comments to english

Author: binary-husky

ajet/context_tracker/base_tracker.py

@@ -1,5 +1,4 @@
-from typing import List, Tuple, Union
-from typing import List, Union, Tuple, Dict, Optional
+from typing import Any, Dict, List, Optional, Tuple, Union
 from ajet.schema.task import WorkflowTask
 
 from ajet.schema.extended_msg import (
@@ -141,7 +140,7 @@ def __init__(self, config, tokenizer, workflow_task: WorkflowTask, **kwargs):
         self.already_mad_flag: bool = False
         self.round_cnt = 0
         self.generation_prompt_token = None
-        self.log_metrics: Optional[Dict[str, Union[float, List[float]]]] = None  # Initialize workflow_metadata to store tool statistics
+        self.log_metrics: Optional[Dict[str, Union[float, List[float], Dict[str, Any]]]] = None  # Initialize workflow_metadata to store tool statistics
 
         assert (
             self.config.ajet.data.max_prompt_length

ajet/task_reader/__init__.py

@@ -62,7 +62,7 @@ def __init__(self, reader_type, reader_config):
         elif task_reader_type == "random_dummy":
             self.task_reader = RandomDummyTaskReader(reader_config)
         elif task_reader_type == "deep_finance":
-            # deep_finance 专用: 数据从 JSON 文件加载并组装 init_messages，工具调用走 env_service
+            # deep_finance: load message from JSON file and assemble init_messages, tool calls go through env_service
             from tutorial.example_deep_finance.deep_finance_reader import DeepFinanceReader
             self.task_reader = DeepFinanceReader(reader_config)
         else:

ajet/task_rollout/resource_keeper.py

@@ -98,18 +98,18 @@ def _initialize_environment_and_messages(self) -> List[dict]:
                     self.env.release_instance(self.workflow_task.episode_uuid)
                 raise e
         elif reader_type == "deep_finance":
-            # deep_finance: 调用 create_instance 注册实例，但使用 reader 组装的 init_messages
+            # deep_finance: call create_instance to register instance, but use init_messages assembled by the reader
             if self.env is None:
                 raise ValueError("Environment client is None but deep_finance type is specified")
             try:
-                # 必须调用 create_instance，让服务端创建实例，后续 step() 才能工作
+                # call create_instance, let the server create an instance, so that subsequent step() can work
                 self.env.create_instance(
                     env_type=self.env_type,
                     task_id=self.task_id,
                     instance_id=self.workflow_task.episode_uuid,
                     params=self.env_params,
                 )
-                # 不使用返回的 state，直接用 reader 组装的 init_messages
+                # Do not use the returned state, directly use the init_messages assembled by the reader
                 task = self.workflow_task.task
                 if task.init_messages:
                     init_messages = task.init_messages

ajet/task_runner/general_runner.py

@@ -54,12 +54,10 @@ def execute(self, workflow_task: WorkflowTask) -> BaseContextTracker:
             )
         else:
             raw_reward, is_success = self.get_judge().compute_reward(workflow_task, workflow_output)
-            # Sync reward_stats from metadata to log_metrics after judge computation
 
-            if "reward_stats" in workflow_output.metadata:
+        if "reward_stats" in workflow_output.metadata:
+            workflow_output.log_metrics["reward_stats"] = workflow_output.metadata["reward_stats"]
 
-                workflow_output.log_metrics["reward_stats"] = workflow_output.metadata["reward_stats"]
-                
         workflow_task.gym_env = None  # clear gym env client reference to avoid serialization issue
 
         assert not isinstance(