Deep Finance Update with New Judge

.gitignore

@@ -158,6 +158,9 @@ tutorial/example_deep_finance/yaml/*
 tutorial/example_deep_finance/config/*
 tutorial/example_deep_finance/scripts/*
 flash_attn-2.8.*.whl
+tutorial/example_deep_finance/prepare_data/*
+tutorial/example_deep_finance/judge/analytical_sufficiency/*
+
 .dockerignore
 benchmark_datasets
 modelscope_cache

ajet/context_tracker/multiagent_tracking.py

@@ -82,6 +82,18 @@ def extract_text_content_from_content_dict(self, msg):
         #        },
         #    ],
         # }
+        # or tool_result format?? not observed yet:
+        # msg = {
+        #    "role": "tool",
+        #    "content": [
+        #        {
+        #           "type": "tool_result",
+        #           "id": "call_xxx",
+        #           "output": "tool output content",
+        #           "name": "tool_name"
+        #        },
+        #    ],
+        # }
 
 
         str_content = ""

ajet/task_runner/general_runner.py

@@ -9,6 +9,7 @@
 from ajet.schema.trajectory import Reward
 from ajet.task_runner.base_runner import BaseAgentRunner
 from ajet.utils.dynamic_import import dynamic_import
+from ajet.utils.metric_helper.reward_metric_helper import populate_reward_metadata_from_stats
 
 
 class GeneralRunner(BaseAgentRunner):
@@ -73,6 +74,10 @@ def execute(self, workflow_task: WorkflowTask) -> BaseContextTracker:
             madness=0,
             description="",
         )
+
+        # Populate reward metadata with deep_finance reward stats if available
+        if "reward_stats" in workflow_output.metadata:
+            populate_reward_metadata_from_stats(reward, workflow_output.metadata["reward_stats"])
         context_tracker.process_reward(reward)
         # generate token before merging
         context_tracker.group_merge()

ajet/utils/metric_helper/reward_metric_helper.py

@@ -11,9 +11,12 @@
 - judge_time/                 Judge time consumption statistics
 """
 
-from typing import List, Dict, Any
+from typing import List, Dict, Any, TYPE_CHECKING
 import numpy as np
 
+if TYPE_CHECKING:
+    from ajet.schema.trajectory import Reward
+
 
 def extract_reward_stats_from_trajectories(trajectories: List[Any]) -> List[Dict[str, Any]]:
     """
@@ -72,22 +75,15 @@ def compute_reward_metrics(reward_stats_list: List[Dict[str, Any]], prefix: str
     metrics[f"{prefix}rewards/penalty_count"] = len(non_zero_penalties)
     metrics[f"{prefix}rewards/penalty_rate"] = len(non_zero_penalties) / n * 100 if n > 0 else 0.0
 
-    # ========== Detect OpenJudge Usage ==========
+    # ========== OpenJudge Metrics (PresentationQualityGrader, GroundingGrader) ==========
     openjudge_enabled_count = sum(1 for rs in reward_stats_list if rs.get('openjudge_enabled', False))
 
     if openjudge_enabled_count > 0:
-        # ========== OpenJudge Metrics ==========
-
-        # Dynamically extract OpenJudge grader fields
-        # Currently supported graders: report_resolution, trajectory_faithfulness,
-        # rubrics_performance, trajectory_comprehensive, information_gain, action_loop
+        # OpenJudge graders: presentation_quality, grounding
         openjudge_graders = [
-            "report_resolution",
-            "trajectory_faithfulness",
-            "rubrics_performance",
-            "trajectory_comprehensive",
-            "information_gain",
-            "action_loop",
+            "presentation_quality",
+            "grounding",
+            "planning"
         ]
 
         for grader_name in openjudge_graders:
@@ -151,3 +147,18 @@ def compute_reward_metrics_from_trajectories(trajectories: List[Any], prefix: st
     reward_stats_list = extract_reward_stats_from_trajectories(trajectories)
     return compute_reward_metrics(reward_stats_list, prefix=prefix)
 
+
+def populate_reward_metadata_from_stats(reward: "Reward", reward_stats: Dict[str, Any]) -> None:
+    """
+    Populate Reward.metadata with all reward statistics.
+
+    Args:
+        reward: The Reward object to populate
+        reward_stats: The reward_stats dictionary from judge
+    """
+    if not reward_stats:
+        return
+
+    # Directly copy all reward_stats into metadata
+    reward.metadata.update(reward_stats)
+

tutorial/example_deep_finance/__init__.py

@@ -0,0 +1 @@
+# tutorial/example_deep_finance package

tutorial/example_deep_finance/deep_finance.sh

@@ -1,29 +1,29 @@
 #!/bin/bash
-set -e
+set -e  
 #===============================================================================
 # 1. 配置区域 - 用户只需修改这里
 #===============================================================================
-SUFFIX="deep_finance"     # 实验后缀，影响所有日志和实验名称
-PREFIX="open"                        # 实验前缀，影响日志和实验所在文件夹
+SUFFIX="newjudge"     # 实验后缀，影响所有日志和实验名称
+PREFIX="ajet_newjudge"                        # 实验前缀，影响日志和实验所在文件夹
 
 # OpenJudge 模型配置
 OPENJUDGE_LLM='qwen-flash'        # OpenJudge 评分模型
 RM_LLM='qwen-max'                 # RM Gallery 评分模型
 JUDGE_CONCURRENCY=10
 
 # 奖励权重配置
-RM_WEIGHT=0.4
-CITATION_AUDIT_WEIGHT=0.2
-REPORT_RESOLUTION_WEIGHT=0.2
-TRAJECTORY_FAITHFULNESS_WEIGHT=0.2
+RM_WEIGHT=0.5
+PRESENTATION_QUALITY_WEIGHT=0.25
+GROUNDING_WEIGHT=0.25
 
 # 训练参数配置
 NUM_REPEAT=4        # group size，每个query rollout NUM_REPEAT次
 TRAIN_BATCH_SIZE=32  # 训练batchsize
 NUM_STEPS=6         # 每个样本step轮数
 DEEPFINANCE_TOOL_RESULT_MAX_CHARS=10000
 
-# 主目录
+# 主目录（需要更改）
+export AJET_ROOT="/mnt/data_cpfs/taoshuchang.tsc/deepresearch/AgentJet_new"
 
 NNODES=${WORLD_SIZE}
 
@@ -46,7 +46,7 @@ fi
 # 2. 动态生成配置文件 (从yaml template生成yaml)
 #===============================================================================
 # 修改：配置文件生成路径，现在动态生成到 yaml 目录下
-CONFIG_TEMPLATE="tutorial/example_deep_finance/yaml_template/deep_finance_template.yaml"
+CONFIG_TEMPLATE="tutorial/example_deep_finance/deep_finance.yaml"
 CONFIG_FILE="${AJET_ROOT}/tutorial/example_deep_finance/yaml/${SUFFIX}.yaml"
 mkdir -p $(dirname ${CONFIG_FILE})
 
@@ -55,12 +55,11 @@ sed -e "s|{{SUFFIX}}|${SUFFIX}|g" \
     -e "s|{{MODEL_PATH}}|${MODEL_PATH}|g" \
     -e "s|{{NNODES}}|${NNODES}|g" \
     -e "s|{{RM_WEIGHT}}|${RM_WEIGHT}|g" \
-    -e "s|{{CITATION_AUDIT_WEIGHT}}|${CITATION_AUDIT_WEIGHT}|g" \
+    -e "s|{{PRESENTATION_QUALITY_WEIGHT}}|${PRESENTATION_QUALITY_WEIGHT}|g" \
+    -e "s|{{GROUNDING_WEIGHT}}|${GROUNDING_WEIGHT}|g" \
     -e "s|{{OPENJUDGE_LLM}}|${OPENJUDGE_LLM}|g" \
     -e "s|{{RM_LLM}}|${RM_LLM}|g" \
     -e "s|{{JUDGE_CONCURRENCY}}|${JUDGE_CONCURRENCY}|g" \
-    -e "s|{{REPORT_RESOLUTION_WEIGHT}}|${REPORT_RESOLUTION_WEIGHT}|g" \
-    -e "s|{{TRAJECTORY_FAITHFULNESS_WEIGHT}}|${TRAJECTORY_FAITHFULNESS_WEIGHT}|g" \
     -e "s|{{NUM_REPEAT}}|${NUM_REPEAT}|g" \
     -e "s|{{NUM_STEPS}}|${NUM_STEPS}|g" \
     -e "s|{{TRAIN_BATCH_SIZE}}|${TRAIN_BATCH_SIZE}|g" \
@@ -72,7 +71,7 @@ sed -e "s|{{SUFFIX}}|${SUFFIX}|g" \
     ${AJET_ROOT}/${CONFIG_TEMPLATE} > ${CONFIG_FILE}
 
 echo "配置文件已生成: ${CONFIG_FILE}"
-echo "参数确认: RM=${RM_WEIGHT}, Citation=${CITATION_AUDIT_WEIGHT}, OpenJudge=${OPENJUDGE_LLM}, RM_LLM=${RM_LLM}"
+echo "参数确认: RM=${RM_WEIGHT}, PresentationQuality=${PRESENTATION_QUALITY_WEIGHT}, Grounding=${GROUNDING_WEIGHT}, OpenJudge=${OPENJUDGE_LLM}, RM_LLM=${RM_LLM}"
 
 #===============================================================================
 # 3. 环境配置
@@ -106,15 +105,15 @@ export DEEPFINANCE_MCP_CONFIG  DEEPFINANCE_TOOL_RESULT_MAX_CHARS
 # 其他服务配置
 HF_ENDPOINT="https://hf-mirror.com"
 ES_HOSTS="http://11.160.132.46:8200"
-export HF_ENDPOINT ES_HOSTS
+export HF_ENDPOINT ES_HOSTS 
 
 # log 文件位置
 CURRENT_TIME=$(date "+%Y%m%d_%H%M%S")
 LOG_DIR="${AJET_ROOT}/logs/${PREFIX}"
 MASTER_IP_FILE="${LOG_DIR}/master-ip_${SUFFIX}.log"
 ENV_SERVICE_LOG="${LOG_DIR}/env_service_${SUFFIX}_${CURRENT_TIME}.log"
 TRAIN_LOG="${LOG_DIR}/train_${SUFFIX}_${CURRENT_TIME}.log"
-
+env_log_prefix="${SUFFIX}__${CURRENT_TIME}"
 # 多机训练参数配置
 GPUS_PER_NODE=8
 EXPECTED_WORKERS=$WORLD_SIZE
@@ -156,6 +155,8 @@ export NCCL_ASYNC_ERROR_HANDLING=1
 
 export PYTHONPATH="${AJET_ROOT}:${PYTHONPATH}"
 export RAY_CLUSTER_MODE="multi_node"
+export DEEPFINANCE_PATH="${ENV_SERVICE_ROOT}" # AgentJet 内部可能使用此路径
+export DEEPFINANCE_SCRIPT="source /mnt/data/taoshuchang.tsc/anaconda3/etc/profile.d/conda.sh && conda activate finworld_1209  && cd ${ENV_SERVICE_ROOT} && DEEPFINANCE_TOOL_RESULT_MAX_CHARS=${DEEPFINANCE_TOOL_RESULT_MAX_CHARS} DEEPFINANCE_MCP_CONFIG=${DEEPFINANCE_MCP_CONFIG} CACHE_TYPE=${CACHE_TYPE} MONGO_URI=${MONGO_URI} MONGO_DB_NAME=${MONGO_DB_NAME} MONGO_COLLECTION_NAME=${MONGO_COLLECTION_NAME} python -m env_service.env_service --env finworld --portal 0.0.0.0 --port 8080"
 
 
 #===============================================================================
@@ -202,11 +203,12 @@ if [[ $HOSTNAME == *"-master-"* ]]; then
 
     # 启动训练任务（最核心）
     python ajet/launcher.py \
+        --with-deepfinance \
         --conf ${CONFIG_FILE} \
         --backbone="verl" \
-        --prefix=${SUFFIX} \
+        --prefix=${env_log_prefix} \
         2>&1 | tee ${TRAIN_LOG}
-
+    
 
 #===============================================================================
 # 6.2 Worker 节点启动流程
@@ -218,4 +220,4 @@ else
     ray stop || true
     ray start --address $MASTER_ADDR:6379 --num-gpus 8
     while true; do sleep 60; done
-fi
+fi

tutorial/example_deep_finance/deep_finance.yaml

@@ -1,27 +1,26 @@
 # ------------------ 主要配置 ------------------
 ajet:
-  project_name: ajet_deep_finance
-  experiment_name: "ajet_deep_finance"
+  project_name: "{{PREFIX}}"
+  experiment_name: "{{SUFFIX}}"
   # Judge 配置（嵌套结构，对应 self.config.ajet.judge.*）
   judge:
-    openjudge_llm: qwen-flash     # OpenJudge 模型
-    rm_llm: qwen-max                   # RM Gallery 模型
-    concurrency: 10   # Judge 并发数
+    openjudge_llm: {{OPENJUDGE_LLM}}     # OpenJudge 模型
+    rm_llm: {{RM_LLM}}                   # RM Gallery 模型
+    concurrency: {{JUDGE_CONCURRENCY}}   # Judge 并发数
     train_ref_ans_path: {{TRAIN_REF_ANS_PATH}}   # 训练集 Reference Answer 路径
     val_ref_ans_path: {{VAL_REF_ANS_PATH}}       # 验证集 Reference Answer 路径
   # OpenJudge 权重配置
-  report_resolution_weight: 0.2         # 报告质量评估
-  trajectory_faithfulness_weight: 0.2   # 事实准确性评估
-  citation_audit_weight: 0.2               # 引用审计评估 (覆盖率 + 真实性)
-  rm_weight: 0.4                                       # RM Gallery 权重
+  presentation_quality_weight: {{PRESENTATION_QUALITY_WEIGHT}}   # 报告呈现质量评估
+  grounding_weight: {{GROUNDING_WEIGHT}}                         # 引用规范性评估
+  rm_weight: {{RM_WEIGHT}}                                       # RM Gallery 权重
   task_judge:
     # 使用本地 DeepFinanceJudge 进行评估（解耦远程 env_service）
     judge_protocol: tutorial.example_deep_finance.deep_finance_judge->DeepFinanceJudgeByOpenJudge
   model:
     # ✨✨✨✨ 设置待训练的模型
     path: {{MODEL_PATH}}
   trainer_common:
-    nnodes: 8
+    nnodes: {{NNODES}}
     n_gpus_per_node: 8
     val_before_train: True
     val_pass_n: 8
@@ -32,44 +31,42 @@ ajet:
   rollout:
     # ✨✨✨✨ 编写并选择Agent
     user_workflow: tutorial.example_deep_finance.deep_finance->ExampleDeepResearchProtocol
-    force_disable_toolcalls: True
+    force_disable_toolcalls: False
     enable_oversample: False
     tensor_model_parallel_size: 8
-    num_repeat: 4
+    num_repeat: {{NUM_REPEAT}}
     max_env_worker: 64  # 增加环境并行数
     max_num_seqs: 64    # 增加VLLM并发序列数
     max_response_length_in_one_turn: 8000
     max_model_len: 50000
     agent_madness_reward: 0.0
     compute_madness_checklist: None
     multi_turn:
-      max_steps: 6
+      max_steps: {{NUM_STEPS}}
   interchange_server:
     interchange_method: 'tcp' # options: 'tcp' (multi-nodes) or  'ipc' (1 node)
   debug:
     debug_max_parallel: 1  # 增加并行任务数，充分利用GPU
     debug_first_n_tasks: 100  # 增加处理的任务数
   data:
-    train_batch_size: 32
+    train_batch_size: {{TRAIN_BATCH_SIZE}}
     max_prompt_length: 8000
     max_response_length: 41000
 
   task_reader:
     type: deep_finance  # 数据从 JSON 加载并组装 init_messages，工具调用走 env_service
     deep_finance:
       training:
-        file_path: {{TRAIN_PATH}}
+        file_path: {{TRAIN_DATA_PATH}}
       validation:
-        file_path: {{VAL_PATH}}
+        file_path: {{VAL_DATA_PATH}}
     # env_service 仍需配置（用于工具调用）
     env_service:
       env_type: "finworld"
       env_url: {{ENV_SERVICE_URL}}
       env_action_preference: code
-
-
 trainer:
-  default_local_dir: {{CKPT_SAVE_PATH}}
+  default_local_dir: "{{CKPT_SAVE_PATH}}/{{PREFIX}}/{{SUFFIX}}"
   # resume_mode: disable  # 禁用自动恢复，从头开始训练
 actor_rollout_ref:
   rollout: