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

Abstract

The present application discloses a model updating method, electronic device, storage medium and program product, which relate to the field of machine learning technology, including: when it is monitored that there is a dialogue model to be updated in a group of dialogue models, obtaining a specified feedback data set, performing feature point matching on a data set feature point sequence of the specified feedback data set and a baseline feature point sequence of a pre-trained model in a group of pre-trained models, and obtaining a feature point matching result; and determining a specified parameter adjustment method from a group of parameter adjustment methods based on the feature point matching result, thereby updating the model parameters of the dialogue model to be updated based on the specified feedback data set and the specified parameter adjustment method, so that the model update is more in line with actual needs, avoids invalid or negative tuning, solves the technical problem of poor model update effect existing in the related art, and achieves the technical effect of stable improvement of dialogue model performance.

Description

Technical Field

[0001] The present application relates to the field of machine learning technology, and in particular to a model updating method, electronic device, storage medium, and program product. Background Art

[0002] With the advent of the artificial intelligence era, the widespread adoption of AI platforms and big data model training, and the widespread availability of integrated training and push machines, the demand for model tuning capabilities on AI platforms has become increasingly critical. During the use of large models in these machines, conversational models must be continuously updated to adapt to changing user needs and improve service quality. However, most existing model update strategies either rely solely on predefined, general datasets, ignoring the value of real-time user feedback; or they are overly simplistic and crude in processing user feedback data, failing to accurately capture subtle changes in user needs. This results in insignificant or even degraded model update performance.

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

[0004] The present application provides a model updating method, an electronic device, a storage medium, and a program product to at least solve the problem of poor model updating effect in related technologies.

[0005] This application provides a model updating method, which is applied to an artificial intelligence platform, wherein the artificial intelligence platform is deployed with a set of dialogue models; the method comprises:

[0006] When it is detected that a dialogue model to be updated exists in a group of dialogue models, obtaining a designated feedback data set, wherein the designated feedback data in the designated feedback data set is feedback data obtained by conducting historical dialogues using the dialogue model to be updated;

[0007] performing feature point matching on a dataset feature point sequence of the specified feedback dataset and a reference feature point sequence of a pre-trained model in a set of pre-trained models to obtain a feature point matching result, wherein the dataset feature point sequence is used to represent attribute information of the specified feedback dataset;

[0008] determining, based on the feature point matching result, a specified parameter adjustment method from a set of parameter adjustment methods, wherein the parameter adjustment method in the set of parameter adjustment methods is a method of adjusting model parameters of a dialogue model in the set of dialogue models;

[0009] The model parameters of the dialogue model to be updated are updated according to the specified feedback data set and the specified parameter adjustment method.

[0010] The present application also provides a model updating device, which is applied to an artificial intelligence platform, wherein the artificial intelligence platform is deployed with a set of dialogue models; the device comprises:

[0011] an acquisition module, configured to, upon detecting that a dialogue model to be updated exists in the group of dialogue models, acquire a specified feedback data set, wherein the specified feedback data in the specified feedback data set is feedback data obtained from historical dialogues using the dialogue model to be updated;

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

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

[0014] An updating module is configured to update 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, comprising: a memory for storing a computer program; and a processor for implementing the steps of any of the above-mentioned 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. When the computer program is executed by a processor, the steps of any of the above-mentioned model updating methods are implemented.

[0017] The present application also provides a computer program product, comprising a computer program, which implements the steps of any of the above-mentioned model updating methods when executed by a processor.

[0018] Through the present application, when it is monitored that there is a dialogue model to be updated in a group of dialogue models, a specified feedback dataset is obtained, and feature point matching is performed on the dataset feature point sequence of the specified feedback dataset and the baseline feature point sequence of the pre-trained model in a group of pre-trained models to obtain a feature point matching result. Based on the feature point matching result, a specified parameter adjustment method is determined from a group of parameter adjustment methods, and model parameters of the dialogue model to be updated are updated according to the specified feedback dataset and the specified parameter adjustment method. By performing feature point matching on the dataset feature point sequence of the specified feedback dataset and the baseline feature point sequence of the pre-trained model in a group of pre-trained models and determining the specified parameter adjustment method based on the feature point matching result, the dialogue model to be updated is updated, so that the model update is more in line with actual needs, avoiding invalid or negative tuning, and to a certain extent ensuring the performance of the updated dialogue model. This solves the technical problem of poor model update effect existing in the related art, and achieves the technical effect of stable improvement of dialogue model performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the embodiments of the present application, the following is a brief introduction to the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0020] Figure 1 A schematic diagram of an application scenario of a model updating method provided in an embodiment of the present application.

[0021] Figure 2 A flowchart of an optional model updating method provided in an embodiment of the present application.

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

[0023] Figure 4 A schematic diagram of an optional model updating method provided in an embodiment of the present application.

[0024] Figure 5 A structural block diagram of an optional model updating device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0025] The following will be combined with the accompanying drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

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

[0027] In order to enable those skilled in the art to better understand the present application, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0028] With the advent of the artificial intelligence era, the widespread adoption of AI platforms and big data model training, and the widespread availability of integrated training and push machines, the demand for model tuning capabilities on AI platforms has become increasingly critical. During the use of large models in these machines, conversational models must be continuously updated to adapt to changing user needs and improve service quality. However, most existing model update strategies either rely entirely on predefined, generalized datasets, ignoring the value of real-time user feedback; or are overly simplistic and crude in processing user feedback data, failing to accurately capture subtle changes in user needs. This results in insignificant or even degraded model update performance. Consequently, these existing technologies suffer from poor model update effectiveness.

[0029] To solve the above problems, an embodiment of the present application provides a model updating method, which updates the dialogue model to be updated by performing feature point matching on a dataset feature point sequence of a specified feedback dataset and a baseline feature point sequence of a pre-trained model in a group of pre-trained models, and determining a specified parameter adjustment method based on the feature point matching results. This makes the model update more in line with actual needs, avoids invalid or negative tuning, and to a certain extent guarantees the performance of the updated dialogue model. It solves the technical problem of poor model update effect existing in related technologies and achieves the technical effect of stable improvement of dialogue model performance.

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

[0031] The aforementioned network may include, but is not limited to, at least one of the following: a wired network and a wireless network. The aforementioned wired network may include, but is not limited to, at least one of the following: a wide area network, a metropolitan area network, or a local area network. The aforementioned wireless network may include, but is not limited to, at least one of the following: wireless fidelity (Wi-Fi) and Bluetooth. The terminal device 102 may be, but is not limited to, a personal computer (PC), a mobile phone, a tablet computer, etc. The server 104 may be, but is not limited to, a cloud server, a server cluster, or other server types.

[0032] The model updating method of the embodiment of the present application can be executed by the server 104, or by the terminal device 102, or by both the server 104 and the terminal device 102. The model updating method of the embodiment of the present application can also be executed by the client installed on the terminal device 102.

[0033] Taking the terminal device 102 as an example to execute the model updating method in this embodiment, the model updating method is applied to an artificial intelligence platform, wherein the artificial intelligence platform deploys a group of dialogue models. Figure 2 is a flow chart of an optional model updating method according to an embodiment of the present application, such as Figure 2 As shown, the process of the method may include the following steps:

[0034] Step S202 , when it is detected that a dialogue model to be updated exists in a group of dialogue models, a designated feedback data set is obtained, wherein the designated feedback data in the designated feedback data set is feedback data obtained from historical dialogues using the dialogue model to be updated.

[0035] It should be noted that the model update method of this embodiment can be applied to an artificial intelligence (AI) platform environment, providing automated and personalized updates for a set of deployed conversational models. An AI platform can refer to a system platform capable of processing large-scale data, training, and running AI models, such as a high-performance training and promotion platform. In this embodiment, the AI ​​platform can deploy, manage, and update a variety of different types of conversational models. Multiple agent applications can be deployed within the AI ​​platform, each of which has a conversational model deployed.

[0036] A group of dialogue models can be multiple AI models deployed on an artificial intelligence platform. A group of dialogue models is mainly used for natural language processing tasks, such as chatbots, voice assistants, etc., which can interact with users through dialogue.

[0037] Optionally, determining whether a conversation model in a group of conversation models requires updating can be done through real-time performance monitoring, identifying metrics such as latency and accuracy fluctuations when the model processes requests. Once a conversation model's feedback score falls below an average threshold or its performance metrics decline, the conversation model is marked as requiring updating.

[0038] Alternatively, if the amount of specified feedback data exceeds a preset threshold, a dialog model is determined to be present in the group of dialog models to be updated. Alternatively, a threshold for the amount of feedback data is set for each dialog model, and the accumulated feedback data received by each dialog model up to the current time is recorded. If the amount of feedback data exceeds the threshold corresponding to the dialog model, the dialog model with the amount of feedback data exceeding the threshold is determined to be the dialog model to be updated.

[0039] Optionally, when it is determined that there is a dialogue model to be updated in a group of dialogue models, a specified feedback data set is obtained, where the specified feedback data set includes multiple specified feedback data. The specified feedback data can be obtained by collecting feedback information from historical dialogues between users and dialogue models in the platform, where the feedback information includes but is not limited to user satisfaction scores for dialogue model responses, correction suggestions, or direct negative feedback.

[0040] Step S204 , performing feature point matching on the dataset feature point sequence of the specified feedback dataset and the baseline feature point sequence of the pre-trained model in a group of pre-trained models to obtain a feature point matching result, wherein the dataset feature point sequence is used to represent the attribute information of the specified feedback dataset.

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

[0042] Optionally, the dataset feature point sequence can be used to characterize attribute information of a specified feedback dataset. This attribute information includes at least one of the following: data statistics, data type information, data complexity information, and dataset application scenario information. Data statistics can include basic statistical data of a specified feedback dataset, including but not limited to data volume, average conversation length, question frequency, user activity, and so on. Data statistics can indicate the scale and structural characteristics of the dataset. Data type information can indicate the types of data contained in the specified feedback dataset, such as text data, voice data, image data, or mixed data. Data complexity information can include information about the dimensionality of the data within the specified feedback dataset and the similarity of the data within the specified feedback dataset. Specifically, data complexity information can describe the organizational hierarchy and diversity of the information within the dataset, including semantic complexity, topic breadth and depth, and emotional fluctuation range. The dataset application scenario information can specifically indicate the actual application scenarios from which the dataset was derived, such as customer service Q&A, medical consultation, and educational counseling.

[0043] The AI ​​platform deploys a model repository, which contains a set of pretrained models. These models are deep learning models pre-trained on large amounts of general data (such as books and news). Pretrained models and conversational models operate on a hierarchical scale, from "basic" to "specialized." Through further fine-tuning, these models are tailored to specific conversational scenarios or tasks, such as customer service conversations and intelligent question-and-answering. Pretrained models provide powerful initial capabilities for conversational models, while conversational models further optimize the performance of pretrained models by learning from conversational data from specific domains or users.

[0044] Each pre-trained model in a set of pre-trained models corresponds to a baseline feature point sequence. This baseline feature point sequence represents a collection of inherent performance metrics for the pre-trained model in a specific domain or task. The baseline feature point sequence includes a set of baseline feature points that can cover various properties of the pre-trained model, such as model complexity (number of layers, number of parameters), data preference (images, text, code, etc.), processing speed, and accuracy.

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

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

[0047] Step S206 : determining a specified parameter adjustment method from a group of parameter adjustment methods based on the feature point matching result, wherein the parameter adjustment method in the group of parameter adjustment methods is a method for adjusting model parameters of a dialogue model in a group of dialogue models.

[0048] It should be noted that a set of parameter adjustment methods can be a preset set of methods for adjusting model parameters, which can include a variety of fine-tuning strategies, such as learning rate adjustment, regularization parameter setting, layered learning rate, low-rank adaptation (LoRA) fine-tuning, etc. Each strategy optimizes different aspects of the model parameters to adapt to specific data sets and task requirements.

[0049] Specifically, when parameter adjustment is performed using the learning rate, the learning rate refers to the factor that controls 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 undertraining. Typically, the platform dynamically adjusts the learning rate, for example, by employing a learning rate decay strategy. Initially setting a high learning rate allows the model to quickly explore the solution space, then gradually reducing it to allow the model to more finely tune parameters and achieve more refined optimization results.

[0050] When regularization is used as a parameter adjustment method, overly complex models are penalized by adding regularization terms to the loss function. During fine-tuning, properly setting the L1 and L2 regularization parameters can control the size of model parameters, prevent the model from overfitting to noisy data, and improve the model's generalization ability. The choice of regularization parameters generally depends on the size and complexity of the dataset, as well as the capacity of the model itself.

[0051] When using layered learning rate as the parameter adjustment method, different learning rates are set based on the characteristics of different model layers. For example, for deep models, the learning rate near the input layer may be set lower because these layers have already learned more general feature representations, while the learning rate near the output layer may be set higher to better adapt to the details of specific tasks. Layered learning rates help balance overall model updates and local optimization, improving training efficiency and model performance.

[0052] When the parameter adjustment method is LoRA fine-tuning, a small number of parameters in the pre-trained model are adjusted to adapt to the new task through low-rank matrix decomposition. This method saves a lot of computing resources and storage space and is particularly suitable for fine-tuning large pre-trained models in resource-constrained environments.

[0053] Optionally, based on the feature point matching results, a specific parameter adjustment method may be determined from a set of parameter adjustment methods. The specific parameter adjustment method may be considered to be most suitable for the current dataset and the dialogue model to be tuned.

[0054] Step S208: updating the model parameters of the dialogue model to be updated according to the specified feedback data set and the specified parameter adjustment method.

[0055] Optionally, the parameters of the dialogue model to be updated are fine-tuned using a specified feedback dataset and a selected specified parameter adjustment method to improve the performance of the model and provide better service.

[0056] In one example, assume a conversational model based on the Transformer architecture is deployed on an artificial intelligence platform. If the platform detects that users frequently give the conversational model responses with low quality scores, or if the amount of feedback data for the conversational model exceeds a threshold, it obtains a designated feedback dataset. The artificial intelligence platform then derives a dataset feature point sequence from the designated feedback dataset. Specifically, the dataset feature point sequence may include the complexity of the conversational context and the distribution of user questions. The dataset feature point sequence is then matched against a set of baseline feature point sequences from pre-trained models to identify the closest model and parameter adjustment method. The dataset feature point sequence most closely matches the baseline feature point sequence of pre-trained model A, which previously used a "layered learning rate" to improve its generalization capability. Therefore, the "layered learning rate" can be used as the designated parameter adjustment method to retrain the conversational model and adjust its parameters. This embodiment allows an artificial intelligence platform to avoid manual intervention, significantly shorten the time period from user feedback to model updates, and improve the platform's responsiveness and service quality.

[0057] According to an embodiment of the present application, when it is detected that a dialogue model to be updated exists in a group of dialogue models, a specified feedback dataset is obtained, and feature point matching is performed on a dataset feature point sequence of the specified feedback dataset and a baseline feature point sequence of a pre-trained model in a group of pre-trained models to obtain a feature point matching result. A specified parameter adjustment method is determined from a group of parameter adjustment methods based on the feature point matching result, and model parameters of the dialogue model to be updated are updated based on the specified feedback dataset and the specified parameter adjustment method. By performing feature point matching on a dataset feature point sequence of the specified feedback dataset and a baseline feature point sequence of a pre-trained model in a group of pre-trained models and determining the specified parameter adjustment method based on the feature point matching result, the dialogue model to be updated is updated, making the model update more in line with actual needs, avoiding invalid or negative tuning, and to a certain extent ensuring the performance of the updated dialogue model. This solves the technical problem of poor model update effect existing in the related art, and achieves the technical effect of steadily improving the performance of the dialogue model.

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

[0059] It should be noted that a dataset feature point sequence can be an ordered set of numerical features for a specific feedback dataset. This dataset feature point sequence can include information such as the size of the feedback dataset, its data type (e.g., text, image, language), and its distribution characteristics (e.g., topic diversity). The dataset feature point sequence can represent the attribute information of the specific feedback dataset.

[0060] A baseline feature point sequence represents a set of inherent performance metrics for a pre-trained model in a specific domain or task. The baseline feature point sequence includes a set of baseline feature points that cover various attributes of the pre-trained model, such as model complexity (number of layers, number of parameters), data preference (images, text, code, etc.), processing speed, and accuracy.

[0061] Optionally, during the feature point matching process, the difference between the feature point sequence of the dataset and the baseline feature point sequence of each pre-trained model in a set 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 ​​constitute a difference value set.

[0062] This embodiment accurately identifies the optimal match between the model and the dataset through feature point matching and calculation of difference value sets, enabling targeted model parameter adjustments and avoiding the risk of performance degradation caused by blind tuning. The updated conversational model can process data with greater accuracy and responsiveness, improving the conversational quality of the conversational model.

[0063] In an exemplary embodiment, before performing feature point matching on the dataset feature point sequence of a specified feedback dataset and the baseline feature point sequence of a pre-trained model in a group of pre-trained models, the above method also includes: extracting an initial feature point set of the specified feedback dataset; sorting 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 a dataset feature point sequence.

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

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

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

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

[0068] In one example, a designated feedback dataset is collected within an AI platform to update a specific conversational model within a set of conversational models, namely, the conversational model to be updated. An initial feature point set is extracted from the designated feedback dataset and sorted. For example, the initial feature points can be sorted based on their impact on model performance. For example, semantic complexity and user satisfaction ratings are identified as the most influential feature points for model tuning and are ranked at the top of the sequence. The top n initial feature points (e.g., the top five) are selected from the sorted dataset and defined as a dataset feature point sequence. This sequence can include key metrics such as semantic complexity, user satisfaction ratings, and conversation length. Finally, the dataset feature point sequence is compared and matched with a baseline feature point sequence from a pre-trained model to identify the closest pre-trained model. This allows the system to determine a corresponding fine-tuning strategy to update the specific conversational model and better respond to user feedback and needs.

[0069] This embodiment extracts an initial set of feature points from a specified feedback dataset before performing feature point matching. These initial feature points are sorted to form a data feature point sequence, selecting a specified number of initial feature points. This focuses on the key feature points of the dataset, allowing model tuning to more accurately identify and address key issues in the dataset, avoiding adjustments to irrelevant features and improving the relevance and effectiveness of the tuning. Furthermore, the most relevant feature points are selected for processing, reducing the waste of computing resources and improving the efficiency of the feature point matching process.

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

[0071] The following matching operations are performed on each of the pre-trained models in the set of pre-trained models as the current pre-trained model to obtain a set of difference values ​​corresponding to the pre-trained models in the set of pre-trained models, where the reference feature point sequence of the current pre-trained model is the current reference feature point sequence:

[0072] Feature point matching is performed on the feature point sequence of the dataset and the current benchmark feature point sequence to obtain a set of current matching pairs, wherein a current matching pair in the set of current matching pairs includes an initial feature point in the feature point sequence of the dataset 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 respectively calculated to obtain the difference value corresponding to the current pre-trained model; wherein the difference value set corresponding to a 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 set of pre-trained models may have different performance characteristics due to differences in architecture, training data, or parameter settings. The current pre-trained model may be a pre-trained model selected from a set of pre-trained models for feature matching. The current pre-trained model may have a pre-set baseline feature point sequence, namely, a current baseline feature point sequence. The baseline feature point sequence of the current pre-trained model can be used to perform feature point matching with the feature point sequence of the dataset to obtain a set of current matching pairs.

[0074] The current matching pair can be a pair consisting of two feature points with a high degree of matching between the dataset feature point sequence and the current baseline feature point sequence of the current pre-trained model. It can be used to reflect the matching degree between the dataset feature point sequence of the specified feedback dataset and the pre-trained model in the feature dimension. Specifically, the baseline feature point sequence can be a series of feature points defined for each pre-trained model based on its own performance and preferences. The baseline feature points in the baseline feature point sequence 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 specific tasks, etc.

[0075] A set of current matching pairs may include an initial feature point in the dataset feature point sequence and a current reference feature point in the reference feature point sequence of the current pre-trained model. By calculating the differences (such as variances or distances) between the feature points in each set of current matching pairs, a numerical set is formed and accumulated to obtain the difference value corresponding to the current pre-trained model, thereby obtaining a set of difference values ​​corresponding to a set of pre-trained models. The set of difference values ​​corresponding to a set of pre-trained models includes the difference value corresponding to each pre-trained model.

[0076] Specifically, the process of calculating the difference value corresponding to the current pre-trained model by specifying the data feature point sequence of the feedback data set and the current baseline feature point sequence of the current pre-trained model can be shown as formula (1):

[0077] ; (1)

[0078] in, is the difference value corresponding to the current pre-training model, also known as the feature point variance, is the number of feature points in the data feature point sequence, Indicates the first of a set of current matching pairs. Current matching pairs, is the first point in the data feature point sequence Initial feature points, is the first point in the current benchmark feature point sequence The initial feature points constitute the current reference feature points of the current matching pair.

[0079] Through this embodiment, a set of current matching pairs between the feature point sequence of the data set and the current benchmark feature point sequence is determined, and the difference value corresponding to the pre-trained model is determined based on the set of current matching pairs. Therefore, according to the difference value corresponding to the pre-trained model, a suitable pre-trained model is screened and the specified parameter adjustment method is determined, thereby improving the accuracy of the model update.

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

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

[0082] Calculating 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 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] The initial feature points in the feature point sequence of the data set are respectively used as the current initial feature points to perform the following determination operations to obtain a set of current matching pairs: determine the current benchmark feature point with the greatest similarity to the current initial feature point in the current benchmark feature point sequence to obtain the current candidate benchmark feature point; when the similarity between the current initial feature point and the current candidate benchmark feature point is greater than or equal to the similarity threshold, determine the current initial feature point and the current candidate benchmark feature point as a current matching pair.

[0084] It should be noted that feature space mapping can be performed by converting the dataset feature point sequence into the feature space of a pre-trained model to ensure that the initial feature points in the dataset feature point sequence and the reference feature points in the reference feature point sequence are in the same feature space, facilitating subsequent similarity calculation. Optionally, feature space mapping can be achieved through a mathematical transformation or machine learning algorithm, such as principal component analysis dimensionality reduction (PCA) or t-Distributed Stochastic Neighbor Embedding Mapping (t-SNE). Of course, the feature extraction layer of the pre-trained model can be used to perform feature space mapping on the dataset feature point sequence to obtain a specified feature point sequence for the specified feedback dataset. The specified feature point sequence can be the mapped dataset feature point sequence and can include multiple initial feature points. The specified feature point sequence and the current reference feature point sequence are in the same feature space to facilitate feature matching and similarity calculation.

[0085] The similarity between the initial feature point in the specified feature point sequence and the current benchmark feature point in the current benchmark feature point sequence is calculated to measure the similarity between the feature point sequence of the dataset and the current benchmark feature point of the pre-trained model. Commonly used similarity calculation methods include cosine similarity, Jaccard similarity, and Pearson correlation coefficient.

[0086] In order to determine whether the initial feature point and the current benchmark feature point can form a matching pair, a numerical value can be pre-set to judge whether the similarity between the initial feature point and the current benchmark feature point is large enough. If this similarity is lower than the similarity threshold, it is determined that there is not enough similarity between the initial feature point and the current benchmark feature point 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 (i.e., the most matching pairs and relatively large similarity) to fine-tune the dialogue model to be updated.

[0088] Optionally, a weight may be assigned to each initial feature point, and the difference value corresponding to the pre-trained model may be calculated based on the assigned weight. Specifically, a first weight may be assigned to each sorted initial feature point based on the order of the initial feature points in the feature point sequence of the dataset. A second weight may then be assigned based on the degree of similarity between the initial feature point and a reference feature point in the reference feature point sequence of the pre-trained model when the initial feature point forms a matching pair. The difference value corresponding to the pre-trained model may be calculated by combining the first and second weights.

[0089] It should be noted that the first weight can be assigned based on the importance of the initial feature point. This can be determined by analyzing the frequency of user feedback, the impact on model performance, and the prevalence of the feature point in the dataset. For example, if a feature point has a significant impact on the model's prediction accuracy, it will receive a higher first weight.

[0090] The second weight is assigned based on the similarity between the initial feature point and the pre-trained model matching pair. Similarity can be calculated using various methods, such as Euclidean distance and cosine similarity. The more similar a feature point is to a model feature point, the closer the model's performance at that feature point is to the dataset requirements, thus receiving a higher second weight. The positive correlation between the second weight and the similarity between the matching pairs enables the fine-tuning strategy to adaptively adjust to better match the dataset characteristics and improve the model's generalization capabilities.

[0091] Through the above-described embodiment, the dual-weight mechanism considers both the intrinsic importance and extrinsic matching of feature points, resulting in more accurate calculated differences that better reflect the fit between the model and the dataset. Feature points that are more important and better match the model receive higher weights when calculating differences, which means the fine-tuning process will favor optimizing these feature points, ultimately improving the model's performance on critical tasks.

[0092] Through this embodiment, the AI ​​platform can more objectively and accurately assess the matching degree between the model and the dataset through feature space mapping and similarity calculation. Similarity calculation can quickly eliminate models that seriously mismatch the dataset, reducing ineffective calculations and resource waste, and improving the overall efficiency of model fine-tuning and updating.

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

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

[0095] The initial feature points in the feature point sequence of the data set are respectively used as the current initial feature points to perform the following determination operations to obtain a set of current matching pairs: search for the current benchmark feature point in the current benchmark feature point sequence whose feature point label information is the same as the feature point label information of the current initial feature point; when the current benchmark feature point whose feature point label information is the same as the feature point label information of the current initial feature point is found, the current initial feature point and the current benchmark feature point whose feature point label information is the same as the feature point label information of the current initial feature point are determined as a current matching pair.

[0096] It should be noted that a dataset feature point sequence can include multiple initial feature points. An initial feature point can represent a feature corresponding to a specific attribute in a given feedback dataset. Feature point label information can provide additional descriptions of the initial feature point, primarily used to classify and identify the type of the initial feature point. Examples include labels such as "emotional polarity" and "topic." Each baseline feature point in the pre-trained model's baseline feature point sequence also includes feature point label information.

[0097] Optionally, a current matching pair can be a matching pair consisting of an initial feature point in the dataset feature point sequence and a current reference feature point in the current reference feature point sequence that has the same feature point label information. Optionally, the feature point label information can be understood as specific identification information of the feature point, which can be represented by a string or the like. An initial feature point and a current reference feature point that have the same specific identification information can constitute a matching pair.

[0098] Optionally, after finding the current benchmark feature point with the same feature point label information, the current initial feature point in the dataset feature point sequence and the current benchmark feature point of the pre-trained model are combined into a current matching pair. This process is repeated until all initial feature points in the dataset feature point sequence find the current benchmark feature point corresponding to the current pre-trained model, or if they are not completely found, the matching process ends.

[0099] This embodiment uses feature point label information for matching, ensuring that the pre-trained model selection is closely aligned with the dataset's properties, thereby improving model matching accuracy. Fine-tuning the model to better match the dataset's characteristics can significantly improve the performance of the updated conversational model in specific scenarios. Furthermore, using feature point label information reduces unnecessary data processing, speeds up the matching process, and improves the efficiency of the tuning process.

[0100] In an exemplary embodiment, at least one type of feedback data exists in the specified feedback data set;

[0101] Extracting an initial feature point set of a specified feedback dataset includes: when the specified feedback dataset includes text data, performing syntactic analysis on the text data in the specified feedback dataset to obtain text feature points of the specified feedback dataset, wherein the initial feature point set includes the text feature points of the specified feedback dataset; when the specified feedback dataset includes image data, performing image feature recognition on the image data in the specified feedback dataset to obtain image feature points of the specified feedback dataset, wherein the initial feature point set includes the image feature points of the specified feedback dataset; when the specified feedback dataset includes audio data, performing audio feature analysis on the audio data in the specified feedback dataset to obtain audio feature points of the specified feedback dataset, wherein the initial feature point set includes the audio feature points of the specified feedback dataset.

[0102] It should be noted that a designated feedback dataset can be collected by an AI platform and is a collection of feedback data generated by users interacting with an agent application within the AI ​​platform. A designated feedback dataset can include various types of information, such as text, images, and audio. The designated feedback data in the designated feedback dataset can be used to train and fine-tune the conversational model used in the agent application.

[0103] Syntactic analysis can refer to the process of extracting meaningful text features by analyzing sentence structure and identifying parts of speech, phrase boundaries, and dependency relationships through natural language processing techniques. Alternatively, AI platforms can leverage natural language processing (NLP) tools such as parsers and named entity recognizers to conduct in-depth analysis of text data, identifying sentence subject-verb-object structure, sentiment polarity, keywords, and more. For example, if user feedback indicates that most conversations involve specialized terminology from a specific field, identifying these terms will become a feature point used to assess the pre-trained model's mastery of relevant domain knowledge.

[0104] Image feature recognition refers to the process of extracting key features from an image, such as color, texture, shape, and objects, using computer vision techniques. Optionally, the AI ​​platform can pre-deploy an image recognition model, such as a convolutional neural network (CNN), to process image data and identify objects, region boundaries, color distribution, and other features within the image to generate image feature points. These extracted image feature points are also incorporated into the initial feature point set for subsequent model fine-tuning. If a large number of images in user feedback contain rich colors and complex backgrounds, color recognition capabilities and background processing techniques will be important features to assess the model's suitability for such image data. For example, when an intelligent agent application receives a user-uploaded food image, the AI ​​platform can use image feature recognition to identify features such as the type of food, color distribution, and image clarity. This information will guide the platform in selecting pre-trained models that excel in food recognition and color perception.

[0105] Audio feature analysis is the process of extracting key features such as voiceprint, pitch, and speaking rate from audio data through signal processing and machine learning techniques. Optionally, the AI ​​platform can deploy speech recognition and audio analysis algorithms, such as Mel Frequency Cepstral Coefficients (MFCC) or Spectral Centroid, to process audio clips and identify the speaker's emotional state, speaking rate, and pitch, thereby forming a set of audio feature points.

[0106] Optionally, the text feature points include at least one of the following: a feature point for representing the emotional tendency of the text data, or a feature point for representing the topic distribution of the text data. The image feature points include at least one of the following: a feature point for representing object information of an object in the image, a feature point for representing the texture of the image, or a feature point for representing the color distribution in the image. The audio feature points include at least one of the following: a feature point for representing the spectral characteristics of the audio, a feature point for representing the emotion of the audio, or a feature point for representing changes in intonation in the audio.

[0107] This embodiment performs feature analysis on different types of designated feedback data, enabling the AI ​​platform to more precisely understand user needs and application scenarios, enabling more accurate adjustment of model parameters and improving the performance of the updated conversational model. Furthermore, during training and tuning, the AI ​​platform comprehensively considers the features of multiple data types, such as text, images, and audio, to enhance the updated conversational model's ability to integrate and learn diverse information, thereby increasing its comprehensive processing capabilities and flexibility.

[0108] In an exemplary embodiment, the dialog model to be updated allows input data types that match the model type of the dialog model to be updated. The method further includes: if the dialog model to be updated is a multimodal model and the initial feature point set includes at least two types of initial feature points, inputting the at least two types of initial feature points into a preset attention network to obtain at least two attention weights, wherein 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, weightedly fusing the at least two types of initial feature points to obtain a set of cross-modal feature points, wherein the initial feature point set includes the at least two types of initial feature points and the set of cross-modal feature points, or updating the at least two types of initial feature points in the initial feature point set to form a set of cross-modal feature points.

[0109] It should be noted that a group of dialogue models can exist in many different types, such as large language models (supporting only text reasoning dialogues), multimodal models (supporting image file text analysis and reasoning dialogues), and deep thinking models (supporting deep thinking and reasoning dialogues), etc.

[0110] The model type of the dialogue model to be updated determines the data types it accepts as input, such as text, images, or audio. Before model tuning, the AI ​​platform can verify whether the dialogue model to be updated supports the data types present in the specified feedback dataset. If the model is multimodal and contains at least two types of data (e.g., text and images), the dialogue model to be updated can be trained and tuned across these modalities.

[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 may include initial feature points such as text feature points, image feature points, and audio feature points.

[0113] When 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. A 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, we can then combine them with different types of feature points to generate a set of cross-modal feature points representing comprehensive information. These cross-modal feature points represent a set of feature points representing comprehensive modal information. These cross-modal feature points will be incorporated into the initial feature point set or replace the original feature point set.

[0115] Through this embodiment, by using a preset attention network to perform weighted fusion of different types of feature points corresponding to multimodal data, a more comprehensive and accurate set of 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, thereby 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, which is used to record the correspondence between the preset difference values ​​corresponding to the pre-trained models in the group of pre-trained models and the parameter adjustment methods in the group 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 group of pre-trained models and the corresponding preset difference value to obtain an absolute difference value set corresponding to the group of pre-trained models; and determining 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.

[0118] It's important to note that the AI ​​platform internally stores a data table, known as a correspondence table. This table is pre-established by the platform to record the expected differences between different models at different feature points, along with the most appropriate parameter adjustments for these situations. This correspondence can be determined through experimentation during early model development or summarized based on historical tuning data. The preset difference values ​​can be standard values ​​corresponding to the parameter adjustment methods. These values ​​can be determined based on the expected performance of the model at different feature points, such as precision and recall.

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

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

[0121] Parameter fine-tuning methods can be used to fine-tune 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 value refers to the absolute difference between the difference value corresponding to each pre-trained model in a set of pre-trained models obtained by the AI ​​platform and the preset difference value recorded in the correspondence table. It measures the degree of deviation between the actual difference of the model and the expected difference. The set of absolute difference values ​​corresponding to a set of pre-trained models includes the absolute difference value corresponding to each pre-trained model in the set of pre-trained models. Based on the set of absolute difference values, the parameter adjustment method corresponding to the pre-trained model corresponding to the minimum difference value is selected as 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 performed.

[0124] In specific practice, Figure 3 As shown, a set of pre-trained models in the AI ​​platform can be stored in a model warehouse. The model warehouse contains n pre-trained models, and each pre-trained model of different types corresponds to a parameter adjustment method (i.e., full parameter fine-tuning method). Based on the dataset, a model verification process is performed in the model warehouse to determine a pre-trained model, i.e., a verification point, and the parameter adjustment method corresponding to the pre-trained model is used as the designated parameter adjustment method.

[0125] It should be noted that the n pre-trained models may include pre-trained models trained using images, pre-trained models trained using text, pre-trained models trained using code, pre-trained models trained using deep thinking content, and pre-trained models trained using hierarchical models, 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 when the parameter adjustment method is basic fine-tuning, the algorithm that can be used is Stochastic Gradient Descent (SGD) / Adam (Adaptive Moment Estimation, Adam); when the parameter adjustment method is layered learning rate, the algorithm that can be used is Layer-wise Optimizers (layer-wise adaptive learning rate scaling); when the parameter adjustment method is adversarial fine-tuning, the algorithm that can be used is Generative Adversarial Networks (GAN) / Adversarial Training (adversarial training); when the parameter adjustment method is regularized fine-tuning, the algorithm that can be used is L2-SP (L2 Regularization with Sparse Penalty, L2 regularization combined with sparse penalty) / DELTA; when the parameter adjustment method is multi-task fine-tuning, the algorithm that can be used is Gradient Norm Balancing (GradNorm); when the parameter adjustment method is progressive fine-tuning, the algorithm that can be used is Curriculum Learning (Course Learning); when the parameter adjustment method is LoRA fine-tuning, the algorithm that can be used is Low-Rank Adaptation; when 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 based on the differences in the model at specific feature points, avoiding blind tuning or waste of resources. 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 candidate feedback data in a candidate feedback data set corresponding to the dialogue model to be updated, classifying the cleaned candidate feedback data according to data type to obtain classified candidate feedback data, wherein the candidate feedback data in the candidate feedback data set is feedback data obtained from historical dialogues using the dialogue model to be updated; inputting the classified candidate feedback data into a preset identification model to obtain data label information of the classified candidate feedback data; and constructing a designated feedback data set based on the classified candidate feedback data and the data label information of the classified candidate feedback data, wherein the designated feedback data in the designated feedback data set is the classified candidate feedback data with data label information.

[0132] Optionally, data cleaning can be used to remove errors, duplications, incomplete information, or other irrelevant elements from the data to improve the quality and reliability of the dataset. For example, if user feedback contains incomplete conversation records or malformed information, the data cleaning process will identify and correct these problems. Based on 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 that can be used to identify and label the attributes, type, or emotional tendency of the feedback data. For example, it can identify positive or negative emotions in a conversation, or label image data as "clear" or "blurry."

[0133] Data label information can be metadata added to the data using a preset identification model to help the dialogue model to be updated better understand and learn the characteristics of the data during training. Optionally, the feature point label information can include data label information.

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

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

[0136] It should be noted that AI platforms can deploy designated containers, which are pre-configured, isolated environments for running specific tasks, such as Docker containers. These containers are equipped with software environments and resources specifically for model training and updating.

[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 is an important indicator for measuring the load of the AI platform. The first resource utilization rate threshold and the second resource utilization rate threshold are 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. Updating resource requirements can refer to updating the specific requirements for the computing resources of the AI platform when the dialogue model to be updated is being trained and optimized. Updating 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 is the resource usage interval configured by the AI platform for the 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 the specified container through a self-triggered method. Optionally, the AI platform can include a tuned mirror file to automatically trigger decompression, loading, and creation of a Docker container, and the corresponding command for its creation can be "gunzip ** / docker load -i *** / docker run --gpu all -itd -p **:** -v{$directory} --cap-add=IPC_LOCK --name=*** --ulimit stack= **** --shm-size=500G {$mirror file}".

[0139] It's important to note that within a specified container, you can run data processing tasks, such as obtaining a specified feedback dataset. Furthermore, within a specified container, you can also trigger the selection of a specific parameter adjustment method. For example, within a specified container, execute the command "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. You can then evaluate the configured LogME pre-trained model, select the result with the highest match, and choose the corresponding full-parameter fine-tuning method, i.e., the specified parameter adjustment method.

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

[0141] In an exemplary embodiment, after updating the model parameters of the dialogue model to be updated according to a specified feedback dataset and a specified parameter adjustment method, the method further includes: performing preliminary verification on the updated dialogue model to be updated, wherein the preliminary verification includes performing a performance test and a stability test on the updated dialogue model to be updated; if the preliminary verification passes, deploying the updated dialogue model to be updated to the artificial intelligence platform; and if it is detected that the performance of the updated dialogue model to be updated has degraded 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 preliminary verification can be a preliminary performance and stability test conducted after the dialogue model to be updated is updated to ensure that the new model can meet the expected performance standards and maintain stable operation before actual deployment.

[0143] Performance testing can be achieved by evaluating the updated conversational model's performance in processing user requests, answering speed, and accuracy. Performance testing can involve the model's response speed and accuracy when handling common queries. Alternatively, performance testing examples can be deployed on the AI ​​platform. For example, the platform simulates user query behavior during peak hours, such as rapid succession of questions, complex queries, and specialized questions in specific areas. The model's response time, accuracy, and resource consumption are recorded and compared with the pre-update model to ensure that the updated model is at least as good as the original model in these metrics.

[0144] Stability testing can be performed by checking the stability of the updated conversation model under long-term operation, high concurrent requests, or special input conditions, to avoid potential problems introduced by the model update that could lead to service interruptions or abnormal responses. Stability testing can also check the reliability of the updated conversation model when facing a large number of requests. Alternatively, a stability test instance can be deployed on the AI ​​platform. For example, after the model is updated, a large number of concurrent requests can be generated to check the model's response time and whether any crashes or errors occur. At the same time, some abnormal input data, such as malicious attacks or special characters, can be simulated to ensure that the model can maintain stable operation under various circumstances.

[0145] Optionally, if performance degradation of the updated model is detected (e.g., 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, through preliminary verification and rollback mechanisms, the platform can ensure that model updates will not degrade service quality or bring unsatisfactory experiences to users, which greatly improves the reliability of the artificial intelligence platform.

[0147] The following is an explanation of the updating method of the model in the embodiment of the present application with reference to optional examples. Figure 4 As shown, in this optional example, the artificial intelligence platform is started on the training and pushing machine, ready to start deploying and tuning the model service.

[0148] At least one basic model (i.e., a dialogue model) can be deployed on an artificial intelligence platform. After the basic model is deployed, the model can be published to the intelligent agent application corresponding to the basic model. Users can interact with it through the intelligent agent application, such as dialogue question and answer, image analysis, or deep thinking functions, to interact with the model.

[0149] During user use, the AI ​​platform can record user feedback data, such as conversation satisfaction, evaluation of image analysis results, or ratings of deep thinking questions. For example, if a user is dissatisfied with the chatbot's response, they can give a negative review. This information will be stored and used as a basis for subsequent optimization.

[0150] The AI ​​platform asks the user whether they agree to participate in the automatic tuning program. 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 tuning to protect the user's privacy.

[0151] The AI ​​platform counts the amount of feedback collected and, when the number of feedback reaches a threshold, generates a designated feedback dataset in a designated container. Specifically, the designated container cleans, categorizes, and annotates the feedback data to generate the designated feedback dataset.

[0152] Based on the specified feedback dataset, a full-parameter fine-tuning method matching and tuning training are performed. Specifically, feature point matching is performed between the dataset feature point sequence of the specified feedback dataset and the baseline feature point sequence of a pre-trained model in a set of pre-trained models. Based on the feature point matching results, a specified parameter adjustment method is determined. Based on the specified feedback dataset and the specified parameter adjustment method, the model parameters of the dialogue model to be updated are updated, thereby generating a model.

[0153] The updated conversational model to be updated is deployed to the model service. After the model service is deployed, the readiness of the updated conversational model to be updated is determined, that is, whether the new model has been trained and is ready for deployment. During the deployment process, it can be determined whether the number of deployment failures has reached the failure threshold. The performance of the updated model in the user application scenario is checked to see if it meets expectations. If the model service fails to meet the predetermined performance standards after multiple verifications, that is, the number of deployment failures reaches the failure threshold, the tuning process ends. If the number of deployment failures does not reach the failure threshold, the model is returned to the designated container for full parameter fine-tuning method matching and application training. If the conversational model to be updated is determined to be ready, the model is released. The platform continues to provide services, and users can interact with the fine-tuned model to obtain more feedback.

[0154] This example demonstrates how to fully utilize the redundant resources of the training and pushing machine by automatically collecting feedback data during model service deployment and initiating model fine-tuning when resources are sufficient. This improves resource utilization and reduces training costs. Fine-tuning the dialogue model to be updated based on real user feedback data allows it to better adapt to user needs, enhance the model's accuracy and responsiveness for specific tasks, and significantly enhance service quality. The updated dialogue model to be updated can more accurately understand and respond to users. Furthermore, the platform's automatic tuning mechanism reduces manual intervention, making the model update and tuning process more automated, reducing operational costs and improving platform efficiency.

[0155] It should be noted that the collected information (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.) involved in this application are information and data authorized by the user or fully authorized by all parties, and the collection, storage, use, processing, transmission, provision, disclosure and application of relevant data comply with relevant laws, regulations and standards, take necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation entrances for users to choose to authorize or refuse.

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

[0157] The embodiment of the present application further provides a model updating device, which is applied to an artificial intelligence platform, wherein the artificial intelligence platform is deployed with a group of dialogue models; Figure 5 As shown, the device includes:

[0158] An acquisition module 502 is configured to acquire a specified feedback data set when a dialogue model to be updated is detected in a group of dialogue models, wherein the specified feedback data in the specified feedback data set is feedback data obtained from historical dialogues using the dialogue model to be updated;

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

[0160] a determination module 506 that determines, based on the feature point matching result, a specified parameter adjustment method from a set of parameter adjustment methods, wherein the parameter adjustment method in the set of parameter adjustment methods is a method for adjusting model parameters of a dialogue model in the set of dialogue models;

[0161] The updating module 508 is 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 the present application, when it is detected that a dialogue model to be updated exists in a group of dialogue models, a specified feedback dataset is obtained, feature point matching is performed on a dataset feature point sequence of the specified feedback dataset with a baseline feature point sequence of a pre-trained model in a group of pre-trained models to obtain a feature point matching result, and a specified parameter adjustment method is determined from a group of parameter adjustment methods based on the feature point matching result. Thus, model parameters of the dialogue model to be updated are updated based on the specified feedback dataset and the specified parameter adjustment method. By performing feature point matching on the dataset feature point sequence of the specified feedback dataset with a baseline feature point sequence of a pre-trained model in a group of pre-trained models and determining the specified parameter adjustment method based on the feature point matching result, the dialogue model to be updated is updated, making the model update more in line with actual needs, avoiding invalid or negative tuning, and to a certain extent ensuring the performance of the updated dialogue model. This solves the technical problem of poor model update effect existing in the related art, and achieves the technical effect of steadily improving the performance of the dialogue model.

[0164] In an exemplary embodiment, the matching module 504 includes: a matching unit for respectively calculating the difference values ​​between the feature point sequence of the data set and the baseline feature point sequence of the pre-trained model in a group of pre-trained models, and obtaining a set of difference values ​​corresponding to a group of pre-trained models, wherein the feature point matching result includes a set of difference values ​​corresponding to a group of pre-trained models.

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

[0166] An extraction module, used to extract an initial feature point set of a specified feedback data set;

[0167] A sorting module is used to sort the initial feature points in the initial feature point set to obtain an initial feature point sequence;

[0168] The sequence determination module is used to determine a subsequence containing the first specified number of initial feature points in the initial feature point sequence as a data set feature point sequence.

[0169] In an exemplary embodiment, a matching unit includes: a matching operation subunit, configured to: perform the following matching operations on pre-trained models in a group of pre-trained models as current pre-trained models, respectively, to obtain a set of difference value sets corresponding to the pre-trained models, wherein the reference feature point sequence of the current pre-trained model is the current reference feature point sequence: perform 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; and respectively calculate the difference value between two feature points in the current matching pairs in the set of current matching pairs to obtain a set of difference value sets corresponding to the current pre-trained models;

[0170] The difference value set corresponding to a group of pre-trained models includes the difference value corresponding to the current pre-trained model in the group of pre-trained models.

[0171] In an exemplary embodiment, the matching operation subunit is further used to: perform feature space mapping on the feature point sequence of the data set to obtain a specified feature point sequence of the specified feedback data set, wherein the specified feature point sequence and the current benchmark feature point sequence are in the same feature space; perform similarity calculation on the initial feature point in the specified feature point sequence and the current benchmark feature point in the current benchmark feature point sequence to obtain the similarity between the initial feature point in the specified feature point sequence and the current benchmark feature point in the current benchmark feature point sequence; perform the following determination operations on the initial feature points in the data set feature point sequence as current initial feature points to obtain a set of current matching pairs: determine the current benchmark feature point in the current benchmark feature point sequence that has the greatest similarity with the current initial feature point to obtain a current candidate benchmark feature point; and when the similarity between the current initial feature point and the current candidate benchmark feature point is greater than or equal to a similarity threshold, determine the current initial feature point and the current candidate benchmark feature point as a current matching pair.

[0172] In an exemplary embodiment, the initial feature points in the dataset feature point sequence and the current benchmark feature points in the current benchmark feature point sequence are both provided with feature point label information; the matching operation sub-unit is further used to: perform the following determination operations on the initial feature points in the dataset feature point sequence as the current initial feature points to obtain a set of current matching pairs: search for the current benchmark feature point in the current benchmark feature point sequence whose feature point label information is the same as the feature point label information of the current initial feature point; when a current benchmark feature point is found whose feature point label information is the same as the feature point label information of the current initial feature point, the current initial feature point and the current benchmark feature point whose feature point label information is the same as the feature point label information of the current initial feature point are determined as a current matching pair.

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

[0174] a first analyzing unit, configured to, when the designated feedback dataset includes text data, perform syntactic analysis on the text data in the designated feedback dataset to obtain text feature points of the designated feedback dataset, wherein the initial feature point set includes the text feature points of the designated feedback dataset;

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

[0176] The second analysis unit is used to perform audio feature analysis on the audio data in the specified feedback data set when the specified feedback data set includes audio data, to obtain audio feature points of the specified feedback data set, wherein the initial feature point set includes the audio feature points of the specified feedback data set.

[0177] In an exemplary embodiment, the dialog model to be updated allows the input data type to match the model type of the dialog model to be updated; the model updating apparatus further includes:

[0178] a weight determination module for, when the dialogue model to be updated is a multimodal model and the initial feature point set includes at least two types of initial feature points, inputting the at least two types of initial feature points into a preset attention network to obtain at least two attention weights, wherein the at least two attention weights correspond one-to-one to the at least two types of initial feature points;

[0179] A fusion module is used to use at least two attention weights to perform weighted fusion on at least two types of initial feature points to obtain a set of cross-modal feature points, wherein 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, the 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, which is used to respectively calculate the absolute value of the difference between the difference value corresponding to the same pre-trained model in a group of pre-trained models and the corresponding preset difference value, to obtain a set of absolute difference values ​​corresponding to a group of pre-trained models; and a determination unit, which is used to 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.

[0182] In an exemplary embodiment, the acquisition module 502 includes: a cleaning unit configured to clean candidate feedback data in a candidate feedback data set corresponding to the dialogue model to be updated, and classify the cleaned candidate feedback data according to data type to obtain classified candidate feedback data, wherein the candidate feedback data in the candidate feedback data set is feedback data obtained from historical dialogues using the dialogue model to be updated;

[0183] an obtaining unit, configured to input the classified candidate feedback data into a preset identification model to obtain data label information of the classified candidate feedback data;

[0184] The construction unit is configured to construct a designated feedback data set based on the classified candidate feedback data and the data label information of the classified candidate feedback data, wherein the designated feedback data in the designated feedback data set is the classified candidate feedback data with the data label information.

[0185] In an exemplary embodiment, the artificial intelligence platform is further deployed with a designated container; and the model updating device further comprises:

[0186] A monitoring module, used to monitor the resource usage of the AI ​​platform before obtaining a specified feedback data set;

[0187] a resource determination module, configured to determine that a specified container is in an available state when the resource usage rate is less than a first resource usage rate threshold;

[0188] a configuration module, configured to configure a container resource range for a specified container according to an update resource requirement of the dialogue model to be updated, wherein a maximum value of the container resource range is less than a second resource usage threshold, and the second resource usage threshold is greater than the first resource usage threshold;

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

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

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

[0192] The deployment module is used to deploy the updated dialogue model to the artificial intelligence platform if it passes the preliminary verification;

[0193] The rollback module is used to roll back the updated dialogue model to be updated to the dialogue model to be updated before the update when it is detected that the performance of the updated dialogue model to be updated is degraded compared with the dialogue model to be updated before the update.

[0194] For the description of the features in the embodiment corresponding to the model updating device, please refer to the relevant description of the embodiment corresponding to the model updating method, which will not be repeated here.

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

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

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

[0198] An embodiment of the present application further provides a computer program product, which includes a computer program. When the computer program is executed by a processor, the steps in any of the above-mentioned model updating method embodiments are implemented.

[0199] An embodiment of the present application also provides another computer program product, including a non-volatile computer-readable storage medium, wherein the non-volatile computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps in any of the above-mentioned model updating method embodiments are implemented.

[0200] Professionals may further appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the above description has generally described the components and steps of each example according to their functions. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians may use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0201] The above is a detailed introduction to the model updating method, electronic device, storage medium and program product provided by the present application. Specific examples are used herein to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method and core ideas of the present application. It should be pointed out that for ordinary technicians in this technical field, without departing from the principles of the present application, several improvements and modifications can be made to the present application, and these improvements and modifications also fall within the scope of protection of the claims of the present application.

Claims

1. A method for updating a model, characterized in that: Applied to an artificial intelligence platform, the artificial intelligence platform is deployed with a set of dialogue models; the method includes: When it is detected that a dialogue model to be updated exists in the group of dialogue models, obtaining a designated feedback data set, wherein the designated feedback data in the designated feedback data set is feedback data obtained from historical dialogues using the dialogue model to be updated; performing feature point matching on a dataset feature point sequence of the specified feedback dataset and a reference feature point sequence of a pre-trained model in a group of pre-trained models to obtain a feature point matching result, wherein the dataset feature point sequence is used to represent attribute information of the specified feedback dataset, and the reference feature point sequence is used to represent attribute information of the pre-trained model; determining, based on the feature point matching result, a specified parameter adjustment method from a set of parameter adjustment methods, wherein the parameter adjustment method in the set of parameter adjustment methods is a method of adjusting model parameters of a dialogue model in the set of dialogue models; updating model parameters of the dialogue model to be updated according to the specified feedback data set and the specified parameter adjustment method; The performing feature point matching on the dataset feature point sequence of the specified feedback dataset and the reference feature point sequence of the pre-trained model in a group of pre-trained models to obtain the feature point matching result includes: respectively calculating the difference values ​​between the dataset feature point sequence and the reference feature point sequence of the pre-trained model in the group of pre-trained models to obtain a set of difference values ​​corresponding to the group of pre-trained models, wherein the feature point matching result includes the set of difference values ​​corresponding to the group of pre-trained models; The 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 the group of pre-trained models and the parameter adjustment methods in the group of parameter adjustment methods; The method of determining a specified parameter adjustment method from a group of parameter adjustment methods based on the feature point matching results includes: calculating the absolute values ​​of the differences between the difference values ​​of the same pre-trained model in the group of pre-trained models and the corresponding preset difference values, and obtaining an absolute difference value set corresponding to the group of pre-trained models; and determining 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.

2. The method according to claim 1, characterized in that Before performing feature point matching on the dataset feature point sequence of the specified feedback dataset and the reference feature point sequence of a pre-trained model in a group of pre-trained models, the method further includes: Extracting an initial feature point set of the specified feedback data set; Sorting the initial feature points in the initial feature point set to obtain an initial feature point sequence; A subsequence containing the first specified number of initial feature points in the initial feature point sequence is determined as the dataset feature point sequence.

3. The method according to claim 2, characterized in that The step of respectively calculating the difference values ​​between the feature point sequence of the dataset and the reference feature point sequence of the pre-trained model in the set of pre-trained models to obtain a set of difference values ​​corresponding to the set of pre-trained models includes: The following matching operations are performed on each of the pre-trained models in the set of pre-trained models as the current pre-trained model to obtain 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; Calculating the difference between 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; The difference value set corresponding to the group of pre-trained models includes the difference value corresponding to the current pre-trained model in the group of pre-trained models.

4. The method according to claim 3, characterized in that The 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 includes: Performing feature space mapping on the feature point sequence of the data set to obtain a designated feature point sequence of the designated feedback data set, wherein the designated feature point sequence and the current reference feature point sequence are in the same feature space; performing similarity calculation on the initial feature point in the designated 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 designated feature point sequence and the current reference feature point in the current reference feature point sequence; The following determination operations are performed on the initial feature points in the feature point sequence of the data set as current initial feature points to obtain the set of current matching pairs: Determine a current reference feature point in the current reference feature point sequence that has the greatest similarity to the current initial feature point to obtain a current candidate reference feature point; In a case where the similarity between the current initial feature point and the current candidate reference feature point is greater than or equal to a similarity threshold, the current initial feature point and the current candidate reference feature point are determined as a current matching pair.

5. The method according to claim 3, characterized in that The initial feature point in the dataset feature point sequence and the current reference feature point in the current reference feature point sequence are both provided with feature point label information; The 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 includes: The following determination operations are performed on the initial feature points in the feature point sequence of the data set as current initial feature points to obtain the set of current matching pairs: Searching for a current reference feature point in the current reference feature point sequence whose feature point label information is the same as the feature point label information of the current initial feature point; When a current reference feature point having the same feature point label information as the current initial feature point is found, the current initial feature point and the current reference feature point having the same feature point label information as the current initial feature point are determined as a current matching pair.

6. The method according to claim 2, characterized in that There is at least one type of specified feedback data in the specified feedback data set; The extracting of the initial feature point set of the specified feedback data set includes: In a case where the designated feedback dataset includes text data, performing syntactic analysis on the text data in the designated feedback dataset to obtain text feature points of the designated feedback dataset, wherein the initial feature point set includes the text feature points of the designated feedback dataset; In a case where the designated feedback dataset includes image data, performing image feature recognition on the image data in the designated feedback dataset to obtain image feature points of the designated feedback dataset, wherein the initial feature point set includes the image feature points of the designated feedback dataset; In a case where the designated feedback data set includes audio data, audio feature analysis is performed on the audio data in the designated feedback data set to obtain audio feature points of the designated feedback data set, wherein the initial feature point set includes the audio feature points of the designated feedback data set.

7. The method according to claim 2, characterized in that The data type allowed to be input by the dialogue model to be updated matches the model type of the dialogue model to be updated; the method further includes: When the dialogue model to be updated is a multimodal model and the initial feature point set includes at least two types of initial feature points, inputting the at least two types of initial feature points into a preset attention network to obtain at least two attention weights, wherein 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, the at least two types of initial feature points are weightedly fused to obtain a set of cross-modal feature points, wherein 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.

8. The method according to claim 1, characterized in that The obtaining of the specified feedback data set includes: performing data cleaning on candidate feedback data in a candidate feedback data set corresponding to the dialogue model to be updated, and classifying the cleaned candidate feedback data according to data type to obtain classified candidate feedback data, wherein the candidate feedback data in the candidate feedback data set is feedback data obtained from historical dialogues using the dialogue model to be updated; Inputting the classified candidate feedback data into a preset identification model to obtain data label information of the classified candidate feedback data; The designated feedback data set is constructed based on the classified candidate feedback data and the data label information of the classified candidate feedback data, wherein the designated feedback data in the designated feedback data set is the classified candidate feedback data with data label information.

9. The method according to claim 1, characterized in that The artificial intelligence platform is also deployed with a designated container; Before obtaining the specified feedback data set, the method further includes: monitoring resource usage of the artificial intelligence platform; When the resource usage rate is less than a first resource usage rate threshold, determining that the designated container is in an available state; configuring a container resource range for the designated container according to the update resource requirement of the to-be-updated dialog model, wherein a maximum value of the container resource range is less than a second resource usage threshold, and the second resource usage threshold is greater than the first resource usage threshold; The updating of the to-be-updated dialogue model is performed in the designated container using resources within a container resource range configured for the designated container.

10. The method according to any one of claims 1 to 9, characterized in that 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 method further includes: Performing preliminary verification on the updated dialogue model to be updated, wherein 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; If the preliminary verification passes, deploying the updated dialogue model to be updated to the artificial intelligence platform; When it is detected that the performance of the updated dialogue model to be updated is degraded compared with the dialogue model to be updated before the update, the updated dialogue model to be updated is rolled back to the dialogue model to be updated before the update.

11. An electronic device, characterized in that: include: Memory for storing computer programs; A processor, configured to implement the steps of the model updating method according to any one of claims 1 to 10 when executing the computer program.

12. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, wherein when the computer program is executed by a processor, the steps of the model updating method according to any one of claims 1 to 10 are implemented.

13. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the model updating method according to any one of claims 1 to 10 are implemented.

Images