掺入20%其他大模型agent

Author: root

examples/werewolf/train.sh

@@ -51,6 +51,6 @@ python -m agentlightning.verl \
     trainer.project_name=${PROJECT_NAME} \
     trainer.experiment_name=${EXPERIMENT_NAME} \
     trainer.nnodes=1 \
-    trainer.save_freq=1 \
+    trainer.save_freq=10 \
     trainer.test_freq=0 \
     trainer.total_epochs=1 $@

examples/werewolf/werewolf_agent.py

@@ -306,32 +306,49 @@ async def create_player(role: str, NAME_TO_ROLE:  dict, ch_names: dict, moderato
     """Create a player with the given name and role."""
     name = get_player_name(candidate_names)
     NAME_TO_ROLE[name] = role
-    agent = ReActAgent(
-        name=name,
-        sys_prompt=Prompts.system_prompt.format(
-            player_name=name,
-            guidance=getattr(Prompts, f"notes_{role}"),
-        ),
-        # model=DashScopeChatModel(
-        #     model_name="qwen3-max-preview",
-        #     api_key=os.environ["DASHSCOPE_API_KEY"],
-        #     enable_thinking=True,
-        # ),
-        # model=OpenAIChatModel(
-        #     model_name="/root/dataDisk/Qwen3-8B",
-        #     client_args={"base_url": "http://127.0.0.1:8000/v1"},
-        #     api_key="xxx",
-        #     stream=False,
-        # ),
-        model=OpenAIChatModel(
-            model_name=llm.model,
-            client_args={"base_url": llm.endpoint},
-            api_key="xxx",
-            stream=False,
-        ),
-        # formatter=DashScopeMultiAgentFormatter(),
-        formatter=OpenAIMultiAgentFormatter(),
-    )
+    #添加外部对抗性
+    import random
+    if random.random() < 0.8:
+        agent = ReActAgent(
+            name=name,
+            sys_prompt=Prompts.system_prompt.format(
+                player_name=name,
+                guidance=getattr(Prompts, f"notes_{role}"),
+            ),
+            # model=DashScopeChatModel(
+            #     model_name="qwen3-max-preview",
+            #     api_key=os.environ["DASHSCOPE_API_KEY"],
+            #     enable_thinking=True,
+            # ),
+            # model=OpenAIChatModel(
+            #     model_name="/root/dataDisk/Qwen3-8B",
+            #     client_args={"base_url": "http://127.0.0.1:8000/v1"},
+            #     api_key="xxx",
+            #     stream=False,
+            # ),
+            model=OpenAIChatModel(
+                model_name=llm.model,
+                client_args={"base_url": llm.endpoint},
+                api_key="xxx",
+                stream=False,
+            ),
+            # formatter=DashScopeMultiAgentFormatter(),
+            formatter=OpenAIMultiAgentFormatter(),
+        )
+    else:
+        agent = ReActAgent(
+            name=name,
+            sys_prompt=Prompts.system_prompt.format(
+                player_name=name,
+                guidance=getattr(Prompts, f"notes_{role}"),
+            ),
+            model=DashScopeChatModel(
+                model_name="qwen3-max-preview",
+                api_key=os.environ["DASHSCOPE_API_KEY"],
+                enable_thinking=True,
+            ),
+            formatter=DashScopeMultiAgentFormatter(),
+        )
     # 获取对应角色的游戏指南
     role_notes = getattr(Prompts, f"notes_{role}", "")