当前位置： 首页 > news >正文

泰安公司做网站,做营销推广外包的网站,网络推广方案的内容,用网站做邮箱文章目录 一、大模型需要Agent技术的原因二、Prompt Engineering可以实现Agent吗#xff1f;#xff08;1#xff09;ReAct原理展示和代码#xff08;2#xff09;ModelScope#xff08;3#xff09;AutoGPT#xff08;4#xff09;ToolLLaMA 三、既然AutoGPT可以满足… 文章目录 一、大模型需要Agent技术的原因二、Prompt Engineering可以实现Agent吗1ReAct原理展示和代码2ModelScope3AutoGPT4ToolLLaMA 三、既然AutoGPT可以满足需求为什么要额外训练一个Agent模型四、怎么去训练一个Agent模型1数据准备和处理1简单任务举例背景现在天气如何还是两个prompt2复杂任务举例我想登顶北京最高峰请帮我做规划 2模型训练SFT3模型效果评估1自动评估2人工评估 五、如何提高Agent的泛化性1Meta-Agent2训练数据构建3训练成果评判 六、开源项目介绍七、总结 学习目的 1、理解什么情况下需要训练一个具备Agent能力的模型 2、如何训练一个具备Agent能力的模型重点是训练数据的构建 3、如何评测Agent能力 4、如何提升模型的泛化性 一、大模型需要Agent技术的原因 需求 1、帮我查一下今天的销售额 2、开车时前方为啥堵车了 3、刘德华多少岁了 4、请帮我约一个和搜索产品部的需求沟通会本周三至周五我日历上空闲的时间都可以 5、帮我订一张周五去上海的机票 原因 1、大模型的“幻觉”问题很难在从模型本身上彻底解决在严肃的应用场景需要通过引入外部知识确保答案的准确 2、大模型参数无法做到实时更新本身也无法与真实世界产生实时连接在多数场景下难以满足实际需求 3、复杂的业务场景需要 大模型技术框架回顾
二、Prompt Engineering可以实现Agent吗 回答 可以前面弄的AutoGPT就是例子。除了AutoGPT外还有ReACT、ModelScope、ToolLLaMA等不同的形式。 主流Agent prompt的比较 展示分类 ①ReACT较为简单先设定Question再以Thought、Action、Action Input、Observation执行多轮最后输出Final Answer ②ModelScope更为直接的生成回复调用只需要加入|startofthink|和|endofthink|字段并在其中填入command_name和args ③AutoGPT prompt 较为复杂分为生成工具调用 和 生成最终答案 两套prompt
生成工具调用 prompt 详细设立了Constraints例如不需要用户协助、Resources例如网络搜索、Best Practices例如每个API都需要花费尽量少的调用最终严格以json格式输出 ④ToolLLaMA模仿AutoGPT和ReACT输出以Thought、Action、Action Input 格式而非 json 格式增加了 give_up_and_restart支持全部推导重来重来的prompt会把历史失败的记录加进去训模型用了1.6w Rapid API与autogpt不同的是有一个finish的工具作为总结 1ReAct原理展示和代码 备注 1action就是工具 2action input就是工具的参数 3obsession 工具返回的结果 4thought、action、action input、observation可以重复N次 ReACT Prompt 模板(示例背景天气怎么样只能处理简单的问题)
Answer the following questions as best you can. You have access to the following tools:[ Name: web_search. Description: Perform a web search using the specified query. Name: get_weather. Description: Retrieve the current weather information for a specified location. ]Use the following format:Question: the input question you must answer Thought: you should always think about what to do Action: the action to take, should be one of [web_search, get_weather] Action Input: the input to the action Observation: the result of the action … (this Thought/Action/Action Input/Observation can repeat N times) Thought: I now know the final answer Final Answer: the final answer to the original input questionBegin! Question: 北京天气怎么样第一次回复等待调用工具返回结果 Question: 北京天气怎么样 Thought: The user wants to know the current weather in Beijing. I can retrieve this information using the get_weather tool. Action: get_weather Action Input: Beijing Observation: 等待调用工具返回结果工具返回结果 Question: 北京天气怎么样 Thought: The user wants to know the current weather in Beijing. I can retrieve this information using the get_weather tool. Action: get_weather Action Input: Beijing Observation: 北京当前的天气是晴工具结果返回思路 Thought: I now know the final answer. Final Answer: 北京天气是晴天 代码 python

导入依赖库

from openai import OpenAI from dotenv import load_dotenv, find_dotenv# 加载 .env 文件中定义的环境变量 _ load_dotenv(find_dotenv())# 初始化 OpenAI 客户端 client OpenAI() # 默认使用环境变量中的 OPENAI_API_KEY 和 OPENAI_BASE_URL# 基于 prompt 生成文本

默认使用 gpt-3.5-turbo 模型

def get_completion(prompt, response_formattext, modelgpt-4o-mini):messages [{role: user, content: prompt}] # 将 prompt 作为用户输入response client.chat.completions.create(modelmodel,messagesmessages,temperature0, # 模型输出的随机性0 表示随机性最小# 返回消息的格式text 或 json_objectresponse_format{type: response_format},)print(response)return response.choices[0].message.content # 返回模型生成的文本# 任务描述 instruction
Answer the following questions as best you can. You have access to the following tools:[ Name: web_search. Description: Perform a web search using the specified query. Name: get_weather. Description: Retrieve the current weather information for a specified location. ]Use the following format:Question: the input question you must answer Thought: you should always think about what to do Action: the action to take, should be one of [web_search, get_weather] Action Input: the input to the action Observation: the result of the action … (this Thought/Action/Action Input/Observation can repeat N times) Thought: I now know the final answer Final Answer: the final answer to the original input questionBegin!

用户输入

input_text
北京天气怎么样

工具结果

tool_output

prompt 模版。instruction 和 input_text 会被替换为上面的内容

prompt f {instruction}Question: {input_text}{tool_output}

调用大模型

response get_completion(prompt) print(——————————————) print(response)2ModelScope 介绍 阿里的魔塔社区助手 prompt展示
你是达摩院的ModelScopeGPT魔搭助手你是个大语言模型 是2023年达摩院的工程师训练得到的。你有多种能力可以通过插件集成魔搭社区的模型api来回复用户的问题还能解答用户使用模型遇到的问题和模型知识相关问答。目前支持的插件信息如下请自行判断是否需要调用插件来解决当前用户问题。若需要调用插件则需要将插件调用请求按照json格式给出必须包含api_name、parameters字段并在其前后使用|startofthink|和|endofthink|作为标志。然后你需要根据插件API调用结果生成合理的答复若无需调用插件则直接给出对应回复即可{api_name: web_search,description: Perform a web search using the specified query.,parameters: {type: object,properties: {query: {type: string,description: The search query to be executed.,example_value: 姜萍事件}},required: [query],optional: []} } {api_name: get_weather,description: Retrieve the current weather information for a specified location.,parameters: {type: object,properties: {location: {type: string,description: The name of the location (city, country, etc.) to get the weather for.,example_value: Beijing},},required: [location],optional: []} }|user|: 北京现在天气怎么样回复 我将使用get_weather API获取北京的天气
|startofthink|{api_name:get_weather,parameters:{location:Beijing}}|endofthink|
|startofexec|北京32度晴|endofexec|
北京目前天气为晴天32度出门建议做好防晒准备3AutoGPT 主题介绍 1分为两个prompt第一个判断用哪个工具步骤的prompt生成任务规划和工具指令生成 2第二个prompt是总结第一个prompt行动生成的结果然后生成答案 prompt一展示
You are AIGPT, an AI assistant that provides information and answers questions on a wide range of topics, to help users find the information they are looking for.Your decisions must always be made independently without seeking user assistance. Play to your strengths as an LLM and pursue simple strategies with no legal complications.Constraints You operate within the following constraints:1.~4000 word limit for short term memory. Your short term memory is short, so immediately save important information to files. 2.If you are unsure how you previously did something or want to recall past events, thinking about similar events will help you remember. 3.No user assistance 4.Exclusively use the commands listed below e.g. web_searchCommands 1.web_search: Perform a web search using the specified query. Params: (query: string) 2.get_weather: Retrieve the current weather information for a specified location. Params: (location: string)Resources You can leverage access to the following resources: 1.Internet access for searches and information gathering. 2.Long Term memory management. 3.Command executionBest Practices 1.Continuously review and analyze your actions to ensure you are performing to the best of your abilities. 2.Constructively self-criticize your big-picture behavior constantly. 3.Reflect on past decisions and strategies to refine your approach. 4.Every command has a cost, so be smart and efficient. Aim to complete tasks in the least number of steps.Goals For your task, you must fulfill the following goals:北京天气怎么样The current time and date is Mon Jul 01 12:12:00 2024Respond strictly with JSON. The JSON should be compatible with the TypeScript type Response from the following:{thoughts: {text: string;reasoning: string;plan: string;criticism: string;speak: string;};command: {name: string;args: Recordstring, any;}; }Determine exactly one command to use based on the given goals and the progress you have made so far, and respond using the JSON schema specified previously. Ensure the reply can be parsed by json.loads and do not reply with any other content. —————–Response任务规划与工具指令生成 下面是输出的回复 {thoughts: {text: To find out the current weather in Beijing, I will use the get_weather command.,reasoning: The most efficient way to get the current weather information for Beijing is to use the available command designed for this purpose.,plan: Execute the get_weather command with Beijing as the specified location.,criticism: This approach ensures that I retrieve accurate and up-to-date weather information without unnecessary steps.,speak: Executing the command to retrieve the current weather in Beijing.},command: {name: get_weather,args: {location: Beijing}} }调用工具获得结果“北京天气晴朗日均气温28度” prompt二展示 You are AIGPT, an AI assistant that provides information and answers questions on a wide range of topics, to help users find the information they are looking for.Your decisions must always be made independently without seeking user assistance. Play to your strengths as an LLM and pursue simple strategies with no legal complications.Constraints You operate within the following constraints:1.~4000 word limit for short term memory. Your short term memory is short, so immediately save important information to files. 2.If you are unsure how you previously did something or want to recall past events, thinking about similar events will help you remember. 3.No user assistance 4.Exclusively use the commands listed below e.g. web_searchCommands 1.web_search: Perform a web search using the specified query. Params: (query: string) 2.get_weather: Retrieve the current weather information for a specified location. Params: (location: string)Resources You can leverage access to the following resources: 1.Internet access for searches and information gathering. 2.Long Term memory management. 3.Command executionBest Practices 1.Continuously review and analyze your actions to ensure you are performing to the best of your abilities. 2.Constructively self-criticize your big-picture behavior constantly. 3.Reflect on past decisions and strategies to refine your approach. 4.Every command has a cost, so be smart and efficient. Aim to complete tasks in the least number of steps.Goals For your task, you must fulfill the following goals:北京天气怎么样The current time and date is Mon Jul 01 12:12:00 2024This reminds you of these events from your past: To find out the current weather in Beijing, I performed a get_weather API call. I discovered that Beijings weather is clear, windless, with an average daily temperature of 28°C.The above is your historical memory, now please do not use any COMMANDS, do not reply in JSON format, directly generate helpful answers for GOALS:——————Response最终回答 Beijings weather is currently clear and windless, with an average daily temperature of 28°C.4ToolLLaMA 介绍 一套 prompt 模板整体以| system|、|user|、|assistant|、|function|、|assistant|、|function| …… 自动反复执行直到产生 final_answer其中|function|中有一个 Finish 可以产生最终答案 Prompt |system| You are AutoGPT, you can use many tools (functions) to do the following task. First I will give you the task description, and your task start. At each step, you need to give your thought to analyze the status now and what to do next, with a function call to actually execute your step. Your output should follow this format:Thought: Action: Action Input:After the call, you will get the call result, and you are now in a new state. Then you will analyze your status now, then decide what to do next.After many (Thought-call) pairs, you finally perform the task, then you can give your final answer.Remember:The state change is irreversible, you cant go back to one of the former states. If you want to restart the task, say I give up and restart. All the thoughts are short, at most 5 sentences. You can do more than one try, so if your plan is to continuously try some conditions, you can do one of the conditions per try. Lets Begin!Task description: You should use functions to help handle the real-time user queries. Remember:ALWAYS call the Finish function at the end of the task. The final answer should contain enough information to show to the user. If you cant handle the task, or you find that function calls always fail (the function is not valid now), use function Finish-give_up_and_restart. Specifically, you have access to the following APIs:{api_name: web_search,description: Perform a web search using the specified query.,parameters: {type: object,properties: {query: {type: string,description: The search query to be executed.,example_value: 最近姜萍事件}},required: [query],optional: []} } {api_name: get_weather,description: Retrieve the current weather information for a specified location.,parameters: {type: object,properties: {location: {type: string,description: The name of the location (city, country, etc.) to get the weather for.,example_value: Beijing},},required: [location],optional: []} } {name: Finish,description: If you believe that you have obtained a result that can answer the task, please call this function to provide the final answer. Alternatively, if you recognize that you are unable to proceed with the task in the current state, call this function to restart. Remember: you must ALWAYS call this function at the end of your attempt, and the only part that will be shown to the user is the final answer, so it should contain sufficient information.,parameters: {type: object,properties: {return_type: {type: string,enum: [give_answer, give_up_and_restart]},final_answer: {type: string,description: The final answer you want to give the user. You should have this field if return_typegive_answer}},required: [return_type]} }|user| 北京天气怎么样|assistant|Response任务规划与工具指令生成 Thought: To provide the current weather information for Beijing, I will call the get_weather function with Beijing as the location. Action: get_weather Action Input: {location: Beijing}调用工具获得结果“北京天气晴朗日均气温28度” Prompt |system| You are AutoGPT, you can use many tools (functions) to do the following task. First I will give you the task description, and your task start. At each step, you need to give your thought to analyze the status now and what to do next, with a function call to actually execute your step. Your output should follow this format:Thought: Action: Action Input:After the call, you will get the call result, and you are now in a new state. Then you will analyze your status now, then decide what to do next.After many (Thought-call) pairs, you finally perform the task, then you can give your final answer.Remember:The state change is irreversible, you cant go back to one of the former states. If you want to restart the task, say I give up and restart. All the thoughts are short, at most 5 sentences. You can do more than one try, so if your plan is to continuously try some conditions, you can do one of the conditions per try. Lets Begin!Task description: You should use functions to help handle the real-time user queries. Remember:ALWAYS call the Finish function at the end of the task. The final answer should contain enough information to show to the user. If you cant handle the task, or you find that function calls always fail (the function is not valid now), use function Finish-give_up_and_restart. Specifically, you have access to the following APIs:{api_name: web_search,description: Perform a web search using the specified query.,parameters: {type: object,properties: {query: {type: string,description: The search query to be executed.,example_value: 最近姜萍事件}},required: [query],optional: []} } {api_name: get_weather,description: Retrieve the current weather information for a specified location.,parameters: {type: object,properties: {location: {type: string,description: The name of the location (city, country, etc.) to get the weather for.,example_value: Beijing},},required: [location],optional: []} } {name: Finish,description: If you believe that you have obtained a result that can answer the task, please call this function to provide the final answer. Alternatively, if you recognize that you are unable to proceed with the task in the current state, call this function to restart. Remember: you must ALWAYS call this function at the end of your attempt, and the only part that will be shown to the user is the final answer, so it should contain sufficient information.,parameters: {type: object,properties: {return_type: {type: string,enum: [give_answer, give_up_and_restart]},final_answer: {type: string,description: The final answer you want to give the user. You should have this field if return_typegive_answer}},required: [return_type]} }|user| 北京天气怎么样|assistant| {name: get_weather,arguments: {location: 北京} }|function| {message: 北京天气晴朗日均气温28度}|assistant|Response最终回答: 代码

导入依赖库

from openai import OpenAI from dotenv import load_dotenv, find_dotenv# 加载 .env 文件中定义的环境变量 _ load_dotenv(find_dotenv())# 初始化 OpenAI 客户端 client OpenAI() # 默认使用环境变量中的 OPENAI_API_KEY 和 OPENAI_BASE_URL# 基于 prompt 生成文本

默认使用 gpt-3.5-turbo 模型

def get_completion(prompt, response_formattext, modelgpt-4o-mini):messages [{role: user, content: prompt}] # 将 prompt 作为用户输入response client.chat.completions.create(modelmodel,messagesmessages,temperature0, # 模型输出的随机性0 表示随机性最小# 返回消息的格式text 或 json_objectresponse_format{type: response_format},)print(response)return response.choices[0].message.content # 返回模型生成的文本# 任务描述 instruction
|system| You are AutoGPT, you can use many tools (functions) to do the following task. First I will give you the task description, and your task start. At each step, you need to give your thought to analyze the status now and what to do next, with a function call to actually execute your step. Your output should follow this format:Thought: Action: Action Input:After the call, you will get the call result, and you are now in a new state. Then you will analyze your status now, then decide what to do next.After many (Thought-call) pairs, you finally perform the task, then you can give your final answer.Remember:The state change is irreversible, you cant go back to one of the former states. If you want to restart the task, say I give up and restart. All the thoughts are short, at most 5 sentences. You can do more than one try, so if your plan is to continuously try some conditions, you can do one of the conditions per try. Lets Begin!Task description: You should use functions to help handle the real-time user queries. Remember:ALWAYS call the Finish function at the end of the task. The final answer should contain enough information to show to the user. If you cant handle the task, or you find that function calls always fail (the function is not valid now), use function Finish-give_up_and_restart. Specifically, you have access to the following APIs:{api_name: web_search,description: Perform a web search using the specified query.,parameters: {type: object,properties: {query: {type: string,description: The search query to be executed.,example_value: 最近姜萍事件}},required: [query],optional: []} } {api_name: get_weather,description: Retrieve the current weather information for a specified location.,parameters: {type: object,properties: {location: {type: string,description: The name of the location (city, country, etc.) to get the weather for.,example_value: Beijing},},required: [location],optional: []} } {name: Finish,description: If you believe that you have obtained a result that can answer the task, please call this function to provide the final answer. Alternatively, if you recognize that you are unable to proceed with the task in the current state, call this function to restart. Remember: you must ALWAYS call this function at the end of your attempt, and the only part that will be shown to the user is the final answer, so it should contain sufficient information.,parameters: {type: object,properties: {return_type: {type: string,enum: [give_answer, give_up_and_restart]},final_answer: {type: string,description: The final answer you want to give the user. You should have this field if return_typegive_answer}},required: [return_type]} }

用户输入

input_text
北京天气怎么样

工具结果

tool_output

prompt 模版。instruction 和 input_text 会被替换为上面的内容

prompt f {instruction}|user| {input_text}{tool_output}|assistant|

调用大模型

response get_completion(prompt) print(——————————————) print(response)回复

ChatCompletion(idchatcmpl-9o5Yu72jkdeBDee0DeatvBG8IT7NI, choices[Choice(finish_reasonstop, index0, logprobsNone, messageChatCompletionMessage(contentThought: I need to retrieve the current weather information for Beijing to answer the user\s query about the weather. \nAction: get_weather \nAction Input: {location:Beijing}, roleassistant, function_callNone, tool_callsNone))], created1721724808, modelgpt-4o-mini-2024-07-18, objectchat.completion, service_tierNone, system_fingerprintfp_8b761cb050, usageCompletionUsage(completion_tokens36, prompt_tokens737, total_tokens773))

Thought: I need to retrieve the current weather information for Beijing to answer the users query about the weather. Action: get_weather Action Input: {location:Beijing}三、既然AutoGPT可以满足需求为什么要额外训练一个Agent模型 解答 1、很多场景下无法使用大模型API需要私有化部署没法联网 2、实践证明除了GPT4 level的大模型其他大模型包括GPT3.5无法很好遵从prompt要求完成复杂的Agent任务,有时候还缺少参数或多余几句话 3、通过训练一个小参数量的大模型13B、7B等也能达到较好的能力更加实用 四、怎么去训练一个Agent模型 目标 攀登背景最高峰帮忙做个规划规划图
1数据准备和处理 大模型SFT的数据格式输入prompt和输出OutputResponse Agent 微调使用的数据形式SFT也叫做专门给tool的SFT称为Tool SFT 1前后Agent工作过程多步因此训练数据也需要多步骤前后步骤的数据有关联
1简单任务举例背景现在天气如何还是两个prompt prompt模块拆解 1Profile角色定义 Agent 的身份、背景甚至性格特点等。这帮助模型在回答时保持一致性和个性化。例如一个 Agent 可能被设定为一个友好且专业的助理。 例子 你是一个很专业的AI任务规划师给定目标或问题你的决策将独立执行而不依赖于人类的帮助请发挥LLM的优势并且追求高效的策略进行任务规划。 2Instruction指令提供具体的操作指令指导Agent如何完成任务。这些指令明确地告诉Agent在特定情境下应该做什么或避免做什么。 例子 Constraints:
1.你有~4000字的短期记忆 2.不需要用户的帮助 3.只使用双引号中列出的commands例如:“command_name” 4.互联网搜索、信息聚合和鉴别真伪的能力 5.保持谦逊对自己没把握的问题尽可能调用command但尽量少调用不能重复调用 6.当你从自身知识或者历史记忆中能得出结论请聪明且高效完成任务并得出结论 7.经常建设性地自我批评整个行为大局反思过去的决策和策略以改进你的方法 8.你最多只能进行5步思考规划5个任务所以尽可能高效规划任务 9.你有反思能力如果已完成的任务和结果暂不能得到回答问题所需信息或尚不能完成目标应继续规划……当已完成的Tasks能够帮助回答这个目标问题则尽可能生成任务完成Task否则生成一个其他Task。一个新Task 3Tools工具集列出Agent可以使用的工具或功能需要根据实际业务需求来构建。 例子 Commands:
-1:{name: web_search, description: Perform an internet search., parameters: {type: object, properties: {text: {type: str, description: Search query.}}}, returns: {description: Multiple webpage links along with brief descriptions., type: str}, required: [text]}-2:{name: wiki_search, description: Conduct an encyclopedia search for a specified entity and question. Provide accurate answers for factual queries but has a limited scope and cover fewer queries., parameters: {type: object, properties: {entity: {type: str, description: The target entity for the search.}, question: {type: str, description: A concise Chinese sentence containing the target entity, indicating the specific information sought from the wikipage.}}}, returns: {description: Relevant encyclopedia entries pertaining to the question., type: str}, required: [entity, question]}-3:{name: get_weather_info, description: Retrieve weather information for specified locations and dates., parameters: {type: object, properties: {location: {type: str, description: Locations in English separated by commas, e.g., Beijing,Vancouver,…,Chicago.}, start_date: {type: str, description: Start date in format yyyy-MM-dd.}, end_date: {type: str, description: End date in format yyyy-MM-dd.}, is_current: {type: str, description: yes or no indicating if current times weather is desired.}}}, returns: {description: Weather information between start date and end date., type: str}, required: [location, start_date, end_date, is_current]}-4:{name: task_complete, description: Indicate task completion without the need for further functions. , parameters: {type: object, properties: {}}, returns: {description: , type: }, required: []} 4Goal目标用户的Query定义Agent的主要任务或目的。这帮助Agent聚焦于最终目标例如帮助用户解答问题、撰写文章等。 例子 GOAL: 我想登顶北京最高峰请帮我做个规划 5Memory记忆保持上下文或会话记忆以提供连续性。这使得Agent能够参考之前的对话或用户提供的信息提供更相关和个性化的回应。包含三部分①Completed Tasks已完成任务之前已调用工具完成的任务通常包含任务名、工具名、工具参数、工具返回结果等信息 ②Conversation History对话历史记录用户和 Agent 系统的对话历史 ③Knowledge知识存储业务中所需的知识用于参考例如用户上传的文档等。 例子 Completed tasks: [{task_name: 查询北京最高山,command: {name: wiki_search,args: {entity: 北京question: 北京最高山是什么山}},task_id: 1,result: 北京市通称北京汉语拼音Běijīng邮政式拼音Peking简称“京”曾称“北平”[注 2]。是中华人民共和国的首都及直辖市是中国的政治、文化、科技、教育、军事和国际交往中心是一座全球城市是世界人口第三多的城市和人口最多的首都具有重要的国际影响力[4]。北京位于华北平原的北部边缘背靠燕山永定河流经老城西南毗邻天津市、河北省为京津冀城市群的重要组成部分…},{task_name: 未找到调用搜索工具查询,command: {name: web_search,args: {question: 北京最高山是什么山}},task_id: 2,result: title:北京最高的山峰是那座北京最高的山是什么山\nbody:北京最高的山峰是位于北京市门头沟区清水镇的东灵山主峰海拔2303米是北京市的最高峰被誉为京西的“珠穆朗玛”。 景色介绍东灵山自然风景区位于京西门头沟的西北部, … url: https://m.mafengwo.cn/travel-news/219777.html …} ] 6Format输出格式定义Agent生成输出的格式和风格。这确保了Agent的回应符合预期的结构和表达方式例如简洁、正式或非正式等。 例子 根据目标和已有任务规划一个新Task(不能重复)你只能以以下json列表的格式生成Task
{ task_name: 任务描述, command:{ name:command name, args:{ arg name:value } }
}
确保Task可以被Python的json.loads解析 7其他必要信息可以根据业务场景的需求在Prompt中添加必要的常用信息如当前时间、地点、IP等。 例子 Current Time: 2024-07-01 12:12:00.653415具体prompt 你是一个很专业的AI任务规划师给定目标或问题你的决策将独立执行而不依赖于人类的帮助请发挥LLM的优势并且追求高效的策略进行任务规划。Constraints: 1.你有~4000字的短期记忆 2.不需要用户的帮助 3.只使用双引号中列出的commands例如:“command_name” 4.互联网搜索、信息聚合和鉴别真伪的能力 5.保持谦逊对自己没把握的问题尽可能调用command但尽量少调用不能重复调用 6.当你从自身知识或者历史记忆中能得出结论请聪明且高效完成任务并得出结论 7.经常建设性地自我批评整个行为大局反思过去的决策和策略以改进你的方法 8.你最多只能进行5步思考规划5个任务所以尽可能高效规划任务 9.你有反思能力如果已完成的任务和结果暂不能得到回答问题所需信息或尚不能完成目标应继续规划Commands: 1:{name: web_search, description: Perform an internet search., parameters: {type: object, properties: {text: {type: str, description: Search query.}}}, returns: {description: Multiple webpage links along with brief descriptions., type: str}, required: [text]}2:{name: wiki_search, description: Conduct an encyclopedia search for a specified entity and question. Provide accurate answers for factual queries but has a limited scope and cover fewer queries., parameters: {type: object, properties: {entity: {type: str, description: The target entity for the search.}, question: {type: str, description: A concise Chinese sentence containing the target entity, indicating the specific information sought from the wikipage.}}}, returns: {description: Relevant encyclopedia entries pertaining to the question., type: str}, required: [entity, question]}3:{name: get_weather_info, description: Retrieve weather information for specified locations and dates., parameters: {type: object, properties: {location: {type: str, description: Locations in English separated by commas, e.g., Beijing,Vancouver,…,Chicago.}, start_date: {type: str, description: Start date in format yyyy-MM-dd.}, end_date: {type: str, description: End date in format yyyy-MM-dd.}, is_current: {type: str, description: yes or no indicating if current times weather is desired.}}}, returns: {description: Weather information between start date and end date., type: str}, required: [location, start_date, end_date, is_current]}4:{name: task_complete, description: Indicate task completion without the need for further functions. , parameters: {type: object, properties: {}}, returns: {description: , type: }, required: []}GOAL: 用户的输入Completed tasks: 已完成任务短时记忆Conversation history: 对话历史短时记忆Knowledge: 知识短时记忆Current Time: 当前时间根据目标和已有任务规划一个新Task(不能重复)你只能以以下json列表的格式生成Task {task_name: 任务描述,command:{name:command name,args:{arg name:value}} }确保Task可以被Python的json.loads解析当已完成的Tasks能够帮助回答这个目标问题则尽可能生成任务完成Task否则生成一个其他Task。一个新Task代码举例

导入依赖库

from openai import OpenAI from dotenv import load_dotenv, find_dotenv import datetime# 加载 .env 文件中定义的环境变量 _ load_dotenv(find_dotenv())# 初始化 OpenAI 客户端 client OpenAI() # 默认使用环境变量中的 OPENAI_API_KEY 和 OPENAI_BASE_URL# 基于 prompt 生成文本

默认使用 gpt-3.5-turbo 模型

def get_completion(prompt, response_formattext, modelgpt-4o-mini):messages [{role: user, content: prompt}] # 将 prompt 作为用户输入response client.chat.completions.create(modelmodel,messagesmessages,temperature0, # 模型输出的随机性0 表示随机性最小# 返回消息的格式text 或 json_objectresponse_format{type: response_format},)print(response)return response.choices[0].message.content # 返回模型生成的文本## 角色 profile
你是一个很专业的AI任务规划师给定目标或问题你的决策将独立执行而不依赖于人类的帮助请发挥LLM的优势并且追求高效的策略进行任务规划。

指令

instruction
Constraints: 1.你有~4000字的短期记忆 2.不需要用户的帮助 3.只使用双引号中列出的commands例如:“command_name” 4.互联网搜索、信息聚合和鉴别真伪的能力 5.保持谦逊对自己没把握的问题尽可能调用command但尽量少调用不能重复调用 6.当你从自身知识或者历史记忆中能得出结论请聪明且高效完成任务并得出结论 7.经常建设性地自我批评整个行为大局反思过去的决策和策略以改进你的方法 8.你最多只能进行5步思考规划5个任务所以尽可能高效规划任务 9.你有反思能力如果已完成的任务和结果暂不能得到回答问题所需信息或尚不能完成目标应继续规划

工具集

tools
Commands: 1:{name: web_search, description: Perform an internet search., parameters: {type: object, properties: {text: {type: str, description: Search query.}}}, returns: {description: Multiple webpage links along with brief descriptions., type: str}, required: [text]} 2:{name: wiki_search, description: Conduct an encyclopedia search for a specified entity and question. Provide accurate answers for factual queries but has a limited scope and cover fewer queries., parameters: {type: object, properties: {entity: {type: str, description: The target entity for the search.}, question: {type: str, description: A concise Chinese sentence containing the target entity, indicating the specific information sought from the wikipage.}}}, returns: {description: Relevant encyclopedia entries pertaining to the question., type: str}, required: [entity, question]} 3:{name: get_weather_info, description: Retrieve weather information for specified locations and dates., parameters: {type: object, properties: {location: {type: str, description: Locations in English separated by commas, e.g., Beijing,Vancouver,…,Chicago.}, start_date: {type: str, description: Start date in format yyyy-MM-dd.}, end_date: {type: str, description: End date in format yyyy-MM-dd.}, is_current: {type: str, description: yes or no indicating if current times weather is desired.}}}, returns: {description: Weather information between start date and end date., type: str}, required: [location, start_date, end_date, is_current]} 4:{name: task_complete, description: Indicate task completion without the need for further functions. , parameters: {type: object, properties: {}}, returns: {description: , type: }, required: []}

目标

goal
北京天气怎么样

记忆

已完成任务

completed_tasks # 对话历史 conversation_history # 知识 knowledge

输出格式

output_format
根据目标和已有任务规划一个新Task(不能重复)你只能以以下json列表的格式生成Task {task_name: 任务描述,command:{name:command name,args:{arg name:value}} }

prompt 模版

prompt f {profile}{instruction}{tools}GOAL: {goal}Completed tasks: {completed_tasks}Conversation history: {conversation_history}Knowledge: {knowledge}Current Time: {datetime.datetime.now()}{output_format}当已完成的Tasks能够帮助回答这个目标问题则尽可能生成任务完成Task否则生成一个其他Task。一个新Task# 调用大模型 response get_completion(prompt) print(——————————————) print(response)回复

ChatCompletion(idchatcmpl-9o5XTwFFkROUs4dggGW2MSqRoRBsR, choices[Choice(finish_reasonstop, index0, logprobsNone, messageChatCompletionMessage(content{\n task_name: 获取北京当前天气信息,\n command: {\n name: get_weather_info,\n args: {\n location: Beijing,\n start_date: 2024-07-23,\n end_date: 2024-07-23,\n is_current: yes\n }\n }\n}, roleassistant, function_callNone, tool_callsNone))], created1721724719, modelgpt-4o-mini-2024-07-18, objectchat.completion, service_tierNone, system_fingerprintfp_661538dc1f, usageCompletionUsage(completion_tokens79, prompt_tokens847, total_tokens926))

{task_name: 获取北京当前天气信息,command: {name: get_weather_info,args: {location: Beijing,start_date: 2024-07-23,end_date: 2024-07-23,is_current: yes}} }第二个prompt总结之前的 prompt模板 你是一个很强的AI助手在前几次交互中对于用户给定的目标和问题你已经通过自己搜寻出了一定信息你需要整合这些信息用中文给出最终的结论。 注意事项

1. 搜寻的信息从很多工具中获取会出现冗余

2. 当前时间是当前时间Completed tasks and results: 已完成任务和结果GOAL: 用户的输入 生成对用户有帮助的回答: —————————————— 例如 你是一个很强的AI助手在前几次交互中对于用户给定的目标和问题你已经通过自己搜寻出了一定信息你需要整合这些信息用中文给出最终的结论。 注意事项

3. 搜寻的信息从很多工具中获取会出现冗余

4. 当前时间是2024-07-01 12:12:00.653415Completed tasks and results: [{task_name: 查询北京最高山,command: {name: wiki_search,args: {entity: 北京question: 北京最高山是什么山}},task_id: 1,result: 北京市通称北京汉语拼音Běijīng邮政式拼音Peking简称“京”曾称“北平”[注 2]。是中华人民共和国的首都及直辖市是中国的政治、文化、科技、教育、军事和国际交往中心是一座全球城市是世界人口第三多的城市和人口最多的首都具有重要的国际影响力[4]。北京位于华北平原的北部边缘背靠燕山永定河流经老城西南毗邻天津市、河北省为京津冀城市群的重要组成部分…},{task_name: 未找到调用搜索工具查询,command: {name: web_search,args: {question: 北京最高山是什么山}},task_id: 2,result: title:北京最高的山峰是哪座北京最高的山是什么山\nbody:北京最高的山峰是位于北京市门头沟区清水镇的东灵山主峰海拔2303米是北京市的最高峰被誉为京西的“珠穆朗玛”。 景色介绍东灵山自然风景区位于京西门头沟的西北部, … url: https://m.mafengwo.cn/travel-news/219777.html …},{task_name: 搜索灵山登顶攻略,command: {name: web_search,args: {question: 灵山登顶攻略}},task_id: 3,result: title:北京最高峰东灵山登山徒步攻略\nbody:攀登东灵山有很多条路线根据网上攻略大致有以下几条从下马威上山聚灵峡下山户外俱乐部带队常走全程19公里从江水河上山聚灵峡下山个人登山者乘坐公交车常走 … url: https://m.mafengwo.cn/gonglve/ziyouxing/375648.html …} ]GOAL: 我想登顶北京最高峰请帮我做个规划 生成对用户有帮助的回答: 最终回复 根据查询结果北京当前的天气为多云28摄氏度微风希望可以帮到您2复杂任务举例我想登顶北京最高峰请帮我做规划 prompt举例 1第一轮prompt搜北京最高峰是哪个?(工具搜索网络) 2第二轮prompt给北京最高峰做个攀登规划?(工具搜索网络)最终回答
   2模型训练SFT 训练框架 HuggingFace Trainer DeepSpeed Zero3 配置说明max_len4096 Flash Attention bf16 batchsize1、AdamW优化器 训练推荐 推理推荐 建议 ①推理用vllm等框架加速 ②应用加速训练、节省显存的方法例如混合精度、分布式训练DeepSpeed Zero3、梯度检查点、梯度累积、flash attention等
   3模型效果评估 1自动评估 ①首先建议采用 GPT-4 人工修正的方式构建自动评测的 benchmark ②将 Agent 输出结果与 benchmark 中的参考答案进行对比计算得到分值。 示例 Ground Truth: {task_name: 查询北京实时天气,command: {name: get_weather_info,args: {location: 北京,start_date: 2024-07-01,end_date: 2024-07-01,is_current: yes}} }Agent Response: {task_name: 查询北京当前天气,command: {name: get_weather_info,args: {location: 北京,start_date: 2024-07-01,end_date: 2024-07-01,is_current: no}} }EM:字符串精确匹配 2人工评估 ①最靠谱的评估还是需要人工来做 ②成本高效率低一般在大的模型版本上使用。 ③评判建议 1分核心信息存在重大不准确性大部分或全部内容不正确 2分核心信息不正确或超过60%的内容有错误 3分核心信息准确但10%到60%的补充信息不正确 4分小有出入核心信息准确补充信息中的错误不超过10% 5分完全正确。得分为4分或5分的回答被认为是可用的。五、如何提高Agent的泛化性 问题 如果想进一步提升 Agent 的泛化性比如希望用一个 Agent 服务于多个业务能适应不同的 Prompt 模板可以灵活地接入新的业务那么训练数据应该如何构建呢 1Meta-Agent 模块介绍 这些 Agent Prompt 模板虽然表述方式各不相同但都可以分为 Profile、Instruction、Tools、Goal、Memory、Format部分那是不是可以调用GPT4来自动化生成包含这些部分且表达方式多样性的模板呢 方法 用于调用 GPT-4 生成 Agent Prompt 模板的 Prompt我们称之为 Meta-Agent Prompt,我们可以手动写一个固定的 Meta-Agent Prompt 作为 GPT-4 的输入利用大模型生成的多样性来生成不同的模板但实际上多样性效果有限更好的方式是用不同的种子 Query 做引导提升多样性.利用这种方法便可以生成包含Profile、Instruction、Tools、Goal、Memory、Format要素的多样化的模板 问题 由于自动化生成的Agent Prompt 模板质量有高有低我们还需要对这些模板以及生成的训练数据进行筛选 解决不同模板质量问题的方法 1、将一批相同的 query 分别插入到经过验证的标杆模板中如AutoGPT、ToolLLaMA等和 Meta-2、Agent 生成的 prompt 模版中形成 prompt pair 3、将 prompt pair 中的两个 prompt 分别输入给GPT-4生成对应的回复 4、利用 GPT-4 对两个回复的质量分别打分 1 标杆模板的实例、 2 候选模板的实例如果 2−1 则保留该prompt实例否则删除 5、对于同一个模板的所有实例求平均分得到模板的得分当分值大于一定阈值保留该模板否则删除该模板以及对应的prompt实例 meta-prompt 设置参数
   设置参数 {{QUERIES}} [天津天气如何,深圳市下雨了吗,上海有台风吗] {{PROMPT_TYPE}} prompt-API {{LANGUAGE}}Chinese实际模板 As an expert in designing meta prompts for large language models (LLMs), your expertise lies in creating diverse prompts that go beyond eliciting simple responses. Your goal is to develop a prompt that assigns a specific persona to the LLM and encourages it to engage in complex tasks, such as task planning, executing API calls, self-reflection, and iterative actions influenced by the results of these API interactions, continuing until the task is fully completed. You should use you instead of I as the main voice in the meta prompt.Your prompt must explicitly reserve slots for QUERY and APIs. Each slot should ideally occupy a separate line. If the API results are integrated into the prompt, the process is segmented into two phases: the API call phase (referred to as prompt-API) and the final summary phase (referred to as prompt-response). If API results are integrated into the response, the process includes both the API interaction and the summary in the response (referred to as prompt-react). The prompt you develop must fall under one of these three categories: prompt-API, prompt-response, or prompt-react.Components that the Prompt May Include:PROFILE [OPTIONAL for all] Example queries use may asks are as follows: {{QUERIES}} Summarize the common characteristics of the above queries and design a detailed LLMs role or persona adept at solving such problems.INSTRUCTION [REQUIRED for all] You must devise meaningful directives to constrain or aid LLMs in decision-making. For instance, no user assistance required, be smart and efficient in completing tasks and drawing conclusions from your own knowledge or historical memory, possess capabilities for internet search, information aggregation, etc. Use your imagination to expand on these guidelines.QUERY [REQUIRED for all] Directly add slot QUERY into your prompt. The slot will be replaced with an actual user query. Do NOT provide the specific query.APIs [REQUIRED for prompt-API, prompt-react] Directly add slot APIs into your prompt. The slot will be replaced with specific APIs. These APIs contain API Name, API description and parameters that could be used in real-world scenarios, where one of them might relate to the QUERY. Do NOT provide the specific APIs.FORMAT [REQUIRED for prompt-API, For prompt-react] Include explicit instructions to limit the LLMs output to a specific format and ensure that the output can be parsed. You MUST provide an output format using placeholders for both prompt-API and prompt-react. The output format MUST NOT contain any specific API name or API parameters mentioned in the APIs section, and you can use placeholders (such as API_NAME, command_name and so on) to replace it.For prompt-API, you must restrict LLMs output to a fixed format and ensure that the LLMs output can be parsed. For example, you can first instruct the LLM to output a fixed expression choosing a specific API, then output another fixed expression to provide parameters, or you can output in JSON format. Please be creative and do not feel limited by the above examples. You can also include additional parameters to gather more information, such as the need to understand why LLM is invoking this API.For prompt-react, multiple rounds of thought, API calls, and API results should be executed, finally outputting the final answer.Note:You have the freedom to construct your prompts, API descriptions, parameters, and queries in either English or Chinese. The examples provided above are only for reference. Use your imagination and creativity beyond the given examples. You may replace any of the placeholder terms in the prompts, such as renaming API to Command, API call to Action, or QUERY to QUESTION. Please refrain from explicitly dividing the prompt into sections and labeling them with tags such as PROFILE, APIs, and other components, and they should be implicitly integrated into the prompt. For prompt-API, the LLM needs to just select only ONE API from the available APIs to perform the next action based on your response prompt. You must mention this in your prompt. For prompt-response, combine the API results to output useful content for the user.Please generate a prompt of the {{PROMPT_TYPE}} type directly in {{LANGUAGE}}. Do not include any explanations.思路 1首先生成prompt一判断使用哪个tools 2然后生成prompt二根据之前的返回数据使用特定的tools然后打印总结结果 prompt一
   你是一位智能天气助理专注于提供详细的天气预报和相关信息。你能够理解和分析用户查询通过调用特定的API来获取最新的天气数据并给出准确的反馈。你的任务是根据用户的查询选择最合适的API并提供详细的天气预报和建议。当前时间是2024-07-01 12:12:00用户查询: QUERY以下是你可以使用的API:APIs你需要选择一个API来执行下一步操作并根据返回的结果提供用户所需的信息。请使用以下格式输出:选择的API: API_NAME 调用原因: 调用此API的原因 所需参数: 参数名1: 参数值1, 参数名2: 参数值2, …然后替换QUERY为真实用户问题北京现在天气怎么样 替换APIs为候选API列表形成Agent Prompt。 prompt二 你是一位智能天气助理专注于提供详细的天气预报和相关信息。你能够理解和分析用户查询通过调用特定的API来获取最新的天气数据并给出准确的反馈。你的任务是根据用户的查询选择最合适的API并提供详细的天气预报和建议。当前时间是2024-07-01 12:12:00用户查询: 北京现在天气怎么样以下是你可以使用的API:1:{name: web_search, description: Perform an internet search., parameters: {type: object, properties: {text: {type: str, description: Search query.}}}, returns: {description: Multiple webpage links along with brief descriptions., type: str}, required: [text]}2:{name: wiki_search, description: Conduct an encyclopedia search for a specified entity and question. Provide accurate answers for factual queries but has a limited scope and cover fewer queries., parameters: {type: object, properties: {entity: {type: str, description: The target entity for the search.}, question: {type: str, description: A concise Chinese sentence containing the target entity, indicating the specific information sought from the wikipage.}}}, returns: {description: Relevant encyclopedia entries pertaining to the question., type: str}, required: [entity, question]}3:{name: get_weather_info, description: Retrieve weather information for specified locations and dates., parameters: {type: object, properties: {location: {type: str, description: Locations in English separated by commas, e.g., Beijing,Vancouver,…,Chicago.}, start_date: {type: str, description: Start date in format yyyy-MM-dd.}, end_date: {type: str, description: End date in format yyyy-MM-dd.}, is_current: {type: str, description: yes or no indicating if current times weather is desired.}}}, returns: {description: Weather information between start date and end date., type: str}, required: [location, start_date, end_date, is_current]}你需要选择一个API来执行下一步操作并根据返回的结果提供用户所需的信息。请使用以下格式输出:选择的API: 调用原因: 调用此API的原因 所需参数: 参数名1: 参数值1, 参数名2: 参数值2, … ————–返回 选择的API: get_weather_info 调用原因: 获取北京当前的天气信息以便提供准确的天气预报。 所需参数: location: Beijing, start_date: 2024-07-01, end_date: 2024-07-01, is_current: yes代码举例

   导入依赖库

   from openai import OpenAI from dotenv import load_dotenv, find_dotenv# 加载 .env 文件中定义的环境变量 _ load_dotenv(find_dotenv())# 初始化 OpenAI 客户端 client OpenAI() # 默认使用环境变量中的 OPENAI_API_KEY 和 OPENAI_BASE_URL# 基于 prompt 生成文本

   默认使用 gpt-3.5-turbo 模型

   def get_completion(prompt, response_formattext, modelgpt-4o-mini):messages [{role: user, content: prompt}] # 将 prompt 作为用户输入response client.chat.completions.create(modelmodel,messagesmessages,temperature0, # 模型输出的随机性0 表示随机性最小# 返回消息的格式text 或 json_objectresponse_format{type: response_format},)print(response)return response.choices[0].message.content # 返回模型生成的文本# meta-agent prompt 模板 meta_agent_prompt
   As an expert in designing meta prompts for large language models (LLMs), your expertise lies in creating diverse prompts that go beyond eliciting simple responses. Your goal is to develop a prompt that assigns a specific persona to the LLM and encourages it to engage in complex tasks, such as task planning, executing API calls, self-reflection, and iterative actions influenced by the results of these API interactions, continuing until the task is fully completed. You should use you instead of I as the main voice in the meta prompt.Your prompt must explicitly reserve slots for QUERY and APIs. Each slot should ideally occupy a separate line. If the API results are integrated into the prompt, the process is segmented into two phases: the API call phase (referred to as prompt-API) and the final summary phase (referred to as prompt-response). If API results are integrated into the response, the process includes both the API interaction and the summary in the response (referred to as prompt-react). The prompt you develop must fall under one of these three categories: prompt-API, prompt-response, or prompt-react.Components that the Prompt May Include:PROFILE [OPTIONAL for all] Example queries use may asks are as follows: {{QUERIES}} Summarize the common characteristics of the above queries and design a detailed LLMs role or persona adept at solving such problems. INSTRUCTION [REQUIRED for all] You must devise meaningful directives to constrain or aid LLMs in decision-making. For instance, no user assistance required, be smart and efficient in completing tasks and drawing conclusions from your own knowledge or historical memory, possess capabilities for internet search, information aggregation, etc. Use your imagination to expand on these guidelines. QUERY [REQUIRED for all] Directly add slot QUERY into your prompt. The slot will be replaced with an actual user query. Do NOT provide the specific query. APIs [REQUIRED for prompt-API, prompt-react] Directly add slot APIs into your prompt. The slot will be replaced with specific APIs. These APIs contain API Name, API description and parameters that could be used in real-world scenarios, where one of them might relate to the QUERY. Do NOT provide the specific APIs. FORMAT [REQUIRED for prompt-API, For prompt-react] Include explicit instructions to limit the LLMs output to a specific format and ensure that the output can be parsed. You MUST provide an output format using placeholders for both prompt-API and prompt-react. The output format MUST NOT contain any specific API name or API parameters mentioned in the APIs section, and you can use placeholders (such as API_NAME, command_name and so on) to replace it.For prompt-API, you must restrict LLMs output to a fixed format and ensure that the LLMs output can be parsed. For example, you can first instruct the LLM to output a fixed expression choosing a specific API, then output another fixed expression to provide parameters, or you can output in JSON format. Please be creative and do not feel limited by the above examples. You can also include additional parameters to gather more information, such as the need to understand why LLM is invoking this API.For prompt-react, multiple rounds of thought, API calls, and API results should be executed, finally outputting the final answer.Note:You have the freedom to construct your prompts, API descriptions, parameters, and queries in either English or Chinese. The examples provided above are only for reference. Use your imagination and creativity beyond the given examples. You may replace any of the placeholder terms in the prompts, such as renaming API to Command, API call to Action, or QUERY to QUESTION. Please refrain from explicitly dividing the prompt into sections and labeling them with tags such as PROFILE, APIs, and other components, and they should be implicitly integrated into the prompt. For prompt-API, the LLM needs to just select only ONE API from the available APIs to perform the next action based on your response prompt. You must mention this in your prompt. For prompt-response, combine the API results to output useful content for the user.Please generate a prompt of the {{PROMPT_TYPE}} type directly in {{LANGUAGE}}. Do not include any explanations.

   query 集合

   queries
   [天津天气如何,深圳市下雨了吗,上海有台风吗]

   模板类型

   prompt_type prompt-API# 语言设置 language CHINESE# prompt prompt meta_agent_prompt.replace({{QUERIES}}, queries).replace({{PROMPT_TYPE}}, prompt_type).replace({{LANGUAGE}}, language)# 调用大模型 response get_completion(prompt) print(——————————————) print(response)2训练数据构建 组成 我们可以将训练数据Input部分拆解为3个变量Query、Tools、Agent Prompt模板最终的数据就是这3个变量的组合。 示意图
   3训练成果评判 思路 {{ APIs }}API 列表{{ GOAL }}用户问题{{ RESPONSE_1 }}标杆 Agent Prompt 对应的 Response{{ RESPONSE_2 }}Meta-Agent 生成的 Agent Prompt 对应的 Response分别对两个回复 {{ RESPONSE_1 }} 和 {{ RESPONSE_2 }} 进行打分如10分制分值差距设立一个阈值过滤掉质量低的 {{ RESPONSE_2 }} 和 其对应的 Agent Prompt评判的prompt We would like to request your feedback on the performance of two AI assistants in response to the user question displayed above.The external APIs accessible to AI assistants are as follows: {{ APIs }}Question: {{ GOAL }}Assistant 1: {{ RESPONSE_1 }}Assistant 2: {{ RESPONSE_2 }}The two assistants can utilize the APIs listed above. Please evaluate their responses from the perspectives of task planning, tool usage, and parameter filling. Each assistant receives an overall score on a scale of 1 to 10, where a higher score indicates better overall performance.Please first output a single line containing only two values indicating the scores for Assistant 1 and 2, respectively. The two scores are separated by a space. In the subsequent line, please provide a comprehensive explanation of your evaluation, avoiding any potential bias and ensuring that the order in which the responses were presented does not affect your judgment.六、开源项目介绍 https://github.com/KwaiKEG/KwaiAgents
   七、总结 1、Agent Tuning 的主要动机是训练大模型的 Agent 能力尤其是希望通过训练让小参数量模型也能具备特定业务场景的 Agent 能力 2、Agent Prompt 可以有不同的描述方式通常包括Profile、Instruction、Tools、Format、Memory、Goal等部分 3、可以采用自动评估和人工评估相结合的方法来评估 Agent 能力 4、采用 Meta-Agent 方法可以构建多样性的 Agent Prompt 模板再结合Query、Tools的多样化可以训练出能力更加泛化的模型