Model updating method, electronic equipment, storage medium and program product

Abstract

The invention discloses a model updating method, electronic equipment, a storage medium and a program product, and relates to the technical field of machine learning, and the method comprises the steps: obtaining a specified feedback data set under the condition of monitoring that a to-be-updated dialogue model exists in a group of dialogue models, performing feature point matching on the data set feature point sequence of the specified feedback data set and the reference feature point sequence of the pre-training model in the group of pre-training models to obtain a feature point matching result, and determining a specified parameter adjustment mode from a group of parameter adjustment modes according to the feature point matching result, therefore, model parameter updating is carried out on the to-be-updated dialogue model according to the specified feedback data set and the specified parameter adjustment mode, so that updating of the model better meets actual requirements, invalid or negative tuning is avoided, the technical problem that the updating effect of the model is poor in the prior art is solved, and the user experience is improved. The technical effect of stably improving the performance of the dialogue model is achieved.

Description

Technical Field

[0001] This application relates to the technical field of machine learning, and particularly to a method for updating a model, an electronic device, a storage medium, and a program product. Background Art

[0002] With the advent of the era of artificial intelligence, along with the popularization of artificial intelligence platforms and big data model training, and the large-scale launch of training and inference integrated machines, the requirement for the model tuning ability of artificial intelligence platforms becomes increasingly important. During the use of large models in training and inference integrated machines, the dialogue model needs to be continuously updated to adapt to changes in user needs and improve service quality. However, in most of the related technology model update strategies, either completely rely on a predefined general dataset, ignoring the value of real-time user feedback; or are too simple and rough when processing user feedback data, unable to accurately capture the subtle changes in user needs, resulting in an unclear performance improvement or even a decline in performance when updating the model.

[0003] Therefore, there is a technical problem in the related technology that the update effect of the model is not good. Summary of the Invention

[0004] This application provides a method for updating a model, an electronic device, a storage medium, and a program product to at least solve the problem that the update effect of the model in the related technology is not good.

[0005] This application provides a method for updating a model, which is applied to an artificial intelligence platform, and a group of dialogue models are deployed on the artificial intelligence platform; the method includes:

[0006] When it is monitored that there is a dialogue model to be updated in a group of dialogue models, obtain a specified feedback dataset, where the specified feedback data in the specified feedback dataset is the feedback data obtained from historical conversations using the dialogue model to be updated;

[0007] Perform feature point matching on the dataset feature point sequence of the specified feedback dataset and the benchmark feature point sequence of a pre-trained model in a group of pre-trained models to obtain a feature point matching result, where the dataset feature point sequence is used to characterize the attribute information of the specified feedback dataset;

[0008] According to the feature point matching result, determine a specified parameter adjustment method from a group of parameter adjustment methods, where the parameter adjustment methods in the group of parameter adjustment methods are methods for adjusting the model parameters of the dialogue models in the group of dialogue models;

[0009] Update the model parameters of the dialogue model to be updated according to the specified feedback dataset and the specified parameter adjustment method.

[0010] The present application also provides an updating device for a model, which is applied to an artificial intelligence platform, and a set of dialogue models are deployed on the artificial intelligence platform; the device includes:

[0011] An acquisition module, configured to acquire a specified feedback data set when it is monitored that there is a dialogue model to be updated in the set of dialogue models, wherein the specified feedback data in the specified feedback data set is the feedback data obtained by using the dialogue model to be updated for historical conversations;

[0012] A matching module, configured to perform feature point matching on the data set feature point sequence of the specified feedback data set and the benchmark feature point sequence of a pre-trained model in a set of pre-trained models to obtain a feature point matching result, wherein the data set feature point sequence is used to characterize the attribute information of the specified feedback data set;

[0013] A determination module, configured to determine a specified parameter adjustment method from a set of parameter adjustment methods according to the feature point matching result, wherein the parameter adjustment method in the set of parameter adjustment methods is a method for adjusting the model parameters of the dialogue models in the set of dialogue models;

[0014] An updating module, configured to update the model parameters of the dialogue model to be updated according to the specified feedback data set and the specified parameter adjustment method.

[0015] The present application also provides an electronic device, including: a memory, configured to store a computer program; a processor, configured to implement the steps of any of the above model updating methods when executing the computer program.

[0016] The present application also provides a computer-readable storage medium, in which a computer program is stored, and when the computer program is executed by a processor, the steps of any of the above model updating methods are implemented.

[0017] The present application also provides a computer program product, including a computer program, and when the computer program is executed by a processor, the steps of any of the above model updating methods are implemented.

[0018] Through the present application, in the case where it is detected that there is a dialogue model to be updated in a set of dialogue models, a specified feedback data set is obtained, and feature point matching is performed between the data set feature point sequence of the specified feedback data set and the benchmark feature point sequence of the pre-trained model in a set of pre-trained models to obtain a feature point matching result. Then, according to the feature point matching result, a specified parameter adjustment method is determined from a set of parameter adjustment methods. Thus, according to the specified feedback data set and the specified parameter adjustment method, the model parameters of the dialogue model to be updated are updated. By performing feature point matching between the data set feature point sequence of the specified feedback data set and the benchmark feature point sequence of the pre-trained model in a set of pre-trained models and determining the specified parameter adjustment method according to the feature point matching result, the dialogue model to be updated is updated, making the update of the model more in line with actual needs, avoiding ineffective or negative tuning, and ensuring the performance of the updated dialogue model to a certain extent. This solves the technical problem of poor update effect of the model in the related art and achieves the technical effect of stable improvement of the performance of the dialogue model. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] To more clearly illustrate the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0020] Figure 1 FIG. is a schematic diagram of an application scenario of a method for updating a model provided by an embodiment of the present application.

[0021] Figure 2 FIG. is a schematic flowchart of an optional method for updating a model provided by an embodiment of the present application.

[0022] Figure 3 FIG. is a schematic diagram of an optional method for determining a specified parameter adjustment method provided by an embodiment of the present application.

[0023] Figure 4 FIG. is a schematic diagram of an optional method for updating a model provided by an embodiment of the present application.

[0024] Figure 5 FIG. is a block diagram of the structure of an optional model update device provided by an embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0025] The following will clearly and completely describe the technical solutions in the embodiments of the present application with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the protection scope of the present application.

[0026] It should be noted that in the description of the present application, the terms "include", "comprise" or any other variant thereof are intended to cover a non-exclusive inclusion, such that a process, method, article or device including a series of elements not only includes those elements but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device. The terms "first", "second", etc. in the present application are used to distinguish similar objects and are not used to describe a specific order or sequence.

[0027] In order to enable those skilled in the art of the present technology to better understand the solution of the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0028] With the advent of the era of artificial intelligence, along with the popularization of artificial intelligence platforms and big data model training, and the large-scale listing of training and inference integrated machines, the requirement for the model optimization ability of artificial intelligence platforms becomes increasingly important. During the use of large models in training and inference integrated machines, the dialogue model needs to be continuously updated to adapt to the changes in user needs and improve service quality. However, in related technologies, most model update strategies either completely rely on predefined general datasets, ignoring the value of real-time user feedback; or are too simple and rough when processing user feedback data, unable to accurately capture the subtle changes in user needs, resulting in insignificant model update performance or even performance degradation. Therefore, there is a problem of poor model update effect in related technologies.

[0029] To solve the above problems, the embodiments of the present application provide a method for updating a model. Optionally, in this embodiment, the above method for updating a model can be but is not limited to being applied to, for example

[0030] According to one aspect of the embodiments of the present application, a method for updating a model is provided. Optionally, in this embodiment, the above method for updating a model can be but is not limited to being applied to, such as Figure 1In the hardware environment including the terminal device 102 and the server 104 as shown. The server 104 can be connected to the terminal device 102 through a network, and can be used to provide services (such as application services, etc.) for the terminal device 102 or the client installed on the terminal device 102. A database can be set up on the server 104 or independently of the server 104 to provide data storage services for the server 104.

[0031] The above network can include but is not limited to at least one of the following: wired network, wireless network. The above wired network can include but is not limited to at least one of the following: wide area network, metropolitan area network, local area network. The above wireless network can include but is not limited to at least one of the following: Wireless Fidelity (WIFI for short), Bluetooth. The terminal device 102 can be but is not limited to a personal computer (PC for short), mobile phone, tablet computer, etc. The server 104 can be but is not limited to a cloud server, server cluster or other server types.

[0032] The model update method of the embodiment of the present application can be executed by the server 104, or can be executed by the terminal device 102, or can also be jointly executed by the server 104 and the terminal device 102. Among them, the execution of the model update method of the embodiment of the present application by the terminal device 102 can also be executed by the client installed on it.

[0033] Taking the execution of the model update method in this embodiment by the terminal device 102 as an example, this model update method is applied to an artificial intelligence platform, where a set of dialogue models are deployed on the artificial intelligence platform. Figure 2 It is a schematic flowchart of an optional model update method according to an embodiment of the present application, as Figure 2 shown, the process of this method can include the following steps:

[0034] Step S202, when it is monitored that there is a dialogue model to be updated in a set of dialogue models, obtain a specified feedback data set, where the specified feedback data in the specified feedback data set is the feedback data obtained from historical conversations using the dialogue model to be updated.

[0035] It should be noted that the model update method in this embodiment can be applied to an Artificial Intelligence (AI) platform environment, which is a platform for automatically and personalized updating a set of deployed dialogue models. The AI platform can refer to a system platform with the ability to process large-scale data, train, and run AI models. For example, a high-performance training and inference integrated machine platform. In this embodiment, the AI platform can deploy, manage, and update various types of dialogue models. Multiple agent applications can be deployed in the AI platform, and each agent application in the multiple agent applications deploys a dialogue model.

[0036] A set of dialogue models can be multiple AI models deployed on the AI platform. A set of dialogue models is mainly used for natural language processing tasks, such as chatbots, voice assistants, etc., and can interact with users in dialogue.

[0037] Optionally, the way to determine whether there is a dialogue model to be updated in a set of dialogue models can identify metrics such as latency and accuracy fluctuations when the model processes requests through real-time performance monitoring. Once it is found that the feedback score of a certain dialogue model is lower than the average threshold or the performance metric drops, the dialogue model is marked as a state to be updated.

[0038] Or, in the case where the data volume of the specified feedback data is greater than the preset threshold, it is determined that there is a dialogue model to be updated in a set of dialogue models. Or, for each dialogue model, a threshold of the feedback data volume is set, and the feedback data received by each dialogue model cumulatively up to the current time is recorded. In the case where the data volume of the feedback data is greater than the quantity threshold corresponding to the dialogue model, the dialogue model with the data volume of the feedback data greater than the quantity threshold corresponding to the dialogue model is determined as the dialogue model to be updated.

[0039] Optionally, in the case where it is determined that there is a dialogue model to be updated in a set of dialogue models, a specified feedback data set is obtained. The specified feedback data set includes multiple specified feedback data. The specified feedback data can be obtained by collecting feedback information during the historical dialogue process between the user and the dialogue model in the platform. The feedback information includes, but is not limited to, the satisfaction score of the user with the reply of the dialogue model, correction suggestions, or direct negative feedback.

[0040] Step S204: Perform feature point matching on the data set feature point sequence of the specified feedback data set and the benchmark feature point sequence of the pre-trained model in a set of pre-trained models to obtain a feature point matching result, where the data set feature point sequence is used to characterize the attribute information of the specified feedback data set.

[0041] It should be noted that the dataset feature point sequence of the specified feedback dataset may include a series of key attribute information extracted for the specified feedback dataset. For example, dialogue length, topic diversity, sentiment tendency, etc., which are used to quantify and characterize the characteristics of the dataset.

[0042] Optionally, the dataset feature point sequence can be used to characterize the attribute information of the specified feedback dataset. The attribute information includes at least one of the following information: data statistical information, data type information, data complexity information, and application scenario information of the dataset. Among them, the data statistical information can be the basic statistical data of the specified feedback dataset, including but not limited to the size of the data volume, the average dialogue length, the question frequency, the user activity, etc. The data statistical information can represent the scale and structural characteristics of the dataset. The data type information can be used to represent the types of data included in the specified feedback dataset. Such as text data, voice data, image data or mixed data, etc. The data complexity information can include the dimension of the data in the specified feedback dataset, the similarity of the data in the specified feedback dataset, etc. Specifically, the data complexity information can describe the index of the organizational level and diversity of the internal information of the dataset, including semantic complexity, topic breadth and depth, sentiment fluctuation range, etc. The application scenario information of the dataset can specifically indicate the actual application scenarios from which the dataset is derived, such as customer service Q&A, medical consultation, education tutoring, etc.

[0043] There is a model repository deployed in the artificial intelligence platform. Among them, a set of pre-trained models are set in the model repository. The pre-trained models can be deep learning models pre-trained on a large amount of general data (such as books, news, etc.). There is a hierarchical relationship of "foundation" to "specialization" between the pre-trained models and the dialogue models. Through further fine-tuning, they are made more suitable for specific dialogue scenarios or tasks, such as customer service conversations, intelligent Q&A, etc. The pre-trained models provide powerful initial capabilities for the dialogue models, while the dialogue models further optimize the performance of the pre-trained models by learning the dialogue data of specific domains or users.

[0044] For each pre-trained model in a set of pre-trained models, there exists a corresponding benchmark feature point sequence. The benchmark feature point sequence represents a set of inherent performance indicators of the pre-trained model in the feature domain or task. The benchmark feature point sequence includes a set of benchmark feature points, and these benchmark feature points can cover various attributes of the pre-trained model, such as the complexity of the model (number of layers, number of parameters), data preference (pictures, texts, codes, etc.), processing speed and accuracy, etc.

[0045] Optionally, perform feature point matching on the dataset feature point sequence of the specified feedback dataset and the benchmark feature point sequence of the pre-trained model in a group of pre-trained models to obtain a feature point matching result. Specifically, extract the attribute information of the specified feedback dataset to obtain the dataset feature point sequence of the specified feedback dataset; perform feature point matching based on the dataset feature point sequence and the benchmark feature point sequence of the pre-trained model in a group of pre-trained models, and find the pre-trained model that is closest to or most matches the specified feedback dataset from the group of pre-trained models.

[0046] The feature point matching result may include multiple groups of matching results. Among them, each matching result in the multiple matching results is the matching result of the specified feedback dataset and a pre-trained model in a group of pre-trained models.

[0047] Step S206, determine the specified parameter adjustment method according to the feature point matching result, where the parameter adjustment method in a group of parameter adjustment methods is a method for adjusting the model parameters of the dialogue model in a group of dialogue models.

[0048] It should be noted that a group of parameter adjustment methods can be a preset group of methods for adjusting model parameters, which can include various different fine-tuning strategies. For example, learning rate adjustment, regularization parameter setting, hierarchical learning rate, Low-Rank Adaptation (LoRA) fine-tuning, etc. Each strategy optimizes different aspects of the model parameters to adapt to specific dataset and task requirements.

[0049] Specifically, in the case where the parameter adjustment method is learning rate adjustment, the learning rate can refer to the control factor of the parameter update step size during model training. During fine-tuning, appropriately adjusting the learning rate can help the model converge to the optimal solution faster, avoiding overfitting or under-training. Usually, the platform dynamically adjusts the learning rate, such as adopting a learning rate decay strategy. Initially, a relatively high learning rate is set so that the model can quickly explore the solution space, and then it is gradually decreased, enabling the model to more finely adjust the parameters and achieve a more refined optimization effect.

[0050] In the case where the parameter adjustment method is regularization parameter setting, the overly complex model is punished by adding a regularization term to the loss function. During the fine-tuning process, reasonably setting the L1 and L2 regularization parameters can control the magnitude of the model parameters, avoid the model from overfitting the noise data, and improve the generalization ability of the model. The selection of the regularization parameter usually depends on the size and complexity of the dataset, as well as the capacity of the model itself.

[0051] In the case where the parameter adjustment method is hierarchical learning rate, different learning rates are set according to the characteristics of different model layers. For example, for a deep model, the learning rate near the input layer may be set lower because this part of the model layer has learned more general feature representations, while the learning rate near the output layer will be higher to better adapt to the details of specific tasks. Hierarchical learning rate helps to balance the overall update and local optimization of the model, improving training efficiency and model performance.

[0052] In the case where the parameter adjustment method is LoRA fine-tuning, a small part of the parameters in the pre-trained model are adjusted to adapt to the new task by means of low-rank matrix factorization. This method saves a large amount of computing resources and storage space, and is particularly suitable for fine-tuning large pre-trained models in resource-constrained environments.

[0053] Optionally, according to the feature point matching result, a specified parameter adjustment method can be determined from a set of parameter adjustment methods. The specified parameter adjustment method can be considered the most suitable for the current dataset and the dialogue model to be optimized.

[0054] Step S208, update the model parameters of the dialogue model to be updated according to the specified feedback dataset and the specified parameter adjustment method.

[0055] Optionally, use the specified feedback dataset and the selected specified parameter adjustment method to fine-tune the parameters of the dialogue model to be updated to improve the performance of the model and make it serve better.

[0056] In one example, assume that a dialogue model based on the Transformer architecture is deployed on an artificial intelligence platform. After the platform monitors that users frequently give low quality scores to the responses of this dialogue model or when the number of feedback data for this dialogue model is greater than a threshold, a specified feedback dataset is obtained. Subsequently, the artificial intelligence platform can obtain a dataset feature point sequence according to the specified feedback dataset. Specifically, the dataset feature point sequence can include the complexity of the dialogue context and the topic distribution of user questions. Then, it is matched with the benchmark feature point sequence in a set of pre-trained models to find the closest model and parameter adjustment method. The dataset feature point sequence is most closely matched with the benchmark feature point sequence of the pre-trained model A, and the pre-trained model A used "hierarchical learning rate" to improve the generalization ability of the model. Therefore, "hierarchical learning rate" can be used as the specified parameter adjustment method to re-train this dialogue model and adjust its parameters. Through this embodiment, it is possible to avoid manual intervention on an artificial intelligence platform, greatly shorten the time cycle from user feedback to model update, and improve the response speed and service quality of the platform.

[0057] Through the embodiments of the present application, when it is detected that there is a dialogue model to be updated in a group of dialogue models, a specified feedback data set is obtained, and feature point matching is performed on the data set feature point sequence of the specified feedback data set and the benchmark feature point sequence of the pre-trained model in a group of pre-trained models to obtain a feature point matching result. Then, according to the feature point matching result, a specified parameter adjustment method is determined from a group of parameter adjustment methods. Thus, according to the specified feedback data set and the specified parameter adjustment method, the model parameters of the dialogue model to be updated are updated. By performing feature point matching on the data set feature point sequence of the specified feedback data set and the benchmark feature point sequence of the pre-trained model in a group of pre-trained models and determining the specified parameter adjustment method according to the feature point matching result, the dialogue model to be updated is updated, making the model update more in line with actual needs, avoiding ineffective or negative tuning, and ensuring the performance of the updated dialogue model to a certain extent. This solves the technical problem of poor update effect of the model in the related art and achieves the technical effect of stable improvement of the performance of the dialogue model.

[0058] In an exemplary embodiment, step S204 includes: calculating the difference values between the data set feature point sequence and the benchmark feature point sequence of the pre-trained model in a group of pre-trained models respectively to obtain a set of difference values corresponding to the group of pre-trained models, where the feature point matching result includes the set of difference values corresponding to the group of pre-trained models.

[0059] It should be noted that the data set feature point sequence can be an ordered set of a series of numerical features for the specified feedback data set. The data set feature point sequence can include the size of the specified feedback data set, data types (such as text, image, language, etc.), distribution characteristics of the data (such as topic diversity), and so on. The data set feature point sequence can represent the attribute information of the specified feedback data set.

[0060] The benchmark feature point sequence can represent a set of inherent performance indicators of the pre-trained model in the feature field or task. The benchmark feature point sequence includes a group of benchmark feature points, and these benchmark feature points can cover various attributes of the pre-trained model, such as the complexity of the model (number of layers, number of parameters), data preference (pictures, text, code, etc.), processing speed, and accuracy.

[0061] Optionally, during the feature point matching process, the differences between the data set feature point sequence and the benchmark feature point sequence of each pre-trained model in a group of pre-trained models can be calculated, such as distance, similarity, or variance, etc. Each pre-trained model corresponds to a difference value, and these difference values form a set of difference values.

[0062] Through this embodiment, by feature point matching and calculation of the difference value set, the best match between the model and the data set can be accurately identified, so as to perform targeted adjustment of model parameters, avoiding the risk of performance degradation caused by blind tuning. The updated dialogue model can show higher accuracy and response speed when processing data, improving the dialogue quality of the dialogue model.

[0063] In an exemplary embodiment, before performing feature point matching on the data set feature point sequence of the specified feedback data set and the benchmark feature point sequence of the pre-trained model in a group of pre-trained models, the above method further includes: extracting the initial feature point set of the specified feedback data set; performing feature point sorting on the initial feature points in the initial feature point set to obtain an initial feature point sequence; and determining a subsequence containing the first specified number of initial feature points in the initial feature point sequence as the data set feature point sequence.

[0064] It should be noted that the initial feature point set can be a series of numerical features initially extracted from the specified feedback data set. The numerical features in the initial feature point set can cover multiple different dimensions in the specified feedback data set, such as data diversity, dialogue length, complexity, topic distribution, sentiment tendency, and so on.

[0065] Optionally, extracting the initial feature point set of the specified feedback data set may include extracting the initial feature points in the specified feedback data set according to a preset feature extraction model. Specifically, different preset feature extraction models can be selected for feature extraction according to the data type in the specified feedback data set.

[0066] Optionally, according to a preset evaluation criterion, the initial feature points in the initial feature point set are sorted to obtain an ordered initial feature point sequence. Specifically, the preset evaluation criterion can be sorted based on the importance of the initial features in different types of dialogue models. Of course, it can also be sorted according to the frequency of occurrence of the initial feature points, the relevance of the initial feature points, and the importance of the initial feature points.

[0067] Optionally, from the initial feature point training, initial feature point screening is performed to obtain a subsequence of the first specified number of initial feature points in the initial feature point sequence, that is, the data set feature point sequence. It should be noted that the subsequence of the first specified number of initial feature points in the initial feature point sequence can be considered the most representative and decisive initial feature points. The size of the specified number can be set based on actual needs and will not be elaborated herein.

[0068] In one example, in an artificial intelligence platform, a specified feedback dataset is collected to update a specific dialogue model in a set of dialogue models, that is, the dialogue model to be updated. An initial set of feature points in the specified feedback dataset is extracted, and the initial feature points in the initial set of feature points are sorted. For example, they can be sorted according to the degree of influence of the initial feature points on the model performance. For example, semantic complexity and user satisfaction scores are identified as the feature points with the greatest influence on model tuning and are ranked at the front of the sequence. The first n initial feature points (e.g., the first 5) after sorting are selected, and these feature points will be determined as the dataset feature point sequence. Key metrics such as semantic complexity, user satisfaction score, and dialogue length can be included in this sequence. Finally, the dataset feature point sequence is compared and matched with the benchmark feature point sequence of the pre-trained model to find the pre-trained model that is the closest, so as to determine the corresponding fine-tuning strategy, thereby updating the specific dialogue model to better respond to user feedback and requirements.

[0069] Through this embodiment, before performing feature point matching, an initial set of feature points in the specified feedback dataset is extracted, and the initial feature points in the initial set of feature points are sorted to select the first specified number of initial feature points to form a data feature point sequence, realizing the focus on the key feature points of the dataset. Model tuning can more accurately identify and solve the main problems in the dataset, avoiding adjustments to irrelevant features, and improving the pertinence and effect of tuning. Moreover, by selecting the most relevant feature points for processing, the waste of computing resources is reduced, and the efficiency of the feature point matching process is improved.

[0070] In an exemplary embodiment, the difference values between the dataset feature point sequence and the benchmark feature point sequences of the pre-trained models in a set of pre-trained models are calculated respectively to obtain a set of difference values corresponding to the set of pre-trained models, including:

[0071] The pre-trained models in a set of pre-trained models are respectively used as the current pre-trained model to perform the following matching operations to obtain a set of difference values corresponding to the pre-trained models in the set of pre-trained models, where the benchmark feature point sequence of the current pre-trained model is the current benchmark feature point sequence:

[0072] Feature point matching is performed on the dataset feature point sequence and the current benchmark feature point sequence to obtain a set of current matching pairs. One current matching pair in the set of current matching pairs includes an initial feature point in the dataset feature point sequence and a current benchmark feature point in the current benchmark feature point sequence; the difference values between the two feature points of the current matching pairs in the set of current matching pairs are calculated respectively to obtain the difference value corresponding to the current pre-trained model; where the set of difference values corresponding to the set of pre-trained models includes the difference value corresponding to the current pre-trained model in the set of pre-trained models.

[0073] It should be noted that each pre-trained model in a group of pre-trained models may have different performance characteristics due to differences in architecture, training data, or parameter settings. The current pre-trained model can be the pre-trained model selected from a group of pre-trained models for feature matching. The current pre-trained model can have a pre-set sequence of benchmark feature points, that is, the current sequence of benchmark feature points. The sequence of benchmark feature points of the current pre-trained model can be used to perform feature point matching with the sequence of dataset feature points to obtain a set of current matching pairs.

[0074] The current matching pair can be a pair composed of two feature points with a high degree of matching in the sequence of dataset feature points and the current sequence of benchmark feature points of the current pre-trained model, and can be used to reflect the matching degree of the sequence of dataset feature points of the specified feedback dataset and the pre-trained model in the feature dimension. Specifically, the sequence of benchmark feature points can be a series of feature points defined by each pre-trained model according to its own performance and preferences. The benchmark feature points in the sequence of benchmark feature points reflect the characteristics of the corresponding pre-trained model when processing different types of data, such as sensitivity to a certain data format, adaptability to a specific task, etc.

[0075] A set of current matching pairs can include an initial feature point in the sequence of dataset feature points and a current benchmark feature point in the sequence of benchmark feature points of the current pre-trained model. By calculating the differences (such as variance or distance) of the feature points in each set of current matching pairs, a numerical set is formed. By accumulating this set, the difference value corresponding to the current pre-trained model can be obtained, and thus a set of difference values corresponding to a group of pre-trained models can be obtained. The set of difference values corresponding to a group of pre-trained models includes the difference values corresponding to each pre-trained model.

[0076] Specifically, for the sequence of data feature points of the specified feedback dataset and the current sequence of benchmark feature points of the current pre-trained model, the process of calculating the difference value corresponding to the current pre-trained model can be as shown in formula (1):

[0077] ; (1)

[0078] Where is the difference value corresponding to the current pre-trained model, also known as the feature point variance, is the number of feature points in the sequence of data feature points, represents the th current matching pair in a set of current matching pairs, is the th initial feature point in the sequence of data feature points, is the current benchmark feature point in the sequence of current benchmark feature points that forms a current matching pair with the th initial feature point.

[0079] In this embodiment, by determining a set of current matching pairs between the feature point sequence of the data set and the current reference feature point sequence, a difference value corresponding to the pre-trained model is determined according to the set of current matching pairs, so as to screen a suitable pre-trained model according to the difference value corresponding to the pre-trained model, and determine the specified parameter adjustment method, thereby improving the accuracy of model update.

[0080] In an exemplary embodiment, feature point matching is performed on the feature point sequence of the data set and the current reference feature point sequence to obtain a set of current matching pairs, including:

[0081] Perform feature space mapping on the feature point sequence of the data set to obtain the specified feature point sequence of the specified feedback data set, where the specified feature point sequence and the current reference feature point sequence are in the same feature space;

[0082] Calculate the similarity between the initial feature point in the specified feature point sequence and the current reference feature point in the current reference feature point sequence to obtain the similarity between the initial feature point in the specified feature point sequence and the current reference feature point in the current reference feature point sequence;

[0083] Take the initial feature points in the feature point sequence of the data set as the current initial feature points respectively and perform the following determination operations to obtain a set of current matching pairs: determine the current reference feature point with the largest similarity to the current initial feature point in the current reference feature point sequence to obtain the current candidate reference feature point; in the case where the similarity between the current initial feature point and the current candidate reference feature point is greater than or equal to the similarity threshold, determine the current initial feature point and the current candidate reference feature point as a current matching pair.

[0084] It should be noted that the feature space mapping can be to transform the sequence of feature points of the data set into the feature space of the pre-trained model, so as to determine that the initial feature points in the sequence of feature points of the data set and the reference feature points in the reference feature point sequence are in the same feature space, which is convenient for subsequent similarity calculation. Optionally, the feature space mapping can be implemented through a certain mathematical transformation or machine learning algorithm, such as principal component analysis dimensionality reduction (abbreviated as PCA), t-SNE mapping (t-Distributed Stochastic Neighbor Embedding Mapping), etc. Of course, the feature extraction layer of the pre-trained model can be used to perform feature space mapping on the sequence of feature points of the data set to obtain the sequence of specified feature points of the specified feedback data set. Among them, the specified feature point sequence can be the sequence of feature points of the mapped data set, and the specified feature point sequence can include multiple initial feature points. The specified feature point sequence and the current reference feature point sequence are in the same feature space for feature matching and similarity calculation.

[0085] Calculate the similarity between the initial feature points in the specified feature point sequence and the current reference feature points in the current reference feature point sequence to measure the similarity between the sequence of feature points of the data set and the current reference feature points of the pre-trained model. Commonly used similarity calculation methods include cosine similarity, Jaccard similarity, Pearson correlation coefficient, etc.

[0086] In order to determine whether the initial feature points and the current reference feature points can form a matching pair, a value can be preset in advance to judge whether the similarity between the initial feature points and the current reference feature points is large enough. If this similarity is lower than the similarity threshold, it is considered that there is not enough similarity between the initial feature points and the current reference feature points and they will not be matched.

[0087] Optionally, the artificial intelligence platform can select the parameter adjustment method corresponding to the pre-trained model with the highest matching degree (that is, the most matching pairs and relatively large similarity) to fine-tune and update the dialogue model to be updated.

[0088] Optionally, by assigning weights to each initial feature point, the difference value corresponding to the pre-trained model can be calculated according to the assigned weights. Specifically, according to the sorting of the initial feature points in the dataset feature point sequence, a first weight is assigned to each sorted initial feature point, and then a second weight is assigned based on the similarity when the initial feature point forms a matching pair with the reference feature point in the reference feature point sequence of the pre-trained model. Combining the first weight and the second weight, the difference value corresponding to the pre-trained model can be calculated.

[0089] It should be noted that the first weight can be assigned according to the importance of the initial feature point, and can be determined by analyzing the frequency of user feedback, the magnitude of the impact on model performance, and the universality of the feature point in the dataset. For example, if a feature point has a significant impact on the prediction accuracy of the model, then it will obtain a higher first weight.

[0090] The second weight can be assigned according to the similarity of the initial feature point and the matching pair of the pre-trained model. The similarity can be calculated in various ways, such as Euclidean distance, cosine similarity, etc. The more similar the feature point is to the model feature point, the closer the performance of the model at this feature point is to the dataset requirements, so a higher second weight can be obtained. The positive correlation between the second weight and the similarity of the matching pair enables the fine-tuning strategy to be adaptively adjusted to better match the dataset characteristics and improve the generalization ability of the model.

[0091] Through the above embodiments, the dual weight mechanism takes into account the intrinsic importance and extrinsic matching degree of the feature points, making the calculated difference degree more accurate and better reflecting the adaptability between the model and the dataset. Feature points with higher importance and better matching with the model obtain higher weights when calculating the difference degree, which means that the fine-tuning process will be more inclined to optimize these feature points, ultimately improving the performance of the model when processing key tasks.

[0092] Through this embodiment, through feature space mapping and similarity calculation, the artificial intelligence platform can more objectively and accurately evaluate the matching degree between the model and the dataset. Similarity calculation can quickly exclude models that do not match the dataset severely, reduce invalid calculations and resource waste, and improve the overall efficiency of model fine-tuning and updating.

[0093] In an exemplary embodiment, both the initial feature points in the dataset feature point sequence and the current reference feature points in the current reference feature point sequence are set with feature point label information;

[0094] Performing feature point matching on the dataset feature point sequence and the current reference feature point sequence to obtain a set of current matching pairs, including:

[0095] Taking the initial feature points in the dataset feature point sequence as the current initial feature points respectively, perform the following determination operations to obtain a set of current matching pairs: Search for the current reference feature points in the current reference feature point sequence whose feature point label information is the same as that of the current initial feature point; in the case of finding the current reference feature point whose feature point label information is the same as that of the current initial feature point, determine the current initial feature point and the current reference feature point whose feature point label information is the same as that of the current initial feature point as a current matching pair.

[0096] It should be noted that the dataset feature point sequence may include multiple initial feature points. The initial feature points can represent the features corresponding to a specific attribute in the specified feedback dataset. The feature point label information can represent the additional description of the initial feature point, mainly used to classify and identify the type of the initial feature point. For example, labels such as "sentiment polarity", "topic", etc. Each reference feature point in the reference feature point sequence of the pre-trained model is also attached with feature point label information.

[0097] Optionally, a current matching pair can be a matching pair composed of an initial feature point in the dataset feature point sequence and a current reference feature point in the current reference feature point sequence with the same feature point label information. Optionally, the feature point label information can be understood as a specific identification information of the feature point, which can be represented by a string, etc. The initial feature point and the current reference feature point with the same specific identification information can form a matching pair.

[0098] Optionally, after finding the current reference feature points with the same feature point label information, form a current matching pair with the current initial feature point in the dataset feature point sequence and the current reference feature point of the pre-trained model, and repeat this process until all the initial feature points in the dataset feature point sequence find the current reference feature points corresponding to the current pre-trained model, or not all are found, but the matching process ends.

[0099] Through this embodiment, matching through the feature point label information ensures that the selection of the pre-trained model is closely related to the attributes of the dataset, improving the accuracy of model matching. Through fine-tuning the model that better matches the dataset characteristics, the performance of the dialogue model to be updated in a specific scenario can be significantly improved. Moreover, using the feature point label information can reduce unnecessary data processing processes, speed up the matching process, and improve the efficiency of the tuning process.

[0100] In an exemplary embodiment, there is at least one type of feedback data in the specified feedback dataset;

[0101] Extract the initial set of feature points from the specified feedback data set, including: when the specified feedback data set includes text data, perform syntactic analysis on the text data in the specified feedback data set to obtain the text feature points of the specified feedback data set, where the initial set of feature points includes the text feature points of the specified feedback data set; when the specified feedback data set includes image data, perform image feature recognition on the image data in the specified feedback data set to obtain the image feature points of the specified feedback data set, where the initial set of feature points includes the image feature points of the specified feedback data set; when the specified feedback data set includes audio data, perform audio feature analysis on the audio data in the specified feedback data set to obtain the audio feature points of the specified feedback data set, where the initial set of feature points includes the audio feature points of the specified feedback data set.

[0102] It should be noted that the specified feedback data set can be collected by the artificial intelligence platform and is a set of feedback data obtained from the interaction process of users using the agent application in the artificial intelligence platform. The specified feedback data set can include various types of information, such as text, images, audio, etc. The specified feedback data in the specified feedback data set can be used to train and optimize the dialogue model used in the agent application.

[0103] Syntactic analysis can refer to the process of analyzing sentence structure, identifying parts of speech, phrase boundaries, dependency relationships, etc. through natural language processing techniques to extract meaningful text feature points. Optionally, the artificial intelligence platform can use natural language processing (NLP) tools, such as parsers, named entity recognizers, etc., to deeply analyze the text data and identify the subject-predicate-object structure, sentiment polarity, keywords, etc. of the sentence. For example, if user feedback shows that most conversations involve professional terms in a specific field, then the recognition of these terms will become one of the feature points for evaluating the pre-trained model's mastery of knowledge in the relevant field.

[0104] Image feature recognition can refer to the process of extracting key feature points from an image through computer vision technology, such as color, texture, shape, objects, etc. Optionally, the artificial intelligence platform can be pre-deployed with an image recognition model, such as a Convolutional Neural Network (CNN for short), to process the image data, mark the objects, region boundaries, color distribution, etc. in the image to generate image feature points. The extracted image feature points will also be integrated into the initial feature point set for subsequent selection of fine-tuning models. If a large number of pictures in the user feedback contain rich colors and complex backgrounds, then color recognition ability and background processing technology will be important feature points for evaluating whether the model is suitable for such image data. For example, when the intelligent agent application receives a food picture uploaded by the user, the artificial intelligence platform can mark feature points such as the food type, color distribution, and image clarity in the picture through image feature recognition. This information will guide the platform to select pre-trained models that perform well in food recognition and color perception.

[0105] Audio feature analysis refers to the process of extracting key feature points such as voiceprint, pitch, and speech rate from audio data through signal processing and machine learning techniques. Optionally, the artificial intelligence platform can be deployed with speech recognition and audio analysis algorithms, such as Mel Frequency Cepstral Coefficients (MFCC for short) or SpectralCentroid (spectral center), to process audio segments, identify the speaker's emotional state, speech rate, pitch, etc. to form an audio feature point set.

[0106] Optionally, the text feature points include at least one of the following feature points: feature points for representing the emotional tendency of text data, feature points for representing the theme distribution of text data. The image feature points include at least one of the following feature points: feature points for representing the object information of the objects existing in the image, feature points for representing the texture of the image, feature points for representing the color distribution in the image. The audio feature points include at least one of the following feature points: feature points for representing the spectral features of the audio, feature points for representing the emotion of the audio, feature points for representing the intonation changes in the audio.

[0107] Through this embodiment, by separately performing feature analysis on different types of specified feedback data, the artificial intelligence platform can more precisely understand the characteristics of user needs and application scenarios, thereby more accurately adjusting the model parameters and improving the performance of the dialogue model to be updated. Moreover, during the training and optimization process, the artificial intelligence platform can comprehensively consider data features of various types such as text, images, and audio, promoting the integrated learning ability of the dialogue model to be updated for different types of information and enhancing the comprehensive processing ability and flexibility of the dialogue model to be updated.

[0108] In an exemplary embodiment, the data types allowed to be input by the dialogue model to be updated match the model type of the dialogue model to be updated. The above method further includes: when the dialogue model to be updated is a multi-modal model and there are at least two types of initial feature points in the initial feature point set, inputting the at least two types of initial feature points into a preset attention network to obtain at least two attention weights, where the at least two attention weights correspond one-to-one to the at least two types of initial feature points; using the at least two attention weights to perform weighted fusion on the at least two types of initial feature points to obtain a set of cross-modal feature points, where the initial feature point set includes at least two types of initial feature points and a set of cross-modal feature points, or at least two types of initial feature points in the initial feature point set are updated to a set of cross-modal feature points.

[0109] It should be noted that there can be multiple different types of a set of dialogue models, such as large language models (only supporting text inference conversations), multi-modal models (supporting text analysis and inference conversations of picture files), and deep thinking models (supporting inference conversations after deep thinking), etc.

[0110] According to the model type of the dialogue model to be updated, the data types it allows to be input can be determined, such as text, images, or audio, etc. Before model optimization, the artificial intelligence platform can verify whether the dialogue model to be updated supports the data types existing in the specified feedback dataset. If the model is multi-modal and contains at least two types of data (such as text and images), then the dialogue model to be updated can perform cross-modal training and optimization.

[0111] A preset attention network can also be deployed in the artificial intelligence platform. The preset attention network can be a neural network structure that enables the model to focus on certain key parts when processing complex inputs.

[0112] The initial feature point set can include initial feature points such as text feature points, image feature points, and audio feature points.

[0113] In the case of processing multimodal data corresponding to a multimodal model, the artificial intelligence platform can input initial feature points such as text feature points and image feature points into a preset attention network to obtain a set of attention weights corresponding to different data types, and the set of attention weights can be used to reflect the importance of different modal data in a specific scenario.

[0114] After obtaining a set of attention weights, the set of attention weights can be weighted and fused with different types of feature points to generate a set of cross-modal feature points representing comprehensive information. The cross-modal feature points can represent a set of feature points integrating modal information, and the cross-modal feature points will be integrated into the initial feature point set or replace the original feature point set.

[0115] Through this embodiment, by using the preset attention network to perform weighted fusion on different types of feature points corresponding to multimodal data, a set of more comprehensive and accurate cross-modal feature points can be generated, enabling the dialogue model to be updated to better understand the multi-dimensional information of user feedback data and improving the accuracy and generalization ability of the model when processing complex tasks.

[0116] In an exemplary embodiment, the artificial intelligence platform is provided with a correspondence table, and the correspondence table is used to record the correspondence between the preset difference values corresponding to the pre-trained models in the set of pre-trained models and the parameter adjustment methods in the set of parameter adjustment methods.

[0117] Step S206 includes: respectively calculating the absolute value of the difference between the difference value corresponding to the same pre-trained model in the set of pre-trained models and the corresponding preset difference value to obtain a set of absolute difference values corresponding to the set of pre-trained models; determining the parameter adjustment method of the pre-trained model corresponding to the minimum absolute difference value in the set of absolute difference values as the specified parameter adjustment method.

[0118] It should be noted that there is a data table stored inside the artificial intelligence platform, that is, the correspondence table. The correspondence table is used for the platform to pre-establish such a relationship table to record the expected difference values of different models at different feature points and the most suitable parameter adjustment methods in these cases. This correspondence can be determined through experiments in the initial stage of model development or can be derived from historical tuning data. The preset difference value can be a standard value corresponding to the parameter adjustment method. These values can be determined based on the performance expectations of the model at different feature points, such as precision, recall rate, etc.

[0119] The preset difference value can be a reasonable target difference value set based on the improvement of the model performance when the pre-trained model is trained using the parameter adjustment method, and it can describe the historical performance relationship between different types of full-parameter fine-tuning methods and the training model on specific performance indicators.

[0120] Each pre-trained model in a group of pre-trained models corresponds to a preset difference value.

[0121] The parameter fine-tuning method can be a method for fine-tuning a pre-trained model to adapt to a specific task or dataset, including but not limited to learning rate adjustment, regularization parameter modification, model architecture change, etc.

[0122] The absolute difference can refer to the absolute gap between the difference value corresponding to each pre-trained model in a group of pre-trained models obtained by an artificial intelligence platform and the preset difference value recorded in the corresponding relationship table. It measures the deviation degree between the actual difference and the expected difference of the model. The set of absolute differences corresponding to a group of pre-trained models includes the absolute differences corresponding to each pre-trained model in the group of pre-trained models. According to the set of absolute differences, the parameter adjustment method corresponding to the pre-trained model with the smallest difference is the specified parameter adjustment method.

[0123] Optionally, a probability value of successful execution of the parameter adjustment method corresponding to the pre-trained model can also be set. After determining the specified parameter adjustment method, the probability value of successful execution corresponding to the specified parameter adjustment method can be obtained and compared with the current standard probability threshold (such as the threshold set by the current administrator). If it is greater than the current standard probability threshold, the model update operation is executed.

[0124] In specific practice, as Figure 3 shown, a group of pre-trained models in an artificial intelligence platform can be stored in a model repository. In the model repository, there are n pre-trained models, and each pre-trained model of different types corresponds to a parameter adjustment method (i.e., the full-parameter fine-tuning method). Through a dataset, model verification processing is performed from the model repository to determine a pre-trained model, that is, the verification point, so as to use the parameter adjustment method corresponding to the pre-trained model as the specified parameter adjustment method.

[0125] It should be noted that the n pre-trained models can include pre-trained models obtained by training with pictures, pre-trained models obtained by training with text, pre-trained models obtained by training with code, pre-trained models obtained by training with in-depth thinking content, and pre-trained models obtained by hierarchical model training, etc.

[0126] Specifically, the parameter adjustment method in this embodiment can be as shown in Table 1.

[0127] Table 1

[0128]

[0129] It should be noted that, in the case where the parameter adjustment method is basic fine-tuning, the algorithms that can be used are Stochastic Gradient Descent (abbreviated as SGD) / Adam (Adaptive Moment Estimation, Adam); in the case where the parameter adjustment method is hierarchical learning rate, the algorithm that can be used can be Layer-wise Optimizers (layer-wise adaptive learning rate scaling); in the case where the parameter adjustment method is adversarial fine-tuning, the algorithms that can be used are Generative Adversarial Networks (abbreviated as GAN) / Adversarial Training; in the case where the parameter adjustment method is regularization fine-tuning, the algorithms that can be used are L2-SP (L2 Regularization with Sparse Penalty, L2 regularization combined with sparse penalty) / DELTA; in the case where the parameter adjustment method is multi-task fine-tuning, the algorithm that can be used is Gradient Norm Balancing (abbreviated as GradNorm); in the case where the parameter adjustment method is progressive fine-tuning, the algorithm that can be used is Curriculum Learning; in the case where the parameter adjustment method is LoRA fine-tuning, the algorithm that can be used is Low-Rank Adaptation; in the case where the parameter adjustment method is contrastive learning fine-tuning, the algorithm that can be used is Contrastive Loss.

[0130] Through this embodiment, using the correspondence table, the artificial intelligence platform can automatically select the most suitable parameter adjustment method according to the differences of the model at specific feature points, avoiding blind tuning or resource waste. At the same time, it also improves the model performance of the dialogue model to be updated to a certain extent.

[0131] In an exemplary embodiment, step S202 includes: performing data cleaning on the candidate feedback data in the candidate feedback data set corresponding to the dialogue model to be updated, classifying the cleaned candidate feedback data according to the data type to obtain the classified candidate feedback data, where the candidate feedback data in the candidate feedback data set is the feedback data obtained by using the dialogue model to be updated for historical conversations; inputting the classified candidate feedback data into a preset identification model to obtain the data label information of the classified candidate feedback data; and constructing a specified feedback data set based on the classified candidate feedback data and the data label information of the classified candidate feedback data, where the specified feedback data in the specified feedback data set is the classified candidate feedback data with data label information.

[0132] Optionally, data cleaning can be used to remove errors, duplicates, incomplete information, or other insignificant elements in the data to improve the quality and reliability of the data set. For example, if the user feedback contains incomplete conversation records or misformatted information, the data cleaning process will identify and correct these problems. According to the nature of the feedback data, the feedback data is divided into different categories, such as text data, image data, audio data, etc. The preset identification model can be a pre-trained model, which can be used to identify and label the attributes, types, or sentiment tendencies of the feedback data, etc. For example, it can identify positive or negative emotions in the conversation, or label the image data as "clear" or "blurry".

[0133] The data label information can be meta-information added to the data through the preset identification model, which is used to help the dialogue model to be updated better understand and learn the characteristics of the data during the training process. Optionally, the feature point label information can include the data label information.

[0134] Through this embodiment, through data cleaning and classification, the platform can avoid using invalid or low-quality data for training, thereby saving computing resources and shortening the training time. By adding data labels, the efficiency and accuracy of the dialogue model update are greatly improved.

[0135] In an exemplary embodiment, the artificial intelligence platform is also deployed with a specified container; before obtaining the specified feedback data set, the above method further includes: monitoring the resource utilization rate of the artificial intelligence platform; in the case where the resource utilization rate is less than the first resource utilization rate threshold, determining that the specified container is in an available state; configuring a container resource range for the specified container according to the update resource requirements of the dialogue model to be updated, where the maximum value of the container resource range is less than the second resource utilization rate threshold, and the second resource utilization rate threshold is greater than the first resource utilization rate threshold; wherein, updating the dialogue model to be updated is performed within the specified container using the resources within the configured container resource range of the specified container.

[0136] It should be noted that a specified container can be deployed in the artificial intelligence platform. The specified container can be a pre-configured isolation environment for running specific tasks, such as a Docker container. The specified container is equipped with the software environment and resources specifically involved in model training and update.

[0137] The resource utilization rate of an AI platform can refer to the proportion of the hardware resources of the AI platform (such as the Central Processing Unit (CPU), Graphics Processing Unit (GPU), memory, and hard disk space) that are currently occupied, which can be an important indicator for measuring the load of the AI platform. The first resource utilization rate threshold and the second resource utilization rate threshold can be boundary values used to determine the platform resource status. Optionally, the first resource utilization rate threshold is relatively low and can be used to determine whether a specified container can be started; the second resource utilization rate threshold is relatively high and is used to limit the upper limit of the resource usage of the specified container to ensure that the overall resources of the platform are not over-occupied. The updated resource requirements can refer to the specific requirements for the computing resources of the AI platform when the dialogue model to be updated is being trained and optimized. The updated resource requirements can include, but are not limited to, the number of CPU cores, the number of GPUs, the memory size, and the hard disk space. The container resource range can be the resource usage interval configured by the AI platform for a specified container, which can include the minimum and maximum resource usage amounts, ensuring that the container runs under sufficient resources while also avoiding resource waste or negative impacts on other services of the platform.

[0138] In specific practices, the AI platform can start a specified container in a self-triggered manner. Optionally, the AI platform can include a tuning image file, automatically trigger decompression, loading, and creation of a Docker container, and its corresponding creation command can be "gunzip ** / docker load -i *** / docker run --gpu all -itd -p **:** -v{$directory} --cap-add=IPC_LOCK --name=*** --ulimit stack= **** --shm-size=500G {$image file}".

[0139] It should be noted that in a specified container, data processing tasks can be run, such as the process of obtaining a specified feedback data set. Of course, in the specified container, the process of triggering the selection of a specified parameter adjustment method can also be carried out. For example, in the specified container, the corresponding command is executed as "python -u / logme / train_example / logme-cnn_lm_v1.01 / scripts / lm.py –model* lvgg16 –feature 20 -1000 –batch_size 256 –data_dir= / targetData)" to obtain model_check_res. Then, through the evaluation of the pre-trained LogME model configured, the result with the highest matching degree can be selected, and the corresponding full-parameter fine-tuning method, that is, the specified parameter adjustment method, can be selected.

[0140] Through this embodiment, by setting the resource utilization rate threshold and adjusting the container resource range, the platform effectively utilizes idle resources, avoids resource waste, and improves resource utilization efficiency on the premise of ensuring service stability.

[0141] In an exemplary embodiment, after updating the model parameters of the dialogue model to be updated according to the specified feedback data set and the specified parameter adjustment method, the above method further includes: performing a preliminary verification on the updated dialogue model to be updated, where the preliminary verification includes performing a performance test on the updated dialogue model to be updated and performing a stability test on the updated dialogue model to be updated; in the case where the preliminary verification passes, deploying the updated dialogue model to be updated to the artificial intelligence platform; in the case where it is monitored that the updated dialogue model to be updated has a performance degradation compared to the dialogue model to be updated before the update, rolling back the updated dialogue model to be updated to the dialogue model to be updated before the update.

[0142] It should be noted that the preliminary verification can be a preliminary performance and stability test performed after the update of the dialogue model to be updated to ensure that the new model can meet the expected performance standards and maintain stable operation before actual deployment.

[0143] The performance test can be achieved by evaluating the performance of the updated dialogue model in aspects such as processing user requests, answering speed, and accuracy. The performance test can involve the response speed and accuracy rate of the model when processing common queries. Optionally, a performance test example can be deployed in the artificial intelligence platform. For example, the platform simulates the query behavior of users during peak hours, such as asking questions quickly and continuously, complex queries, and professional questions in specific fields, records the response time, accuracy rate, and resource consumption of the model, and compares them with the model before the update to ensure that the updated model is at least on par with the original model in these metrics.

[0144] The stability test can be achieved by checking the stability of the updated dialogue model under long - term operation, high - concurrency requests, or special input conditions, avoiding service interruptions or abnormal responses caused by potential problems introduced by model updates. The stability test can examine the reliability of the updated dialogue model when facing a large number of requests. Optionally, a stability test instance can be deployed in the artificial intelligence platform. For example, after model updates, by generating a large number of concurrent requests, the response time of the model and whether there are crashes or errors can be checked. At the same time, some abnormal input data, such as malicious attacks or special characters, are simulated to ensure that the model can operate stably under various conditions.

[0145] Optionally, in the case of detecting a decline in the performance of the updated model (such as longer response time, lower accuracy), the model version before the update can be automatically or manually switched back to ensure that the service quality and user experience are not negatively affected.

[0146] Through this embodiment, with the preliminary verification and rollback mechanism, the platform can ensure that model updates will not reduce service quality or bring an unsatisfactory experience to users, which greatly improves the reliability of the artificial intelligence platform.

[0147] The following explains the model update method in the embodiments of the present application with reference to optional examples. As Figure 4 shown, in this optional example, the artificial intelligence platform is launched on a training and inference integrated machine, and preparations are made to start the deployment and optimization of model services.

[0148] At least one basic model (i.e., dialogue model) can be deployed on the artificial intelligence platform. After the deployment of the basic model is completed, model publishing can be carried out to publish it to the intelligent agent application corresponding to the basic model. Users can interact with it through the intelligent agent application, such as functions like dialogue Q&A, image analysis, or in - depth thinking, to interact with the model.

[0149] During the user's usage process, the artificial intelligence platform can record the user's dialogue feedback data. For example, dialogue satisfaction, evaluation of image analysis results, or scoring of in - depth thinking questions. For example, if a user is not satisfied with the answer of the chatbot, a negative review can be given, and this information will be stored as the basis for subsequent optimization.

[0150] The artificial intelligence platform asks the user whether they agree to participate in the automatic optimization plan. If the user agrees, the platform proceeds to the next step; if not, the platform does not collect the user's feedback data for automatic optimization to protect the user's privacy.

[0151] The artificial intelligence platform counts the number of feedback messages collected. When the number of feedback messages reaches a threshold, a specified feedback dataset is generated in a specified container. Specifically, the specified container cleans, classifies, and labels the feedback data to obtain the specified feedback dataset.

[0152] Based on the specified feedback dataset, a full-parameter fine-tuning method matching and tuning training are performed. Specifically, the dataset feature point sequence of the specified feedback dataset is matched with the benchmark feature point sequence of a pre-trained model in a group of pre-trained models to obtain a feature point matching result. According to the feature point matching result, a specified parameter adjustment method is determined, and the model parameters of the to-be-updated dialogue model are updated according to the specified feedback dataset and the specified parameter adjustment method, thereby generating a model.

[0153] The updated to-be-updated dialogue model is deployed for model services. After the model service deployment, it is judged whether the updated to-be-updated dialogue model is ready, that is, whether the new model has been trained and is waiting for deployment. During the deployment process, it can be determined whether the number of deployment failures reaches a failure threshold. By checking whether the performance of the updated model in the user application scenario meets the expectations, if the model service fails to meet the predetermined performance standard in multiple validations, that is, the number of deployment failures reaches the failure threshold, the tuning process ends. When the number of deployment failures does not reach the failure threshold, return to the specified container to perform full-parameter fine-tuning method matching and application training; when it is determined that the to-be-updated dialogue model is ready, return to model release. The platform continues to provide services, and users can use the fine-tuned model to interact to obtain more feedback messages.

[0154] Through this example, by automatically collecting feedback data during the model service deployment and starting model fine-tuning when resources are sufficient, the redundant resources of the training and inference integrated machine are fully utilized, the resource utilization rate is improved, and the training cost is reduced. The to-be-updated dialogue model is fine-tuned based on the real feedback data of users, which can better meet the needs of users, improve the accuracy and response speed of the model on specific tasks, and thus significantly improve the service quality. The updated to-be-updated dialogue model can understand and respond to users more accurately. Moreover, the automatic tuning mechanism of the platform reduces manual intervention, makes the model update and tuning process more automated, reduces the operation and maintenance cost, and improves the operation efficiency of the platform.

[0155] It should be noted that the information collected in this application (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for display, data for analysis, etc.) are information and data that have been authorized by the user or fully authorized by all parties. Moreover, for the processing of relevant data such as collection, storage, use, processing, transmission, provision, disclosure, and application, all comply with relevant laws, regulations, and standards, necessary confidentiality measures are taken, it does not violate public order and good customs, and corresponding operation entrances are provided for users to choose to authorize or refuse.

[0156] Through the description of the above embodiments, those skilled in the art can clearly understand that the method according to the above embodiments can be implemented by means of software plus a necessary general hardware platform. Of course, it can also be implemented by hardware, but in many cases, the former is a better implementation method.

[0157] An embodiment of the present application further provides an update device for a model, which is applied to an artificial intelligence platform, and a set of dialogue models are deployed on the artificial intelligence platform; as Figure 5 shown, the device includes:

[0158] An acquisition module 502, configured to acquire a specified feedback data set when it is detected that there is a dialogue model to be updated in a set of dialogue models, wherein the specified feedback data in the specified feedback data set is feedback data obtained by using the dialogue model to be updated for historical conversations;

[0159] A matching module 504, configured to perform feature point matching on the data set feature point sequence of the specified feedback data set and the benchmark feature point sequence of the pre-trained model in a set of pre-trained models, to obtain a feature point matching result, wherein the data set feature point sequence is used to characterize the attribute information of the specified feedback data set;

[0160] A determination module 506, configured to determine a specified parameter adjustment method from a set of parameter adjustment methods according to the feature point matching result, wherein the parameter adjustment method in the set of parameter adjustment methods is a method for adjusting the model parameters of the dialogue models in a set of dialogue models;

[0161] An update module 508, configured to update the model parameters of the dialogue model to be updated according to the specified feedback data set and the specified parameter adjustment method.

[0162] It should be noted that the acquisition module 502 in this embodiment can be used to execute the above step S202, the matching module 504 in this embodiment can be used to execute the above step S204, the determination module 506 in this embodiment can be used to execute the above step S206, and the update module 508 in this embodiment can be used to execute the above step S208.

[0163] Through the embodiments provided in this application, when it is detected that there is a dialogue model to be updated in a set of dialogue models, a specified feedback data set is obtained, and feature point matching is performed on the data set feature point sequence of the specified feedback data set and the benchmark feature point sequence of the pre-trained model in a set of pre-trained models to obtain a feature point matching result. Then, according to the feature point matching result, a specified parameter adjustment method is determined from a set of parameter adjustment methods. Thus, according to the specified feedback data set and the specified parameter adjustment method, the model parameters of the dialogue model to be updated are updated. By performing feature point matching on the data set feature point sequence of the specified feedback data set and the benchmark feature point sequence of the pre-trained model in a set of pre-trained models and determining the specified parameter adjustment method according to the feature point matching result, the dialogue model to be updated is updated, making the model update more in line with actual needs, avoiding ineffective or negative tuning, and ensuring the performance of the updated dialogue model to a certain extent. This solves the technical problem of poor update effect of the model in the related art and achieves the technical effect of stable improvement of the dialogue model performance.

[0164] In an exemplary embodiment, the matching module 504 includes: a matching unit configured to calculate the difference values between the data set feature point sequence and the benchmark feature point sequence of the pre-trained model in a set of pre-trained models respectively, to obtain a set of difference values corresponding to the set of pre-trained models, wherein the feature point matching result includes the set of difference values corresponding to the set of pre-trained models.

[0165] In an exemplary embodiment, the model update device further includes:

[0166] An extraction module configured to extract an initial feature point set of the specified feedback data set;

[0167] A sorting module configured to perform feature point sorting on the initial feature points in the initial feature point set to obtain an initial feature point sequence;

[0168] A sequence determination module configured to determine a subsequence including the first specified number of initial feature points in the initial feature point sequence as the data set feature point sequence.

[0169] In an exemplary embodiment, the matching unit includes: a matching operation sub-unit for: using the pre-trained models in a set of pre-trained models as the current pre-trained model respectively to perform the following matching operations, obtaining a set of difference values corresponding to the set of pre-trained models, wherein the reference feature point sequence of the current pre-trained model is the current reference feature point sequence: performing feature point matching on the dataset feature point sequence and the current reference feature point sequence to obtain a set of current matching pairs, wherein one current matching pair in the set of current matching pairs includes an initial feature point in the dataset feature point sequence and a current reference feature point in the current reference feature point sequence; respectively calculating the difference values between the two feature points of the current matching pairs in the set of current matching pairs to obtain the set of difference values corresponding to the current pre-trained model;

[0170] Wherein, the set of difference values corresponding to the set of pre-trained models includes the difference values corresponding to the current pre-trained model in the set of pre-trained models.

[0171] In an exemplary embodiment, the matching operation sub-unit is further configured to: perform feature space mapping on the dataset feature point sequence to obtain a specified feature point sequence of a specified feedback dataset, wherein the specified feature point sequence and the current reference feature point sequence are in the same feature space; calculate the similarity between the initial feature point in the specified feature point sequence and the current reference feature point in the current reference feature point sequence; use the initial feature points in the dataset feature point sequence as the current initial feature points respectively to perform the following determination operations to obtain a set of current matching pairs: determining the current reference feature point in the current reference feature point sequence with the greatest similarity to the current initial feature point to obtain the current candidate reference feature point; in the case where the similarity between the current initial feature point and the current candidate reference feature point is greater than or equal to the similarity threshold, determining the current initial feature point and the current candidate reference feature point as a current matching pair.

[0172] In an exemplary embodiment, both the initial feature points in the dataset feature point sequence and the current reference feature points in the current reference feature point sequence are set with feature point label information; the matching operation sub-unit is further configured to: use the initial feature points in the dataset feature point sequence as the current initial feature points respectively to perform the following determination operations to obtain a set of current matching pairs: searching for the current reference feature points in the current reference feature point sequence with the same feature point label information as the current initial feature point; in the case where the current reference feature points with the same feature point label information as the current initial feature point are found, determining the current initial feature point and the current reference feature points with the same feature point label information as the current initial feature point as a current matching pair.

[0173] In an exemplary embodiment, there is at least one type of specified feedback data in the specified feedback data set; the extraction module further includes:

[0174] A first analysis unit, configured to perform syntactic analysis on the text data in the specified feedback data set to obtain text feature points of the specified feedback data set when the specified feedback data set includes text data, where the initial feature point set includes the text feature points of the specified feedback data set;

[0175] A first recognition unit, configured to perform image feature recognition on the image data in the specified feedback data set to obtain image feature points of the specified feedback data set when the specified feedback data set includes image data, where the initial feature point set includes the image feature points of the specified feedback data set;

[0176] A second analysis unit, configured to perform audio feature analysis on the audio data in the specified feedback data set to obtain audio feature points of the specified feedback data set when the specified feedback data set includes audio data, where the initial feature point set includes the audio feature points of the specified feedback data set.

[0177] In an exemplary embodiment, the data types allowed to be input by the dialogue model to be updated match the model type of the dialogue model to be updated; the model update device further includes:

[0178] A weight determination module, configured to input at least two types of initial feature points into a preset attention network to obtain at least two attention weights when the dialogue model to be updated is a multi-modal model and there are at least two types of initial feature points in the initial feature point set, where the at least two attention weights correspond one-to-one to the at least two types of initial feature points;

[0179] A fusion module, configured to perform weighted fusion on at least two types of initial feature points by using the at least two attention weights to obtain a set of cross-modal feature points, where the initial feature point set includes at least two types of initial feature points and a set of cross-modal feature points, or at least two types of initial feature points in the initial feature point set are updated to a set of cross-modal feature points.

[0180] In an exemplary embodiment, an artificial intelligence platform is provided with a correspondence table, which is used to record the correspondence between the preset difference values corresponding to the pre-trained models in a group of pre-trained models and the parameter adjustment methods in a group of parameter adjustment methods.

[0181] The determination module 506 includes: a calculation unit for calculating, respectively, the absolute value of the difference between the difference value corresponding to the same pre-trained model in a set of pre-trained models and the corresponding preset difference value, to obtain a set of absolute difference values corresponding to the set of pre-trained models; a determination unit for determining the parameter adjustment method of the pre-trained model corresponding to the minimum absolute difference value in the set of absolute difference values as the specified parameter adjustment method.

[0182] In an exemplary embodiment, the acquisition module 502 includes: a cleaning unit for performing data cleaning on the candidate feedback data in the candidate feedback data set corresponding to the dialogue model to be updated, classifying the cleaned candidate feedback data according to the data type, to obtain the classified candidate feedback data, where the candidate feedback data in the candidate feedback data set is the feedback data obtained by using the dialogue model to be updated for historical conversations;

[0183] An obtaining unit for inputting the classified candidate feedback data into a preset identification model to obtain the data label information of the classified candidate feedback data;

[0184] A construction unit for constructing a specified feedback data set based on the classified candidate feedback data and the data label information of the classified candidate feedback data, where the specified feedback data in the specified feedback data set is the classified candidate feedback data with data label information.

[0185] In an exemplary embodiment, the artificial intelligence platform is also deployed with a specified container; the model update device further includes:

[0186] A monitoring module for monitoring the resource utilization rate of the artificial intelligence platform before obtaining the specified feedback data set;

[0187] A resource determination module for determining that the specified container is in an available state when the resource utilization rate is less than the first resource utilization rate threshold;

[0188] A configuration module for configuring a container resource range for the specified container according to the update resource requirements of the dialogue model to be updated, where the maximum value of the container resource range is less than the second resource utilization rate threshold, and the second resource utilization rate threshold is greater than the first resource utilization rate threshold;

[0189] Wherein, updating the dialogue model to be updated is performed within the specified container using the resources within the configured container resource range of the specified container.

[0190] In an exemplary embodiment, the model update device further includes:

[0191] A verification module, configured to perform a preliminary verification on the to-be-updated dialogue model after updating the model parameters of the to-be-updated dialogue model according to a specified feedback data set and a specified parameter adjustment method, where the preliminary verification includes performing a performance test on the updated to-be-updated dialogue model and performing a stability test on the updated to-be-updated dialogue model;

[0192] A deployment module, configured to deploy the updated to-be-updated dialogue model to an artificial intelligence platform when the preliminary verification is passed;

[0193] A rollback module, configured to roll back the updated to-be-updated dialogue model to the to-be-updated dialogue model before update when it is detected that the performance of the updated to-be-updated dialogue model has deteriorated compared to the to-be-updated dialogue model before update.

[0194] For the description of the features in the corresponding embodiments of the model update device, reference can be made to the relevant descriptions in the corresponding embodiments of the model update method, which will not be elaborated here one by one.

[0195] An embodiment of the present application further provides an electronic device, including a memory and a processor. A computer program is stored in the memory, and the processor is configured to run the computer program to execute the steps in any one of the above-mentioned model update method embodiments.

[0196] An embodiment of the present application further provides a computer-readable storage medium, in which a computer program is stored, where the computer program is configured to execute the steps in any one of the above-mentioned model update method embodiments when running.

[0197] In an exemplary embodiment, the above-mentioned computer-readable storage medium may include, but is not limited to: USB flash drive, read-only memory (ROM for short), random access memory (RAM for short), mobile hard disk, magnetic disk or optical disc, etc., various media that can store computer programs.

[0198] An embodiment of the present application further provides a computer program product, where the computer program product includes a computer program, and the computer program implements the steps in any one of the above-mentioned model update method embodiments when executed by a processor.

[0199] An embodiment of the present application further provides another computer program product, including a non-volatile computer-readable storage medium, and the non-volatile computer-readable storage medium stores a computer program, and the computer program implements the steps in any one of the above-mentioned model update method embodiments when executed by a processor.

[0200] Those skilled in the art may further realize that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be implemented by electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described according to functions in the above description. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered as exceeding the scope of this application.

[0201] The above has introduced in detail a method for updating a model, an electronic device, a storage medium, and a program product provided by this application. Specific examples are used herein to elaborate on the principle and implementation manner of this application. The description of the above embodiments is only used to help understand the method and its core idea of this application. It should be noted that for those of ordinary skill in the art, without departing from the principle of this application, several improvements and modifications can be made to this application, and these improvements and modifications also fall within the protection scope of the claims of this application.

Claims

1. A method for updating a model, characterized in that, Applied to an artificial intelligence platform, a set of dialogue models is deployed on the artificial intelligence platform; the method includes: When it is detected that there is a dialogue model to be updated in the set of dialogue models, obtain a specified feedback data set, where the specified feedback data in the specified feedback data set is the feedback data obtained from historical conversations using the dialogue model to be updated; Perform feature point matching on the data set feature point sequence of the specified feedback data set and the benchmark feature point sequence of a pre-trained model in a set of pre-trained models to obtain a feature point matching result, where the data set feature point sequence is used to characterize the attribute information of the specified feedback data set; According to the feature point matching result, determine a specified parameter adjustment method from a set of parameter adjustment methods, where the parameter adjustment method in the set of parameter adjustment methods is a method for adjusting the model parameters of the dialogue models in the set of dialogue models; Update the model parameters of the dialogue model to be updated according to the specified feedback data set and the specified parameter adjustment method.

2. The method according to claim 1, characterized in that The performing feature point matching on the data set feature point sequence of the specified feedback data set and the benchmark feature point sequence of a pre-trained model in a set of pre-trained models to obtain a feature point matching result includes: Calculate the difference values between the data set feature point sequence and the benchmark feature point sequences of the pre-trained models in the set of pre-trained models respectively to obtain a set of difference values corresponding to the set of pre-trained models, where the feature point matching result includes the set of difference values corresponding to the set of pre-trained models.

3. The method according to claim 2, wherein Before the performing feature point matching on the data set feature point sequence of the specified feedback data set and the benchmark feature point sequence of a pre-trained model in a set of pre-trained models, the method further includes: Extract the initial feature point set of the specified feedback data set; Perform feature point sorting on the initial feature points in the initial feature point set to obtain an initial feature point sequence; Determine the subsequence containing the first specified number of initial feature points in the initial feature point sequence as the data set feature point sequence.

4. The method according to claim 3, wherein The calculating the difference values between the data set feature point sequence and the benchmark feature point sequences of the pre-trained models in the set of pre-trained models respectively to obtain a set of difference values corresponding to the set of pre-trained models includes: Use the pre-trained models in the set of pre-trained models as the current pre-trained model respectively to perform the following matching operations to obtain a set of difference values corresponding to the set of pre-trained models, where the benchmark feature point sequence of the current pre-trained model is the current benchmark feature point sequence: Perform feature point matching on the data set feature point sequence and the current benchmark feature point sequence to obtain a set of current matching pairs, where one current matching pair in the set of current matching pairs includes an initial feature point in the data set feature point sequence and a current benchmark feature point in the current benchmark feature point sequence; Calculate the difference values between the two feature points of the current matching pairs in the set of current matching pairs respectively to obtain the difference value corresponding to the current pre-trained model; Among them, the set of difference values corresponding to the set of pre-trained models includes the difference value corresponding to the current pre-trained model in the set of pre-trained models.

5. The method according to claim 4, wherein The feature point matching of the dataset feature point sequence and the current reference feature point sequence to obtain a set of current matching pairs includes: Performing feature space mapping on the dataset feature point sequence to obtain a specified feature point sequence of the specified feedback dataset, where the specified feature point sequence and the current reference feature point sequence are in the same feature space; Calculating the similarity between the initial feature points in the specified feature point sequence and the current reference feature points in the current reference feature point sequence to obtain the similarity between the initial feature points in the specified feature point sequence and the current reference feature points in the current reference feature point sequence; Using the initial feature points in the dataset feature point sequence as the current initial feature points respectively to perform the following determination operations to obtain the set of current matching pairs: Determining the current reference feature point in the current reference feature point sequence with the largest similarity to the current initial feature point to obtain the current candidate reference feature point; In the case where the similarity between the current initial feature point and the current candidate reference feature point is greater than or equal to the similarity threshold, determining the current initial feature point and the current candidate reference feature point as a current matching pair.

6. The method according to claim 4, characterized in that Both the initial feature points in the dataset feature point sequence and the current reference feature points in the current reference feature point sequence are set with feature point label information; The feature point matching of the dataset feature point sequence and the current reference feature point sequence to obtain a set of current matching pairs includes: Using the initial feature points in the dataset feature point sequence as the current initial feature points respectively to perform the following determination operations to obtain the set of current matching pairs: Searching for the current reference feature points in the current reference feature point sequence with the same feature point label information as the current initial feature point; In the case where the current reference feature points with the same feature point label information as the current initial feature point are found, determining the current initial feature point and the current reference feature points with the same feature point label information as the current initial feature point as a current matching pair.

7. The method according to claim 3, characterized in that, There is at least one type of specified feedback data in the specified feedback dataset; The extraction of the initial feature point set of the specified feedback dataset includes: In the case where the specified feedback dataset includes text data, performing syntactic analysis on the text data in the specified feedback dataset to obtain the text feature points of the specified feedback dataset, where the initial feature point set includes the text feature points of the specified feedback dataset; In the case where the specified feedback dataset includes image data, performing image feature recognition on the image data in the specified feedback dataset to obtain the image feature points of the specified feedback dataset, where the initial feature point set includes the image feature points of the specified feedback dataset; When the specified feedback data set includes audio data, perform audio feature analysis on the audio data in the specified feedback data set to obtain audio feature points of the specified feedback data set, where the initial feature point set includes the audio feature points of the specified feedback data set.

8. The method according to claim 3, wherein The data types allowed to be input by the dialogue model to be updated match the model type of the dialogue model to be updated; the method further includes: When the dialogue model to be updated is a multi-modal model and there are at least two types of initial feature points in the initial feature point set, input the at least two types of initial feature points into a preset attention network to obtain at least two attention weights, where the at least two attention weights correspond one-to-one to the at least two types of initial feature points; Use the at least two attention weights to perform weighted fusion on the at least two types of initial feature points to obtain a set of cross-modal feature points, where the initial feature point set includes the at least two types of initial feature points and the set of cross-modal feature points, or the at least two types of initial feature points in the initial feature point set are updated to the set of cross-modal feature points.

9. The method according to claim 2, wherein The artificial intelligence platform is provided with a correspondence table for recording the correspondence between the preset difference values corresponding to the pre-trained models in the set of pre-trained models and the parameter adjustment methods in the set of parameter adjustment methods; Determining a specified parameter adjustment method from a set of parameter adjustment methods according to the feature point matching result includes: Calculate the absolute value of the difference between the difference value of the same pre-trained model in the set of pre-trained models and the corresponding preset difference value respectively to obtain an absolute difference value set corresponding to the set of pre-trained models; Determine the parameter adjustment method of the pre-trained model corresponding to the minimum absolute difference value in the absolute difference value set as the specified parameter adjustment method.

10. The method according to claim 1, wherein Obtaining the specified feedback data set includes: Perform data cleaning on the candidate feedback data in the candidate feedback data set corresponding to the dialogue model to be updated, and classify the cleaned candidate feedback data according to the data type to obtain the classified candidate feedback data, where the candidate feedback data in the candidate feedback data set is the feedback data obtained by using the dialogue model to be updated for historical conversations; Input the classified candidate feedback data into a preset identification model to obtain the data label information of the classified candidate feedback data; Based on the classified candidate feedback data and the data label information of the classified candidate feedback data, construct the specified feedback data set, where the specified feedback data in the specified feedback data set is the classified candidate feedback data with data label information.

11. The method according to claim 1, characterized in that, The artificial intelligence platform is also deployed with a specified container; Before obtaining the specified feedback data set, the method further includes: Monitor the resource utilization rate of the artificial intelligence platform; When the resource utilization rate is less than the first resource utilization rate threshold, determine that the specified container is in an available state; Configure a container resource range for the specified container according to the update resource requirements of the to-be-updated dialogue model, where the maximum value of the container resource range is less than a second resource utilization rate threshold, and the second resource utilization rate threshold is greater than the first resource utilization rate threshold; Among them, the update of the to-be-updated dialogue model is executed within the specified container using the resources within the container resource range configured for the specified container.

12. The method according to any one of claims 1 to 11, characterized in that, After the model parameters of the to-be-updated dialogue model are updated according to the specified feedback data set and the specified parameter adjustment method, the method further includes: Perform a preliminary verification on the updated to-be-updated dialogue model, where the preliminary verification includes performing a performance test on the updated to-be-updated dialogue model and performing a stability test on the updated to-be-updated dialogue model; In the case where the preliminary verification passes, deploy the updated to-be-updated dialogue model to the artificial intelligence platform; In the case where it is monitored that the updated to-be-updated dialogue model has a performance degradation compared to the to-be-updated dialogue model before the update, roll back the updated to-be-updated dialogue model to the to-be-updated dialogue model before the update.

13. An electronic device, characterized in that, Include: A memory for storing a computer program; A processor for implementing the steps of the model update method as described in any one of claims 1 to 12 when executing the computer program.

14. A computer-readable storage medium, characterized in that, A computer program is stored in the computer-readable storage medium, where the computer program, when executed by a processor, implements the steps of the model update method as described in any one of claims 1 to 12.

15. A computer program product, comprising a computer program, characterized in that, The computer program, when executed by a processor, implements the steps of the model update method as described in any one of claims 1 to 12.

Images