Interaction method, electronic device, storage medium and computer program product

Abstract

The invention provides an interaction method, electronic equipment, a storage medium and a computer program product. The method comprises the following steps: acquiring an input interaction context; the interaction context and a preset system cue word template are input into a large language model, at least one candidate PEFT module is obtained, and the system cue word template is used for indicating the large language model to select and output the candidate PEFT module from a plurality of preset PEFT modules according to the interaction context; determining a target PEFT module according to the candidate PEFT module, and loading the target PEFT module to the large language model to obtain a large language model after the module is loaded; and inputting the interaction context into the large language model loaded with the module to obtain an interaction response. According to the method and the device, automatic selection and loading of the PEFT module can be realized, and the interaction performance and expression of a large language model for the current interaction context can be improved.

Description

Technical Field

[0001] This application relates to the field of communication technology, and in particular to an interaction method, electronic device, computer-readable storage medium, and computer program product. Background Technology

[0002] To improve the interactive performance of Large Language Models (LLMs) in specific domains, the industry has proposed Parameter-Efficient Fine-Tune (PEFT) technology. PEFT is an efficient fine-tuning technique that extends the capabilities of a basic LLM. Its principle is to freeze most of the LLM's model parameters during the tuning phase, optimizing and updating only a small subset of parameters. This not only significantly reduces the computational cost and time of model tuning but also reduces the amount of data required for training. However, in practical applications, due to numerous demands, there are often a large number of available PEFT modules. How to automatically load the appropriate PEFT module for the LLM when faced with user interaction requests is a technical problem that needs to be solved. Summary of the Invention

[0003] This application provides an interaction method, electronic device, computer-readable storage medium, and computer program product that can automatically load an adapted PEFT module for an LLM based on the user-input interaction context.

[0004] In a first aspect, embodiments of this application provide an interaction method, the method comprising:

[0005] Obtain the interaction context of the input;

[0006] The interaction context and the preset system prompt word template are input into a large language model to obtain at least one candidate PEFT module. The system prompt word template is used to instruct the large language model to select and output the candidate PEFT module from a plurality of preset PEFT modules according to the interaction context.

[0007] The target PEFT module is determined based on the candidate PEFT modules, and the target PEFT module is loaded into the large language model to obtain the large language model after loading the module;

[0008] The interaction context is input into the large language model after the module is loaded to obtain the interaction response.

[0009] Secondly, embodiments of this application provide an electronic device, including:

[0010] One or more processors;

[0011] A memory having stored one or more programs that, when executed by one or more processors, cause the one or more processors to implement the interaction method described in the first aspect above.

[0012] Thirdly, embodiments of this application provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the interaction method described in the first aspect above.

[0013] Fourthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the interaction method described in the first aspect above.

[0014] In the interaction method provided in this application embodiment, the interaction context and a preset system prompt word template are input into a large language model. The large language model, leveraging its powerful natural language processing capabilities, selects candidate PEFT modules from a set of preset PEFT modules that closely match the interaction context, based on the instructions of the system prompt word template. The selected candidate PEFT modules are then output, and the target PEFT module to be loaded into the large language model is determined based on these candidate PEFT modules, thus achieving automatic selection and loading of PEFT modules. Furthermore, since the target PEFT module is determined based on the input interaction context, it can adapt well to the current interaction context scenario and provide customized interaction responses for different interaction contexts, improving the interaction performance and characteristics of the large language model. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of a system architecture provided in an embodiment of this application;

[0016] Figure 2 This is a flowchart of an interaction method provided in an embodiment of this application;

[0017] Figure 3 This is a schematic diagram of an interactive system architecture provided in an embodiment of this application;

[0018] Figure 4 This is a schematic diagram of the data structure of a system prompt word template provided in an embodiment of this application;

[0019] Figure 5 This is a schematic diagram of a system prompt word template provided in an embodiment of this application;

[0020] Figure 6 This is a schematic diagram of a system module information table provided in an embodiment of this application;

[0021] Figure 7This is a schematic diagram of an interactive system architecture provided in an embodiment of this application;

[0022] Figure 8 This is a schematic diagram of the architecture of an enhanced interaction device provided in an embodiment of this application;

[0023] Figure 9 This is a schematic diagram of a new PEFT module processing flow provided in an embodiment of this application;

[0024] Figure 10 This is a schematic diagram of a process for removing a PEFT module provided in an embodiment of this application;

[0025] Figure 11 This is a schematic diagram of a process for updating PEFT module information provided in an embodiment of this application;

[0026] Figure 12 This is a schematic diagram of an enhanced interactive process provided in an embodiment of this application;

[0027] Figure 13 This is a schematic diagram illustrating an interactive example of a product promotion scenario provided in an embodiment of this application;

[0028] Figure 14 This is a schematic diagram of the device structure of an electronic device provided in one embodiment of this application. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0030] It should be understood that in the description of the embodiments of this application, the use of terms such as "first" and "second" is only for the purpose of distinguishing technical features and should not be construed as indicating or implying relative importance, or implicitly indicating the number of technical features indicated, or implicitly indicating the sequential relationship of the technical features indicated. "At least one" means one or more, and "more" means two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent the cases where A exists alone, A and B exist simultaneously, or B exists alone. A and B can be singular or plural. The character " / " generally indicates that the related objects before and after are in an "or" relationship. "At least one of the following" and similar expressions refer to any group of these items, including any group of singular or plural items. For example, at least one of a, b, and c can represent: a, b, c, a and b, a and c, b and c, or a and b and c, where a, b, and c can be single or multiple.

[0031] Furthermore, the technical features involved in the various embodiments of this application described below can be combined with each other as long as they do not conflict with each other.

[0032] To facilitate the description of the embodiments of this application, the following introduces concepts related to the embodiments of this application.

[0033] LLM (Learning Model for Natural Languages) is a deep learning model trained on massive amounts of text data. It can not only generate natural language text, but also deeply understand the meaning of text and handle various natural language tasks, such as text summarization, question answering, and translation.

[0034] PEFT is a technique for fine-tuning large language models (LLMs). PEFT fine-tunes only a small subset of the model parameters while freezing most of the pre-trained network. By minimizing the number of fine-tuned parameters and computational complexity, it improves the performance of pre-trained models on new tasks, thereby alleviating the training cost of large pre-trained models.

[0035] Large-scale language models (LLMs), such as Generative Pre-Trained Transformers (GPT) and Large Language Models Meta AI (LLAMA), have made significant progress in multiple fields and are widely used in text generation, natural language understanding, machine translation, chatbots, content recommendation, and other scenarios. By analyzing and learning from large-scale text data, they can generate coherent and natural text, provide accurate information retrieval and summarization, and even engage in direct human conversation. Furthermore, these models have demonstrated enormous potential in specialized fields such as sentiment analysis, speech recognition, and automatic programming, significantly driving innovation in scientific research, education, and healthcare.

[0036] However, while LLMs are powerful, they also have significant limitations and shortcomings. First, training an LLM from scratch is astronomically costly, requiring massive amounts of training data, expensive hardware, and high power consumption. This limits the capabilities of only a very few companies to build and customize a model from scratch. Second, the power of an LLM is often directly proportional to its parameter size (more powerful models have larger parameter sizes), resulting in high deployment and operating costs and limiting its application scale and breadth. Finally, LLMs rely primarily on publicly available datasets for training, meaning they lack direct access to private or specialized domain data. This leads to insufficient effectiveness and specialization in specific professional fields, failing to fully meet the needs of particular industries.

[0037] To address the aforementioned issues, the PEFT method has been proposed, a highly efficient tuning technique for extending the capabilities of a basic LLM (Low-Level Model). Its principle is to freeze most of the LLM's model parameters during the tuning phase, optimizing and updating only a small subset of parameters. This significantly reduces the computational cost and time of model tuning, as well as the amount of data required for training. This allows users to obtain small, enhanced modules through PEFT tuning with minimal cost and a small amount of specialized data. These modules work in conjunction with the LLM to improve its interactive performance in specific domains. PEFT technology enables users to quickly adapt LLM to specific tasks or domains at a relatively low cost, thereby greatly enhancing the flexibility of LLM in the face of constantly changing real-world needs.

[0038] In current application scenarios, due to numerous demands, a large number of PEFT modules are often available. When faced with a user's interaction request, determining which PEFT module to load for the LLM is primarily done manually. This is suitable for scenarios with fewer specialized areas and lower requirements for automation efficiency. However, in scenarios with many specialized areas and high efficiency requirements, manual methods are insufficient. How to automatically load the appropriate PEFT module for the LLM based on the current interaction request is a technical problem that needs to be solved.

[0039] In view of this, embodiments of this application provide an interaction method, an electronic device, a computer-readable storage medium, and a computer program product that can automatically load an adapted PEFT module for an LLM based on the user-input interaction context.

[0040] The interaction method provided in this application, based on user-inputted interaction context, inputs the interaction context and a preset system prompt word template into a large language model. The large language model then utilizes its powerful natural language processing capabilities to select candidate PEFT modules from a set of preset PEFT modules that highly match the interaction context, according to the instructions of the system prompt word template. The selected candidate PEFT modules are then output, and the target PEFT module to be loaded into the large language model is determined based on these candidate PEFT modules, achieving automatic selection and loading of PEFT modules. Furthermore, since the target PEFT module is determined based on the input interaction context, it can adapt well to the current interaction context scenario and provide customized interaction responses for different interaction contexts, improving the interaction performance and characteristics of the large language model.

[0041] Figure 1 A schematic diagram of an exemplary system architecture to which the technical solutions of the embodiments of this application can be applied is shown. For example... Figure 1As shown, the system architecture may include terminal devices and servers. In practical applications, terminal devices and servers can be connected directly or indirectly via wired or wireless communication, and this application does not impose any restrictions.

[0042] The server in this application embodiment can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as 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 big data and artificial intelligence platforms.

[0043] The terminal device in this application embodiment can be a smartphone, tablet computer, laptop computer, PDA, personal computer, smart TV, smartwatch, in-vehicle device, wearable device, smart voice interaction device, smart home appliance, aircraft, etc., but is not limited to these. The terminal device can run a client, which can run on the terminal device in the form of a browser, or in the form of a standalone application (APP), etc. The specific form of the client is not limited here.

[0044] When an interaction method provided in this application is applied to Figure 1 In the system architecture shown, a process can be as follows: First, the user inputs an interaction context (e.g., "Please recommend restaurants near location A") on the client app of the terminal device. The terminal device sends this user context to the server. After receiving the user's input interaction context, the server inputs the interaction context and a preset system prompt word template into a large language model deployed on the server. Guided by the system prompt word template, the large language model selects one or more candidate PEFT modules associated with "restaurants near location A" from a preset set of multiple PEFT modules based on the user's input interaction context, and outputs the labels of the candidate PEFT modules. Then, the target PEFT module is determined from the candidate PEFT modules. Next, the target PEFT module deployed on the server is called and loaded into the large language model to obtain a large language model with the module loaded. The interaction context is then input into the large language model with the module loaded, and the large language model with the module loaded outputs an interactive response carrying information about restaurants near location A. Finally, the server sends the interactive response to the terminal device.

[0045] It should be understood that Figure 1The number of terminal devices and servers shown are merely illustrative. Depending on implementation needs, there can be any number of terminal devices and any number of servers; that is, there can be multiple terminal devices and multiple servers.

[0046] It should be understood that the interaction methods provided in the embodiments of this application are generally executed by a server, and correspondingly, the interaction devices are generally located in the server. However, in other embodiments of this application, the terminal device may also have similar functions to the server, thereby executing the interaction methods provided in the embodiments of this application.

[0047] It should be understood that in the specific implementation of this application, data such as user attribute information is involved. When the above embodiments of this application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0048] Figure 2 A flowchart of an interaction method provided according to an embodiment of this application is shown, such as... Figure 2 As shown, the interaction method includes at least the following steps:

[0049] S110. Obtain the interaction context of the input.

[0050] In this embodiment of the application, the interaction context can be text data, image data, or voice data input by the user. The specific modality of the interaction context is not limited here.

[0051] S120. Input the interaction context and the preset system prompt word template into the large language model to obtain at least one candidate parameter for efficient fine-tuning of the PEFT module. The system prompt word template is used to instruct the large language model to select and output the candidate PEFT module from multiple preset PEFT modules according to the interaction context.

[0052] In this embodiment, the system prompt word template can be a preset template. The system prompt word template can include a PEFT module list and an output instruction, so that when a large language model receives the system prompt word template, it selects candidate PEFT modules from the PEFT module list according to the current input interaction context, and outputs a result carrying candidate PEFT module information according to the format of the output instruction.

[0053] The number of candidate PEFT modules output by a large language model can be one or more.

[0054] S130. Determine the target PEFT module based on the candidate PEFT modules, and load the target PEFT module into the large language model to obtain the large language model after loading the module.

[0055] In this embodiment of the application, the target PEFT module to be finally loaded into the large language model is determined based on the candidate PEFT modules output by the large language model.

[0056] For example, if the number of candidate PEFT modules output by the large language model is one, then the candidate PEFT module can be directly used as the target PEFT module; if the number of candidate PEFT modules output by the large language model is multiple, then the target PEFT module can be selected according to the weight coefficients corresponding to each candidate PEFT module.

[0057] S140. Input the interaction context into the large language model after loading the module to obtain the interaction response.

[0058] In this embodiment, after loading the target PEFT module into the large language model, fine-tuning of interactive capabilities is achieved. Furthermore, since the loaded target PEFT module is determined based on the interaction context, the large language model after loading the module has a stronger expressive capability for the current interaction context. Inputting the interaction context into the large language model after loading the module yields the expected interactive response. The interactive response is returned to the user in response to the user's input interaction context.

[0059] Figure 3 This application provides a schematic diagram of an interactive system architecture, as illustrated in an embodiment. Figure 3 As shown, the architecture of this interactive system involves the following four objects:

[0060] (1) The main control unit is responsible for the full-process implementation of user interaction, the generation and information maintenance of PEFT modules, and the updating and maintenance of system prompt word templates;

[0061] (2) System prompt word template, which includes a list of PEFT modules and output instructions, can affect the output content and format of the LLM when provided to it; the content of the system prompt word template is maintained and updated by the main control unit; the system prompt word template is sent to the LLM by the main control unit to control the output behavior of the LLM in a specific process;

[0062] (3) Large Language Model (LLM) provides the system with natural language processing and interaction capabilities. It can control its output behavior through additional system prompts. The main control unit sends the interaction context and obtains the response by calling its interface. The main control unit can load the PEFT module to it, so that it can obtain additional knowledge and capabilities, thereby affecting its response content in specific tasks.

[0063] (4) The PEFT module library is a centralized storage point for PEFT module files. PEFT modules are generated by the main control unit and sent here for storage. When the main control unit needs to load a specific PEFT module for the LLM, it reads the module file from this library and loads it. The main control unit can delete specific PEFT module files from this library as needed.

[0064] The following is combined Figure 3 The system architecture shown is used to exemplarily describe the interaction method provided in the embodiments of this application.

[0065] 11. The main control unit receives the interaction context from the external user;

[0066] 12. The main control unit reads the system prompt word template;

[0067] 13. The main control unit attaches the interaction context to the system prompt word template and sends it to the LLM;

[0068] 14. Influenced by the system prompt words, the LLM outputs one or more matching candidate PEFT module tags according to the interaction context and returns them to the main control unit. The main control unit obtains the weight and file path of the relevant modules from the internally maintained PEFT module information list based on the candidate PEFT module tags returned by the LLM, and then uses a selection algorithm to determine the final target PEFT module based on the weight of each module.

[0069] 15. The main control unit extracts the target PEFT module file from the PEFT module library;

[0070] 16. Load the LLM module based on the target PEFT module file;

[0071] 17. Once the target PEFT module is loaded, the main control unit will send the user's interaction context to the LLM again. At this time, the interaction context does not need to be attached to the system prompt word template.

[0072] 18. The LLM with the target PEFT module loaded outputs an expected interactive response based on the user's interaction context.

[0073] In the above process, the main control unit receives the user's interaction context and appends it to the system prompt word template before sending it to the LLM. Based on the system prompt word template, the LLM determines candidate PEFT modules according to the interaction context and returns candidate PEFT module tags to the main control unit. The main control unit determines the final target PEFT module from the candidate PEFT module tags returned by the LLM and loads the module into the LLM, thus achieving automatic loading of the PEFT module. The PEFT module loaded by the LLM can adapt to the needs of the interaction context, enabling the LLM to output more expressive interactive responses for that interaction context after loading the module.

[0074] According to an embodiment of this application, inputting the interaction context and a preset system prompt word template into a large language model to obtain at least one candidate PEFT module may include: embedding the interaction context into the system prompt word template to obtain a system prompt word template with embedded context; inputting the system prompt word template with embedded context into the large language model to obtain the output result of the large language model, wherein the output result includes module labels corresponding to at least one candidate PEFT module.

[0075] Figure 4 This diagram illustrates the data structure of a system prompt word template provided in an embodiment of this application. For example... Figure 4 As shown, the system prompt word template can include two parts: a PEFT module list and interaction context processing prompts. The PEFT module list includes module tags and context keywords corresponding to multiple PEFT modules. The interaction context processing prompts instruct the large language model to determine at least one candidate PEFT module based on the degree of matching between the interaction context and the context keywords corresponding to each PEFT module, and output the module tags corresponding to the candidate PEFT module according to a preset format.

[0076] In this embodiment, specific interaction processing and output format requirements are set for the LLM through system prompt word templates. Because the LLM itself possesses powerful natural language processing and understanding capabilities, it will process the current interaction context according to the requirements indicated by the system prompt word templates and output results that meet the format requirements.

[0077] Figure 5 This illustration shows a schematic diagram of a system prompt word template provided in an embodiment of this application. For example... Figure 5As shown, the system's prompt template provides interaction context handling prompts, requiring the LLM to extract only object names relevant to the current interaction context from the object list based on the template's current interaction context, and output a list of relevant object names. Here, the object list is the PEFT module list, the object names in the object list are module tags, and the keywords in the object list are context keywords. The user-input interaction context can be obtained from the "Current Interaction Context" field of the template. The LLM, based on this template, extracts object names relevant to the current context from the object list and outputs a list of relevant object names. For example... Figure 5 In the example interaction context [My facial skin hasn't been looking good lately, what are some ways to improve it?], LLM extracts the relevant object names from the object list based on this interaction context and outputs a list of relevant object names ["<Advertiser B Module>", "<Advertiser D Module>"].

[0078] According to an embodiment of this application, before inputting the interaction context and the preset system prompt word template into the large language model, the method may further include: obtaining module description information of the PEFT module; determining the context keywords corresponding to the PEFT module based on the module description information; and writing the module tag and context keywords corresponding to the PEFT module into the system prompt word template.

[0079] In this embodiment, the context keywords corresponding to the PEFT module can be extracted from the module description information of the PEFT module, which can be obtained by user input.

[0080] For example, a custom user can initiate a new PEFT module operation through the client and input the module description information for the new PEFT module. This module description information includes the context keywords of the newly added PEFT module. After receiving the context keywords of the newly added PEFT module, the main system adds a PEFT module information item to the PEFT module list in the system prompt word template, and writes the module tag and context keywords of the newly added PEFT module into the new PEFT module information item. A custom user can also initiate an update PEFT module information operation through the client. This operation updates the module description information of an existing PEFT module. The custom user inputs the updated module description information, which includes updated context keywords, and replaces the original context keywords in the template with the updated context keywords.

[0081] According to an embodiment of this application, determining a target PEFT module based on candidate PEFT modules may include: obtaining the weight coefficients corresponding to each candidate PEFT module; and determining the target PEFT module based on a preset module selection algorithm and the weight coefficients corresponding to the candidate PEFT modules.

[0082] It should be understood that each PEFT module is configured with a corresponding weight coefficient. Generally speaking, the higher the weight coefficient, the higher the probability of being selected as the target PEFT module. By configuring a weight coefficient for each PEFT module, the embodiments of this application can improve the control over the selection of PEFT modules and the flexibility of adaptation to different application scenarios.

[0083] In this embodiment, a preset module selection algorithm is used to calculate the weight coefficients corresponding to each candidate PEFT module, and the target PEFT module is determined based on the calculation results. The specific method for calculating the weight coefficients corresponding to each candidate PEFT module to determine the target PEFT module may depend on the actual business scenario requirements. For example, in some scenarios, a probabilistic selection algorithm (such as a random weighted selection algorithm) may be needed, in which case the weight coefficients can be used as probability factors in the calculation; while in other scenarios, a deterministic selection algorithm (such as a simple method of taking the maximum or minimum value) may be needed, in which case the weight coefficients can simply be used as the sorting criterion.

[0084] According to an embodiment of this application, before obtaining the weight coefficients corresponding to each candidate PEFT module, the method may further include: obtaining module description information of the input PEFT module; and determining the weight coefficients corresponding to the PEFT module based on the module description information.

[0085] For example, the module description information of a PEFT module can include priority information. The weight coefficient of the PEFT module is determined based on its priority information; generally, the higher the priority of the PEFT module, the higher its corresponding weight coefficient. As another example, the module description information of a PEFT module can also include popularity information, which represents the popularity value of the knowledge that the PEFT module focuses on. The weight coefficient of the PEFT module is determined based on its popularity information; generally, the higher the popularity of the knowledge that the PEFT module focuses on, the higher its corresponding weight coefficient. Furthermore, for PEFT modules applied in recommendation scenarios, the module description information can include recommendation revenue information. The weight coefficient of the PEFT module is determined based on the recommendation revenue information; generally, the higher the recommendation revenue of the PEFT module, the higher its corresponding weight coefficient.

[0086] According to an embodiment of this application, after determining the weight coefficient corresponding to the PEFT module based on the module description information, the method may further include: writing the weight coefficient corresponding to the PEFT module into a preset system module information table.

[0087] In one possible implementation of this application, the system module information table contains module information for all PEFT modules in the system. This module information may include the PEFT module's label and weight coefficients. The system module information table can be maintained and updated by the system's main control unit.

[0088] According to the embodiments of this application, obtaining the weight coefficients corresponding to each candidate PEFT module may include: searching for the weight coefficients corresponding to each candidate PEFT module from the system module information table.

[0089] When it is necessary to determine the target PEFT module from multiple candidate PEFT modules, the weight coefficients corresponding to each candidate PEFT module can be found in the system module information table. Then, the preset module selection algorithm calculates the weight coefficients corresponding to each candidate PEFT module and determines the target PEFT module based on the calculation results.

[0090] According to an embodiment of this application, loading a target PEFT module into a large language model may include: obtaining file path information corresponding to the target PEFT module; finding the module file corresponding to the target PEFT module based on the file path information; and loading the module file corresponding to the target PEFT module into the large language model.

[0091] After identifying the target PEFT module, the relevant module file can be found based on the file path information corresponding to the target PEFT module, and then the module file can be loaded into the LLM.

[0092] According to an embodiment of this application, before obtaining the file path information corresponding to the target PEFT module, the method may further include: obtaining customized training data of the input PEFT module; training the PEFT module based on the customized training data to obtain the module file of the PEFT module; saving the module file of the PEFT module and generating the file path information corresponding to the PEFT module.

[0093] PEFT modules can be trained using custom training data. Custom training data can incorporate knowledge specific to a particular task or domain, enabling the trained PEFT module to improve the performance of the LLM in that specific task or domain after being loaded into the LLM.

[0094] Customized training data can be input by the user on the client. For example, when a user initiates the operation of adding a new PEFT module through the client, they upload customized training data for the new PEFT module. The main control unit trains the new PEFT module based on the customized training data. After training, the module file of the new PEFT module is obtained and saved in the module library. At the same time, the file path information corresponding to the PEFT module is generated to facilitate finding the module file of the PEFT module based on the file path information.

[0095] In one possible scenario, when it is necessary to add knowledge points to the PEFT module (such as knowledge points related to new products or knowledge points related to hot issues), the custom user can initiate an update operation of the PEFT module through the client and upload the updated custom training data. The main control unit updates and trains the PEFT module based on the updated custom training data.

[0096] According to an embodiment of this application, after generating the file path information corresponding to the PEFT module, the method further includes: writing the file path information corresponding to the PEFT module into a preset system module information table.

[0097] After saving the module file corresponding to the PEFT module, the file path information of the PEFT module is written into the system module information table to facilitate the retrieval of the file path information corresponding to the target PEFT module through the system module information table.

[0098] According to an embodiment of this application, obtaining the file path information corresponding to the target PEFT module includes: searching for the file path information corresponding to the target PEFT module from the system module information table.

[0099] In one possible implementation of this application, the system module information table contains the tags and file path information of all PEFT modules in the system. This system module information table can be maintained and updated by the system's main control unit. After determining the tag of the target PEFT module, the corresponding file path information is searched based on the target PEFT module's tag, and then the module file corresponding to the target PEFT module is obtained based on the searched file path information.

[0100] Figure 6 A schematic diagram of a system module information table provided in an embodiment of this application is shown. Figure 6 As shown, the system module information table can include module information for multiple PEFT modules. This information includes, but is not limited to, module tags, weight coefficients, file path information, and context keywords. By maintaining the module information for all PEFT modules in the system through this table, it becomes easy to find PEFT module information and perform related operations.

[0101] It should be understood that when the PEFT modules in the system change (new PEFT modules or information on existing PEFT modules needs to be updated), the PEFT module information in the system module information table and system prompt word template involved in this application embodiment needs to be updated synchronously. The following is in conjunction with... Figure 7 The system architecture shown is used to exemplarily describe the PEFT module information update process provided in the embodiments of this application.

[0102] 21. The main control unit receives module-related data input by an external user. This module-related data includes customized training data and module description information for the PEFT module that the user needs to add.

[0103] 22. The main control unit assigns a module label (name or identifier ID, etc.) to the newly added PEFT module, and then uses customized training data to fine-tune PEFT based on LLM to obtain the PEFT module file. At the same time, it determines the relevant information of the module based on the module description information, such as weight coefficients and context keywords. Based on the module label, weight coefficients, file path information and context keywords of the newly added PEFT module, it updates the system module information table maintained internally by the system.

[0104] 23. Send the PEFT module file to the PEFT module library for storage. After the PEFT module file is saved, the file path information can be obtained. The system module information table maintained internally by the system is updated based on the file path information.

[0105] 24. Based on the module tags and context keywords of the newly added PEFT modules, update the PEFT module list information in the system prompt word template.

[0106] It should be understood that if it is necessary to update information of an existing PEFT module, such as updating weight coefficients or context keyword information, a similar process to the above can be used, which will not be elaborated here.

[0107] It should be understood that the solutions in this application embodiment may involve the following three user roles: service provider, responsible for implementing and managing the entire augmented interaction system; customized client, providing the system with customized training data and related descriptive information, based on which the system generates new PEFT modules and updates the system prompt word data and internal PEFT module information list, and finally the interactive response returned by the system to ordinary users will contain customized content that conforms to the current interactive context, which satisfies the requirements of customized clients; and ordinary user, who initiates interaction requests to the augmented interaction system and obtains responses to meet their daily life or work needs.

[0108] The solutions in this application are applicable to business scenarios where LLM is used to provide interactive services to a specific number of users, including:

[0109] The LLM used is a general capability model, which lacks professional knowledge and information specific to users. The embodiments of this application automatically select the PEFT module to fine-tune the LLM, thereby improving the performance of the LLM in specific tasks or fields.

[0110] There are multiple customers with customization needs. These customers want to provide customized interactive responses for different interactive contexts. The embodiments of this application realize customized PEFT modules and can automatically select the corresponding PEFT module for loading according to the input interactive context, so that the final output response of LLM can meet the customer's customization needs.

[0111] Each custom client can customize a PEFT module to meet their own needs by providing a small amount of training data related to their customized content, as well as relevant matching context information.

[0112] Service providers can meet the customization needs of each customized client with limited investment, providing customized interactive content that meets the client's expectations for different interactive contexts.

[0113] According to the solution of this application embodiment, there is no need to retrain or modify the basic LLM on a large scale, only a small PEFT tuning cost is required; it can make full use of the inherent capability of LLM to automatically select and load the appropriate PEFT module according to the interaction context, thereby improving the efficiency of the user interaction process and the response quality of the final output; by using the weight parameters of the PEFT module, the decision result of selecting the final PEFT module from the matching module label list can be easily interfered with by modifying the module configuration, without the need for retraining or tuning.

[0114] Figure 8 A schematic diagram of the architecture of an enhanced interaction device provided in an embodiment of this application is shown. Figure 8 As shown, the device includes: an interface submodule, a business logic submodule, a PEFT module management submodule, a system prompt word management submodule, an LLM interface call submodule, a PEFT fine-tuning training submodule, and a PEFT module loading submodule.

[0115] The interface submodule may include several API interfaces for the terminal device to interact with the interactive device.

[0116] The business logic submodule maintains the business logic flow of this application embodiment and can be used to drive other submodules to complete specific actions, such as the generation and maintenance of PEFT modules, the updating of system prompt word data, and the initiation of requests to LLM, etc.

[0117] The PEFT module management submodule is responsible for persistently saving and maintaining the PEFT module information list. This list stores information such as tags, weights, and file paths for all PEFT modules and dynamically updates its content based on requests from the business logic submodule. This module also saves the PEFT module files generated by the PEFT fine-tuning training submodule to the PEFT module library. After the business logic submodule drives the PEFT fine-tuning training submodule to train and obtain the PEFT module files, it informs the PEFT module management submodule of the temporary generation path of the module files. The PEFT module management submodule then copies the module files from the temporary generation path to the final PEFT module library and updates the PEFT module information list with the final storage path (i.e., file path information) of the modules.

[0118] The system prompt word management submodule stores system prompt word templates and updates them as needed based on requests from the business logic submodule. The system prompt word template consists of two main parts: a PEFT module list and interaction context handling prompts. The interaction context handling prompts generally remain unchanged, while the PEFT module list may change frequently. The business logic submodule sends the updated PEFT module list to the system prompt word management submodule, which then updates and saves the system prompt word template.

[0119] The LLM interface call submodule contains the interface for calling a specific LLM. The business logic submodule uses this interface to send interactive data to the LLM and receive responses.

[0120] The PEFT fine-tuning training submodule is used to generate PEFT modules. The business logic submodule sends the customized training data provided by the customer to the PEFT fine-tuning training submodule. Based on this training data, the PEFT fine-tuning training submodule uses PEFT technology to fine-tune the LLM to obtain the PEFT module file, and informs the business logic submodule of the temporary generation path of the module file.

[0121] The PEFT module loading submodule is used to load specified PEFT modules into the LLM. The business logic submodule retrieves the save path of the module file from the PEFT module management submodule based on the PEFT module tag to be loaded, and then informs the PEFT module loading submodule of the path. The PEFT module loading submodule then retrieves the corresponding module file from the PEFT module library based on the path and loads it into the LLM.

[0122] The following is based on Figure 8 The architecture of the enhanced interaction device shown is illustrated in the exemplary description of the processing flow of the newly added PEFT module provided in the embodiments of this application, such as... Figure 9As shown, in one possible embodiment of this application, the process for adding a PEFT module includes:

[0123] S301. The system's business logic submodule obtains the customized training data and module description information sent by the terminal device from the interface submodule.

[0124] S302. The business logic submodule assigns a PEFT module label (name or ID) to the newly added PEFT module and determines the context keywords, weights and other parameters of the module based on the module description information provided by the customer.

[0125] S303. The business logic submodule sends the customized training data to the PEFT fine-tuning training submodule to obtain the PEFT module file.

[0126] S304. The business logic submodule sends a first request to the PEFT module management submodule. The first request carries parameters such as the tag, weight and temporary file path of the newly added PEFT module.

[0127] After receiving the first request, the S305 PEFT module management submodule first copies the module file from the temporary storage path, stores the module file in the PEFT module library, determines the final file path (file path information), and then inserts the entry of the new module into the PEFT module information table it maintains based on parameters such as the module tag, weight and file path information of the new PEFT module, thus obtaining the updated PEFT module information table.

[0128] S306. The business logic submodule sends a second request to the system prompt word management submodule. The second request carries parameters such as the module label and context keywords of the newly added PEFT module.

[0129] S307. After receiving the second request, the system prompt word management submodule inserts the entry of the new module into the PEFT module list in the system prompt word template it maintains, and obtains the updated system prompt word template.

[0130] It should be noted that, as Figure 9 As shown, some of the above steps can be executed in parallel; for example, S304 and S306 can be executed in parallel. Through the above process, the creation of a PEFT module can be achieved.

[0131] The following is based on Figure 8 The architecture of the enhanced interaction device shown is illustrated in the exemplary description of the PEFT module removal process provided in the embodiments of this application, such as... Figure 10 As shown, in one possible embodiment of this application, the process for removing the PEFT module includes:

[0132] S401. The business logic submodule receives the PEFT module tag to be removed from the interface submodule.

[0133] S402. The business logic submodule sends a third request to the PEFT module management submodule, which carries the PEFT module tag to be removed.

[0134] After receiving the third request, the S403 PEFT module management submodule queries the path of the module file based on the tag, then finds the module file based on the path and deletes the found module file from the PEFT module library.

[0135] S404. After receiving a third request, the PEFT module management submodule removes the entry corresponding to the module tag from the PEFT module information table it maintains.

[0136] S405. The business logic submodule sends a fourth request to the system prompt word management submodule, which carries the PEFT module tag to be removed.

[0137] S406. After receiving the fourth request, the system prompt word management submodule removes the entry corresponding to the tag from the PEFT module list in the system prompt word template it maintains.

[0138] It should be noted that, as Figure 10 As shown, some of the above steps can be executed in parallel. For example, steps S402 and S405 can be executed in parallel, and steps S403 and S404 can also be executed in parallel. Through the above process, it is possible to respond to the user's command to remove the PEFT module.

[0139] The following is based on Figure 8 The architecture of the enhanced interaction device shown is illustrated in the following exemplary description of the process for updating PEFT module information provided in the embodiments of this application: Figure 11 As shown, in one possible embodiment of this application, the process for updating PEFT module information includes:

[0140] S501, The business logic submodule obtains the update data of the existing specific PEFT module from the interface submodule;

[0141] S502. Determine whether the training data of a specific module has changed based on the updated data. If so, execute S503 and S504.

[0142] S503, the PEFT fine-tuning training submodule obtains the updated PEFT module file based on the new training data;

[0143] S504, The PEFT module management submodule updates the PEFT module library files and the PEFT module information table;

[0144] S505. Determine whether the context description information of a specific PEFT module has changed based on the updated data. If so, execute S506.

[0145] S506. The System Prompt Word Management submodule updates the content of the corresponding entries in the PEFT module list in the System Prompt Word Template using the new context keywords;

[0146] S507. Determine whether the weight of a specific PEFT module has changed based on the updated data. If so, execute S508.

[0147] The S508 PEFT module management submodule updates the data of the corresponding entries in the PEFT module information list it maintains using the new weighting coefficients.

[0148] It should be noted that, as Figure 10 As shown, some of the above steps can be performed in parallel. Users can update the information of customized PEFT modules, including updating PEFT module files, updating PEFT module weights, and updating PEFT module context keywords; after receiving the user's update data, the system updates the PEFT module information maintained by the system through the above process.

[0149] The following is based on Figure 8 The architecture of the enhanced interaction device shown is illustrated in the example description of the enhanced interaction process provided in the embodiments of this application, such as... Figure 12 As shown, in one possible embodiment of this application, the enhanced interaction process includes:

[0150] S601. The business logic submodule receives an interaction request initiated by a regular user from the interface submodule.

[0151] S602. The business logic submodule obtains system prompts from the system prompt management submodule;

[0152] S603, the business logic submodule appends the user's original input to the system prompt and then sends it to the LLM (by calling the submodule through the LLM interface);

[0153] S604 and LLM output a list of candidate PEFT module tags based on the user's original input and the information contained in the system prompts, and return it to the business logic sub-module (which calls the sub-module through the LLM interface).

[0154] S605. The business logic submodule retrieves the specific parameters (including weight and file path) of the candidate modules from the PEFT module management submodule based on the candidate PEFT module tag list.

[0155] S606, The business logic submodule determines whether there is only one element in the candidate list or whether the weight parameter does not exist in the information of the PEFT module;

[0156] If yes, execute S607; otherwise, execute S608.

[0157] S607. Directly select the target enhancement module (choose unique or random selection);

[0158] S608. After evaluating each candidate module based on its weight, select the final PEFT module to be loaded. The evaluation algorithm can be customized according to specific business needs (e.g., using a probabilistic selection algorithm or a deterministic selection algorithm based on the weight value).

[0159] S609. After the business logic submodule selects the PEFT module to be loaded, it sends the selected module information to the PEFT module loading submodule, requesting it to load the PEFT module with the specified file path for the LLM.

[0160] S610. The business logic submodule sends the user's original input to the LLM loaded with the PEFT module through the LLM interface call submodule (note that the system prompt words will no longer be attached at this time);

[0161] S611. The business logic submodule obtains the final interactive response from the LLM that has loaded the PEFT module state through the LLM interface call submodule, and returns it to the ordinary user through the interface submodule.

[0162] The interaction method of this application embodiment can be applied to scenarios such as machine translation, question-answering systems, product promotion, or map navigation, and has wide applicability.

[0163] The following example of a product promotion scenario illustrates the interaction method provided in this application.

[0164] In a product promotion scenario, advertisers and AI (Artificial Intelligence) interaction service providers pre-agree on the payment price and promotional data format. Each advertiser then provides the service provider with relevant promotional data for their product (including but not limited to: product type, product name, product description, company name, product contextual keywords, etc.) and a quote. The system uses the promotional data as training data for PEFT modules and the quote as the weights for the PEFT modules. Based on the promotional data provided by the advertisers, the service provider uses LoRA (Low-Rank Adaptation, a PEFT technique) to fine-tune the LLM (Lower Rank Model), obtaining the PEFT module file corresponding to the advertiser and saving it to the PEFT module library. Upon obtaining the new PEFT module file and related information, the service provider synchronously updates the registered PEFT module information and system prompt keyword data.

[0165] When an external user initiates an AI interaction request, the system adds a system prompt before sending it to the LLM. The LLM processes the user's original input based on the information contained in the system prompt and outputs a list of candidate PEFT module tags. After obtaining the candidate PEFT module tag list from the LLM, the system searches and extracts the weights of all candidate modules from the internally maintained PEFT module information list, and then selects the final PEFT module to be loaded using a random weighted selection algorithm (the higher the weight, the higher the probability of selection). After selecting the PEFT module, the system extracts the corresponding module file path from the internally maintained PEFT module information list and loads it into the LLM, then directly passes the original user input to the LLM with the PEFT module loaded. The LLM with the PEFT module loaded analyzes and processes the original user input and returns the final output to the user. Due to the influence of the loaded PEFT module, the LLM's output will naturally include promotional content provided by a specific advertiser. New advertisers can join the AI ​​interaction service platform at any time, and existing advertisers can leave the platform at any time. In addition, the promotional content and pricing of each advertiser can change at any time. Therefore, AI interaction service providers continuously receive new promotional data, constantly update PEFT module files, weights, contextual keywords, and system prompts, and dynamically adjust the LLM's response behavior to different users to promptly meet the promotional expectations of all advertisers. For example... Figure 13 The image shown is an interactive example of a product promotion scenario provided in an embodiment of this application. Figure 13 As can be seen from the examples, the method provided in the embodiments of this application can naturally promote products to users during interactions, thereby increasing the probability of successful product promotion.

[0166] This application also provides an electronic device, such as... Figure 14 As shown, the electronic device 1400 includes:

[0167] One or more processors 1410;

[0168] Memory 1420, which stores one or more programs, causes one or more processors 1410 to perform the following when the one or more programs are executed:

[0169] Such as the interaction method applied to the first node of the second network.

[0170] Memory 1420, as a non-transitory network system, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory 1420 may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory 1420 may optionally include remotely located memories 1420 relative to processor 1410, which can be connected to processor 1410 via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0171] The memory 1420 can be implemented as a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM). The memory 1420 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 1420 and is called and executed by the processor 1410.

[0172] The processor 1410 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application.

[0173] In some embodiments, the electronic device further includes:

[0174] Input / output interfaces are used to implement information input and output;

[0175] The communication interface is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0176] The bus transmits information between various components of the device (such as processor 1410, memory 1420, input / output interfaces, and communication interfaces);

[0177] The processor 1410, memory 1420, input / output interface, and communication interface can communicate with each other within the device via a bus.

[0178] One embodiment of this application also provides a computer-readable storage medium storing computer-executable instructions for executing interactive methods that implement any embodiment of this application.

[0179] An embodiment of this application also provides a computer program product, including a computer program or computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer program or computer instructions from the computer-readable storage medium and executes the computer program or computer instructions, causing the computer device to perform an interactive method that implements any embodiment of this application.

[0180] The system architecture and application scenarios described in this application are intended to more clearly illustrate the technical solutions of this application and do not constitute a limitation on the technical solutions provided in this application. Those skilled in the art will understand that as system architectures evolve and new application scenarios emerge, the technical solutions provided in this application are also applicable to similar technical problems.

[0181] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0182] It will be understood by those skilled in the art that all or some of the steps and systems in the methods disclosed above can be implemented as software, firmware, hardware, and suitable combinations thereof. Some or all of the physical components can be implemented as software executed by a processor, such as a central processing unit, digital signal processor, or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit. Such software can be distributed on a computer-readable medium, which can include computer storage media (or non-transitory media) and communication media (or transient media). As is known to those skilled in the art, the term computer storage media includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data). Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disc (DVD) or other optical disc storage, magnetic cartridges, magnetic tape, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and is accessible to a computer. Furthermore, as is known to those skilled in the art, communication media typically contain computer-readable instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms, and may include any information delivery medium.

[0183] The above description, with reference to the accompanying drawings, illustrates some embodiments of this application, but does not limit the scope of this application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of this application shall be within the scope of this application.

Claims

1. An interaction method, the method comprising: obtaining an input interaction context; inputting the interaction context and a preset system prompt word template into a large language model to obtain at least one candidate parameter efficient fine-tuning (PEFT) module, wherein the system prompt word template is used to instruct the large language model to select and output the candidate PEFT module from a plurality of preset PEFT modules according to the interaction context; determining a target PEFT module according to the candidate PEFT module, and loading the target PEFT module into the large language model to obtain a large language model after loading the module; inputting the interaction context into the large language model after loading the module to obtain an interaction response.

2. The method of claim 1, wherein, The determining of the target PEFT module according to the candidate PEFT module comprises: obtaining a weight coefficient corresponding to each candidate PEFT module; determining the target PEFT module according to a preset module selection algorithm and the weight coefficient corresponding to the candidate PEFT module.

3. The method of claim 2, wherein, Before obtaining the weight coefficient corresponding to each candidate PEFT module, the method further comprises: obtaining module description information of the input PEFT module; determining the weight coefficient corresponding to the PEFT module according to the module description information.

4. The method of claim 3, wherein, After determining the weight coefficient corresponding to the PEFT module according to the module description information, the method further comprises: writing the weight coefficient corresponding to the PEFT module into a preset system module information table; 5. The method of claim 1, wherein, The obtaining of the weight coefficient corresponding to each candidate PEFT module comprises: finding the weight coefficient corresponding to each candidate PEFT module from the system module information table. The loading of the target PEFT module into the large language model comprises: obtaining file path information corresponding to the target PEFT module; 6. The method of claim 5, wherein, finding a module file corresponding to the target PEFT module according to the file path information; loading the module file corresponding to the target PEFT module into the large language model. Before obtaining the file path information corresponding to the target PEFT module, the method further comprises: obtaining customized training data of the input PEFT module; 7. The method of claim 6, wherein, training the PEFT module based on the customized training data to obtain a module file of the PEFT module; saving the module file of the PEFT module and generating file path information corresponding to the PEFT module.

8. The method of claim 1, wherein, After generating the file path information corresponding to the PEFT module, the method further comprises: writing the file path information corresponding to the PEFT module into a preset system module information table; The obtaining of the file path information corresponding to the target PEFT module comprises: finding the file path information corresponding to the target PEFT module from the system module information table. The inputting of the interaction context and the preset system prompt word template into the large language model to obtain at least one candidate PEFT module comprises: embedding the interaction context into the system prompt word template to obtain a system prompt word template with embedded context; inputting the embedded context system prompt word template into the large language model to obtain an output result of the large language model, the output result including a module label corresponding to at least one candidate PEFT module.

9. The method of claim 1, wherein, The system prompt word template includes a PEFT module list and interactive context processing prompt information. The PEFT module list includes module labels and context keywords corresponding to a plurality of PEFT modules. The interactive context processing prompt information is used to instruct the large language model to determine at least one candidate PEFT module according to a matching degree of the interactive context and the context keywords corresponding to each PEFT module, and output the module label corresponding to the candidate PEFT module in a preset format.

10. The method of claim 9, wherein, Before inputting the interactive context and the preset system prompt word template into the large language model, the method further includes: obtaining module description information of the PEFT module; determining the context keywords corresponding to the PEFT module according to the module description information; writing the module label and the context keywords corresponding to the PEFT module into the system prompt word template. 11.An electronic device, comprising: one or more processors; a memory having one or more programs stored thereon, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the interactive method of any one of claims 1-10. 12.A computer-readable storage medium having a computer program stored thereon, wherein when the program is executed by a processor, the program implements the interactive method of any one of claims 1-10. 13.A computer program product, comprising a computer program, wherein when the computer program is executed by a processor, the computer program implements the interactive method of any one of claims 1-10.

Images