Model training data construction method, device, equipment and computer program product

Abstract

The invention provides a model training data construction method, apparatus and device, and a computer program product, and relates to the technical field of computers and Internet. The method comprises the steps that a first large language model obtains a business analysis task construction request and business analysis task construction prompt information; the first large language model constructs a first business analysis task according to the business analysis task construction request and the business analysis task construction prompt information, wherein the first business analysis task comprises a plurality of to-be-analyzed data dimensions; the first large language model obtains data construction prompt information; the first large language model obtains synthetic business data under multiple to-be-analyzed data dimensions according to the data construction prompt information; and the first large language model performs service analysis on the synthetic service data according to the first service analysis task to obtain a service analysis result, so that the first service analysis task and the service analysis result are used as training data of model training. According to the embodiment of the invention, the model training data can be automatically and efficiently constructed.

Description

Technical Field

[0001] The present disclosure relates to the field of computer and Internet technology, and in particular to a method, device, electronic device, computer-readable storage medium, and computer program product for constructing model training data. Background Art

[0002] This section is intended to provide a background or context to the embodiments of the present disclosure that are recited in the claims. No statement herein is admitted to be prior art by virtue of its inclusion in this section.

[0003] A standard AI (Artificial Intelligence) model development process includes data collection and construction, model training, model testing, and iteration. For certain analytical scenarios, data collection involves both internal and external sources. Internal data can be directly queried from internal systems, while external data, including but not limited to open source data and financial reports, requires web searches.

[0004] However, whether it is internal data or external data, data collection is limited by the size of the data, the original data storage format, the data collection tools, and the time required for data collection, resulting in high data collection costs. Summary of the Invention

[0005] The purpose of the present disclosure is to provide a method, device, electronic device, computer-readable storage medium and computer program product for constructing model training data, which can quickly and efficiently construct model training data.

[0006] Other features and advantages of the present disclosure will become apparent from the following detailed description, or may be learned in part by practice of the present disclosure.

[0007] The present disclosure provides a method for constructing model training data, comprising: obtaining, by a first large language model, a business analysis task construction request and business analysis task construction prompt information; constructing, by the first large language model, a first business analysis task according to the business analysis task construction request and the business analysis task construction prompt information, wherein the first business analysis task includes multiple data dimensions to be analyzed; and obtaining, by the first large language model, data construction prompt information;

[0008] The first large language model obtains the synthetic business data under the multiple data dimensions to be analyzed according to the data construction prompt information; the first large language model performs business analysis on the synthetic business data according to the first business analysis task to obtain a business analysis result, so as to use the first business analysis task and the business analysis result as training data for model training.

[0009] In some embodiments, the data construction prompt information comprises: legal data dimension information, legality judgment prompt information, and prior distribution generation prompt information; wherein obtaining the synthesized business data under the plurality of to-be-analyzed data dimensions according to the data construction prompt information comprises: determining whether each data dimension in the plurality of to-be-analyzed data dimensions is legal according to the legal data dimension information in the data construction prompt information; in the case where it is determined that the plurality of to-be-analyzed data dimensions are legal, designing a prior distribution for each to-be-analyzed data dimension according to the prior distribution generation prompt information; and constructing synthesized business data for each to-be-analyzed data dimension according to the prior distribution corresponding to each to-be-analyzed data dimension.

[0010] In some embodiments, the first business analysis task comprises a legal data dimension, which is used to describe a legal data dimension in the first business analysis task; wherein the first large language model obtains data construction prompt information, which comprises: the first large language model obtains to-be-filled data construction prompt information; the first large language model fills the legal data dimension in the first business analysis task into the to-be-filled data construction prompt information to generate the data construction prompt information, so as to determine whether each data dimension in the plurality of to-be-analyzed data dimensions is legal according to the legal data dimension.

[0011] In some embodiments, the first business analysis task comprises a first analysis level and a second analysis level, and the plurality of to-be-analyzed data dimensions comprises a first data dimension under the first analysis level and a second data dimension under the second analysis level; wherein obtaining the synthesized business data under the plurality of to-be-analyzed data dimensions according to the data construction prompt information comprises: constructing synthesized business data under the first data dimension in combination with the data construction prompt information; wherein the first large language model performs business analysis on the synthesized business data according to the first business analysis task to obtain a business analysis result, which comprises: the first large language model performs business analysis on the synthesized business data under the first data dimension according to the first business analysis task to obtain a first analysis result in the business analysis result; determining that a target data range under the second data dimension has analysis value according to the first analysis result; wherein constructing the synthesized business data under the plurality of to-be-analyzed data dimensions in combination with the data construction prompt information further comprises: constructing synthesized business data within the target data range under the second data dimension in combination with the data construction prompt information, so that the first large language model continues to perform business analysis on the synthesized business data within the target data range under the second data dimension according to the first business analysis task.

[0012] In some embodiments, the first large language model obtains the synthetic business data under the plurality of data dimensions to be analyzed according to the data construction prompt information, including: the first large language model constructs the synthetic business data under the plurality of data dimensions to be analyzed in combination with the data construction prompt information; or the first large language model sends the data construction prompt information and dimension information corresponding to the plurality of data dimensions to be analyzed to a second large language model, so that the second large language model constructs the synthetic business data under the plurality of data dimensions to be analyzed in combination with the data construction prompt information; and the first large language model receives the synthetic business data under the plurality of data dimensions to be analyzed from the second large language model.

[0013] In some embodiments, the second large language model includes a data query large model and a data search large model, and the synthetic business data includes search business data and query business data; wherein the second large language model constructs the synthetic business data under the plurality of data dimensions to be analyzed in combination with the data construction prompt information, including: the data search large model constructs the search business data in combination with the data construction prompt information; the data query large model constructs the query business data in combination with the data construction prompt information; wherein the data query large model and the data search large model share a prior distribution when constructing the business data of the first business analysis task, so that the search business data and the query business data have consistent data distribution.

[0014] The embodiments of the present disclosure provide a model training data construction device, including: a request obtaining module, a task construction module, a construction prompt information obtaining module, a data obtaining module, and an analysis module.

[0015] The request obtaining module is configured to obtain, by the first large language model, a business analysis task construction request and business analysis task construction prompt information; the task construction module is configured to construct, by the first large language model, a first business analysis task according to the business analysis task construction request and the business analysis task construction prompt information, wherein the first business analysis task includes a plurality of data dimensions to be analyzed; the construction prompt information obtaining module is configured to obtain, by the first large language model, data construction prompt information; the data obtaining module is configured to obtain, by the first large language model, synthetic business data under the plurality of data dimensions to be analyzed according to the data construction prompt information; and the analysis module is configured to perform, by the first large language model, business analysis on the synthetic business data according to the first business analysis task to obtain a business analysis result, so as to use the first business analysis task and the business analysis result as training data for model training.

[0016] The embodiment of the present disclosure provides an electronic device, which comprises a memory and a processor; the memory is used for storing computer program instructions; the processor invokes the computer program instructions stored in the memory, and is used for implementing the model training data construction method.

[0017] The embodiment of the present disclosure provides a computer readable storage medium, which stores computer program instructions, and implements the model training data construction method.

[0018] The embodiment of the present disclosure provides a computer program product or a computer program, which comprises computer program instructions stored in a computer readable storage medium. The computer program instructions are read from the computer readable storage medium, and a processor executes the computer program instructions to implement the model training data construction method.

[0019] The model training data construction method, device, electronic device, computer readable storage medium and computer program product provided by the embodiment of the present disclosure can first construct a first business analysis task according to a business analysis task construction request, then obtain the synthetic business data corresponding to the first business analysis task according to data construction prompt information, and finally analyze the first business analysis task according to the synthetic business data to obtain an analysis result. The above scheme can automatically and efficiently generate a business analysis task and corresponding business analysis result data, and can solve the problem of insufficient or difficult-to-obtain real business data. On the other hand, the diversity of the business analysis task and the business analysis result can be expanded through the probabilistic random generation mode of the large model, and the business edge cases that are difficult to enumerate can be covered.

[0020] It should be understood that the above general description and the following detailed description are only exemplary and do not limit the present disclosure. BRIEF DESCRIPTION OF DRAWINGS

[0021] The accompanying drawings, which are incorporated into and form part of the specification, illustrate embodiments consistent with the present disclosure and, together with the specification, serve to explain the principles of the present disclosure. Obviously, the drawings in the following description are only some embodiments of the present disclosure, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0022] Figure 1 A scene schematic diagram of the model training data construction method or the model training data construction device that can be applied to the embodiment of the present disclosure is shown.

[0023] Figure 2 is a flowchart of a model training data construction method according to an exemplary embodiment.

[0024] Figure 3 is a task expansion schematic diagram according to an example embodiment.

[0025] Figure 4 is an analysis task synthesis schematic diagram according to an example embodiment.

[0026] Figure 5 is a flow chart of a synthetic business data generation method according to an example embodiment.

[0027] Figure 6 is a flow chart of a data construction prompt information generation method according to an example embodiment.

[0028] Figure 7 is a flow chart of a synthetic business data construction method according to an example embodiment.

[0029] Figure 8 is a flow chart of a synthetic business data construction method according to an example embodiment.

[0030] Figure 9 is a block diagram of a model training data construction apparatus according to an example embodiment.

[0031] Figure 10 shows a structural schematic diagram of an electronic device suitable for implementing the embodiments of the present disclosure. DETAILED DESCRIPTION

[0032] Example embodiments now will be described more fully hereinafter with reference to the accompanying drawings. Example embodiments can, however, be implemented in many different forms and should not be construed as limited to the embodiments set forth herein; rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the concept of example embodiments to those skilled in the art. Like reference numerals refer to like elements throughout the figures, and thus description of the same will be omitted.

[0033] Those skilled in the art know that the embodiments of the present disclosure can be a system, apparatus, device, method or computer program product. Therefore, the present disclosure can be embodied in the form of a complete hardware, complete software (including firmware, resident software, microcode, etc.), or a combination of hardware and software.

[0034] The features, structures or characteristics described in the present disclosure can be incorporated in any suitable manner in one or more embodiments. In the following description, numerous specific details are provided to give a thorough understanding of embodiments of the present disclosure. One skilled in the relevant art will recognize, however, that the techniques of the present disclosure can be practiced without one or more of the specific details, or with other methods, components, devices, steps, etc. In other instances, well-known methods, devices, implementations, or operations are not shown or described in detail to avoid obscuring aspects of the present disclosure.

[0035] In the embodiments of the present disclosure, the term "module" or "unit" refers to a computer program or a part of a computer program with a predetermined function, and works together with other related parts to achieve a predetermined target, and can be implemented entirely or partially by using software, hardware (such as a processing circuit or a memory), or a combination thereof. Similarly, one processor (or multiple processors or memories) can be used to implement one or more modules or units. In addition, each module or unit can be a part of an integral module or unit that includes the functions of the module or unit.

[0036] The accompanying drawings are merely schematic illustrations of the present disclosure, in which the same reference numerals refer to the same or similar parts, and thus repeated description thereof will be omitted. Some of the block diagrams shown in the drawings do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in the form of software, or in one or more hardware modules or integrated circuits, or in different networks and / or processor devices and / or microcontroller devices.

[0037] The flowcharts shown in the drawings are merely exemplary illustrations, and do not necessarily include all contents and steps, nor are they necessarily executed in the order described. For example, some steps can be further divided, and some steps can be combined or partially combined, so the actual execution order can be changed according to the actual situation.

[0038] In the description of the present disclosure, unless otherwise specified, " / " means "or", for example, A / B can mean A or B. "And / or" in this document is merely a description of the association relationship between the associated objects, which means that there can be three relationships, for example, A and / or B, which means that there are three cases of A alone, A and B together, and B alone. In addition, "one or more" means one or more, and "multiple" means two or more. "First", "second", etc. do not limit the quantity and execution order, and "first", "second", etc. do not necessarily mean different; the terms "include", "comprise" and "have" are used to mean open inclusion and mean that in addition to the listed elements / components / etc. there can be other elements / components / etc.

[0039] In order to enable more clear understanding of the above-mentioned objects, features and advantages of the present application, the present application is further described in detail below in conjunction with the accompanying drawings and specific embodiments, it should be noted that the embodiments of the present disclosure and the features in the embodiments can be combined with each other without conflict.

[0040] It should be noted that in the technical solutions of the present disclosure, the collection, collection, updating, analysis, processing, use, transmission, storage and the like of user personal information are in line with the relevant legal regulations, are used for legal purposes, and do not violate public order and good customs. Necessary measures are taken for user personal information to prevent illegal access to user personal information data and maintain user personal information security and network security.

[0041] In the related art, the data required by the business analysis scene can be taken from the underlying data table based on SQL (Structured Query Language), and the AI model is required to reasonably and accurately interpret the complex data table. However, the data collection and construction method in the prior art has the following problems.

[0042] (I) Internal data problems

[0043] 1. Small inventory data, poor diversity: the amount of accumulated data is always limited, such as new business data from last year, such as holiday back and forth, there are only a few (data fluctuation or richness is not high), such as inventory data, the change trend or abnormal phenomenon worth analyzing is back and forth, there are only a few (all changes cannot be exhausted). If used as training data for AI models, the data seen by the AI model will be very limited, and the diversity of data changes will be small, which will not be conducive to the generalization of the AI model.

[0044] 2. Tool constraints: the company's SQL data extraction tool is subject to performance constraints, and the time for data aggregation may take up to ten minutes to return the results each time; similarly, web search tools are also subject to performance limitations, and cannot perform web searches in large quantities at low cost (either need to pay or have the risk of being anti-crawled). Such tool constraints make it impossible to carry out some advanced AI training algorithms (such as reinforcement learning algorithms involving tool calls).

[0045] 3. High privacy: as an algorithm development, it is necessary to apply for permission or sign a confidentiality agreement layer by layer to obtain these limited data. The acquisition process is very complex.

[0046] (II) External data problems

[0047] 1. Format mismatch: external business data is mostly in the form of financial reports and news. There is a fundamental difference from the underlying data table, so it is difficult to match our business analysis scene.

[0048] 2. Security risk: Large-scale acquisition of external financial reports, even some unconventional channels, may have greater legal risks.

[0049] In summary, in the current business analysis scenario, AI model training faces the dual challenges of data scarcity and tool constraints: on the one hand, the amount of data obtained by the underlying SQL is small and lacks diversity (such as short business data, limited holiday samples, and single trend changes), which limits the model's generalization ability; on the other hand, due to the performance bottleneck of the SQL data extraction tool (long aggregation time) and the cost / risk control restrictions of web search, it is difficult to efficiently build large-scale training datasets, and the data requires complex approval processes due to high confidentiality levels, further hindering the development and application of advanced algorithms such as reinforcement learning.

[0050] To address the above problems, the present application provides a model training data construction method.

[0051] The example embodiments of the present disclosure will be described in detail below with reference to the accompanying drawings.

[0052] Figure 1 A scenario diagram of a model training data construction method or a model training data construction device that can be applied to the embodiments of the present disclosure is shown.

[0053] Reference is made to Figure 1 which shows a schematic diagram of an implementation environment provided by an example embodiment of the present disclosure.

[0054] As shown in Figure 1 , the system architecture 100 can include terminal devices 101, 102, 103, a network 104, and a server 105. The network 104 serves as a medium to provide communication links between the terminal devices 101, 102, 103, and the server 105. The network 104 can include various connection types, such as wired, wireless communication links, or optical fiber cables, etc.

[0055] A user can use the terminal devices 101, 102, 103 to interact with the server 105 through the network 104 to receive or send messages, etc. Among them, the terminal devices 101, 102, 103 can be various electronic devices with display screens and support for web browsing, including but not limited to smartphones, tablet computers, laptop computers, desktop computers, wearable devices, virtual reality devices, smart home devices, etc.

[0056] The server 105 can be a server that provides various services, such as a background management server that supports the operation of the device performed by the user using the terminal devices 101, 102, 103. The background management server can analyze and process the received request data, etc., and feed back the processing results to the terminal device.

[0057] The server can be a standalone physical server, a server cluster or a distributed system composed of multiple physical servers, a cloud server providing cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and basic cloud computing services such as big data and artificial intelligence platforms, and the like, and the present disclosure does not limit the same.

[0058] The server 105 can obtain a business analysis task construction request and business analysis task construction prompt information through the first large language model, for example; the server 105 can construct a first business analysis task according to the business analysis task construction request and the business analysis task construction prompt information through the first large language model, for example, wherein the first business analysis task includes a plurality of data dimensions to be analyzed; the server 105 can obtain data construction prompt information through the first large language model, for example; the server 105 can obtain synthetic business data under a plurality of data dimensions to be analyzed according to the data construction prompt information through the first large language model, for example; the server 105 can perform business analysis on the synthetic business data according to the first business analysis task through the first large language model to obtain a business analysis result, so as to use the first business analysis task and the business analysis result as training data for model training.

[0059] It should be understood that Figure 1 The number of terminal devices, networks and servers in the above-mentioned embodiments is only illustrative, and the server 105 can be a server of one entity, or can be composed of multiple servers. According to actual needs, the server 105 can have any number of terminal devices, networks and servers.

[0060] Figure 2 FIG. 1 is a flowchart of a method for constructing model training data according to an example embodiment. The method provided by the embodiments of the present disclosure can be executed by any electronic device with computing processing capability, for example, the method can be executed by the server or the terminal device in the above-mentioned embodiments, or can be executed by the server and the terminal device together. In the following embodiments, the server is taken as an example for illustration, but the present disclosure is not limited thereto. Figure 1

[0061] With reference to Figure 2 The method for constructing model training data provided by the embodiments of the present disclosure can include the following steps.

[0062] In step S202, the first large language model obtains a business analysis task construction request and business analysis task construction prompt information.

[0063] ​The business analysis task can refer to a task of analyzing a business in a certain scenario. For example, a holiday operation analysis in a take-out scenario can be analyzed; for another example, a shared bicycle operation can be analyzed; for another example, a group purchase business can also be analyzed. The present application does not limit the scenario of the business analysis task (such as a take-out scenario, a bicycle scenario, or an online shopping scenario, etc.), nor does it limit the specific business analysis content. As long as it is a task of analyzing a specific business, it can be a business analysis task in the present application.

[0064] Among them, take-out refers to a service mode in which consumers order food, goods or services from catering stores or other merchants through channels such as telephone, mobile application or website, and the goods are delivered to the designated place by the merchant or third-party delivery personnel.

[0065] Shared bicycles are an Internet-based short-distance travel solution. Users unlock bicycles by scanning codes, pay for usage time or distance, and park in compliance areas after reaching the destination.

[0066] Group purchase is a business model that acquires goods or services at a lower price by aggregating the purchasing needs of a large number of consumers.

[0067] The above business analysis task construction request can be a request for requesting the construction of a business analysis task.

[0068] The above business analysis task can be an analysis task for a business in any scenario, such as an analysis task in a take-out scenario, an analysis task in a shared bicycle operation scenario, or an analysis task in a drone management scenario. The present application does not limit this.

[0069] The business analysis task construction prompt information can prompt how to construct a business analysis task. The task scenario to be constructed, the task construction example template, or the task analysis path, etc. can be included in the business analysis task construction prompt information, and the present application does not limit the specific content in the business analysis task construction prompt information.

[0070] In some embodiments, a task example can be given in the business analysis task construction request so that the large language model can mimic the task example to construct the task. The task example can be a real original task.

[0071] It should be noted that the business analysis task construction request and the business analysis task construction prompt information can exist in the same information, or they can be independent of each other, and the present application does not limit this.

[0072] In some embodiments, the business analysis task construction prompt information described above can be deployed in the large language model in advance, for example, as a system prompt word in the large language model; the business analysis task described above can also be input into the large language model separately each time it is used; or can be injected at the beginning of the dialogue, which is not limited in the present application.

[0073] In the interaction of the large language model (LLM, Large Language Model), unlike the question information (Prompt) input by the user to the large language model, the system prompt word (System Prompt) can be a global instruction, which can be used to set the model behavior rules, background information or output format, and can be injected at the beginning of the dialogue. By inputting specific control instructions or information to the large language model, the output content of the large language model is structurally constrained and regulated, and the large language model is guided to respond to subsequent queries in a specific way; it can also be directly deployed in the large language model.

[0074] The embodiment of the present disclosure controls the output format of the numbers output by the large language model through the system prompt word when starting the generation training through the large language model, so as to realize the training of the number output ability of the large language model.

[0075] Step S204, the first large language model constructs a first business analysis task according to the business analysis task construction request and the business analysis task construction prompt information, wherein the first business analysis task includes a plurality of data dimensions to be analyzed.

[0076] In some embodiments, the first large language model can construct the first business analysis task according to the business analysis task construction request and the business analysis task construction prompt information.

[0077] In some embodiments, the business analysis task construction prompt information can include task analysis paths (such as the path limits the analysis path of "overall analysis -> drilling down analysis -> cross analysis" for analysis, which can include overall analysis), task analysis prompts, etc.

[0078] In some embodiments, the business analysis task construction prompt information can include analysis task examples, which can be analyzed according to the task analysis path described above.

[0079] The overall analysis can be an observation of the summary of the data from a global perspective, focusing on core indicators and overall trends. Aggregated data (such as comprehensive, average, proportion, etc.) can be used for analysis. For example, a certain data dimension is analyzed by day, month or quarter.

[0080] Drill-down analysis can refer to gradually deepening from overall data to detailed data, layer by layer, to locate problems or opportunities. For example, after discovering that "sales are declining", drill down to see the performance of specific provinces, stores or goods; or decompose from "user churn rate is rising" to different user groups or behavior paths.

[0081] Cross analysis is a method of mining data correlation through multiple dimension combinations to reveal hidden relationships or patterns between variables.

[0082] In some embodiments, the first business analysis task described above can include multiple data analysis dimensions to be analyzed. For example, at least one data analysis dimension to be analyzed can be included at the overall analysis level; for another example, at least one data analysis dimension to be analyzed can be included at the drill-down analysis level; for another example, at least one data analysis dimension to be analyzed can be included at the cross analysis level.

[0083] In step S206, the first large language model obtains data construction prompt information.

[0084] In some embodiments, the data construction prompt information described above can be global information for limiting or suggesting the format, content, input and output method, and focus content (such as analysis perspective) of the current data to be synthesized, so that the synthesized data has stability and reasonableness.

[0085] In some embodiments, the data construction prompt information described above can refer to a system prompt (System Prompt) or the like that can be deployed in a large language model, for example, that can be deployed in a large language model. The data construction prompt information described above can also be obtained from the outside (outside the large language model), for example, input to the large model before synthesizing business data, which is not limited by the present application.

[0086] In step S208, the first large language model obtains synthesized business data under multiple data analysis dimensions to be analyzed according to the data construction prompt information.

[0087] In some embodiments, the first large language model can construct synthesized business data under multiple data analysis dimensions to be analyzed in combination with the data construction prompt information.

[0088] In some embodiments, the first large language model can also send the data construction prompt information and dimension information corresponding to the multiple data analysis dimensions to be analyzed to the second large language model, so that the second large language model constructs synthesized business data under multiple data analysis dimensions to be analyzed in combination with the data construction prompt information; and the first large language model receives synthesized business data under multiple data analysis dimensions to be analyzed from the second large language model.

[0089] That is, the above synthetic business data can be synthesized by the first large language model itself, or can be obtained by requesting the second large language model through data construction prompt information (synthesized by the second large language model).

[0090] In some other embodiments, the first large language model and the second large language model can be the same large language model, or can be different large language models.

[0091] In step S210, the first large language model performs business analysis on the synthetic business data according to the first business analysis task to obtain a business analysis result, so as to take the first business analysis task and the business analysis result as training data for model training.

[0092] The technical scheme provided by the embodiment can first construct a first business analysis task according to a business analysis task construction request, then obtain synthetic business data corresponding to the first business analysis task according to data construction prompt information, and finally analyze the first business analysis task according to the synthetic business data to obtain an analysis result. The above scheme can automatically and efficiently generate a business analysis task and corresponding business analysis result data, and can solve the problem of insufficient or difficult-to-obtain real business data. On the other hand, the diversity of business analysis tasks and business analysis results can be expanded through the probabilistic random generation mode of a large model, and business edge cases that are difficult to enumerate can be covered.

[0093] Through the above steps 202-204, multiple analysis tasks in multiple scenarios can be expanded, and multiple analysis tasks in each scenario can also be expanded. For example, Figure 3 As shown in the figure, a task prompt (which can include a business analysis task construction request and a business analysis task construction prompt information) can be input to the large language model, and then the artificial intelligence model can automatically expand the task based on the task prompt, such as a first business analysis task.

[0094] The purpose of the generation step of the first business analysis task is to expand a variety of other analysis tasks from the original analysis task (an example of an analysis task). In theory, thanks to the probabilistic random generation mode of the LLM (Large Language Model), the multi-task expansion here can expand an infinite number of tasks.

[0095] This step can require the LLM to generate similar new tasks based on the existing task prompt (such as business analysis task construction prompt information). For example,

[0096] As shown in the figure, the following content can be input to the model through the task prompt 1. Figure 3

[0097] ​Below is a business analysis scenario: <Character> You are a L7 business analysis expert with clear logic and good process analysis skills, and you are good at combining tools for business analysis.

[0098] <Analysis>

[0099] <Task Name> Holiday Analysis< / Task Name>

[0100] <Analysis Method>

[0101] <Standard Path>

[0102] Step 1. [Overall analysis + business line analysis] (fixed analysis approach): Focus on tool calls related to overall and business line.

[0103] 1.1 Overall grouping by day: [contrast analysis] + [day trend analysis].

[0104] 1.2 Day average grouping by BU (Business Unit): [contrast analysis] + [average analysis].

[0105] Step 2. [Drill-down analysis]: Focus on tool calls related to drill-down (concentrate analysis on tool calls or technical implementations that support data drill-down functions). After completing common drill-down, at most, perform an irregular drill-down (free play).

[0106] 2.1 Single dimension drill-down: [drill-down analysis] + [contrast analysis].

[0107] 2.2 Partial dimension drill-down: [drill-down analysis] + [contrast analysis].

[0108] 2.3 Full dimension drill-down: [drill-down analysis] + [contrast analysis].

[0109] Step 4. [Report writing]: Focus on content summarization and report writing.

[0110] < / Standard Path>

[0111] <Analysis Perspective>

[0112] * In terms of overall business performance, focus on key indicators such as total real payment transaction amount, payment order volume, and real payment average price.

[0113] * In terms of user traffic, focus on DAU (Daily Active Users), DTU (Daily Time User), and visit-to-purchase rate, as well as first-level category, city level, and local analysis.

[0114] *In terms of supply capacity, we focus on observing transaction online merchant numbers, dynamic sales merchant numbers, dynamic sales rates, and other supply theme indicators. We attribute transaction growth to supply in combination with city levels and other dimensions.

[0115] *In terms of regional markets, we focus on observing payment order quantities, real payment order average prices, and other core indicators. We analyze the trend changes of different city level indicators during the festival period, and combine with the first category and other dimensions to disassemble the driving factors and interpret the flow of population between different regions and their consumption characteristics during the festival period.

[0116] *In terms of subsidies, we focus on observing merchant subsidy amounts, subsidy order quantities, and subsidy rates. We combine the first category, city level, and other dimensions to disassemble the causes.

[0117] < / Analysis Perspective>

[0118] < / Analysis Method

[0119] < / Analysis

[0120] <Tool>

[0121] <Get Data>

[0122] < / Precise Indicators

[0123] <Scope of Data>

[0124] <Scope of Data - Supported Indicator Dimension Cross-Method>

[0125] -DAU: Business BU, Date, Province, Whether Low-Income City, City Level, Native or Not (Alias: Whether Local), City.

[0126] -DTU: Business BU, Date, Whether Live Channel, Secondary Category, Age Group, Primary Category, Gender, Whether Special Price Group Purchase Channel, Province, Whether Marketing Scene Channel, Whether New Store, Whether Low-Income City, Whether New Customer, City Level.

[0127] -Visit Purchase Rate: Business BU, Date, Province, Whether Low-Income City, Native or Not (Alias: Whether Local), City.

[0128] -Supply Merchant Number: Business BU, Date, Secondary Category, Primary Category, Province, Whether New Store, Whether Low-Income City, City Level, Tertiary Category, City.

[0129] (A series of indicator definitions)

[0130] <Unified Indicator Specification>(The specification of a certain unified indicator, calculation method and data source. It ensures that different people understand and calculate the same indicator consistently, avoiding data ambiguity or misuse).

[0131] *Average order value: Paid GTV (Gross Transaction Value) / DTU (Daily Transaction Units) in the same period

[0132] *Paid average order value: Paid GTV / GTV order quantity.

[0133] *Dynamic sales rate: Number of dynamic sales merchants / number of online transaction merchants.

[0134] *Single store output GTV order quantity: GTV order quantity / number of dynamic sales merchants. *Single store paid GTV: Paid GTV / number of dynamic sales merchants.

[0135] *User transaction frequency: GTV order quantity / DTU.

[0136] *Transaction user ARPU: Paid GTV / DTU.

[0137] *Merchant subsidy rate: Merchant subsidy amount / paid GTV.

[0138] < / Precise indicators>

[0139] < / Data range

[0140] < / Tools

[0141] The above can be a business analysis task construction prompt information.

[0142] As Figure 3 shown, you can also input the following content to the large language model through task prompts (such as task prompt 3).

[0143] Next, please strictly follow the definition of [step-output (step-output)] based on the following analysis request to expand your analysis. Now, as an L9-level senior business analysis expert, you need to issue a new comprehensive analysis task for L7 experts. Please imitate the above business analysis scene and write a new shared bicycle business data analysis scene with a similar standard path (overall -> drilling down -> cross). In the new analysis scene, the steps in the standard path also need to be defined in the drilling down and cross manner. (The above can be a business analysis task construction request)

[0144] Input the above business analysis task construction prompt information and business analysis task construction request into the large language model, and the large language model may output the following content.

[0145] <role> You are an L7-level business data analysis expert, focusing on the operation optimization, efficiency improvement and user experience improvement of shared bicycle business, and are good at discovering business opportunity points through multi-dimensional data analysis.< / role>

[0146] <Analysis>

[0147] <Task Name> Shared Bike Business Analysis

[0148] <Analysis Method>

[0149] <Standard Path>

[0150] Step 1. [Overall Business Performance Analysis]: Focus on macro indicators and trends.

[0151] 1.1 Core Indicator Overview: [Time Series Analysis] + [Same Ring Analysis] + [Indicator Correlation Analysis].

[0152] 1.2 City Dimension Analysis: [City Comparison Analysis] + [City Stratification Analysis] + [City Contribution Analysis].

[0153] Step 2. [Drill-down Analysis]: Focus on sub-dimensions and problem positioning.

[0154] 2.1 User Dimension Drill-down: [User Type Analysis] + [User Behavior Analysis] + [User Value Analysis].

[0155] 2.2 Vehicle Dimension Drill-down: [Vehicle Type Analysis] + [Vehicle Life Cycle Analysis] + [Vehicle Efficiency Analysis].

[0156] 2.3 Regional Dimension Drill-down: [Regional Type Analysis] + [Hotspot Area Analysis] + [Regional Efficiency Analysis].

[0157] 2.4 Time Dimension Drill-down: [Time Period Analysis] + [Periodicity Analysis] + [Special Period Analysis].

[0158] Step 3. [Multi-dimensional Cross Analysis]: Focus on complex relationships and deep insights.

[0159] 3.1 User-Region Cross: [Behavior Differences of Users in Different Regions] + [Regional User Structure Differences].

[0160] 3.2 Time-Region Cross: [Regional Time Heat Map] + [Tide Phenomenon Analysis] + [Regional Flow Analysis].

[0161] 3.3 Operation-Effect Cross: [Scheduling Strategy Effect Analysis] + [Pricing Strategy Effect Analysis] + [Promotion Effect Analysis].

[0162] Step 4. [Problem Diagnosis and Opportunity Identification]: Focus on bottleneck breakthrough and growth points.

[0163] 4.1 Problem Diagnosis: [Efficiency Bottleneck Analysis] + [Cost Pain Point Analysis] + [User Experience Problem Analysis].

[0164] 4.2 Opportunity Identification: [High Potential Area Identification] + [User Growth Opportunity] + [Revenue Improvement Opportunity].

[0165] Step 5. [Policy Suggestions and Action Plan]: Focus on executable optimization solutions.

[0166] < / Analysis Pathway>

[0167] <Analysis Perspective>

[0168] * In terms of overall business performance, focus on key indicators such as the number of ride orders, ride revenue, and bike turnover rate to grasp business development trends and periodic characteristics.

[0169] * In terms of operational efficiency, focus on analyzing indicators such as dispatch response time, maintenance cycle, and single vehicle maintenance cost to optimize operational resource allocation and processes.

[0170] < / Analysis Perspective>

[0171] < / Analysis Method>

[0172] < / Analysis

[0173] <Tools>

[0174] <Acquire Data>

[0175] < / Accurate Indicators>

[0176] <Data Range>

[0177] "Data Range - Supported Indicator Dimension Cross-Method"

[0178] - Daily Active Users: Date, City, Region Type, User Type, Age Group, Gender, Membership Status, Application Version.

[0179] - New User Number: Date, City, Region Type, Channel Source, Age Group, Gender, Customer Acquisition Cost Interval.

[0180] - … (A series of indicator definitions)

[0181] "Statistical Indicator Scope"

[0182] * Average Daily Ride Frequency per Bike: Ride Order Volume / Vehicle Deployment Number.

[0183] * Average Daily Revenue per Bike: Ride Revenue / Vehicle Deployment Number.

[0184] < / Accurate Indicators>

[0185] < / Data Range>

[0186] < / Tools>

[0187] With such a "fully circumvented real data acquisition cost" rich and diverse multi-task data, the subsequent training of the business analysis model can greatly improve its generalization.

[0188] The above scheme can prompt the large language model to design a new business analysis task based on a small amount of original task samples.

[0189] Through the above method, on the one hand, the business analysis task can be automatically generated, reducing the degree of human participation and reducing costs; on the other hand, different scene analysis tasks can be automatically expanded to cover many edge cases, improving the coverage and depth of analysis and avoiding missing potential analysis directions.

[0190] As shown in Figure 4 After generating the business analysis task, the LLM data retrieval tool and the LLM search tool can be used to simulate business data obtained from internal system queries and external data obtained through web searches for business analysis. The above data retrieval tool and search tool can be a large language model (such as the first large language model or the second large language model), and the LLM data retrieval tool and the LLM search tool can be the same large language model or different large language models. The present application does not limit this.

[0191] The LLM data retrieval tool can be used to simulate business data obtained from internal system queries. The LLM search tool can be used to simulate business data obtained from external systems through web searches.

[0192] In the following, the present application will introduce how to synthesize business data in combination with specific embodiments.

[0193] Figure 5 is a flowchart of a method for synthesizing business data according to an exemplary embodiment.

[0194] In some embodiments, the data construction prompt information can include at least one of legal data dimension information, legality judgment prompt information, and prior distribution generation prompt information.

[0195] The legal data dimension information can be used to determine whether the data dimension of the current request is legal. Specifically, it is determined whether the data dimension to be analyzed is legal.

[0196] Referring to Figure 5 The above method for synthesizing business data can include the following steps.

[0197] Step S502, according to the legal data dimension information in the data construction prompt information, determine whether each data dimension in the plurality of data dimensions to be analyzed is legal.

[0198] Step S504, in the case of determining that the plurality of to-be-analyzed data dimensions are legal, generating prompt information according to the prior distribution is to design a prior distribution for each of the plurality of to-be-analyzed data dimensions.

[0199] Step S506, constructing a synthetic business data for each to-be-analyzed data dimension according to the prior distribution corresponding to each to-be-analyzed data dimension.

[0200] Through the above method, on the one hand, it can be judged whether the to-be-analyzed data dimension to be synthesized is legal, and only in the case of legal to-be-analyzed data dimension, the data under the to-be-analyzed data dimension will be synthesized; on the other hand, the prior distribution can be designed for the to-be-analyzed data dimension according to the characteristics of the to-be-analyzed data dimension itself, which can avoid the blindness of pure data-driven.

[0201] Figure 6 is a flowchart of a data construction prompt information generation method according to an example embodiment.

[0202] Reference Figure 6 The above data construction prompt information generation method can include the following steps.

[0203] In some embodiments, the first business analysis task can include a legal data dimension, which is used to describe the legal data dimension in the first business analysis task.

[0204] Step S602, the first large language model obtains to-be-filled data construction prompt information.

[0205] The to-be-filled data construction prompt information and the data construction prompt information can be basically the same, and the difference between them can be whether there is a legal data dimension.

[0206] Step S604, the first large language model fills the legal data dimension in the first business analysis task into the to-be-filled data construction prompt information to generate the data construction prompt information, so as to judge whether each data dimension in the plurality of to-be-analyzed data dimensions is legal according to the legal data dimension.

[0207] In some embodiments, the first business analysis task can include a legal data dimension (such as "data range: DAU, DTU, visit purchase rate, supplier number, etc." in the above example).

[0208] The legal data dimension in the above first business analysis task can be filled into the to-be-filled data construction prompt information to generate the data construction prompt information.

[0209] Through the above method, the legal data dimension in the data construction prompt information can come from the first business analysis task, so that the to-be-analyzed data dimension can be more accurately judged.

[0210] Figure 7 The figure is a flowchart of a method for constructing synthetic business data according to an exemplary embodiment.

[0211] The synthetic business data construction method provided in this embodiment can be executed by the first largest language model or the second largest language model, and this application does not impose any restrictions on this.

[0212] In some embodiments, the large language model (which can be the first large language model or the second large language model) can perform analysis on the first business analysis task through multiple rounds of interaction. For example, the large language model may first analyze a data dimension in the first business analysis task. If it determines that the data dimension has analytical value, it may continue to drill down or perform cross-analysis on the data dimension.

[0213] Next, this application will be Figure 7 The above multi-round interaction process is described in detail.

[0214] In some embodiments, the first business analysis task may include a first analysis level (e.g., an overall analysis level) and a second analysis level (e.g., a drill-down analysis level), and the multiple data dimensions to be analyzed include a first data dimension at the first analysis level and a second data dimension at the second analysis level. The first analysis level may be a level above the second analysis level.

[0215] The above-mentioned first business analysis task may also include more analysis levels (such as cross-analysis levels), and this application does not impose any limitation on this.

[0216] refer to Figure 7 , the above-mentioned synthetic business data construction method may include the following steps.

[0217] Step S702: construct synthetic business data under the first data dimension in combination with the data construction prompt information.

[0218] In some embodiments, the large language model may first synthesize business data under a first data dimension under a first analysis level.

[0219] Step S704 : The first large language model performs business analysis on the synthetic business data under the first data dimension according to the first business analysis task to obtain a first analysis result in the business analysis results.

[0220] Step S706: Determine, based on the first analysis result, whether the target data range under the second data dimension has analysis value.

[0221] In different application scenarios, whether a certain data range has analysis value is different, and the present application does not limit this, which can be set according to the actual scene. For example, for the shared bicycle operation scene, the area with high frequency of use can facilitate personnel analysis and may have more analysis value; for example, for the shared bicycle operation scene, the place with low frequency of use may need further deployment of bicycles, and may also have more analysis value. In short, which data range has more analysis value needs to be judged in combination with the specific application scene and business content, and the present application does not limit this.

[0222] In some embodiments, the first data dimension under the first analysis level can be analyzed first to determine the data range with analysis value in the first data dimension.

[0223] Step S708, in combination with data, the target data range under the second data dimension is constructed to construct the synthesized business data, so that the first large language model continues to perform business analysis on the synthesized business data in the target data range under the second data dimension according to the first business analysis task.

[0224] For example, the first data dimension can be a "regional data dimension", and when analyzing the regional data, it is found that 1XX and 2XX have analysis value, then 1XX and 2XX data under the second analysis level can be constructed to analyze the 1XX and 2XX data.

[0225] Through the above method, on the one hand, the preliminary analysis of the first data dimension (such as the regional dimension) quickly identifies the high-value target range (such as 1XX / 2XX), avoids the waste of resources of full data construction, forms an iterative analysis chain of "preliminary screening→precise construction", and makes the computing resources concentrated in the key data dimension; on the other hand, the hierarchical construction strategy (first / second analysis level) significantly reduces the processing complexity of the large language model, and realizes the gradual refinement of the analysis granularity through the dimension progression (region→specific location), which conforms to the real business analysis logic.

[0226] As can be seen, the model can directly output a brand new, hypothetical analysis task scene. The subsequent steps synthesize analysis process data in this scene, which can realize complete separation from real data and completely no longer need any real data.

[0227] Figure 8 is a flowchart of a synthesized business data construction method according to an example embodiment.

[0228] Reference Figure 8 The above synthesized business data construction method can include the following steps.

[0229] In some embodiments, the second large language model can include at least one, for example, one, two, three, or four, etc.

[0230] In some embodiments, the second large language model described above can include a data query large model and a data search large model. The synthetic business data includes search business data and query business data.

[0231] The data query large model described above can be used to synthesize query business data, which can be used to simulate business data obtained by querying internal systems (i.e., internal data).

[0232] The data search large model described above can be used to synthesize search business data, which can be used to simulate business data obtained by network search (i.e., external data).

[0233] Step S802, the data search large model combines data to construct prompt information to construct search business data.

[0234] Step S804, the data query large model combines data to construct prompt information to construct query business data.

[0235] It should be noted that in order to ensure that the obtained business data has consistent data distribution, at least one second large language model can share the prior distribution, for example, the data query large model and the data search large model can share the prior distribution when constructing the business data of the first business analysis task.

[0236] The above scheme can make the business data synthesized by different large language models have consistent data distribution.

[0237] In the following, the present application will combine specific embodiments to explain how to synthesize business data.

[0238] This step can simulate any scene data taking tool (such as simulating data taking tool by query large model), search tool (such as simulating search tool by search large model) based on large language, and requires the stability and rationality of the synthesized data.

[0239] In theory, thanks to the probabilistic random generation mode of the large language model, the large language model simulation tool here can generate an infinite number of data distributions, even under the same task, it can generate completely different data distributions (such as different abnormal phenomena, different trends).

[0240] This step only needs to design a set of system prompt words containing step-by-step instructions for the large language, and then let the large language model continuously provide synthetic data for a certain analysis task in multiple rounds of interaction.

[0241] The following will be illustrated with the system prompt words of the data query tool simulated by the large language model (a kind of data construction prompt information, or a kind of data construction prompt information to be filled in).

[0242] You are a professional business data analyst, and you are providing the data needed for business analysis to the intern.

[0243] The following is the legal data query system: {data range}.

[0244] You will receive data query requests from interns, please simulate the data needed by the interns based on the following steps.

[0245] Step 1: Determine whether the data query is legal. For example: whether there are non-existent query dimensions.

[0246] Step 2: Design the schema of the data;

[0247] -Requirements: Each piece of data is a neat row table, with the first column as a specific (sub) category, project, and the remaining columns as indicators or dimensions.

[0248] Step 3: Randomly design the prior distribution of the data: for example: "the order of magnitude or value range of xxx indicator is xxx", "in the query date range, design xxx dimension to reach the maximum value on the third day, and the overall first increase and then decrease", "in the 10 business units of xxx, xxx, etc., design xxx, xxx as the first two dimensions, and introduce an anomaly for the xxx dimension to increase the difficulty of analysis for the intern."

[0249] Step 4: Simulate real data;

[0250] -Fill in all the required numbers, do not allow nan or omissions;

[0251] -Adopt XX format.

[0252] Please strictly follow the above steps to output your data provision process in XX, and your output format is as follows:

[0253] <Step 1> [Intern's query legality description, if not legal, please list the illegal places and modify the legal query statement, continue to the next step] Step 1;

[0254] <Step 2> [Design of data schema] Step 2;

[0255] <Step 3> [Design of data prior distribution] Step 3;

[0256] <Step 4> Output the data table in a format.

[0257] (Note: {Data range} in the above data retrieval tool system prompt word is a placeholder, which will be replaced by the actual data range content in the first business analysis task during use, which describes the legal queryable data dimensions in the analysis task.)

[0258] System prompt word of step-by-step instruction: The advantage of this design is that the large language model can finally synthesize data step by step, so that the synthesis process has rationality (such as designing a prior distribution).

[0259] Multi-round interaction method to synthesize data: The advantage of this design is that it can make the data synthesized by LLM in front and back multiple times have uniformity; if not (i.e. each data synthesis is a single round of interaction and terminates), it will easily lead to unreasonable context of synthesized data. (For example, the first time, the indicators of A city are the highest; but the second time, related data is involved, because LLM does not know the previous context, it may generate data with B indicators being the highest, leading to contradictions in the results of the analysis task.)

[0260] The synthesis process of synthesized business data will be described below in combination with specific embodiments.

[0261] Example query: "Query the fourth quarter of 2023, the overall monthly summary of the number of new users, the number of ride orders, ride revenue, and ride person times, and horizontally correlate the YoY (Year-over-Year) growth rate of each indicator in the fourth quarter of 2022……."

[0262] Example of simulating data synthesis.

[0263] <Step 1>

[0264] Query is legal. All requested indicators (number of new users, number of ride orders, ride revenue, and ride person times) are within the supported range, and the time dimension (monthly summary) and comparison dimension (YoY, month-on-month) conform to the statistical logic.

[0265] Correction note: No correction items are needed.

[0266] < / Step 1>

[0267] <Step 2>

[0268] **Table design**

[0269] The table design can be designed according to actual needs, which will not be shown here.

[0270] < / Step 2>

[0271] <Step 3>

[0272] **Prior distribution design**

[0273] 1. Order of magnitude:

[0274] - New user number: 100,000-300,000 per month;

[0275] - Ride order volume: 1-5 million per month;

[0276] - Ride income: 5-20 million per month (unit: yuan);

[0277] - Ride person times: 800,000-4 million per month.

[0278] 2. Trend design:

[0279] - YoY growth: 10-25% overall YoY growth in Q4 2023 (driven by winter promotion), with a slight decrease in growth in December due to the influence of cold waves;

[0280] - Abnormal point: November ride income YoY abnormally high (+35%), due to membership fee rate adjustment.

[0281] 3. Top dimension:

[0282] - Ride income burst coefficient is the highest (1.5 times), and ride person times is the lowest (1.2 times);

[0283] < / Step 3>

[0284] <Step 4>

[0285] Output in a certain format. Explanation: The technical personnel in this field can output according to the actual demand, and this place will not be repeated

[0286] < / Step 4>

[0287] Similarly, the search tool simulated by LLM is also designed in the form of multi-round interaction in steps, only need to modify the system prompt words (i.e. data construction prompt information) slightly, such as:

[0288] <Step 1> [Intern's query legality description, if not legal, please list the illegal places and modify the query sentence to be legal, continue to the next step]< / Step 1>;

[0289] <Step 2> [Design of prior distribution of data]< / Step 2>;

[0290] <Step 3>

Generate a 300-word article (the content of the article conforms to the prior distribution)

[0291] Note: Since the data extraction tool and the search tool are jointly used as an analysis task tool, the two simulation tools should share the prior distribution, and both should have consistent data simulation distribution.

[0292] In addition, the data extraction tool can be a query large model in the application, and the search tool can be a search large model in the application.

[0293] With the "endless" new analysis task and "endless" data extraction and search tools, any diverse analysis data can be constructed based thereon (such as constructing a link of "data extraction -> analysis thinking -> re-data extraction -> re-analysis thinking -> re-search -> writing a final analysis report").

[0294] The above scheme can realize reasonable simulation of LLM data extraction tools and LLM search tools based on large language models; that is, the large language model can be used to generate a step-by-step paradigm, which can provide reasonable and diverse data for any business analysis task scenario, thereby completely avoiding the acquisition restrictions of real business data and constructing any rich and diverse business analysis training data.

[0295] It should be particularly pointed out that each step in each embodiment of the above model training data construction method can be crossed, replaced, added, deleted. Therefore, these reasonable permutations and combinations of the model training data construction method should also belong to the protection scope of the present disclosure, and the protection scope of the present disclosure should not be limited to the embodiments.

[0296] Based on the same inventive concept, the present disclosure also provides a model training data construction device, as follows. Since the principles of the device embodiments solve problems similar to the above method embodiments, the implementation of the device embodiments can be referred to the implementation of the above method embodiments, and the repeated parts will not be repeated.

[0297] Figure 9 is a block diagram of a model training data construction device according to an exemplary embodiment. Referring to Figure 9 The model training data construction device 900 provided by the embodiments of the present disclosure can include a request acquisition module 901, a task construction module 902, a construction prompt information acquisition module 903, a data acquisition module 904, and an analysis module 905.

[0298] The request obtaining module 901 can be configured to obtain, by the first large language model, a business analysis task construction request and business analysis task construction prompt information; the task construction module 902 can be configured to construct, by the first large language model, a first business analysis task according to the business analysis task construction request and the business analysis task construction prompt information, wherein the first business analysis task includes a plurality of to-be-analyzed data dimensions; the construction prompt information obtaining module 903 can be configured to obtain, by the first large language model, data construction prompt information; the data obtaining module 904 can be configured to obtain, by the first large language model, synthetic business data under the plurality of to-be-analyzed data dimensions according to the data construction prompt information; and the analysis module 905 can be configured to perform business analysis on the synthetic business data according to the first business analysis task to obtain a business analysis result, so as to use the first business analysis task and the business analysis result as training data for model training.

[0299] It should be noted that the request obtaining module 901, the task construction module 902, the construction prompt information obtaining module 903, the data obtaining module 904, and the analysis module 905 correspond to S202-S206 in the method embodiment, and the above modules have the same examples and application scenarios as the corresponding steps, but are not limited to the disclosure of the above method embodiments. It should be noted that the above modules as part of the device can be executed in a computer system such as a group of computer executable instructions.

[0300] In some embodiments, the data construction prompt information includes legal data dimension information, legality judgment prompt information, and prior distribution generation prompt information; and the data obtaining module 904 can include a legality judgment sub-module, a prior distribution design sub-module, and a data construction sub-module.

[0301] The legality judgment sub-module can be configured to determine, according to the legal data dimension information in the data construction prompt information, whether each data dimension in the plurality of to-be-analyzed data dimensions is legal; the prior distribution design sub-module can be configured to, in a case where it is determined that the plurality of to-be-analyzed data dimensions are legal, design a prior distribution for each of the plurality of to-be-analyzed data dimensions according to the prior distribution generation prompt information; and the data construction sub-module can be configured to construct synthetic business data for each to-be-analyzed data dimension according to the prior distribution corresponding to the to-be-analyzed data dimension.

[0302] In some embodiments, the first business analysis task includes a legal data dimension, and the legal data dimension is used to describe a legal data dimension in the first business analysis task; and the construction prompt information obtaining module 903 can include a to-be-filled information obtaining sub-module and a filling sub-module.

[0303] The to-be-filled information obtaining sub-module can be configured to obtain, by the first large language model, to-be-filled data construction prompt information.

[0304] In some embodiments, the first business analysis task includes a first analysis level and a second analysis level, and the plurality of to-be-analyzed data dimensions include a first data dimension at the first analysis level and a second data dimension at the second analysis level; the data obtaining module 904 can include a first synthesis sub-module and a second synthesis sub-module.

[0305] The first synthesis sub-module can be configured to synthesize business data at the first data dimension based on the data construction prompt information.

[0306] The analysis module 905 can include a first analysis result obtaining sub-module and an analysis value determining sub-module.

[0307] The first analysis result obtaining sub-module can be configured to obtain, by the first large language model, a first analysis result of the business analysis result based on the first business analysis task and the synthesized business data at the first data dimension; and the analysis value determining sub-module can be configured to determine that the target data range at the second data dimension has an analysis value based on the first analysis result.

[0308] The second synthesis sub-module can be configured to synthesize business data in the target data range at the second data dimension based on the data construction prompt information, so that the first large language model continues to perform business analysis on the synthesized business data in the target data range at the second data dimension based on the first business analysis task.

[0309] In some embodiments, the data obtaining module 904 can include a construction sub-module and a construction data request sub-module.

[0310] The construction sub-module can be configured to synthesize, by the first large language model, synthesized business data at the plurality of to-be-analyzed data dimensions based on the data construction prompt information.

[0311] The construction data request sub-module can be configured to send, by the first large language model, the data construction prompt information and dimension information corresponding to the plurality of to-be-analyzed data dimensions to the second large language model, so that the second large language model synthesizes the synthesized business data at the plurality of to-be-analyzed data dimensions based on the data construction prompt information; and the first large language model receives the synthesized business data at the plurality of to-be-analyzed data dimensions from the second large language model.

[0312] In some embodiments, the second large language model comprises a data query large model and a data search large model, and the synthesized business data comprises search business data and query business data; wherein the data request construction sub-module can comprise a search business data determination unit and a query business data determination unit.

[0313] The search business data determination unit can be configured to construct the search business data by combining the data search large model with data construction prompt information, and the query business data determination unit can be configured to construct the query business data by combining the data query large model with data construction prompt information; wherein the data query large model and the data search large model share a prior distribution when constructing the business data of the first business analysis task, so that the search business data and the query business data have consistent data distribution.

[0314] Since the functions of the apparatus 900 have been described in detail in the corresponding method embodiments, the present disclosure will not be repeated here.

[0315] The modules and / or sub-modules and / or units described in the embodiments of the present disclosure can be implemented in the form of software or in the form of hardware. The described modules and / or sub-modules and / or units can also be arranged in a processor. In some cases, the names of these modules and / or sub-modules and / or units do not constitute a limitation on the modules and / or sub-modules and / or units themselves.

[0316] The flowcharts and block diagrams in the accompanying drawings illustrate the possible implementation architectures, functions and operations of the systems, methods and computer program products according to various embodiments of the present disclosure. In this regard, each block in the flowcharts or block diagrams can represent a module or a part of a program segment containing one or more executable instructions for implementing the specified logical functions. It should also be noted that in some alternative implementations, the functions noted in the blocks can occur in different orders from those shown in the accompanying drawings. For example, two blocks that are shown in succession can actually be executed substantially in parallel, and sometimes in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams or flowcharts, and combinations of blocks in the block diagrams or flowcharts, can be implemented by a dedicated hardware-based system that performs the specified functions or operations, or can be implemented by a combination of dedicated hardware and computer program instructions.

[0317] In addition, the above-described accompanying drawings are only schematic illustrations of the processes included in the method according to the exemplary embodiments of the present disclosure, and are not for limitation purposes. It is easy to understand that the processes shown in the above-described accompanying drawings do not indicate or limit the time sequence of these processes. In addition, it is also easy to understand that these processes can be executed synchronously or asynchronously, for example, in multiple modules.

[0318] Figure 10 A structural diagram of an electronic device suitable for implementing the embodiments of the present disclosure is shown. It should be noted that Figure 10 The electronic device 1000 shown is only one example and should not be taken as limiting the scope of the functionality and use of the embodiments of the present disclosure.

[0319] As Figure 10 shown, the electronic device 1000 includes a central processing unit (CPU) 1001 that can perform various appropriate actions and processes in accordance with a program stored in a read only memory (ROM) 1002 or a program loaded from a storage section 1008 into a random access memory (RAM) 1003. In the RAM 1003, various programs and data required for the operation of the electronic device 1000 are also stored. The CPU 1001, the ROM 1002, and the RAM 1003 are connected to each other through a bus 1004. An input / output (I / O) interface 1005 is also connected to the bus 1004.

[0320] The following components are connected to the I / O interface 1005: an input section 1006 including a keyboard, a mouse, etc.; an output section 1007 including a display such as a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 1008 including a hard disk, etc.; and a communication section 1009 including a network interface card such as a LAN card, a modem, etc. The communication section 1009 performs communication processing via a network such as the Internet. A drive 1010 is also connected to the I / O interface 1005 as necessary. A removable media 1011 such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc. is attached to the drive 1010 as necessary, so that a computer program read therefrom is installed into the storage section 1008 as necessary.

[0321] In particular, the processes described above with reference to the flowcharts can be implemented as a computer software program in accordance with the embodiments of the present disclosure. For example, the embodiments of the present disclosure include a computer program product including a computer program carried on a computer-readable storage medium, the computer program containing computer program instructions for executing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network by the communication section 1009, and / or installed from the removable media 1011. When the computer program is executed by the central processing unit (CPU) 1001, the above-described functions defined in the system of the present disclosure are executed.

[0322] It should be noted that the computer-readable storage medium shown in the present disclosure can be a computer-readable signal medium or a computer-readable storage medium or any combination of the above two. The computer-readable storage medium may, for example, be but is not limited to an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or component, or any combination of the above. More specific examples of computer-readable storage media can include but are not limited to: an electrical connection having one or more wires, a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present disclosure, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, device or component. In the present disclosure, the computer-readable signal medium can include a data signal carried in a baseband or as part of a carrier wave, which carries computer-readable computer program instructions. Such a propagated data signal can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. The computer-readable signal medium can also be any computer-readable storage medium other than the computer-readable storage medium, which can send, propagate or transmit programs for use by or in conjunction with an instruction execution system, device or component. The computer program instructions contained on the computer-readable storage medium can be transmitted in any suitable medium, including but not limited to: wireless, wire, optical cable, RF, etc., or any suitable combination of the above.

[0323] As another aspect, the present disclosure also provides a computer-readable storage medium, which can be contained in the device described in the above embodiments; or can exist separately without being assembled into the device. The above computer-readable storage medium carries one or more programs, which, when executed by the device, enable the device to implement functions including: the first large language model obtains a business analysis task construction request and business analysis task construction prompt information; the first large language model constructs a first business analysis task according to the business analysis task construction request and the business analysis task construction prompt information, wherein the first business analysis task includes a plurality of data dimensions to be analyzed; the first large language model obtains data construction prompt information; the first large language model obtains synthetic business data under a plurality of data dimensions to be analyzed according to the data construction prompt information; the first large language model performs business analysis on the synthetic business data according to the first business analysis task to obtain a business analysis result, so as to take the first business analysis task and the business analysis result as training data for model training.

[0324] According to one aspect of the present disclosure, a computer program product or computer program is provided, which includes computer program instructions stored in a computer readable storage medium. The computer program instructions are read from the computer readable storage medium and executed by a processor to implement the method provided in various optional implementation manners of the above-described embodiments.

[0325] Through the above description of the embodiments, those skilled in the art can easily understand that the example embodiments described herein can be implemented by software, or by software in combination with necessary hardware. Therefore, the technical solutions of the embodiments of the present disclosure can be embodied in the form of a software product. The software product can be stored in a non-volatile storage medium (which can be a CD-ROM, a U disk, a mobile hard disk, etc.), and includes a plurality of computer program instructions to make an electronic device (which can be a server or a terminal device, etc.) execute the method according to the embodiments of the present disclosure.

[0326] Other embodiments of the present disclosure will be readily apparent to those skilled in the art upon considering the description of the specification and practicing the present disclosure as disclosed herein. The present disclosure is intended to cover any variations, uses, or adaptations of the present disclosure following the general principles thereof and including other known ins and customary technical means not specifically disclosed in the present disclosure. The specification and embodiments are only considered exemplary, and the true scope and spirit of the present disclosure are indicated by the claims.

[0327] It should be understood that the present disclosure is not limited to the detailed structure, drawing manner or implementation method shown herein, but rather, the present disclosure is intended to cover various modifications and equivalent arrangements included in the spirit and scope of the appended claims.

Claims

1. A method for constructing model training data, characterized in that: include: The first language model obtains business analysis task construction requests and business analysis task construction prompt information; The first language model constructs a first business analysis task according to the business analysis task construction request and the business analysis task construction prompt information, wherein the first business analysis task includes a plurality of data dimensions to be analyzed; The first language model obtains data construction prompt information; The first language model acquires the synthetic business data under the multiple data dimensions to be analyzed according to the data construction prompt information; The first large language model performs business analysis on the synthesized business data according to the first business analysis task to obtain a business analysis result, so that the first business analysis task and the business analysis result are used as training data for model training.

2. The method according to claim 1, characterized in that The data construction prompt information includes: legal data dimension information and prior distribution generation prompt information; wherein, obtaining the synthetic business data under the multiple data dimensions to be analyzed according to the data construction prompt information includes: Determining whether each of the multiple data dimensions to be analyzed is legal according to the legal data dimension information in the data construction prompt information; When it is determined that the multiple dimensions of the data to be analyzed are legal, generating prompt information based on the prior distribution and designing prior distributions for the multiple dimensions of the data to be analyzed respectively; Construct synthetic business data for each data dimension to be analyzed based on the prior distribution corresponding to each data dimension to be analyzed.

3. The method according to claim 1, characterized in that The first business analysis task includes a legal data dimension, and the legal data dimension is used to describe the legal data dimension in the first business analysis task; wherein the first language model obtains data construction prompt information, including: The first language model obtains prompt information of constructing data to be filled; The first large language model fills the legal data dimensions in the first business analysis task into the data construction prompt information to be filled, and generates the data construction prompt information, so as to judge whether each data dimension in the multiple data dimensions to be analyzed is legal based on the legal data dimensions.

4. The method according to claim 1, characterized in that The first business analysis task includes a first analysis level and a second analysis level, and the multiple data dimensions to be analyzed include a first data dimension at the first analysis level and a second data dimension at the second analysis level. Acquiring synthetic business data under the multiple data dimensions to be analyzed according to the data construction prompt information includes: constructing synthetic business data under the first data dimension in combination with the data construction prompt information; The first language model performs business analysis on the synthesized business data according to the first business analysis task to obtain a business analysis result, including: The first language model performs business analysis on the synthetic business data under the first data dimension according to the first business analysis task to obtain a first analysis result in the business analysis results; Determining, based on the first analysis result, that the target data range under the second data dimension has analysis value; The step of constructing the synthetic business data under the multiple data dimensions to be analyzed in combination with the data construction prompt information further includes: In combination with the data construction prompt information, synthetic business data is constructed within the target data range under the second data dimension, so that the first large language model continues to perform business analysis on the synthetic business data within the target data range under the second data dimension according to the first business analysis task.

5. The method according to claim 1, characterized in that: The first language model obtains the synthetic business data under the multiple data dimensions to be analyzed according to the data construction prompt information, including: The first language model is combined with the data construction prompt information to construct the synthetic business data under the multiple data dimensions to be analyzed; or The first large language model sends the data construction prompt information and the dimension information corresponding to the multiple data dimensions to be analyzed to the second large language model, so that the second large language model constructs the synthetic business data under the multiple data dimensions to be analyzed in combination with the data construction prompt information; and the first large language model receives the synthetic business data under the multiple data dimensions to be analyzed from the second large language model.

6. The method according to claim 5, characterized in that The second largest language model includes a data query model and a data search model, and the synthesized business data includes search business data and query business data. The second largest language model constructs the synthesized business data under the multiple data dimensions to be analyzed in combination with the data construction prompt information, including: The data search big model combines the data construction prompt information to construct the search business data; The data query big model combines the data construction prompt information to construct the query business data; The data query big model and the data search big model share a prior distribution when constructing the business data of the first business analysis task, so that the search business data and the query business data have consistent data distribution.

7. A device for constructing model training data, characterized in that: include: The request acquisition module is used for the first language model to obtain the business analysis task construction request and business analysis task construction prompt information; a task construction module, configured to use the first language model to construct a first business analysis task according to the business analysis task construction request and the business analysis task construction prompt information, wherein the first business analysis task includes a plurality of data dimensions to be analyzed; A construction prompt information acquisition module, used for acquiring data from the first language model to construct prompt information; A data acquisition module, configured to use the first language model to acquire, according to the data construction prompt information, synthetic business data under the multiple data dimensions to be analyzed; An analysis module is used for the first language model to perform business analysis on the synthetic business data according to the first business analysis task to obtain a business analysis result, so as to use the first business analysis task and the business analysis result as training data for model training.

8. An electronic device, characterized in that: include: memory and processor; The memory is used to store computer program instructions; the processor calls the computer program instructions stored in the memory to implement the method for constructing model training data as described in any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method for constructing model training data as described in any one of claims 1 to 6 is implemented.

10. A computer program product comprising computer program instructions stored in a computer-readable storage medium, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images