Large model adjustment method and question and answer processing method and device based on large model technology

Abstract

The invention provides a large model adjustment method and a question and answer processing method and device based on a large model technology, relates to the field of artificial intelligence, in particular to the technical fields of large models, natural language processing, deep learning and the like, and can be applied to the application fields of generative search, intelligent assistants, intelligent customer service and the like. According to the specific implementation scheme, a reference question is generated by adopting a first large model according to description information of a target tool, and the reference question needs to call the target tool to generate a corresponding reply in a question and answer dialogue; according to the reference problem, generating a tool calling request of the target tool by adopting a second large model; and under the condition that the tool calling request successfully calls the target tool, generating a training sample for training the third model according to the tool calling request and the corresponding reference problem.

Description

Technical Field

[0001] The present disclosure relates to the field of artificial intelligence technology, specifically to technical fields such as big models, natural language processing, and deep learning, and can be used in application fields such as generative search, intelligent assistants, and intelligent customer service. In particular, it relates to a big model adjustment method, a question-answering processing method and device based on big model technology. Background Art

[0002] In recent years, although large language models (LLMs) have demonstrated excellent performance in various application scenarios, the tool calling capabilities of large models still need to be improved. In order to improve the tool calling capabilities of large models, related technologies propose to obtain training samples for targeted training of large models. However, these technologies are highly dependent on manual intervention when obtaining training samples, and the efficiency of obtaining training samples is low. Summary of the invention

[0003] The present invention provides a large model adjustment method, a question and answer processing method and device based on large model technology.

[0004] According to a first aspect of the present disclosure, a method for acquiring training samples of a large model is provided, comprising: based on description information of a target tool, using a first large model to generate a reference question, wherein the reference question requires calling the target tool to generate a corresponding reply in a question-and-answer dialogue; based on the reference question, using a second large model to generate a tool call request for the target tool; and when the tool call request successfully calls the target tool, generating training samples for training a third large model based on the tool call request and the corresponding reference question.

[0005] According to a second aspect of the present disclosure, a large model adjustment method is provided, comprising: obtaining training samples of the large model, wherein the training samples are obtained based on the large model training sample acquisition method described in the first aspect; and adjusting model parameters of the large model based on the training samples.

[0006] According to a third aspect of the present disclosure, a question-answering processing method based on a big model technology is provided, comprising: obtaining a question to be answered; inputting the question into a big model to obtain a reply output by the big model to the question, wherein the big model is obtained based on the big model adjustment method described in the second aspect.

[0007] According to a fourth aspect of the present disclosure, a training sample acquisition device for a large model is provided, comprising: a first generation module, for generating a reference question using a first large model according to description information of a target tool, wherein the reference question requires calling the target tool to generate a corresponding reply in a question-and-answer dialogue; a second generation module, for generating a tool call request for the target tool using a second large model according to the reference question; and a third generation module, for generating training samples for training the third large model according to the tool call request and the corresponding reference question when the tool call request successfully calls the target tool.

[0008] According to a fifth aspect of the present disclosure, a large model adjustment device is provided, comprising: a sample acquisition module, used to acquire training samples of the large model, wherein the training samples are obtained based on the large model training sample acquisition method described in the first aspect; and an adjustment module, used to adjust model parameters of the large model based on the training samples.

[0009] According to the sixth aspect of the present disclosure, a question and answer processing device based on the big model technology is provided, including: a question acquisition module, used to obtain questions to be answered; a reply module, used to input the questions into the big model, and obtain a reply output by the big model for the questions, wherein the big model is obtained based on the big model adjustment method described in the second aspect.

[0010] According to a seventh aspect of the present disclosure, there is provided an electronic device, including:

[0011] at least one processor; and

[0012] a memory communicatively connected to the at least one processor; wherein,

[0013] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the training sample acquisition method of the large model as described in the first aspect, or execute the large model adjustment method as described in the second aspect, or execute the question and answer processing method based on the large model technology as described in the third aspect.

[0014] According to an eighth aspect of the present disclosure, a non-transitory computer-readable storage medium storing computer instructions is provided, wherein the computer instructions are used to enable the computer to execute the large model training sample acquisition method as described in the first aspect, or to execute the large model adjustment method as described in the second aspect, or to execute the question and answer processing method based on the large model technology as described in the third aspect.

[0015] According to the ninth aspect of the present disclosure, a computer program product is provided, comprising computer instructions, which, when executed by a processor, implement the method for acquiring training samples of the large model as described in the first aspect, or implement the method for adjusting the large model as described in the second aspect, or implement the steps of the question and answer processing method based on the large model technology as described in the third aspect.

[0016] The present disclosure provides a large model adjustment method, a question-answering processing method and a device based on the large model technology, which have the following beneficial effects:

[0017] According to the description information of the target tool, the first model is used to generate reference questions. The reference questions need to call the target tool to generate corresponding replies in the question-answering dialogue; according to the reference questions, the second model is used to generate a tool call request for the target tool; when the tool call request successfully calls the target tool, a training sample for training the third model is generated according to the tool call request and the corresponding reference question. The present disclosure uses the first model to automatically generate reference questions according to the description information of the target tool, and uses the second model to automatically generate tool call requests according to the reference questions, and then automatically generates training samples according to the tool call requests and the corresponding reference questions. The entire process does not require manual intervention, which improves the efficiency of obtaining training samples.

[0018] It should be understood that the content described in this section is not intended to identify the key or important features of the embodiments of the present disclosure, nor is it intended to limit the scope of the present disclosure. Other features of the present disclosure will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The accompanying drawings are used to better understand the present solution and do not constitute a limitation of the present disclosure.

[0020] Figure 1 It is a flowchart of a method for obtaining training samples of a large model provided in accordance with an embodiment of the present disclosure;

[0021] Figure 2 is a flowchart of a method for obtaining training samples of a large model provided in accordance with another embodiment of the present disclosure;

[0022] Figure 3 is a flowchart of a method for obtaining training samples of a large model provided in accordance with another embodiment of the present disclosure;

[0023] Figure 4 is a flowchart of a method for obtaining training samples of a large model provided in accordance with another embodiment of the present disclosure;

[0024] Figure 5 is a flow chart of a large model adjustment method provided according to an embodiment of the present disclosure;

[0025] Figure 6 It is a flowchart of a question-answering processing method based on a large model technology provided according to an embodiment of the present disclosure;

[0026] Figure 7 A schematic diagram of the structure of a training sample acquisition device for a large model provided according to an embodiment of the present disclosure;

[0027] Figure 8 is a flow chart of a large model adjustment device provided according to an embodiment of the present disclosure;

[0028] Fig. 9 It is a flowchart of a question-answering processing device based on a large model technology provided according to an embodiment of the present disclosure;

[0029] Fig.10 It is a block diagram of an electronic device used to implement the large model training sample acquisition method or large model adjustment method or large model technology-based question and answer processing method of the embodiment of the present disclosure. DETAILED DESCRIPTION

[0030] The following is a description of exemplary embodiments of the present disclosure in conjunction with the accompanying drawings, including various details of the embodiments of the present disclosure to facilitate understanding, which should be considered as merely exemplary. Therefore, it should be recognized by those of ordinary skill in the art that various changes and modifications may be made to the embodiments described herein without departing from the scope and spirit of the present disclosure. Similarly, for the sake of clarity and conciseness, descriptions of well-known functions and structures are omitted in the following description.

[0031] In the technical solution of the present disclosure, the collection, storage, use, processing, transmission, provision and disclosure of user personal information are all carried out with the user's consent, comply with the relevant laws and regulations, and do not violate public order and good morals.

[0032] In order to improve the tool calling capability of large models, related technologies propose a training method based on standard fine-tuning (SFT) or a training method based on preset prompts to perform targeted training on large models. However, the SFT-based training method requires manual annotation of the data set, and the Prompt-based training method requires manual design of prompts to construct training samples. Both methods are overly dependent on manual operations, resulting in low efficiency in obtaining training samples.

[0033] Therefore, in order to improve the efficiency of obtaining training samples, the embodiments of the present disclosure propose a large model adjustment method, a question and answer processing method and device based on the large model technology.

[0034] The following describes the large model adjustment method, the question and answer processing method and device based on the large model technology in the embodiments of the present disclosure with reference to the accompanying drawings.

[0035] Among them, it should be noted that the executor of the large model training sample acquisition method of this embodiment is the large model training sample acquisition device, and the executor of the large model adjustment method is the large model adjustment device. The large model training sample acquisition device or the large model adjustment device can be implemented by software and / or hardware, and can be configured in an electronic device, which may include but is not limited to a terminal, a server, etc.

[0036] Figure 1 It is a flowchart of a method for obtaining training samples of a large model provided according to an embodiment of the present disclosure.

[0037] like Figure 1 As shown, the training sample acquisition method of the large model includes:

[0038] Step 101, based on the description information of the target tool, a first large model is used to generate reference questions, and the reference questions need to call the target tool to generate corresponding replies in the question-answering dialogue.

[0039] The target tool may refer to any tool that can be called by the big model, such as a weather query tool, a search tool, etc. The description information of the target tool may refer to information such as a tool identifier, tool function description information, and tool parameter description information. It should be noted that the tool identifier may refer to the tool name.

[0040] The reference questions generated by the first model may refer to questions that require the target tool to be called when generating corresponding responses to the reference questions. It should be noted that the reference questions may refer to questions based on natural language expressions, and the number of reference questions generated by the first model may be one or more.

[0041] For example, if the target tool is a weather query tool, the reference questions can be "Tell me the weather in city A tomorrow", "I plan to travel to city B next week, can you help me check the weather in city B for the next ten days?" and "I want to know the weather in city C for the next three days, do you have any suggestions?"

[0042] Step 102: Generate a tool call request for a target tool using the second largest model according to the reference question.

[0043] The second largest model and the first largest model may refer to the same large model or different large models.

[0044] The tool call request may include the tool identifier and tool input parameters of the target tool. For example, if the target tool is a weather query tool and the reference question is "Tell me the weather of city A tomorrow", the tool call request is [tool identifier: get_city_weather, tool input parameters: {"city / city": A, "date range / date_range": 1}].

[0045] Step 103, when the tool call request successfully calls the target tool, generate training samples for training the third largest model according to the tool call request and the corresponding reference question.

[0046] The third largest model, the first largest model, and the second largest model may refer to the same large model or different large models. It should be noted that if the third largest model, the first largest model, and the second largest model refer to the same large model, the large model is a pre-trained model, and after obtaining the training samples, the large model can be fine-tuned based on the training samples.

[0047] The training samples include tool call requests and corresponding reference questions.

[0048] It should be noted that if the tool call request fails to successfully call the target tool, it means that the generated tool call request may be inaccurate, and inaccurate training samples will affect the training effect of the large model. Based on this, in order to ensure the training effect of the large model, the present disclosure generates training samples based on the tool call request that successfully calls the target tool and the corresponding reference questions.

[0049] In the disclosed embodiment, according to the description information of the target tool, the first large model is used to generate reference questions. The reference questions need to call the target tool to generate corresponding replies in the question-answering dialogue; according to the reference questions, the second large model is used to generate a tool call request for the target tool; when the tool call request successfully calls the target tool, a training sample for training the third large model is generated according to the tool call request and the corresponding reference question. The disclosed embodiment adopts the first large model to automatically generate reference questions according to the description information of the target tool, adopts the second large model to automatically generate tool call requests according to the reference questions, and then automatically generates training samples according to the tool call requests and the corresponding reference questions. The whole process does not require manual intervention, which improves the efficiency of obtaining training samples.

[0050] Figure 2 It is a flowchart of a method for obtaining training samples for a large model provided in accordance with another embodiment of the present disclosure.

[0051] like Figure 2 As shown, the training sample acquisition method of the large model includes:

[0052] Step 201, based on the description information of the target tool, a reference question is generated using the first large model.

[0053] Step 202: Obtain a first prompt template, wherein the first prompt template is used to prompt the second largest model to generate a tool call request based on the question.

[0054] The first prompt template includes instructions for generating a tool call request based on a task, a question example, and a tool call request example.

[0055] Exemplarily, the first prompt template may be:

[0056] In all subsequent conversations, you can use external tools to answer questions. If you need to use external tools, you must generate a tool call request in the following format. The tool call request contains only 2 lines and does not need to output any explanation or other irrelevant content:

[0057] Tool ID / Action: The name of the tool used

[0058] Tool input parameters / Action Input: Parameters used by the tool, in json format

[0059] If you do not need to use external tools, you do not need to output Action and Action Input, just output your answer directly.

[0060] Here is an example:

[0061] Question: Tell me the weather in city A tomorrow.

[0062] Tool call request: Action: get_city_weather

[0063] Action Input: {"city": "A", "date_range": 1}

[0064] In addition to prompting the second largest model to generate a tool call request based on the question, the first prompt template can also be used to prompt the second largest model to generate a reply to the question based on the tool processing result.

[0065] Exemplarily, the first prompt template may be:

[0066] In all subsequent conversations, you can use external tools to answer questions. If you need to use external tools, you must generate a tool call request in the following format. The tool call request contains only 2 lines and does not need to output any explanation or other irrelevant content:

[0067] Tool ID / Action: The name of the tool used

[0068] Tool input parameters / Action Input: Parameters used by the tool, in json format

[0069] In addition, when you obtain the tool processing results based on the tool call request, you can determine whether you need to further use other tools based on the tool processing results, or give your response based on the tool processing results.

[0070] If you do not need to use external tools, you do not need to output Action and Action Input, just output your answer directly.

[0071] Here is an example:

[0072] Question: Tell me the weather in city A tomorrow.

[0073] Tool call request: Action: get_city_weather

[0074] Action Input: {"city": A, "date_range": 1}

[0075] Tool processing results: "date": 2023-10-11, "weather": cloudy, "high": 26, "low": 17

[0076] Reply: The weather in city A tomorrow will be cloudy, with the highest temperature at 26℃ and the lowest temperature at 17℃.

[0077] Step 203: Fill the reference question into the first prompt template to obtain the first prompt word.

[0078] Filling the reference question into the first prompt template may refer to replacing the question example in the first prompt template with the reference question to obtain the first prompt word, or may refer to filling the reference question into the set question slot in the first prompt template to obtain the first prompt word.

[0079] Step 204: input the first prompt word into the second largest model to obtain a tool call request of the target tool output by the second largest model.

[0080] The first prompt word is input into the second largest model so that the second largest model outputs a corresponding tool call request for the reference question.

[0081] Step 205 , when the tool call request successfully calls the target tool, a training sample for training the third largest model is generated according to the tool call request and the corresponding reference question.

[0082] Among them, when the first prompt template is also used to prompt the second largest model to generate a response to the question based on the tool processing result, the response output by the second largest model to the reference question can be obtained; based on the reference question, the corresponding tool call request and the response, a training sample is generated.

[0083] It should be noted that the explanation of step 201 and step 205 can be found in the relevant description in any embodiment of the present disclosure, and will not be repeated here.

[0084] In the embodiment of the present disclosure, a first prompt template is obtained, wherein the first prompt template is used to prompt the second largest model to generate a tool call request based on a question; a reference question is filled into the first prompt template to obtain a first prompt word; the first prompt word is input into the second largest model to obtain a tool call request for the target tool output by the second largest model. Through the first prompt template, the first prompt word corresponding to the reference question can be quickly generated, thereby improving the efficiency of obtaining the prompt word; in addition, the versatility and extensibility of the first prompt template enable the present disclosure to adapt to a wider range of question types; in addition, based on the guidance of the first prompt word, the second largest model can generate a tool call request corresponding to the reference question more quickly, which not only improves the efficiency of obtaining training samples, but also avoids human intervention and reduces the error rate in the process of generating training samples.

[0085] Figure 3 It is a flowchart of a method for obtaining training samples for a large model provided in accordance with another embodiment of the present disclosure.

[0086] like Figure 3 As shown, the training sample acquisition method of the large model includes:

[0087] Step 301, based on the description information of the target tool, a reference question is generated using the first large model.

[0088] Step 302: Generate a tool call request for a target tool using the second largest model according to the reference question.

[0089] Step 303, when the tool call request successfully calls the target tool, generate training samples for training the third largest model according to the tool call request and the corresponding reference question.

[0090] Step 304, when there are multiple target tools, the fourth model is used to combine the reference questions corresponding to the target tools to obtain a compound question, wherein the compound question requires calling at least two target tools to generate corresponding replies in the question-and-answer dialogue.

[0091] Among them, the fourth largest model, the first largest model, the second largest model, and the third largest model may refer to the same large model or different large models.

[0092] In one possible implementation, the fourth model can be used to combine questions based on the tool category corresponding to the target tool and reference questions, so that the combined compound question needs to call at least two target tools to generate corresponding replies in the question-and-answer dialogue.

[0093] It should be noted that the present disclosure only describes the situation where a compound question in a question-and-answer dialogue requires calling at least two target tools to generate corresponding replies. In actual application, the reference questions can also be randomly combined. In this case, the compound question obtained needs to call one or at least two target tools to generate corresponding replies in the question-and-answer dialogue.

[0094] In a possible implementation, the reference questions corresponding to the target tools can be combined to obtain a composite question in the following manner: obtaining a second prompt template, wherein the second prompt template is used to prompt the fourth largest model to combine the questions; filling the reference questions corresponding to the target tools into the second prompt template to obtain a second prompt word; inputting the second prompt word into the fourth largest model to obtain the composite question output by the fourth largest model.

[0095] The second prompt template may include instructions for combining questions to obtain a compound question, question examples, and compound question examples.

[0096] Exemplarily, the second prompt template may be:

[0097] Your task is to combine the five simple questions to get two compound questions. The combination rules are as follows:

[0098] 1. The input is a json (object notation) array containing N strings, representing N simple questions.

[0099] 2. The output is also a json array, containing 2 objects, representing 2 compound questions. The object has 3 fields: source1, source2, result. Source1 and source2 need to be obtained from the input array, and source1 and source2 are combined to generate result.

[0100] 3. You can add details / background as appropriate, and the combined compound question needs to be as coherent as possible.

[0101] Here is an example:

[0102] Example 1:

[0103] Input array: ["Tell me what are the most popular movies in 2021", "Please provide an abstract of a paper on machine learning", "Can you help me find the recent news in Italy?", "What is the relative strength index of the stock symbol of company D from January 1, 2022 to December 31, 2022?", "What are the recent football games?"]

[0104] Output array: [{"source1":"What is the relative strength index of D company's stock code from January 1, 2022 to December 31, 2022?","source2":"Please provide an abstract of a paper on machine learning","result":"I am studying stocks and machine learning. Please tell me the relative strength index of D company's stock code from January 1, 2022 to December 31, 2022, and then give me an abstract of a paper on machine learning that processes stock data."},{"source1":"Tell me what are the most popular movies in 2021","source2":"Can you help me find the recent news in Italy?","result":"I want to know what are the most popular movies in 2021. In addition, please help me find the recent news in Italy"}]

[0105] Example 2:

[0106] Input array: ["Tell me the weather in city A tomorrow.", "How is the air quality in city A?", "I am going to city B today and want to know what the weather in city B will be like in the next few days.", "Is there any news about city B recently?", "Please tell me what the weather in city C will be like in the next week?"]

[0107] Output array: [{"source1":"I am going to city B today, and would like to know what the weather will be like in the next few days.","source2":"Is there any news about city B recently?","result":"I am going to city B today. Please tell me the weather forecast for city B in the next few days, and provide some recent news about city B so that I can better understand the local situation."}, {"source1":"How is the air quality in city A?","source2":"Please tell me what the weather will be like in city C in the next week?","result":"I would like to know the air quality in city A, and please provide the weather forecast for city C in the next week so that I can make relevant arrangements."}].

[0108] It should be noted that, in the above example, there are two weather query tools corresponding to City A and City C.

[0109] The second prompt template contains instructions or examples to guide the fourth model to understand how to combine multiple independent questions into a more complex question. By filling the reference questions corresponding to each target tool into the second prompt template, the fourth model can receive a more complete and structured input, namely the second prompt word, which helps the fourth model generate more reasonable and accurate compound questions. In addition, the present disclosure automatically completes the question combination task by the large model, which reduces the manual burden and improves the efficiency of question combination.

[0110] In a possible implementation, candidate reference questions for the target tool corresponding to the target tool that is successfully called by the tool call request are screened from the reference questions corresponding to the target tools; the candidate reference questions are filled into the second prompt template to obtain a second prompt word.

[0111] Since the tool call request generated by the second largest model is not necessarily accurate, the tool call based on the tool call request will most likely fail. Therefore, in order to ensure the quality of training samples, improve the training effect of the large model, and reduce the occurrence rate of large model call failure, the present invention first screens the reference questions and the corresponding tool call requests, and retains the tool call requests that successfully make tool calls and the corresponding reference questions for question combination.

[0112] Step 305: Generate tool call requests for at least two target tools using the fifth model according to the complex problem.

[0113] Among them, the fifth largest model, the first largest model, the second largest model, the third largest model, and the fourth largest model may refer to the same large model or different large models.

[0114] Among them, the compound question can be filled into the first prompt template to obtain the fourth prompt word, and the fourth prompt word can be input into the fifth largest model to obtain the tool call request of at least two target tools output by the fifth largest model.

[0115] Step 306, when the tool call request successfully calls at least two target tools, generate training samples for training the sixth model according to the tool call request and the corresponding composite question.

[0116] Among them, the sixth largest model, the first largest model, the second largest model, the third largest model, the fourth largest model, and the fifth largest model may refer to the same large model or different large models.

[0117] It should be noted that after large-scale question combinations, there will be a certain number of compound questions in the generated training samples, corresponding to at least two target tools that have a certain order of calling. Based on these compound questions, large-model training can enable the large model to learn and generalize the ability to autonomously decompose problems and autonomously plan the order of tool calling, thereby improving the large model's multi-tool calling capabilities, enhancing the large model's flexibility and adaptability, and enabling the large model to perform more complex dialogue tasks.

[0118] It should be noted that the explanation of steps 301 to 303 can be found in the relevant description in any embodiment of the present disclosure and will not be repeated here.

[0119] In the embodiment of the present disclosure, when there are multiple target tools, the fourth largest model is used to combine the reference questions corresponding to each target tool to obtain a compound question, wherein the compound question requires calling at least two target tools to generate a corresponding reply in a question-and-answer dialogue; according to the compound question, the fifth largest model is used to generate a tool call request for at least two target tools; when the tool call request successfully calls at least two target tools, a training sample for training the sixth largest model is generated according to the tool call request and the corresponding compound question. Compound questions can be used to improve the multi-tool calling capability of the large model in a single-round dialogue. The present disclosure can automatically generate compound questions that require calling at least two target tools to generate corresponding replies in a question-and-answer dialogue based on the large model, which is highly efficient and enriches the training samples of the large model.

[0120] Figure 4 It is a flowchart of a method for obtaining training samples for a large model provided in accordance with another embodiment of the present disclosure.

[0121] like Figure 4 As shown, the training sample acquisition method of the large model includes:

[0122] Step 401, obtaining initial information obtained by collecting information on a target tool.

[0123] The initial information may refer to the most original information obtained by collecting information about the target tool, and may include tool identification, initial function description information, tool input parameters, tool calling uniform resource locator (URL) and calling method.

[0124] It should be noted that the initial information directly collected corresponds to a language category. For example, the language category corresponding to the initial information may be Chinese, English, or the like.

[0125] Step 402: When the language category of the initial information is not the target language category, the initial information is converted into a language based on the target language category to obtain description information of the target tool.

[0126] The target language category is related to the main language environment of the large model. For example, if the main language environment of the large model is Chinese, the target language category is the Chinese category.

[0127] The language category of the initial information is different from the main language environment of the large model, and it is not easy for the large model to understand the information. Therefore, when the language category of the initial information is not the target language category, language conversion of the initial information can enable the first large model to understand and process the information more accurately.

[0128] In one possible implementation, based on the target language category, the initial information is converted into a language to obtain target information, which includes initial functional description information of the target tool; the initial functional description information is expanded to obtain target functional description information; and based on the target functional description information, the description information of the target tool is obtained.

[0129] Usually, the initial functional description information collected is relatively short and not easy for the model to understand. Therefore, expanding the initial functional description information can not only help the large model to capture the key points of information more accurately and improve the understanding and processing capabilities of the large model, but also guide the large model to generate more specific and targeted reference questions.

[0130] Step 403, obtaining a third prompt template, wherein the third prompt template is used to prompt the first large model to generate questions that require calling the tool to generate corresponding replies in the question-and-answer dialogue based on the description information of the tool.

[0131] The third prompt template may include instructions for generating relevant questions based on the description information of the tool, description information examples and question examples.

[0132] Exemplarily, the third prompt template may be:

[0133] Your task is to generate 5 relevant requests described in natural language for the given interface, according to the following rules:

[0134] 1. The description information of a tool consists of a tool identifier, function description information, and tool input parameters. The tool input parameters are a JSON object that conforms to the JSON schema format.

[0135] 2. The generated questions need to contain parameters described in natural language as much as possible.

[0136] 3. The generated questions must be in natural language format, while maintaining diversity and complexity, and be as colloquial and close to real humans as possible.

[0137] 4. When a tool input parameter requires a field like an identity document (ID), include the randomly generated id in the generated question.

[0138] 5. The generated request is a JSON array with 5 strings corresponding to 5 questions. Please generate the JSON array directly and do not generate irrelevant explanations or other information.

[0139] Here is an example:

[0140] Example 1:

[0141] Description of the tool:

[0142] Tool ID: baidu_search

[0143] Function description information: Baidu Search is a general search engine that can be used to access the Internet, query encyclopedia knowledge, understand what is happening now, etc.

[0144] Tool input parameters: {{"type":"object","properties":{{"json":{{"properties":{{"query":{{"type":"string","description":"Search for a keyword or phrase."}}}},"type":"object"}}}}}}

[0145] Generated questions: ["Introduce some must-read books for beginners in economics", "Where is the highest peak in the world?", "Compare the two movies "The Wandering Earth 2" and "Avatar 2", "What news is there in City A and City B today?", "What news is there about the 2023 Asian Games? Please write a news report in the style of Application E"]

[0146] Example 2:

[0147] Description of the tool:

[0148] Tool ID: get_city_weather

[0149] Function description: Query the weather in a specified city.

[0150] Tool input parameters: {"type":"object","properties":{"city":{"type":"string","description":"city name"},"date_range":{"type":"integer","description":"date range, hope to query the weather for the next N days"}},"required":["city"]}

[0151] Generated questions: ["Tell me the weather forecast for city A tomorrow.", "I plan to travel to city B next week. Can you help me check the weather in city B for the next ten days?", "I am going to city D today and want to know what the weather in city D will be like in the next few days.", "Please tell me what the weather in city F will be like in the next week?", "I want to know the weather in city H for the next three days. Do you have any suggestions?"]

[0152] Step 404: Fill the description information of the target tool into the third prompt template to obtain a third prompt word.

[0153] Filling the description information of the target tool into the third prompt template may refer to the related description of filling the reference question into the first prompt template in the above embodiment, which will not be repeated here.

[0154] Step 405: input the third prompt word into the first large model to obtain a reference question generated by the first large model.

[0155] Among them, after receiving the third prompt word, the first large model can parse and process the third prompt word to output a reference question corresponding to the target tool.

[0156] Step 406: Generate a tool call request for the target tool using the second largest model according to the reference question.

[0157] Step 407, when the tool call request successfully calls the target tool, generate training samples for training the third largest model according to the tool call request and the corresponding reference question.

[0158] It should be noted that the explanation of step 406 and step 407 can be found in the relevant description of any embodiment of the present disclosure, and will not be repeated here.

[0159] In the disclosed embodiment, initial information obtained by collecting information on the target tool is obtained; when the language category of the initial information is not the target language category, based on the target language category, the initial information is converted into a language to obtain the description information of the target tool; a third prompt template is obtained, and the description information of the target tool is filled into the third prompt template to obtain a third prompt word; the third prompt word is input into the first large model to obtain a reference question generated by the first large model. When the language category of the initial information is not the target language category, the initial information is converted into a language, so that the first large model can understand and process the information more accurately. In addition, based on the third prompt template, a large-scale reference question can be automatically generated for the target tool, and then a large-scale training sample can be obtained, so that the large model trained based on the training sample can learn and generalize the ability to adapt to various problems, thereby enhancing the flexibility and adaptability of the large model.

[0160] Figure 5 It is a flowchart of a large model adjustment method provided according to an embodiment of the present disclosure.

[0161] like Figure 5 As shown, the large model adjustment method includes:

[0162] Step 501, obtain training samples of the large model.

[0163] The training samples are obtained based on the training sample acquisition method of the large model in the above embodiment.

[0164] Step 502: Adjust model parameters of the large model based on the training samples.

[0165] In the disclosed embodiment, the disclosed method automatically generates training samples based on the description information of the target tool, which is highly efficient, and the training samples include reference questions / compound questions and corresponding tool call requests, and the questions are rich. Therefore, the trained large model can accurately extract and convert key information from the questions, thereby improving the accuracy of the generated responses. In addition, the traditional large model integration method may require writing a large amount of code to adapt to the interfaces and protocols of different tools, while the disclosed method provides standardized tool description information, and the trained large model can directly call these tools without writing additional adaptation code, which greatly simplifies the integration process and reduces the difficulty of integration.

[0166] Figure 6 It is a flowchart of a question-answering processing method based on a large model technology provided according to an embodiment of the present disclosure.

[0167] like Figure 6 As shown, the question-answering processing method based on the large model technology includes:

[0168] Step 601, obtaining questions to be answered.

[0169] The question to be answered may refer to a question input by a user.

[0170] Step 602: Input the question into the big model and obtain the response output by the big model to the question.

[0171] The large model is obtained based on the large model adjustment method in the above embodiment.

[0172] In the disclosed embodiment, since the training samples include reference questions / compound questions and the questions are rich, the large model obtained based on the large model adjustment method in the above embodiment can adapt to various complex questions. In addition, the large model obtained based on the large model adjustment method in the above embodiment can accurately extract and convert key information from the question, and the generated response has high accuracy.

[0173] Figure 7 It is a structural diagram of a training sample acquisition device for a large model provided according to an embodiment of the present disclosure.

[0174] like Figure 7 As shown, the training sample acquisition device of the large model includes:

[0175] The first generation module 701 is used to generate reference questions using the first large model according to the description information of the target tool. The reference questions need to call the target tool to generate corresponding replies in the question-answering dialogue;

[0176] A second generating module 702 is used to generate a tool call request of a target tool using a second large model according to a reference problem;

[0177] The third generating module 703 is used to generate training samples for training the third largest model according to the tool calling request and the corresponding reference questions when the tool calling request successfully calls the target tool.

[0178] As a possible implementation of the embodiment of the present disclosure, the second generating module 702 is used to:

[0179] Obtaining a first prompt template, wherein the first prompt template is used to prompt the second largest model to generate a tool call request based on the question;

[0180] Fill the reference question into the first prompt template to obtain the first prompt word;

[0181] The first prompt word is input into the second largest model to obtain a tool calling request of the target tool output by the second largest model.

[0182] As a possible implementation of the embodiment of the present disclosure, the device further includes:

[0183] A combination module is used to combine the reference questions corresponding to the target tools using the fourth model to obtain a compound question when there are multiple target tools, wherein the compound question needs to call at least two target tools to generate corresponding replies in the question-answering dialogue;

[0184] a fourth generation module, for generating tool call requests for at least two target tools using the fifth model according to the composite problem;

[0185] The fifth generation module is used to generate training samples for training the sixth model according to the tool call request and the corresponding composite question when the tool call request successfully calls at least two target tools.

[0186] As a possible implementation of the embodiment of the present disclosure, a combined module is used to:

[0187] Obtaining a second prompt template, wherein the second prompt template is used to prompt the fourth model to combine questions;

[0188] Fill the reference questions corresponding to each target tool into the second prompt template to obtain a second prompt word;

[0189] The second prompt word is input into the fourth model to obtain the compound question output by the fourth model.

[0190] As a possible implementation of the embodiment of the present disclosure, a combined module is used to:

[0191] From the reference questions corresponding to each target tool, select candidate reference questions for which the tool call request successfully calls the corresponding target tool;

[0192] Fill the candidate reference question into the second prompt template to obtain the second prompt word.

[0193] As a possible implementation of the embodiment of the present disclosure, the first generating module 701 is used to:

[0194] Obtaining a third prompt template, wherein the third prompt template is used to prompt the first large model to generate questions that require calling the tool to generate corresponding responses in the question-and-answer dialogue based on the description information of the tool;

[0195] Filling the description information of the target tool into the third prompt template to obtain a third prompt word;

[0196] The third prompt word is input into the first large model to obtain a reference question generated by the first large model.

[0197] As a possible implementation of the embodiment of the present disclosure, the device further includes:

[0198] An information acquisition module is used to acquire initial information obtained by collecting information from a target tool;

[0199] The conversion module is used to convert the language of the initial information based on the target language category to obtain the description information of the target tool when the language category of the initial information is not the target language category.

[0200] As a possible implementation of the embodiment of the present disclosure, the conversion module is used to

[0201] Based on the target language category, the initial information is converted into a language to obtain the target information, which includes the initial functional description information of the target tool;

[0202] The initial function description information is expanded to obtain target function description information, and based on the target function description information, description information of the target tool is obtained.

[0203] It should be noted that the aforementioned explanation of the method for obtaining training samples for a large model is also applicable to the device for obtaining training samples for a large model of this embodiment, and will not be repeated here.

[0204] Figure 8 It is a structural schematic diagram of a large model adjustment device provided according to an embodiment of the present disclosure.

[0205] like Figure 8 As shown, the large model adjustment device includes:

[0206] A sample acquisition module 801 is used to acquire a training sample of a large model, where the training sample is obtained based on the large model training sample acquisition method described in the first aspect;

[0207] The adjustment module 802 is used to adjust the model parameters of the large model based on the training samples.

[0208] It should be noted that the aforementioned explanation of the large model adjustment method is also applicable to the large model adjustment device of this embodiment and will not be repeated here.

[0209] Fig. 9 It is a structural diagram of a question and answer processing device based on large model technology provided according to an embodiment of the present disclosure.

[0210] like Fig. 9 As shown, the question-answering processing device based on the large model technology includes:

[0211] The question acquisition module 901 is used to acquire questions to be answered;

[0212] The reply module 902 is used to input the question into the big model to obtain the reply output by the big model to the question. The big model is obtained based on the big model adjustment method described in the second aspect.

[0213] It should be noted that the aforementioned explanation of the question and answer processing method based on the big model technology is also applicable to the question and answer processing device based on the big model technology of this embodiment, and will not be repeated here.

[0214] According to an embodiment of the present disclosure, the present disclosure also provides an electronic device, a readable storage medium and a computer program product.

[0215] Fig.10 A schematic block diagram of an example electronic device 1000 that can be used to implement an embodiment of the present disclosure is shown. The electronic device 1000 is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processing, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present disclosure described and / or required herein.

[0216] like Fig.10 As shown, the electronic device 1000 includes a computing unit 1001, which can perform various appropriate actions and processes according to a computer program stored in a ROM (Read-Only Memory) 1002 or a computer program loaded from a storage unit 1008 to a RAM (Random Access Memory) 1003. In the RAM 1003, various programs and data required for the operation of the electronic device 1000 can also be stored. The computing unit 1001, the ROM 1002, and the RAM 1003 are connected to each other via a bus 1004. An I / O (Input / Output) interface 1005 is also connected to the bus 1004.

[0217] Multiple components in the electronic device 1000 are connected to the I / O interface 1005, including: an input unit 1006, such as a keyboard, a mouse, etc.; an output unit 1007, such as various types of displays, speakers, etc.; a storage unit 1008, such as a disk, an optical disk, etc.; and a communication unit 1009, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 1009 allows the electronic device 1000 to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks.

[0218] The computing unit 1001 may be a variety of general and / or special processing components with processing and computing capabilities. Some examples of the computing unit 1001 include, but are not limited to, a CPU (Central Processing Unit), a GPU (Graphic Processing Units), various dedicated AI (Artificial Intelligence) computing chips, various computing units that run machine learning model algorithms, a DSP (Digital Signal Processor), and any appropriate processor, controller, microcontroller, etc. The computing unit 1001 performs the various methods and processes described above, such as a large model training sample acquisition method, a large model adjustment method, and a question-answering processing method based on a large model technology. For example, in some embodiments, a large model training sample acquisition method, a large model adjustment method, and a question-answering processing method based on a large model technology may be implemented as a computer software program, which is tangibly contained in a machine-readable medium, such as a storage unit 1008. In some embodiments, part or all of the computer program may be loaded and / or installed on the electronic device 1000 via the ROM 1002 and / or the communication unit 1009. When the computer program is loaded into RAM 1003 and executed by the computing unit 1001, one or more steps of the large model training sample acquisition method, large model adjustment method, and question-answering processing method based on large model technology described above may be executed. Alternatively, in other embodiments, the computing unit 1001 may be configured to execute the large model training sample acquisition method or large model adjustment method or question-answering processing method based on large model technology in any other appropriate manner (e.g., by means of firmware).

[0219] Various embodiments of the systems and techniques described above herein may be implemented in digital electronic circuit systems, integrated circuit systems, FPGAs (Field Programmable Gate Arrays), ASICs (Application-Specific Integrated Circuits), ASSPs (Application Specific Standard Products), SOCs (System On Chips), CPLDs (Complex Programmable Logic Devices), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include: being implemented in one or more computer programs that may be executed and / or interpreted on a programmable system including at least one programmable processor that may be a dedicated or general-purpose programmable processor that may receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0220] The program code for implementing the method of the present disclosure may be written in any combination of one or more programming languages. These program codes may be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device, so that the program code, when executed by the processor or controller, enables the functions / operations specified in the flow chart and / or block diagram to be implemented. The program code may be executed entirely on the machine, partially on the machine, partially on the machine and partially on a remote machine as a stand-alone software package, or entirely on a remote machine or server.

[0221] In the context of the present disclosure, a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, device, or equipment. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or device, or any suitable combination of the foregoing. More specific examples of machine-readable storage media may include electrical connections based on one or more lines, portable computer disks, hard disks, RAM, ROM, EPROM (Electrically Programmable Read-Only-Memory) or flash memory, optical fiber, CD-ROM (Compact Disc Read-Only Memory), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0222] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device (e.g., a CRT (Cathode-Ray Tube) or LCD (Liquid Crystal Display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user can provide input to the computer. Other types of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic input, voice input, or tactile input).

[0223] The systems and techniques described herein may be implemented in a computing system that includes backend components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes frontend components (e.g., a user computer with a graphical user interface or a web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such backend components, middleware components, or frontend components. The components of the system may be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: LAN (Local Area Network), WAN (Wide Area Network), the Internet, and blockchain networks.

[0224] A computer system may include a client and a server. The client and the server are generally remote from each other and usually interact through a communication network. The relationship between the client and the server is generated by computer programs running on the corresponding computers and having a client-server relationship with each other. The server may be a cloud server, also known as a cloud computing server or cloud host, which is a host product in the cloud computing service system to solve the defects of difficult management and weak business scalability in traditional physical hosts and VPS services ("Virtual Private Server", or "VPS" for short). The server may also be a server of a distributed system, or a server combined with a blockchain.

[0225] It should be noted that artificial intelligence is a discipline that studies how computers can simulate certain human thought processes and intelligent behaviors (such as learning, reasoning, thinking, planning, etc.), and includes both hardware-level and software-level technologies. Artificial intelligence hardware technologies generally include technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, and big data processing; artificial intelligence software technologies mainly include computer vision technology, speech recognition technology, natural language processing technology, as well as machine learning / deep learning, big data processing technology, knowledge graph technology, and other major directions.

[0226] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps recorded in this disclosure can be executed in parallel, sequentially or in different orders, as long as the desired results of the technical solutions disclosed in this disclosure can be achieved, and this document does not limit this.

[0227] The above specific implementations do not constitute a limitation on the protection scope of the present disclosure. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present disclosure shall be included in the protection scope of the present disclosure.

Claims

1. A method for obtaining training samples of a large model, comprising: Based on the description information of the target tool, a first model is used to generate reference questions, wherein the reference questions need to call the target tool to generate corresponding responses in the question-answering dialogue; generating a tool call request for the target tool using a second large model according to the reference problem; In the case where the tool calling request successfully calls the target tool, a training sample for training the third largest model is generated according to the tool calling request and the corresponding reference question.

2. The method according to claim 1, wherein: The step of generating a tool call request for the target tool using the second largest model according to the reference question includes: Obtain a first prompt template, wherein the first prompt template is used to prompt the second large model to generate a tool call request based on the question; Filling the reference question into the first prompt template to obtain a first prompt word; The first prompt word is input into the second large model to obtain a tool calling request of the target tool output by the second large model.

3. The method according to claim 1, wherein: The method further comprises: In the case where there are multiple target tools, the fourth model is used to combine the reference questions corresponding to the target tools to obtain a compound question, wherein the compound question needs to call at least two target tools to generate corresponding replies in the question-answering dialogue; generating tool call requests for the at least two target tools using a fifth model according to the composite problem; In a case where the tool calling request successfully calls the at least two target tools, a training sample for training the sixth largest model is generated according to the tool calling request and the corresponding composite problem.

4. The method according to claim 3, wherein: The fourth model is used to combine the reference questions corresponding to the target tools to obtain a composite question, including: Acquire a second prompt template, wherein the second prompt template is used to prompt the fourth model to combine questions; Filling the reference questions corresponding to the target tools into the second prompt template to obtain a second prompt word; The second prompt word is input into the fourth model to obtain the compound question output by the fourth model.

5. The method according to claim 4, wherein: The step of filling the reference questions corresponding to each of the target tools into the second prompt template to obtain a second prompt word includes: From the reference questions corresponding to each of the target tools, screening the candidate reference questions of the target tool corresponding to the successful invocation of the tool invocation request; Fill the candidate reference question into the second prompt template to obtain the second prompt word.

6. The method according to claim 1, wherein: The method of generating reference questions using the first model according to the description information of the target tool includes: Obtaining a third prompt template, wherein the third prompt template is used to prompt the first large model to generate a question that requires calling the tool to generate a corresponding answer in a question-and-answer dialogue based on the description information of the tool; Filling the description information of the target tool into the third prompt template to obtain a third prompt word; The third prompt word is input into the first large model to obtain the reference question generated by the first large model.

7. The method according to claim 1, wherein: Before generating the reference question using the first large model according to the description information of the target tool, the method further includes: Acquiring initial information obtained by collecting information from the target tool; When the language category of the initial information is not the target language category, the initial information is converted into a language based on the target language category to obtain description information of the target tool.

8. The method according to claim 7, wherein: The converting the initial information into a language based on the target language category to obtain the description information of the target tool includes: Based on the target language category, the initial information is converted into a language to obtain target information, wherein the target information includes initial function description information of the target tool; The initial function description information is expanded to obtain target function description information, and based on the target function description information, description information of the target tool is acquired.

9. A large model adjustment method, comprising: Obtaining a training sample of a large model, wherein the training sample is obtained based on the method described in any one of claims 1 to 8; Based on the training samples, model parameters of the large model are adjusted.

10. A question-answering processing method based on large model technology, comprising: Get questions waiting for answer; The question is input into a large model to obtain a response output by the large model to the question, wherein the large model is obtained based on the method described in claim 9.

11. A large model training sample acquisition device, comprising: A first generation module is used to generate reference questions using a first large model according to the description information of the target tool, wherein the reference questions need to call the target tool to generate corresponding replies in the question-answering dialogue; A second generating module, configured to generate a tool calling request for the target tool using a second large model according to the reference problem; The third generating module is used to generate training samples for training the third large model according to the tool calling request and the corresponding reference question when the tool calling request successfully calls the target tool.

12. The device according to claim 11, wherein The second generating module is used for: Obtain a first prompt template, wherein the first prompt template is used to prompt the second large model to generate a tool call request based on the question; Filling the reference question into the first prompt template to obtain a first prompt word; The first prompt word is input into the second large model to obtain a tool calling request of the target tool output by the second large model.

13. The device according to claim 11, wherein: The device also includes: A combination module, for combining the reference questions corresponding to the target tools using the fourth model to obtain a compound question when there are multiple target tools, wherein the compound question needs to call at least two target tools to generate corresponding replies in the question-answering dialogue; a fourth generating module, configured to generate tool call requests for the at least two target tools using the fifth model according to the composite problem; The fifth generating module is used to generate training samples for training the sixth model according to the tool calling request and the corresponding compound problem when the tool calling request successfully calls the at least two target tools.

14. The device according to claim 13, wherein: The combined module is used for: Acquire a second prompt template, wherein the second prompt template is used to prompt the fourth model to combine questions; Filling the reference questions corresponding to the target tools into the second prompt template to obtain a second prompt word; The second prompt word is input into the fourth model to obtain the compound question output by the fourth model.

15. The device according to claim 14, wherein: The combined module is used for: From the reference questions corresponding to each of the target tools, screening the candidate reference questions of the target tool corresponding to the successful invocation of the tool invocation request; Fill the candidate reference question into the second prompt template to obtain the second prompt word.

16. The device according to claim 11, wherein The first generating module is used for: Obtaining a third prompt template, wherein the third prompt template is used to prompt the first large model to generate a question that requires calling the tool to generate a corresponding answer in a question-and-answer dialogue based on the description information of the tool; Filling the description information of the target tool into the third prompt template to obtain a third prompt word; The third prompt word is input into the first large model to obtain the reference question generated by the first large model.

17. The device according to claim 11, wherein: The device also includes: An information acquisition module, used to acquire initial information obtained by collecting information from the target tool; The conversion module is used to perform language conversion on the initial information based on the target language category to obtain description information of the target tool when the language category of the initial information is not the target language category.

18. The device according to claim 17, wherein: The conversion module is used to perform language conversion on the initial information based on the target language category to obtain target information, wherein the target information includes initial function description information of the target tool; The initial function description information is expanded to obtain target function description information, and based on the target function description information, description information of the target tool is acquired.

19. A large model adjustment device, comprising: A sample acquisition module, used to acquire training samples of a large model, wherein the training samples are obtained based on the method described in any one of claims 1 to 8; An adjustment module is used to adjust model parameters of the large model based on the training samples.

20. A question-answering processing device based on large model technology, comprising: A question acquisition module is used to obtain questions to be answered; A reply module is used to input the question into a large model to obtain a reply output by the large model to the question, wherein the large model is obtained based on the method described in claim 9.

21. An electronic device, comprising: at least one processor; as well as a memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the method of any one of claims 1 to 8, or the method of claim 9, or the method of claim 10.

22. A non-transitory computer-readable storage medium storing computer instructions, wherein: The computer instructions are used to enable the computer to execute the method of any one of claims 1 to 8, or the method of claim 9, or the method of claim 10.

23. A computer program product, comprising a computer program, wherein when the computer program is executed by a processor, the computer program implements the method of any one of claims 1 to 8, or the method of claim 9, or the method of claim 10.

Images