Merge pull request #10 from modelscope/dev/shuchang_newjudge

# Pull Request: Deep Finance Judge System Enhancement

Author: TaoShuchang

.gitattributes

@@ -0,0 +1,38 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bin.* filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zstandard filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text
+# Audio files - uncompressed
+*.pcm filter=lfs diff=lfs merge=lfs -text
+*.sam filter=lfs diff=lfs merge=lfs -text
+*.raw filter=lfs diff=lfs merge=lfs -text
+# Audio files - compressed
+*.aac filter=lfs diff=lfs merge=lfs -text
+*.flac filter=lfs diff=lfs merge=lfs -text
+*.mp3 filter=lfs diff=lfs merge=lfs -text
+*.ogg filter=lfs diff=lfs merge=lfs -text
+*.wav filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -160,6 +160,8 @@ tutorial/example_deep_finance/scripts/*
 flash_attn-2.8.*.whl
 tutorial/example_deep_finance/prepare_data/*
 tutorial/example_deep_finance/judge/analytical_sufficiency/*
+tutorial/example_deep_finance/output_report/*
+dataset_gsm8k/*
 
 .dockerignore
 benchmark_datasets

ajet/utils/metric_helper/reward_metric_helper.py

@@ -83,7 +83,10 @@ def compute_reward_metrics(reward_stats_list: List[Dict[str, Any]], prefix: str
         openjudge_graders = [
             "presentation_quality",
             "grounding",
-            "planning"
+            "planning",
+            "audit",
+            "traceability",
+            "cgcv"
         ]
 
         for grader_name in openjudge_graders:

tutorial/example_deep_finance/deep_finance.md

@@ -9,7 +9,7 @@
 
 
 # 创建信号量，允许同时12个线程运行
-sem = threading.Semaphore(30)
+sem = threading.Semaphore(60)
 
 class ExampleDeepResearchProtocol(Workflow):
 
@@ -125,9 +125,9 @@ async def execute(
             if info:
                 if 'tool_stats' in info:
                     latest_tool_stats = info['tool_stats']
-                    if latest_tool_stats.get('total_calls', 0) > 0:
-                        logger.info(f"步骤 {step + 1} 工具统计: 调用={latest_tool_stats.get('total_calls', 0)}, "
-                                    f"成功率={latest_tool_stats.get('success_rate', 0):.1f}%")
+                    # if latest_tool_stats.get('total_calls', 0) > 0:
+                    #     logger.info(f"步骤 {step + 1} 工具统计: 调用={latest_tool_stats.get('total_calls', 0)}, "
+                    #                 f"成功率={latest_tool_stats.get('success_rate', 0):.1f}%")
                 if 'reward_stats' in info:
                     latest_reward_stats = info['reward_stats']
                     # 累加工具调用时间

tutorial/example_deep_finance/deep_finance.py

@@ -15,13 +15,20 @@ JUDGE_CONCURRENCY=10
 RM_WEIGHT=0.5
 PRESENTATION_QUALITY_WEIGHT=0.25
 GROUNDING_WEIGHT=0.25
+CGCV_WEIGHT=0.0                   # 不使用 CGCV，设为 0
+AUDIT_WEIGHT=0.0                  # 不使用 Audit，设为 0
+TRACEABILITY_WEIGHT=0.0           # 不使用 Traceability，设为 0
+EBTU_WEIGHT=0.0                   # 不使用 EBTU，设为 0
 
 # 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
 
+# Env Service URL 配置
+ENV_SERVICE_URL="http://127.0.0.1:8080"  # 环境服务地址
+
 # 主目录（需要更改）
 export AJET_ROOT="/mnt/data_cpfs/taoshuchang.tsc/deepresearch/AgentJet_new"
 
@@ -57,6 +64,10 @@ sed -e "s|{{SUFFIX}}|${SUFFIX}|g" \
     -e "s|{{RM_WEIGHT}}|${RM_WEIGHT}|g" \
     -e "s|{{PRESENTATION_QUALITY_WEIGHT}}|${PRESENTATION_QUALITY_WEIGHT}|g" \
     -e "s|{{GROUNDING_WEIGHT}}|${GROUNDING_WEIGHT}|g" \
+    -e "s|{{CGCV_WEIGHT}}|${CGCV_WEIGHT}|g" \
+    -e "s|{{AUDIT_WEIGHT}}|${AUDIT_WEIGHT}|g" \
+    -e "s|{{TRACEABILITY_WEIGHT}}|${TRACEABILITY_WEIGHT}|g" \
+    -e "s|{{EBTU_WEIGHT}}|${EBTU_WEIGHT}|g" \
     -e "s|{{OPENJUDGE_LLM}}|${OPENJUDGE_LLM}|g" \
     -e "s|{{RM_LLM}}|${RM_LLM}|g" \
     -e "s|{{JUDGE_CONCURRENCY}}|${JUDGE_CONCURRENCY}|g" \
@@ -68,10 +79,11 @@ sed -e "s|{{SUFFIX}}|${SUFFIX}|g" \
     -e "s|{{TRAIN_REF_ANS_PATH}}|${TRAIN_REF_ANS_PATH}|g" \
     -e "s|{{VAL_REF_ANS_PATH}}|${VAL_REF_ANS_PATH}|g" \
     -e "s|{{CKPT_SAVE_PATH}}|${CKPT_SAVE_PATH}|g" \
+    -e "s|{{ENV_SERVICE_URL}}|${ENV_SERVICE_URL}|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}, PresentationQuality=${PRESENTATION_QUALITY_WEIGHT}, Grounding=${GROUNDING_WEIGHT}, OpenJudge=${OPENJUDGE_LLM}, RM_LLM=${RM_LLM}"
+echo "参数确认: RM=${RM_WEIGHT}, PresentationQuality=${PRESENTATION_QUALITY_WEIGHT}, Grounding=${GROUNDING_WEIGHT}, CGCV=${CGCV_WEIGHT}, Audit=${AUDIT_WEIGHT}, Traceability=${TRACEABILITY_WEIGHT}, EBTU=${EBTU_WEIGHT}, OpenJudge=${OPENJUDGE_LLM}, RM_LLM=${RM_LLM}"
 
 #===============================================================================
 # 3. 环境配置

tutorial/example_deep_finance/deep_finance.sh

@@ -37,7 +37,7 @@ ajet:
     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
+    max_model_len: 40960
     agent_madness_reward: 0.0
     compute_madness_checklist: None
     multi_turn:

tutorial/example_deep_finance/deep_finance.yaml

@@ -15,7 +15,7 @@
 
 from openjudge.models.openai_chat_model import OpenAIChatModel
 from openjudge.runner.grading_runner import GraderConfig, GradingRunner
-from tutorial.example_deep_finance.judge import PresentationQualityGrader, GroundingGrader
+from tutorial.example_deep_finance.judge import PresentationQualityGrader, GroundingGrader, CGCVGrader, AuditGrader, TraceabilityRewardGrader, EBTUTraceabilityGrader
 
 
 
@@ -103,7 +103,11 @@ def _setup_weights(self):
         self.w = {
             "rm": getattr(cfg, "rm_weight", 1.0) if cfg else 1.0,  # RM Gallery 权重
             "presentation_quality": getattr(cfg, "presentation_quality_weight", 0.25) if cfg else 0.25,
-            "grounding": getattr(cfg, "grounding_weight", 0.25) if cfg else 0.25,
+            "grounding": getattr(cfg, "grounding_weight", 0.0) if cfg else 0.0,  # 引用规范性评估
+            "cgcv": getattr(cfg, "cgcv_weight", 0.25) if cfg else 0.25,  # Citation-Grounded Claim Verification
+            "audit": getattr(cfg, "audit_weight", 0.0) if cfg else 0.0,  # Audit Grader: audit reward 引用逻辑审计
+            "traceability": getattr(cfg, "traceability_weight", 0.0) if cfg else 0.0,  # 可追溯性/可核验性审计 (TVR)
+            "ebtu": getattr(cfg, "ebtu_weight", 0.0) if cfg else 0.0,  # Audit Grader: audit reward EBTU证据优先可追溯性审计
         }
 
         # 归一化（注意：action_loop 是惩罚项，不参与归一化；rm 需要参与归一化）
@@ -256,6 +260,26 @@ def extract_report_content(data: Dict) -> str:
                 grader=GroundingGrader(model=model),
                 mapper=lambda data: {"traj": data},
             ),
+            # CGCV: Citation-Grounded Claim Verification - 引用锤定的断言验证
+            "cgcv": GraderConfig(
+                grader=CGCVGrader(model=model),
+                mapper=lambda data: {"traj": data},
+            ),
+            # Audit: 引用逻辑审计 - 验证引用是否严格符合逻辑蕴含原则
+            "audit": GraderConfig(
+                grader=AuditGrader(model=model),
+                mapper=lambda data: {"traj": data},
+            ),
+            # Traceability: 可追溯性/可核验性审计 - 验证报告断言是否有证据锚点支撑
+            "traceability": GraderConfig(
+                grader=TraceabilityRewardGrader(model=model),
+                mapper=lambda data: {"traj": data},
+            ),
+            # Audit Grader: audit reward EBTU证据优先可追溯性审计 - Evidence-Backed Trace Units
+            "ebtu": GraderConfig(
+                grader=EBTUTraceabilityGrader(model=model),
+                mapper=lambda data: {"traj": data},
+            ),
         }
 
     def compute_reward(self, workflow_task: WorkflowTask, workflow_output: WorkflowOutput) -> Tuple[float, bool]:

tutorial/example_deep_finance/deep_finance_judge.py

@@ -15,17 +15,30 @@ JUDGE_CONCURRENCY=10
 RM_WEIGHT=0.5
 PRESENTATION_QUALITY_WEIGHT=0.25
 GROUNDING_WEIGHT=0.25
+CGCV_WEIGHT=0.0                   # 不使用 CGCV，设为 0
+AUDIT_WEIGHT=0.0                  # 不使用 Audit，设为 0
+TRACEABILITY_WEIGHT=0.0           # 不使用 Traceability，设为 0
+EBTU_WEIGHT=0.0                   # 不使用 EBTU，设为 0
 
 # 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
 
+# Env Service URL 配置
+ENV_SERVICE_URL="http://127.0.0.1:8080"  # 环境服务地址
+
 # 主目录（需要更改）
 export AJET_ROOT="/mnt/data_cpfs/taoshuchang.tsc/deepresearch/AgentJet_new"
 
-NNODES=${WORLD_SIZE}
+# 单机调试配置（默认值）
+NNODES=${WORLD_SIZE:-1}
+GPUS_PER_NODE=8
+CURRENT_TIME=$(date "+%Y%m%d_%H%M%S")
+LOG_DIR="${AJET_ROOT}/logs/${PREFIX}"
+TRAIN_LOG="${LOG_DIR}/train_${SUFFIX}_${CURRENT_TIME}.log"
+mkdir -p ${LOG_DIR}
 
 # 涉密的配置（API_KEY以及模型、数据位置）从.env读取
 cd ${AJET_ROOT}
@@ -42,6 +55,9 @@ else
     echo -e "\033[31m警告: 找不到 .env 文件: $ENV_FILE\033[0m"
 fi
 
+export MODEL_PATH="/mnt/data_cpfs/taoshuchang.tsc/models/Qwen3-8B"
+
+
 #===============================================================================
 # 2. 动态生成配置文件 (从yaml template生成yaml)
 #===============================================================================
@@ -57,6 +73,10 @@ sed -e "s|{{SUFFIX}}|${SUFFIX}|g" \
     -e "s|{{RM_WEIGHT}}|${RM_WEIGHT}|g" \
     -e "s|{{PRESENTATION_QUALITY_WEIGHT}}|${PRESENTATION_QUALITY_WEIGHT}|g" \
     -e "s|{{GROUNDING_WEIGHT}}|${GROUNDING_WEIGHT}|g" \
+    -e "s|{{CGCV_WEIGHT}}|${CGCV_WEIGHT}|g" \
+    -e "s|{{AUDIT_WEIGHT}}|${AUDIT_WEIGHT}|g" \
+    -e "s|{{TRACEABILITY_WEIGHT}}|${TRACEABILITY_WEIGHT}|g" \
+    -e "s|{{EBTU_WEIGHT}}|${EBTU_WEIGHT}|g" \
     -e "s|{{OPENJUDGE_LLM}}|${OPENJUDGE_LLM}|g" \
     -e "s|{{RM_LLM}}|${RM_LLM}|g" \
     -e "s|{{JUDGE_CONCURRENCY}}|${JUDGE_CONCURRENCY}|g" \
@@ -68,10 +88,11 @@ sed -e "s|{{SUFFIX}}|${SUFFIX}|g" \
     -e "s|{{TRAIN_REF_ANS_PATH}}|${TRAIN_REF_ANS_PATH}|g" \
     -e "s|{{VAL_REF_ANS_PATH}}|${VAL_REF_ANS_PATH}|g" \
     -e "s|{{CKPT_SAVE_PATH}}|${CKPT_SAVE_PATH}|g" \
+    -e "s|{{ENV_SERVICE_URL}}|${ENV_SERVICE_URL}|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}, PresentationQuality=${PRESENTATION_QUALITY_WEIGHT}, Grounding=${GROUNDING_WEIGHT}, OpenJudge=${OPENJUDGE_LLM}, RM_LLM=${RM_LLM}"
+echo "参数确认: RM=${RM_WEIGHT}, PresentationQuality=${PRESENTATION_QUALITY_WEIGHT}, Grounding=${GROUNDING_WEIGHT}, CGCV=${CGCV_WEIGHT}, Audit=${AUDIT_WEIGHT}, Traceability=${TRACEABILITY_WEIGHT}, EBTU=${EBTU_WEIGHT}, OpenJudge=${OPENJUDGE_LLM}, RM_LLM=${RM_LLM}"
 
 
 #===============================================================================
@@ -115,15 +136,16 @@ export RAY_CLUSTER_MODE="multi_node"
 #===============================================================================
 # 6. 主流程
 #===============================================================================
-log "节点数: ${NNODES}, 每节点GPU数: ${GPUS_PER_NODE}"
-mkdir -p ${LOG_DIR}
-mkdir -p $(dirname ${CONFIG_FILE})
+log "单机调试模式: NNODES=${NNODES}, GPUS_PER_NODE=${GPUS_PER_NODE}"
 
 #===============================================================================
 #  6.1 Master 节点启动流程
 #===============================================================================
 # 启动训练任务（最核心）
+# 请注意只有单节点需要--with-ray 多节点应该删除
 python ajet/launcher.py \
     --conf ${CONFIG_FILE} \
+    --with-deepfinance \
+    --with-ray \
     --backbone="debug" \
     2>&1 | tee ${TRAIN_LOG}

tutorial/example_deep_finance/deep_finance_single.sh

@@ -1,11 +1,15 @@
 # 使得可以通过 from judge import PresentationQualityGrader 直接引用
 from .grounding.grader import GroundingGrader
 from .presentation_quality.grader import PresentationQualityGrader
+from .cgcv.grader import CGCVGrader
+from .audit.grader import AuditGrader
+from .traceability.grader import TraceabilityRewardGrader
+from .ebtu.grader import EBTUTraceabilityGrader
 # from .research_depth.grader import ResearchDepthGrader
 # from .research_breadth.grader import ResearchBreadthGrader
 
 # 以后添加了其他 grader 也可以加在这里
 # from .grounding.grader import GroundingGrader
 # from .research_breadth.grader import ResearchBreadthGrader
 # __all__ = ["PresentationQualityGrader", "GroundingGrader", "ResearchDepthGrader", "ResearchBreadthGrader"]
-__all__ = ["PresentationQualityGrader", "GroundingGrader"]
+__all__ = ["PresentationQualityGrader", "GroundingGrader", "CGCVGrader", "AuditGrader", "TraceabilityRewardGrader", "EBTUTraceabilityGrader"]