumped_dict)
+
 
 def codeact_user_response(
     state: State,

evaluation/webarena/run_infer.py

@@ -42,7 +42,7 @@ def process_instance(
     reset_logger: bool = True,
 ):
     # Create the agent
-    agent = Agent.get_cls(metadata.agent_class)(llm=LLM(llm_config=metadata.llm_config))
+    agent = Agent.get_cls(metadata.agent_class)(llm=LLM(config=metadata.llm_config))
     env_id = instance.id
     # Setup the logger properly, so you can run multi-processing to parallelize the evaluation
     if reset_logger: