Model fusion method, data processing method, system and storage medium

Abstract

This application discloses a model fusion method, data processing method, system, and storage medium, relating to large model technology and data processing. The method includes: dividing multiple models to be fused into multiple sub-modules according to the granularity of the target model decomposition; performing nonlinear evaluation on the multiple sub-modules to obtain evaluation results; and, in response to the evaluation results meeting preset conditions, performing model fusion on the multiple sub-modules at the target model decomposition granularity to generate a target fused model. This application solves the technical problems of low fusion efficiency and high resource consumption in related technologies when performing model fusion.

Description

Technical Field

[0001] This application relates to large model technology and data processing, specifically to a model fusion method, a data processing method, a system, and a storage medium. Background Technology

[0002] With the rapid development of artificial intelligence technology, the scale of large language models is constantly expanding, and the computational resources and training data required for training these models are also increasing. Pre-trained large language models, after fine-tuning on multiple tasks, can achieve cross-task performance improvements. Model fusion technology can combine the capabilities of multiple existing single-task models to obtain a model with balanced performance across specific tasks. Compared to training multi-task capabilities, this effectively saves model development costs and reduces the demand for training data. In related technologies, multi-model fusion can be performed by calculating the difference in parameters before and after model fine-tuning and then weighting the models. However, the selection of weights relies on grid search, which is time-consuming and has low accuracy. Related technologies can also improve the overall linearity of the model by modifying the training paradigm, thereby improving the model fusion effect. However, this method requires retraining the model, which not only consumes additional resources but also depends on difficult-to-obtain complete training data. Therefore, related technologies suffer from low fusion efficiency and high resource consumption when performing model fusion, further restricting the performance and application flexibility of the fused model.

[0003] There is currently no effective solution to the above problems. Summary of the Invention

[0004] This application provides a model fusion method, data processing method, system, and storage medium to at least solve the technical problems of low fusion efficiency and high resource consumption in related technologies when performing model fusion.

[0005] According to one aspect of the embodiments of this application, a model fusion method is provided, comprising: dividing multiple models to be fused into multiple sub-modules according to a target model decomposition granularity, wherein the target model decomposition granularity is selected from multiple model decomposition granularities, and the multiple model decomposition granularities are used to divide the models to be fused using different model division methods; performing nonlinear evaluation on the multiple sub-modules to obtain an evaluation result, wherein the evaluation result is used to measure the degree of nonlinearity of the multiple sub-modules at the target model decomposition granularity; and, in response to the evaluation result satisfying a preset condition, performing model fusion on the multiple sub-modules at the target model decomposition granularity to generate a target fused model, wherein the target fused model is used to perform task analysis on task data to obtain task processing results.

[0006] According to another aspect of the embodiments of this application, a data processing method is also provided, including: acquiring task data; performing task analysis on the task data using a target fusion model to obtain task processing results; wherein the target fusion model is generated according to any one of the model fusion methods in the embodiments of this application.

[0007] According to another aspect of the embodiments of this application, a data processing method is also provided, including: acquiring programming task data; performing task analysis on the programming task data using a target fusion model to obtain programming task processing results; wherein the target fusion model is generated according to any one of the model fusion methods in the embodiments of this application.

[0008] According to another aspect of the embodiments of this application, a data processing method is also provided, comprising: obtaining a data processing request through a first application programming interface, wherein the request data carried in the data processing request includes: task data; and returning a data processing response through a second application programming interface, wherein the response data carried in the data processing response includes: task processing result, wherein the task processing result is obtained by performing task analysis on the task data using a target fusion model, and the target fusion model is generated according to any one of the model fusion methods in the embodiments of this application.

[0009] According to another aspect of the embodiments of this application, a data processing method is also provided, comprising: acquiring a currently input data processing dialogue request, wherein the request data carried in the data processing dialogue request includes: task data; responding to the data processing dialogue request, returning a data processing dialogue response, wherein the information carried in the data processing dialogue response includes: task processing result, the task processing result being obtained by performing task analysis on the task data using a target fusion model, the target fusion model being generated according to any one of the model fusion methods in the embodiments of this application; and displaying the task processing result in a graphical user interface.

[0010] According to another aspect of the embodiments of this application, a data processing method is also provided, including: displaying task data on the operation interface in response to an input command applied to the operation interface; and displaying task processing results on the operation interface in response to a processing command applied to the operation interface; wherein the task processing results are obtained by performing task analysis on the task data using a target fusion model, and the target fusion model is generated according to any one of the model fusion methods in the embodiments of this application.

[0011] According to another aspect of the embodiments of this application, a data processing system is also provided, including: a client for sending task data; a server connected to the client for performing task analysis on the task data using a target fusion model to obtain task processing results; the client is also used to output the task processing results; wherein the target fusion model is generated according to any one of the model fusion methods in the embodiments of this application.

[0012] According to another aspect of the embodiments of this application, an electronic device is also provided, including: a memory storing an executable program; and a processor for running the program, wherein the program executes the methods in various embodiments of this application when it runs.

[0013] According to another aspect of the embodiments of this application, a computer-readable storage medium is also provided, the computer-readable storage medium including a stored executable program, wherein, when the executable program is running, it controls the device where the computer-readable storage medium is located to perform the methods of various embodiments of this application.

[0014] According to another aspect of the embodiments of this application, a computer program product is also provided, including a computer program that, when executed by a processor, implements the methods of various embodiments of this application.

[0015] According to another aspect of the embodiments of this application, a computer program product is also provided, including a non-volatile computer-readable storage medium storing a computer program that, when executed by a processor, implements the methods in various embodiments of this application.

[0016] According to another aspect of the embodiments of this application, a computer program is also provided, which, when executed by a processor, implements the methods of the various embodiments of this application.

[0017] In this embodiment, multiple models to be fused are divided into multiple sub-modules according to the target model decomposition granularity. Then, nonlinear evaluation is performed on these sub-modules to obtain evaluation results, which measure the degree of nonlinearity of the sub-modules at the target model decomposition granularity. Finally, in response to the evaluation results meeting preset conditions, model fusion is performed on the multiple sub-modules at the target model decomposition granularity to generate a target fused model. This target fused model is used to perform task analysis on task data to obtain task processing results. Therefore, based on the nonlinear evaluation and fusion of sub-modules, finer-grained control can be provided for model fusion, enabling more efficient and accurate model fusion, thereby reducing development costs and improving the flexibility and scalability of the target fused model. By evaluating the degree of nonlinearity of the sub-modules and performing model fusion on the multiple sub-modules at the target model decomposition granularity when the nonlinear evaluation results meet preset conditions, the computational complexity of the fusion process is reduced, thus improving model fusion efficiency. Furthermore, after determining the target decomposition granularity, only a small amount of task-related data is needed for nonlinear evaluation, which can avoid the high resource consumption of retraining or grid search in traditional model fusion methods, greatly reducing resource requirements, and thus solving the technical problems of low fusion efficiency and high resource consumption in related technologies when performing model fusion.

[0018] It is worth noting that the general description above and the detailed description that follow are merely for illustrative purposes and do not constitute a limitation on this application. Attached Figure Description

[0019] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0020] Figure 1 This is a schematic diagram illustrating an application scenario of a data processing method according to an embodiment of this application;

[0021] Figure 2 This is a flowchart of a model fusion method according to an embodiment of this application;

[0022] Figure 3 This is a flowchart of a data processing method according to an embodiment of this application;

[0023] Figure 4 This is a flowchart of another data processing method according to an embodiment of this application;

[0024] Figure 5 This is a flowchart of another data processing method according to an embodiment of this application;

[0025] Figure 6 This is a flowchart of another data processing method according to an embodiment of this application;

[0026] Figure 7 This is a flowchart of another data processing method according to an embodiment of this application;

[0027] Figure 8 This is a structural block diagram of a model fusion apparatus according to an embodiment of this application;

[0028] Figure 9 This is a structural block diagram of a data processing apparatus according to an embodiment of this application;

[0029] Figure 10 This is a structural block diagram of another data processing apparatus according to an embodiment of this application;

[0030] Figure 11 This is a structural block diagram of another data processing apparatus according to an embodiment of this application;

[0031] Figure 12 This is a structural block diagram of another data processing apparatus according to an embodiment of this application;

[0032] Figure 13 This is a structural block diagram of another data processing apparatus according to an embodiment of this application;

[0033] Figure 14 This is a structural block diagram of a computing device according to an embodiment of this application. Detailed Implementation

[0034] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0035] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0036] The technical solution provided in this application is mainly implemented using large-scale model technology. Here, large-scale model refers to a deep learning model with a large number of model parameters, typically containing hundreds of millions, tens of billions, hundreds of billions, trillions, or even tens of trillions of model parameters. Large-scale models can also be called foundation models. They are pre-trained on large-scale unlabeled corpora to produce pre-trained models with hundreds of millions of parameters. Such models can adapt to a wide range of downstream tasks and have good generalization ability. Examples include large language models (LLM) and multi-modal pre-training models.

[0037] It should be noted that, in practical applications, large models can be fine-tuned using a small number of samples to adapt them to different tasks. For example, large models can be widely used in Natural Language Processing (NLP), computer vision, and speech processing. Specifically, they can be applied to computer vision tasks such as Visual Question Answering (VQA), Image Captioning (IC), and Image Generation, as well as NLP tasks such as text-based sentiment classification, text summarization, and machine translation. Therefore, the main application scenarios for large models include, but are not limited to, digital assistants, intelligent robots, search, online education, office software, e-commerce, and intelligent design. In this embodiment, data processing using a target fusion model in a task data processing scenario is used as an example for explanation.

[0038] First, some nouns or terms that appear in the description of the embodiments of this application shall be interpreted as follows:

[0039] Model merging is a technique to improve the overall performance of a model by fusing the parameters of multiple models to obtain a single model, thereby leveraging the strengths of each model being merged.

[0040] Attention layer: A specific module in a Transformer network. The attention layer contains multiple attention heads, each of which assigns different weights to different parts of the input data, allowing the model to focus on important information and thus improve its ability to understand and model sequential data.

[0041] Multilayer Perceptron (MLP): A specific module in a Transformer network structure that enhances the model's ability to express nonlinearities in input data by independently applying two linear transformations and a nonlinear activation function to the input information.

[0042] Submodule: In a large neural network, a submodule is a smaller, independent module or component that together constitutes the overall model.

[0043] Linearity: In mathematics and statistics, linearity refers to a relationship where the output is a linear combination of the inputs. Linear models reduce complexity, making derivation and computation simpler.

[0044] Fine-tuning: Fine-tuning is a method of further training a pre-trained model to adapt it to a specific dataset or task.

[0045] Merging Weights: These refer to the weights used in the actual weighted fusion of each model (or sub-module) during the model fusion process.

[0046] Closed-form solution: A closed-form solution is an explicit solution to a mathematical problem that can be obtained through a finite number of calculations without relying on an iterative solution process.

[0047] Related techniques, such as task arithmetic fusion, use a weighted combination of parameter differences for model fusion. However, determining the weights relies on computationally intensive grid search, which is not only inefficient but also cannot guarantee accuracy. Furthermore, applying the same weights to the entire model ignores the differences between sub-modules. Another fusion method, task arithmetic in tangent space, can improve the linearity of the model and optimize fusion by modifying the training paradigm, but it requires retraining the model and obtaining complete training data. In practice, this is often limited by resource and data unavailability, increasing the technical and economic costs of fusion. Therefore, these methods have significant shortcomings in determining fusion weights and resource efficiency.

[0048] According to an embodiment of this application, a model fusion method is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0049] Considering the large number of model parameters in large models and the limited computing resources of mobile terminals, the method provided in this application embodiment can be applied to, for example, Figure 1 The application scenarios shown are not limited to these. In, for example... Figure 1 In the application scenario shown, the large model is deployed on server 10. Server 10 can connect to one or more client devices 20 via a local area network (LAN), wide area network (WAN), internet connection, or other types of data network. These client devices 20 may include, but are not limited to, smartphones, tablets, laptops, PDAs, personal computers, smart home devices, and in-vehicle devices. Client devices 20 can interact with users through a graphical user interface to access the large model, thereby implementing the method provided in this embodiment.

[0050] In this embodiment, the system consisting of a client device and a server can perform the following steps: the server connects to the client, the client can send task data to the server, after receiving the task data, the server uses a target fusion model to perform task analysis on the task data, obtains the task processing result, and returns the task processing result to the client, and finally the client outputs the task processing result through a graphical user interface. The target fusion model is generated according to any model fusion method in this embodiment.

[0051] It should be noted that with the rapid development of high-performance computing units, the methods provided in this application embodiment can also be applied to model-in-machine systems in other application scenarios. In one optional embodiment, the model-in-machine system has multiple built-in models, and users can select one model to adjust as needed to obtain their own model. The high-performance computing unit built into the model-in-machine system can then directly call the adjusted model to execute the methods provided in this application embodiment. In another optional embodiment, the large model-in-machine system has a pre-trained model built-in, and the high-performance computing unit built into the model-in-machine system can then directly call that model to execute the methods provided in this application embodiment.

[0052] Furthermore, when users need to train their own models, they can upload their own datasets via the client. These datasets are then sent to the server, allowing the server to adjust the pre-trained model using the dataset to obtain the user's customized model, which can then be deployed to the production environment. To facilitate users' model adjustment needs, the server provides complete adjustment tools, development frameworks, and processes, supporting multiple adjustment strategies. This allows the adjusted model to better adapt to different application domains and achieve a high degree of customization.

[0053] Under the aforementioned operating environment, this application provides the following: Figure 2 The model fusion method shown. Figure 2 This is a flowchart of a model fusion method according to an embodiment of this application. Figure 2 As shown, the method may include the following steps:

[0054] Step S21: Divide the multiple models to be fused into multiple sub-modules according to the target model decomposition granularity. The target model decomposition granularity is selected from multiple model decomposition granularities. The multiple model decomposition granularities are used to divide the models to be fused using different model division methods.

[0055] Step S22: Perform nonlinear evaluation on multiple sub-modules to obtain evaluation results, wherein the evaluation results are used to measure the degree of nonlinearity of multiple sub-modules at the target model decomposition granularity.

[0056] Step S23: In response to the evaluation result meeting the preset conditions, at the target model decomposition granularity, multiple sub-modules are fused to generate a target fusion model, wherein the target fusion model is used to perform task analysis on task data to obtain task processing results.

[0057] The aforementioned multiple models to be fused can be pre-trained models that have been fine-tuned for specific tasks. Each pre-trained model focuses on solving one or a few specific tasks, such as text translation, sentiment analysis, and text generation. Pre-trained models typically have different fine-tuning parameters and task specializations.

[0058] The aforementioned multiple model decomposition granularities refer to methods of dividing the model to be fused into different structural levels. These granularities include, but are not limited to, layer granularity, module granularity, and attention head granularity. For example, a common large language model typically contains a multi-layer Transformer structure, with each Transformer layer containing an attention layer and an MLP layer. The attention layer can be further divided into different attention heads. At the layer granularity, the model to be fused can be decomposed into different layers, such as each layer in a Transformer. This is a common decomposition method, with each layer containing an attention mechanism and a feedforward neural network. At the module granularity, the decomposition can be further refined, treating the attention mechanism and feedforward neural network within a layer as independent sub-modules. For example, an attention layer can contain multiple attention heads, each of which can be considered an independent module. At the attention head granularity, each attention head in the attention layer can be decomposed separately, allowing for more precise analysis and fusion of the various parts of the model.

[0059] After selecting the target model decomposition granularity from multiple model decomposition granularities, the model to be fused can be divided into multiple sub-modules according to the target model decomposition granularity. A sub-module refers to a structural unit in the model to be fused that can be analyzed and fused independently. For example, in layer-level decomposition, each Transformer layer is a sub-module; in module-level decomposition, the attention mechanism and the feedforward network each constitute a sub-module; and in attention head-level decomposition, each head in the attention layer is an independent sub-module.

[0060] Nonlinear evaluation is performed on multiple sub-modules to obtain evaluation results, which are then used to measure the degree of nonlinearity of the multiple sub-modules at the target model decomposition granularity. Nonlinear evaluation refers to the process of quantitatively analyzing the linearity of the relationship between the changes in output and model parameters after multiple sub-modules process input data. In linear models, the change in output is a direct linear function of the changes in input or parameters, while in nonlinear models, this relationship may be more complex and no longer follow a simple linear proportion.

[0061] The nonlinearity of multiple sub-modules is evaluated at the granularity of the target model decomposition. The evaluation focuses on specific structural units within the model, i.e., sub-modules, such as attention mechanisms or feedforward networks in Transformer layers, or even finer-grained attention heads. The goal of the nonlinearity evaluation is to determine whether the relationship between parameter changes and output feature changes in multiple sub-modules before and after fine-tuning is approximately linear.

[0062] Specifically, parameter changes refer to the changes in model parameters during fine-tuning, which are reflected in parameters such as weights and biases of submodules. Output feature changes indicate that the processing results of submodules on the same input data will differ before and after fine-tuning; output feature changes can be quantified and measured.

[0063] The above evaluation results are determined based on the nonlinearity score of each submodule at the target decomposition granularity, and serve as an important basis for subsequent decisions on whether and how to merge multiple submodules. The nonlinearity score can be based on the relationship between the parameter difference and the output feature difference before and after fine-tuning, specifically the degree of deviation between the two. If the deviation is small, it means that the relationship between the change in output features and the change in parameters is relatively linear; conversely, it indicates a high degree of nonlinearity.

[0064] In model fusion scenarios, nonlinear evaluation of submodules is crucial, directly impacting the efficiency and effectiveness of the fusion process. Submodules with higher linearity can be more easily fused through weighted combination without requiring large amounts of data or retraining, thus saving computational resources and time. Furthermore, selecting appropriate model decomposition granularity and the degree of nonlinearity of submodules helps to more accurately control the model fusion process and improve the performance of the fused model.

[0065] When the evaluation results meet the preset conditions, multiple sub-modules are fused at the target model decomposition granularity to generate a target fusion model. This target fusion model is expected to have balanced performance across different tasks. The target fusion model is then used to perform task analysis on task data to obtain task processing results. The task data can be any input dataset related to a specific task, used to evaluate the performance of the target fusion model on that task. For example, if the task is text classification, the task data could include a series of labeled text samples, each containing text content and a corresponding classification label. If the task is machine translation, the task data could be multiple pairs of source and target language sentences.

[0066] For example, when using a target fusion model to perform task analysis on task data, the task data can first be preprocessed, such as cleansing and formatting, to meet the requirements of the model input. Then, the preprocessed task data is input into the target fusion model to obtain its output, which may include classification labels, translated text, generated text, etc. Based on the model output and the ground truth of the task data, appropriate evaluation metrics are used to assess the performance of the target fusion model on a specific task. Based on the model evaluation results, it can be determined whether the target fusion model successfully exhibits the expected performance on the target task, whether it achieves a balanced performance across the fusion target task, and whether the target fusion model has a significant performance improvement compared to a single fine-tuned model. Through this process, model fusion not only integrates the capabilities of multiple specialized models into a single model but also validates the performance of the target fusion model through efficient analysis of a small amount of task data, thus providing a high-performance and cost-effective model fusion method for practical applications.

[0067] Based on steps S21 to S23 above, multiple models to be fused are divided into multiple sub-modules according to the target model decomposition granularity. Then, nonlinear evaluation is performed on these sub-modules to obtain evaluation results, which measure the degree of nonlinearity of the sub-modules at the target model decomposition granularity. Finally, in response to the evaluation results meeting preset conditions, model fusion is performed on the multiple sub-modules at the target model decomposition granularity to generate a target fused model. This target fused model is used to perform task analysis on task data to obtain task processing results. Therefore, based on the nonlinear evaluation and fusion of sub-modules, finer-grained control can be provided for model fusion, enabling more efficient and accurate model fusion, thereby reducing development costs and improving the flexibility and scalability of the target fused model. By evaluating the degree of nonlinearity of the sub-modules and performing model fusion on the multiple sub-modules at the target model decomposition granularity when the nonlinear evaluation results meet preset conditions, the computational complexity of the fusion process is reduced, thus improving the efficiency of model fusion. Furthermore, after determining the target decomposition granularity, only a small amount of task-related data is needed for nonlinear evaluation, which can avoid the high resource consumption of retraining or grid search in traditional model fusion methods, greatly reducing resource requirements, and thus solving the technical problems of low fusion efficiency and high resource consumption in related technologies when performing model fusion.

[0068] The model fusion method in the embodiments of this application will be further described below.

[0069] In an optional embodiment, in step S22, a nonlinear evaluation is performed on multiple sub-modules to obtain the evaluation results, including:

[0070] Step S221: Perform nonlinear evaluation on multiple sub-modules and obtain the nonlinear scores of multiple sub-modules;

[0071] Step S222: Determine the evaluation result based on the nonlinear scores of multiple sub-modules.

[0072] Specifically, by performing nonlinear evaluation on multiple sub-modules, the degree of nonlinearity between the parameter changes and output characteristic changes before and after fine-tuning can be calculated. The resulting nonlinearity score is a quantitative indicator that can reflect the degree of nonlinearity of the sub-module's output characteristics to its parameter changes.

[0073] Furthermore, the evaluation result is determined based on the nonlinear scores of multiple sub-modules. For example, the mean of the nonlinear scores of multiple sub-modules is calculated to obtain the nonlinear average score, and the evaluation result is determined based on the nonlinear average score. In addition to calculating the nonlinear average score, the following strategies can also be used to comprehensively analyze and determine the evaluation result: median calculation, standard deviation analysis, distribution visualization, weighted average, setting nonlinear score threshold combinations, cluster analysis, and statistical tests.

[0074] Specifically, calculating the median of the nonlinearity scores for all submodules better reflects the central tendency of the dataset, especially when outliers exist in the data distribution. Calculating the standard deviation of the nonlinearity scores reveals the volatility of the nonlinearity levels of the submodules. A low standard deviation indicates relatively stable nonlinearity scores and more consistent high linearity; a high standard deviation indicates significant nonlinearity differences within the model, requiring further decomposition or selective fusion. For distribution visualization, histograms or box plots can be used to visualize the distribution of nonlinearity scores across multiple submodules, helping to intuitively understand the range of linearity distributions and the existence of extreme cases, thus guiding adjustments to the fusion strategy. Considering the potentially varying importance of different submodules in the model, a weighted average of the nonlinearity scores can be used, with weights determined based on the submodule's contribution to the model or its importance in a specific task. By setting a combination of nonlinearity score thresholds, not only can an average or median threshold be calculated, but a threshold range can also be set. For example, the model might be considered suitable for fusion only when the nonlinearity scores of most (e.g., 80%) submodules are below a certain threshold. In cluster analysis, clustering the nonlinear scores of submodules identifies submodules with similar linear properties. This allows for the selective application of different fusion strategies to different submodule groups to optimize the fusion effect. Statistical tests, such as t-tests or ANOVA, can be used to determine whether the nonlinear scores are significantly lower than the expected value under a certain linear model assumption and whether there are significant differences in historical data, thus scientifically determining the feasibility of model fusion.

[0075] By employing the aforementioned analysis strategies targeting nonlinear scores, we can gain a more comprehensive and in-depth understanding of the nonlinear characteristics of model submodules, thereby producing more accurate evaluation results to guide the model fusion process. For example, even if the nonlinear average score meets the requirements, a high standard deviation indicates significant variations in the linear properties within the model, necessitating finer granularity during fusion. Alternatively, even if the average score is substandard, fusion is still suitable if the nonlinear scores of most submodules are close to ideal values, requiring only special handling strategies for a few highly nonlinear submodules.

[0076] Based on the above optional embodiments, by performing nonlinear evaluation on multiple sub-modules, obtaining the nonlinear scores of multiple sub-modules, and then determining the evaluation result based on the nonlinear scores of multiple sub-modules, it is possible to accurately determine whether a linear fusion strategy is suitable by evaluating the linear properties of the sub-modules, thereby more accurately controlling and optimizing the model fusion process and improving the overall performance and fusion efficiency of the target fusion model.

[0077] In an optional embodiment, step S221, performing a nonlinear evaluation on multiple sub-modules and obtaining the nonlinear scores of the multiple sub-modules includes:

[0078] Step S2211: Obtain the first model weight and the second model weight, wherein the first model weight is the weight used by multiple transformation layers of the pre-trained model, and the second model weight is the weight used by multiple transformation layers of different target sub-modules in multiple sub-modules, and the multiple sub-modules are obtained by fine-tuning the pre-trained model respectively.

[0079] Step S2212: Based on the first model weight and the second model weight, perform feature space distance measurement to obtain the nonlinear scores of multiple sub-modules.

[0080] The first model weights θ0 mentioned above are the weights of multiple transformation layers in the pre-trained model. In large language models, transformation layers are usually part of the Transformer architecture, responsible for handling the self-attention mechanism and feedforward neural network for sequential data. The first model weights can be obtained through self-supervised learning on large-scale unlabeled data, and they contain general language understanding and generation capabilities.

[0081] For example, when obtaining the first model weights, a specified version of the pre-trained model can be used to ensure that all transformation layer weights are included when the model is loaded. This typically involves downloading the model file from the corresponding model library or loading the pre-trained model from a storage location. Further, the multiple transformation layers of the pre-trained model are traversed, and the weight parameters of each layer are extracted, including the weights of the attention layer, the weights of the feedforward layer, and the biases. These weight parameters are usually stored in specific layers of the pre-trained model and can be obtained through the application programming interface of the pre-trained model or access methods provided by the framework.

[0082] The second model weights θ mentioned above are the weights of multiple transformation layers in the target submodule after fine-tuning the pre-trained model to suit a specific task. Fine-tuning the weights can reflect the special patterns and knowledge learned by the model on task-specific data, and is usually updated on a smaller, well-labeled training dataset.

[0083] For example, when obtaining the weights of the second model, the basic structure of the pre-trained model is first loaded. Based on the pre-trained model, fine-tuning training is performed using a dataset specific to the task, updating the model's parameters to adapt to the new task. After fine-tuning, the weights of each transformation layer in any target submodule are saved. This can be done by adding logic to save weights during training or by calling a function separately to save the model state after fine-tuning. Similar to obtaining the weights of the first model, multiple transformation layers of the target submodule are traversed, and the weight parameters of each transformation layer are extracted.

[0084] Furthermore, a feature space distance metric is performed based on the weights of the first and second models to obtain nonlinear scores for multiple sub-modules. This feature space distance metric involves calculating the distance between the pre-trained model and the target sub-module in the feature space. The feature space refers to the high-dimensional space of intermediate representations or feature vectors generated by the model after processing the input data. The distance metric can be Euclidean distance, cosine similarity, or other suitable methods to quantify the difference in output features between two models under the same input.

[0085] The result of the feature space distance metric will be used to calculate the nonlinear score, which typically involves comparing the relationship between the changes in feature output and parameter changes of the target submodule relative to the pre-trained model. If the output changes and parameter changes show a good linear relationship, the nonlinear score will be low; conversely, if the sensitivity of the output changes to parameter changes is not linear, the score will be high.

[0086] Based on the above optional embodiments, by obtaining the first model weight and the second model weight, and then performing a feature space distance measurement based on the first model weight and the second model weight, the nonlinear scores of multiple sub-modules can be obtained. Thus, by evaluating at the sub-module level, the linear part of the model can be identified more efficiently and accurately, which helps to reduce the computational overhead in the model fusion process, while improving the performance and fusion efficiency of the target fusion model.

[0087] In an optional embodiment, in step S2212, the nonlinear scores of multiple sub-modules are obtained by performing a feature space distance metric based on the first model weights and the second model weights, including:

[0088] The task vector of the target submodule is determined by using the first model weight and the second model weight, wherein the task vector is used to represent the model weight difference between the first model weight and the second model weight;

[0089] Based on the training task data, the weights of the first model, the weights of the second model, and the task vector, a feature space distance metric is performed to obtain nonlinear scores for multiple sub-modules.

[0090] Specifically, the task vector can be determined based on the model weight difference τ between the weights of the first model and the weights of the second model. The task vector represents the changes in parameters of the pre-trained model during the fine-tuning process of a specific task, that is, the new knowledge and patterns learned by the pre-trained model for the task. The task vector is the basis for subsequent analysis of the linear properties of the model.

[0091] Furthermore, based on the training task data, the first model weights, the second model weights, and the task vector, a feature space distance metric is performed to obtain the nonlinear scores of multiple sub-modules. For example, a small number of data points x, such as 5-30, are randomly sampled from a dataset X for a specific task as training task data, which is then used to evaluate the linear properties of the target sub-module.

[0092] Based on the above optional embodiments, the task vector of the target sub-module is determined by using the first model weight and the second model weight. Then, based on the training task data, the first model weight, the second model weight, and the task vector, the feature space distance is measured, thereby quickly obtaining the nonlinear scores of multiple sub-modules. Thus, by evaluating at the sub-module level, the linear part of the model can be identified more efficiently and accurately, which helps to reduce the computational overhead in the model fusion process and improve the performance and fusion efficiency of the target fusion model.

[0093] In one optional embodiment, obtaining nonlinear scores for multiple sub-modules by performing a feature space distance metric based on training task data, first model weights, second model weights, and task vectors includes:

[0094] Based on the training task data, the first model weights, the task vector, and the first preset weight coefficients, a first output feature vector is obtained, and based on the training task data, the first model weights, the task vector, and the second preset weight coefficients, a second output feature vector is obtained.

[0095] The third output feature vector is obtained based on the training task data and the first model weights, and the fourth output feature vector is obtained based on the training task data and the second model weights.

[0096] The first output feature vector, the second output feature vector, the third output feature vector, and the fourth output feature vector are used to measure the distance in the feature space and obtain the nonlinear scores of multiple sub-modules; wherein, the first preset weight coefficient is different from the second preset weight coefficient.

[0097] Specifically, taking a first preset weight coefficient of α and a second preset weight coefficient of β as an example, the first output feature vector f(x; θ0+ατ) is obtained based on the training task data x, the first model weight θ0, the task vector τ, and the first preset weight coefficient α. The first output feature vector f(x; θ0+βτ) is obtained based on the training task data x, the first model weight θ0, the task vector τ, and the second preset weight coefficient β.

[0098] The third output feature vector f(x; θ0) is obtained based on the training task data x and the first model weight θ0, and the fourth output feature vector f(x; θ) is obtained based on the training task data x and the second model weight θ.

[0099] Furthermore, the first output feature vector f(x; θ0+ατ), the second output feature vector f(x; θ0+βτ), the third output feature vector f(x; θ0), and the fourth output feature vector f(x; θ) are used to measure the distance in the feature space and obtain the nonlinear scores of multiple sub-modules.

[0100] Based on the above optional embodiments, a first output feature vector is obtained based on training task data, a first model weight, a task vector, and a first preset weight coefficient; a second output feature vector is obtained based on training task data, a first model weight, a task vector, and a second preset weight coefficient; a third output feature vector is obtained based on training task data and the first model weight; and a fourth output feature vector is obtained based on training task data and the second model weight. Finally, the first, second, third, and fourth output feature vectors are used to perform feature space distance measurement, thereby quickly obtaining the nonlinear scores of multiple sub-modules for evaluating the feasibility of model fusion. This reduces the computational overhead in the model fusion process and improves the performance and fusion efficiency of the target fusion model.

[0101] In one optional embodiment, the nonlinear scores of multiple sub-modules are obtained by using the first output feature vector, the second output feature vector, the third output feature vector, and the fourth output feature vector to perform a feature space distance metric.

[0102] The first output feature vector and the second output feature vector are subjected to feature space distance measurement to obtain the first measurement result, and the third output feature vector and the fourth output feature vector are subjected to feature space distance measurement to obtain the second measurement result;

[0103] Based on the first and second metric results, the nonlinear scores of multiple sub-modules are obtained.

[0104] Specifically, a feature space distance measurement is performed on the first output feature vector f(x; θ0+ατ) and the second output feature vector f(x; θ0+βτ) to obtain a first measurement result D(f(x; θ0+ατ), f(x; θ0+βτ)), and a feature space distance measurement is performed on the third output feature vector f(x; θ0) and the fourth output feature vector f(x; θ) to obtain a second measurement result D(f(x; θ), f(x; θ0)).

[0105] Based on the first metric result D(f(x;θ0+ατ),f(x;θ0+βτ)) and the second metric result D(f(x;θ),f(x;θ0)), the nonlinearity scores of multiple sub-modules can be obtained using the following formula:

[0106]

[0107] Based on the above optional embodiments, a first measurement result is obtained by measuring the feature space distance between the first output feature vector and the second output feature vector, and a second measurement result is obtained by measuring the feature space distance between the third output feature vector and the fourth output feature vector. Based on the first measurement result and the second measurement result, the nonlinear scores of multiple sub-modules can be quickly obtained to evaluate the feasibility of model fusion, thereby reducing the computational overhead in the model fusion process and improving the performance and fusion efficiency of the target fusion model.

[0108] In an optional embodiment, the model fusion method in this application further includes:

[0109] The average nonlinear score is calculated by averaging the nonlinear scores of multiple sub-modules. The nonlinear average score is then compared with a preset threshold to obtain a comparison result. If the comparison result indicates that the nonlinear average score is less than the preset threshold, the evaluation result is determined to meet the preset conditions.

[0110] Specifically, the nonlinearity scores of multiple sub-modules are used to calculate the average nonlinearity score of all sub-modules, i.e., the nonlinearity average score S. A preset threshold T is set, which can be determined based on previous knowledge or experimental results, to determine whether the linearity of the model sub-modules is good enough for model fusion. For example, the preset threshold T can be an empirical value of 0.5.

[0111] The nonlinear average score S is compared with a preset threshold T. If the nonlinear average score S is less than the preset threshold T, it indicates that most sub-modules of the model exhibit good linear behavior during fine-tuning, which will serve as the basis for determining whether the evaluation results meet the preset conditions. Furthermore, when the evaluation results meet the preset conditions, it can be determined that the sub-modules are suitable for a fusion strategy based on linear properties.

[0112] Based on the above optional embodiments, the nonlinear average score is obtained by averaging the nonlinear scores of multiple sub-modules. The nonlinear average score is then compared with a preset threshold to obtain a comparison result. Finally, in response to the comparison result showing that the nonlinear average score is less than the preset threshold, it is determined that the evaluation result meets the preset conditions. This not only helps to identify the feasibility of model fusion, but also ensures the efficiency of the fusion process.

[0113] In an optional embodiment, in step S23, at the target model decomposition granularity, model fusion is performed on multiple sub-modules to generate a target fused model, including:

[0114] Step S231: At the target model decomposition granularity, obtain the target fusion weights corresponding to multiple sub-modules;

[0115] Step S232: Linearly fuse multiple sub-modules based on the target fusion weights to generate the target fusion model.

[0116] Specifically, the target model decomposition granularity is a decomposition granularity with linear conditions. When performing model fusion at the decomposition granularity with linear conditions, the target fusion weight of each sub-module of the model to be fused is calculated. The target fusion weight can control the degree of contribution of different sub-modules to the target fused model.

[0117] For example, during the linear fusion of multiple sub-modules based on the target fusion weights, a new model with the same structure as the pre-trained model can be created as the basis for the target fusion model. The target fusion weights of each sub-module are then integrated into the fusion model, replacing the original pre-trained weights, thus constructing the target fusion model. Finally, the generated target fusion model undergoes preliminary performance validation using a small amount of task-related data to check whether the model maintains good performance after fusion and possesses the ability to handle multiple tasks.

[0118] Based on the above optional embodiments, by obtaining the target fusion weights corresponding to multiple sub-modules at the target model decomposition granularity, and then linearly fusing the multiple sub-modules based on the target fusion weights to generate a target fusion model, the target fusion model not only inherits the advantages of multiple sub-modules on specific tasks, but also maintains the universality of the pre-trained model in the overall architecture, and can exhibit balanced performance on multiple tasks, thereby further improving the model fusion efficiency and fusion effect.

[0119] In an optional embodiment, in step S231, obtaining the target fusion weights corresponding to multiple sub-modules at the target model decomposition granularity includes:

[0120] At the target model decomposition granularity, the closed-form solution of the fusion weight is calculated based on the first model weight and the second model weight to obtain the target fusion weights corresponding to multiple sub-modules. The first model weight is the weight used by multiple transformation layers of the pre-trained model, and the second model weight is the weight used by multiple transformation layers of different target sub-modules in the multiple sub-modules. The multiple sub-modules are obtained by fine-tuning the pre-trained model.

[0121] Specifically, the closed-form solution calculation of fusion weights refers to obtaining the exact analytical expression of the fusion weights directly through mathematical analysis and algebraic transformations in mathematical optimization problems, without relying on iterative algorithms or numerical approximation methods. In the context of model fusion, closed-form solutions allow for the direct calculation of the target fusion weights when multiple sub-modules are fused together, without needing to try and determine the target fusion weights one by one using time-consuming methods such as grid search.

[0122] In this embodiment, the closed-loop solution calculation is based on the linear property of the submodules. The linear property means that a small change in the model parameters will lead to a linear change in the model output features. This property allows the target fusion weight combination during model fusion to be found through derivation and analysis, so as to achieve suitable overall performance.

[0123] Based on the above optional embodiments, by calculating the fusion weights using a closed-form solution based on the first model weights and the second model weights at the target model decomposition granularity, the target fusion weights corresponding to multiple sub-modules can be quickly obtained. Thus, by using the closed-form solution to calculate the target fusion weights, the entire process avoids time-consuming grid search, significantly reducing the demand for computing resources. At the same time, because only a small amount of data is needed to complete model fusion, data dependency and processing costs are greatly reduced.

[0124] In one optional embodiment, the closed-form solution calculation of the fusion weights based on the first model weights and the second model weights, to obtain the target fusion weights corresponding to multiple sub-modules, includes:

[0125] The fifth feature vector is calculated based on the training task data and the weights of the first model, and the sixth feature vector is calculated based on the training task data and the weights of the second model.

[0126] Calculate the eigenvector difference between the fifth and sixth eigenvectors;

[0127] The closed-form solution of the fusion weights is calculated based on the difference in feature vectors to obtain the target fusion weights corresponding to multiple sub-modules.

[0128] Specifically, the fifth feature vector is calculated based on the training task data x and the first model weights θ0. And calculate the sixth feature vector based on the training task data x and the weights θ of the second model. Calculate the fifth eigenvector With the sixth eigenvector eigenvector difference between Based on the difference of eigenvectors Perform closed-form calculation of the fusion weights to obtain the target fusion weights corresponding to multiple sub-modules.

[0129] Based on the derivation of the linear properties of the model, the closed-form solution of the target fusion weights can be obtained through the following process:

[0130]

[0131] in, A and b are the target fusion weights of the t-th model in the i-th layer submodule during model fusion. A and b are matrices and vectors calculated based on the output features of the submodule and the difference in model parameters. -1 Let A be the inverse matrix.

[0132] Where A∈R T×T b∈R T ,

[0133]

[0134] After obtaining the target fusion weights, during the linear fusion of multiple sub-modules based on these weights, the input includes: the weights used by multiple transformation layers of the T sub-modules to be fused. Each submodule is part of the first model weights in the pre-trained model. Based on this, it was fine-tuned for different tasks. The datasets related to each task are D1, D2, ..., D... T This is used to evaluate the linearity of the model and calculate the fusion weights. The target fusion weights are...

[0135] For each submodule's task t, from its associated dataset D t A small amount of data is randomly sampled to form dataset d. t Its size is N, where N is a small number, such as 30 data points, used to calculate the non-linear scores of the submodules and the target fusion weights. The dataset d is processed using the first model weights θ0 of the pre-trained model. t For each sample x in the dataset, obtain the input feature set. For computational preparation, 'i' represents the submodule number. Each submodule in each layer undergoes the same processing to obtain its input features. The target fusion weights are calculated for each layer 'i' of the model. The target fusion weights will be used to weight and fuse the parameters of sub-modules from different models to form the fused model.

[0136] Based on the calculated target fusion weights, linear fusion is performed on the sub-modules of each layer i. The fusion process is implemented using the following formula:

[0137]

[0138] in, To achieve the target weighting, These are the weights of the i-th layer in the pre-trained model, and It is the weight difference between submodule t and the pre-trained model at layer i.

[0139] Based on the above optional embodiments, by calculating the fifth feature vector based on the training task data and the first model weights, and calculating the sixth feature vector based on the training task data and the second model weights, and then calculating the feature vector difference between the fifth and sixth feature vectors, and finally calculating the closed-form solution of the fusion weights based on the feature vector difference, the target fusion weights corresponding to multiple sub-modules can be obtained. Thus, by using the closed-form solution to calculate the target fusion weights, the entire process avoids time-consuming grid search, significantly reduces the demand for computing resources, and also greatly reduces data dependence and processing costs because only a small amount of data is needed to complete model fusion.

[0140] Figure 3 This is a flowchart of a data processing method according to an embodiment of this application, such as... Figure 3 As shown, the method may include the following steps:

[0141] Step S31: Obtain task data;

[0142] Step S32: The target fusion model is used to perform task analysis on the task data to obtain the task processing result; wherein, the target fusion model is generated according to the model fusion method of any one of the embodiments of this application.

[0143] Based on the above steps S31 to S32, by acquiring task data and then using the target fusion model to perform task analysis on the task data, the task processing results can be obtained, thereby further improving the processing efficiency of the target fusion model for task data.

[0144] By dividing multiple models to be fused into sub-modules according to the granularity of the target model decomposition, and then performing nonlinear evaluation on these sub-modules to obtain evaluation results, the nonlinearity of each sub-module at the target model decomposition granularity is measured. Finally, in response to the evaluation results meeting preset conditions, the multiple sub-modules are fused at the target model decomposition granularity to generate a target fused model. This target fused model is used to perform task analysis on task data to obtain task processing results. Therefore, based on the nonlinear evaluation and fusion of sub-modules, finer-grained control can be provided for model fusion, enabling more efficient and accurate model fusion, thereby reducing development costs and improving the flexibility and scalability of the target fused model. By evaluating the nonlinearity of the sub-modules and performing model fusion at the target model decomposition granularity when the nonlinear evaluation results meet preset conditions, the computational complexity of the fusion process is reduced, thus improving the efficiency of model fusion. Furthermore, after determining the target decomposition granularity, only a small amount of task-related data is needed for nonlinear evaluation, which can avoid the high resource consumption of retraining or grid search in traditional model fusion methods, greatly reducing resource requirements, and thus solving the technical problems of low fusion efficiency and high resource consumption in related technologies when performing model fusion.

[0145] Figure 4 This is a flowchart of another data processing method according to an embodiment of this application, such as... Figure 4 As shown, the method may include the following steps:

[0146] Step S41: Obtain programming task data;

[0147] Step S42: The programming task data is analyzed using a target fusion model to obtain the programming task processing result; wherein, the target fusion model is generated according to any model fusion method in any of the embodiments of this application.

[0148] Based on the above steps S41 to S42, by acquiring programming task data, and then using the target fusion model to perform task analysis on the programming task data, the programming task processing results can be obtained, thereby further improving the processing efficiency of the target fusion model for programming task data.

[0149] By dividing multiple models to be fused into sub-modules according to the granularity of the target model decomposition, and then performing nonlinear evaluation on these sub-modules to obtain evaluation results, the nonlinearity of each sub-module at the target model decomposition granularity is measured. Finally, in response to the evaluation results meeting preset conditions, the multiple sub-modules are fused at the target model decomposition granularity to generate a target fused model. This target fused model is used to perform task analysis on programming task data to obtain the programming task processing results. Therefore, based on the nonlinear evaluation and fusion of sub-modules, finer-grained control can be provided for model fusion, enabling more efficient and accurate model fusion, thereby reducing development costs and improving the flexibility and scalability of the target fused model. By evaluating the nonlinearity of the sub-modules and fusing them at the target model decomposition granularity when the nonlinear evaluation results meet preset conditions, the computational complexity of the fusion process is reduced, thus improving the efficiency of model fusion. Furthermore, after determining the target decomposition granularity, only a small amount of task-related data is needed for nonlinear evaluation, which can avoid the high resource consumption of retraining or grid search in traditional model fusion methods, greatly reducing resource requirements, and thus solving the technical problems of low fusion efficiency and high resource consumption in related technologies when performing model fusion.

[0150] Figure 5 This is a flowchart of another data processing method according to an embodiment of this application, such as... Figure 5 As shown, the method may include the following steps:

[0151] Step S51: Obtain a data processing request through the first application programming interface, wherein the request data carried in the data processing request includes: task data;

[0152] Step S52: Return a data processing response through the second application programming interface. The response data carried in the data processing response includes: task processing result, which is obtained by performing task analysis on the task data using a target fusion model. The target fusion model is generated according to any of the model fusion methods in the embodiments of this application.

[0153] Based on the above steps S51 to S52, a data processing request is obtained through the first application programming interface, wherein the request data carried in the data processing request includes: task data. Then, a data processing response is returned through the second application programming interface, wherein the response data carried in the data processing response includes: task processing result. The task processing result is obtained by performing task analysis on the task data using the target fusion model, thereby further improving the processing efficiency of the target fusion model for task data.

[0154] By dividing multiple models to be fused into sub-modules according to the granularity of the target model decomposition, and then performing nonlinear evaluation on these sub-modules to obtain evaluation results, the nonlinearity of each sub-module at the target model decomposition granularity is measured. Finally, in response to the evaluation results meeting preset conditions, the multiple sub-modules are fused at the target model decomposition granularity to generate a target fused model. This target fused model is used to perform task analysis on task data to obtain task processing results. Therefore, based on the nonlinear evaluation and fusion of sub-modules, finer-grained control can be provided for model fusion, enabling more efficient and accurate model fusion, thereby reducing development costs and improving the flexibility and scalability of the target fused model. By evaluating the nonlinearity of the sub-modules and performing model fusion at the target model decomposition granularity when the nonlinear evaluation results meet preset conditions, the computational complexity of the fusion process is reduced, thus improving the efficiency of model fusion. Furthermore, after determining the target decomposition granularity, only a small amount of task-related data is needed for nonlinear evaluation, which can avoid the high resource consumption of retraining or grid search in traditional model fusion methods, greatly reducing resource requirements, and thus solving the technical problems of low fusion efficiency and high resource consumption in related technologies when performing model fusion.

[0155] Figure 6 This is a flowchart of another data processing method according to an embodiment of this application, such as... Figure 6 As shown, the method may include the following steps:

[0156] Step S61: Obtain the currently input data processing dialogue request, wherein the request data carried in the data processing dialogue request includes: task data;

[0157] Step S62, in response to the data processing dialogue request, a data processing dialogue response is returned, wherein the information carried in the data processing dialogue response includes: task processing result, which is obtained by performing task analysis on the task data using a target fusion model, and the target fusion model is generated according to any of the model fusion methods in the embodiments of this application;

[0158] Step S63: Display the task processing results within the graphical user interface.

[0159] Based on the above steps S61 to S63, by obtaining the currently input data processing dialogue request, the request data carried in the data processing dialogue request includes: task data, and then responding to the data processing dialogue request, a data processing dialogue reply is returned, the information carried in the data processing dialogue reply includes: task processing result, the task processing result is obtained by performing task analysis on the task data using the target fusion model, and finally the task processing result is displayed in the graphical user interface, thereby further improving the processing efficiency of the target fusion model for task data.

[0160] By dividing multiple models to be fused into sub-modules according to the granularity of the target model decomposition, and then performing nonlinear evaluation on these sub-modules to obtain evaluation results, the nonlinearity of each sub-module at the target model decomposition granularity is measured. Finally, in response to the evaluation results meeting preset conditions, the multiple sub-modules are fused at the target model decomposition granularity to generate a target fused model. This target fused model is used to perform task analysis on task data to obtain task processing results. Therefore, based on the nonlinear evaluation and fusion of sub-modules, finer-grained control can be provided for model fusion, enabling more efficient and accurate model fusion, thereby reducing development costs and improving the flexibility and scalability of the target fused model. By evaluating the nonlinearity of the sub-modules and performing model fusion at the target model decomposition granularity when the nonlinear evaluation results meet preset conditions, the computational complexity of the fusion process is reduced, thus improving the efficiency of model fusion. Furthermore, after determining the target decomposition granularity, only a small amount of task-related data is needed for nonlinear evaluation, which can avoid the high resource consumption of retraining or grid search in traditional model fusion methods, greatly reducing resource requirements, and thus solving the technical problems of low fusion efficiency and high resource consumption in related technologies when performing model fusion.

[0161] Figure 7 This is a flowchart of another data processing method according to an embodiment of this application, such as... Figure 7 As shown, the method may include the following steps:

[0162] Step S71: In response to the input command applied to the operation interface, display the task data on the operation interface;

[0163] Step S72: In response to the processing command applied to the operation interface, the task processing result is displayed on the operation interface; wherein, the task processing result is obtained by performing task analysis on the task data using a target fusion model, and the target fusion model is generated according to any of the model fusion methods in the embodiments of this application.

[0164] Based on the above steps S71 to S72, by responding to the input commands applied to the operation interface, the task data is displayed on the operation interface, and then by responding to the processing commands applied to the operation interface, the task processing result is displayed on the operation interface. The task processing result is obtained by performing task analysis on the task data using the target fusion model, thereby further improving the processing efficiency of the target fusion model for task data.

[0165] By dividing multiple models to be fused into sub-modules according to the granularity of the target model decomposition, and then performing nonlinear evaluation on these sub-modules to obtain evaluation results, the nonlinearity of each sub-module at the target model decomposition granularity is measured. Finally, in response to the evaluation results meeting preset conditions, the multiple sub-modules are fused at the target model decomposition granularity to generate a target fused model. This target fused model is used to perform task analysis on task data to obtain task processing results. Therefore, based on the nonlinear evaluation and fusion of sub-modules, finer-grained control can be provided for model fusion, enabling more efficient and accurate model fusion, thereby reducing development costs and improving the flexibility and scalability of the target fused model. By evaluating the nonlinearity of the sub-modules and performing model fusion at the target model decomposition granularity when the nonlinear evaluation results meet preset conditions, the computational complexity of the fusion process is reduced, thus improving the efficiency of model fusion. Furthermore, after determining the target decomposition granularity, only a small amount of task-related data is needed for nonlinear evaluation, which can avoid the high resource consumption of retraining or grid search in traditional model fusion methods, greatly reducing resource requirements, and thus solving the technical problems of low fusion efficiency and high resource consumption in related technologies when performing model fusion.

[0166] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of the relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation portals are provided for users to choose to authorize or refuse.

[0167] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0168] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods according to the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, they can also be implemented by hardware. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0169] According to an embodiment of this application, a model fusion apparatus for implementing the above-described model fusion method is also provided. Figure 8 This is a structural block diagram of a model fusion apparatus according to an embodiment of this application, such as... Figure 8 As shown, the device includes:

[0170] The partitioning module 801 is used to divide multiple models to be fused into multiple sub-modules according to the target model decomposition granularity. The target model decomposition granularity is selected from multiple model decomposition granularities, and the multiple model decomposition granularities are used to partition the models to be fused using different model partitioning methods.

[0171] Evaluation module 802 is used to perform nonlinear evaluation on multiple sub-modules and obtain evaluation results, wherein the evaluation results are used to measure the degree of nonlinearity of multiple sub-modules at the decomposition granularity of the target model.

[0172] The fusion module 803 is used to perform model fusion on multiple sub-modules at the target model decomposition granularity in response to the evaluation results meeting preset conditions, and generate a target fusion model. The target fusion model is used to perform task analysis on task data to obtain task processing results.

[0173] Optionally, the evaluation module 802 is further configured to: perform nonlinear evaluation on multiple sub-modules, obtain nonlinear scores of multiple sub-modules, and determine the evaluation result based on the nonlinear scores of multiple sub-modules.

[0174] Optionally, the evaluation module 802 is further configured to: obtain a first model weight and a second model weight, wherein the first model weight is the weight used by multiple transformation layers of the pre-trained model, and the second model weight is the weight used by multiple transformation layers of different target sub-modules in multiple sub-modules, and the multiple sub-modules are obtained by fine-tuning the pre-trained model respectively; and perform feature space distance measurement based on the first model weight and the second model weight to obtain nonlinear scores of multiple sub-modules.

[0175] Optionally, the evaluation module 802 is further configured to: determine the task vector of the target submodule using the first model weight and the second model weight, wherein the task vector is used to characterize the model weight difference between the first model weight and the second model weight; and perform feature space distance measurement based on the training task data, the first model weight, the second model weight and the task vector to obtain nonlinear scores for multiple submodules.

[0176] Optionally, the evaluation module 802 is further configured to: obtain a first output feature vector based on training task data, first model weights, task vectors, and a first preset weight coefficient; obtain a second output feature vector based on training task data, first model weights, task vectors, and a second preset weight coefficient; obtain a third output feature vector based on training task data and first model weights; obtain a fourth output feature vector based on training task data and second model weights; and use the first output feature vector, second output feature vector, third output feature vector, and fourth output feature vector to perform feature space distance measurement to obtain nonlinear scores for multiple sub-modules; wherein the first preset weight coefficient is different from the second preset weight coefficient.

[0177] Optionally, the evaluation module 802 is further configured to: measure the feature space distance between the first output feature vector and the second output feature vector to obtain a first measurement result, and measure the feature space distance between the third output feature vector and the fourth output feature vector to obtain a second measurement result; and obtain the nonlinear scores of multiple sub-modules based on the first measurement result and the second measurement result.

[0178] Optionally, the model fusion device further includes: a calculation module 804, used to calculate the mean of the nonlinear scores of multiple sub-modules to obtain a nonlinear average score; a comparison module 805, used to compare the nonlinear average score with a preset threshold to obtain a comparison result; and a determination module 806, used to determine that the evaluation result meets the preset conditions in response to the comparison result showing that the nonlinear average score is less than the preset threshold.

[0179] Optionally, the fusion module 803 is further configured to: obtain the target fusion weights corresponding to multiple sub-modules at the target model decomposition granularity; and perform linear fusion of the multiple sub-modules based on the target fusion weights to generate a target fusion model.

[0180] Optionally, the fusion module 803 is further configured to: calculate the closed-form solution of the fusion weights based on the first model weights and the second model weights at the target model decomposition granularity, and obtain the target fusion weights corresponding to multiple sub-modules, wherein the first model weights are the weights used by multiple transformation layers of the pre-trained model, the second model weights are the weights used by multiple transformation layers of different target sub-modules in the multiple sub-modules, and the multiple sub-modules are obtained by fine-tuning the pre-trained model respectively.

[0181] Optionally, the fusion module 803 is further configured to: calculate a fifth feature vector based on the training task data and the weights of the first model, and calculate a sixth feature vector based on the training task data and the weights of the second model; calculate the feature vector difference between the fifth feature vector and the sixth feature vector; and perform closed-form solution calculation of the fusion weights based on the feature vector difference to obtain the target fusion weights corresponding to multiple sub-modules.

[0182] It should be noted that the division module 801, evaluation module 802, and fusion module 803 correspond to steps S21 to S23 in the above embodiments. The three modules and their corresponding steps implement the same instances and application scenarios, but are not limited to the content disclosed in the above embodiments. It should be noted that the above modules or units can be hardware or software components stored in memory and processed by one or more processors. The above modules can also run as part of the device in the server provided in the above embodiments.

[0183] Figure 9 This is a structural block diagram of a data processing apparatus according to an embodiment of this application, such as... Figure 9 As shown, the device includes:

[0184] Module 901 is used to acquire task data;

[0185] The analysis module 902 is used to perform task analysis on task data using a target fusion model to obtain task processing results; wherein the target fusion model is generated according to any one of the model fusion methods in the embodiments of this application.

[0186] It should be noted that the acquisition module 901 and analysis module 902 correspond to steps S31 to S32 in the above embodiments. The two modules and their corresponding steps implement the same instances and application scenarios, but are not limited to the content disclosed in the above embodiments. It should be noted that the above modules or units can be hardware or software components stored in memory and processed by one or more processors. The above modules can also be part of a device and run in the server provided in the above embodiments.

[0187] Figure 10 According to another structural block diagram of a data processing apparatus according to an embodiment of this application, such as... Figure 10 As shown, the device includes:

[0188] Module 1001 is used to acquire programming task data;

[0189] The analysis module 1002 is used to perform task analysis on programming task data using a target fusion model to obtain programming task processing results; wherein, the target fusion model is generated according to any one of the model fusion methods in the embodiments of this application.

[0190] It should be noted that the acquisition module 1001 and analysis module 1002 correspond to steps S41 to S42 in the above embodiments. The two modules and their corresponding steps implement the same instances and application scenarios, but are not limited to the content disclosed in the above embodiments. It should be noted that the above modules or units may be hardware or software components stored in memory and processed by one or more processors. The above modules may also be part of a device and run in the server provided in the above embodiments.

[0191] Figure 11 This is a structural block diagram of another data processing apparatus according to an embodiment of this application, such as... Figure 11 As shown, the device includes:

[0192] The acquisition module 1101 is used to acquire a data processing request through a first application programming interface, wherein the request data carried in the data processing request includes: task data;

[0193] The return module 1102 is used to return a data processing response through a second application programming interface. The response data carried in the data processing response includes: task processing results, which are obtained by performing task analysis on the task data using a target fusion model. The target fusion model is generated according to any one of the model fusion methods in the embodiments of this application.

[0194] It should be noted that the acquisition module 1101 and the return module 1102 correspond to steps S51 to S52 in the above embodiments. The two modules and the corresponding steps implement the same instances and application scenarios, but are not limited to the content disclosed in the above embodiments. It should be noted that the above modules or units can be hardware components or software components stored in memory and processed by one or more processors. The above modules can also be part of the device and run in the server provided in the above embodiments.

[0195] Figure 12 This is a structural block diagram of another data processing apparatus according to an embodiment of this application, such as... Figure 12 As shown, the device includes:

[0196] The acquisition module 1201 is used to acquire the currently input data processing dialogue request, wherein the request data carried in the data processing dialogue request includes: task data;

[0197] Return module 1202 is used to respond to a data processing dialogue request and return a data processing dialogue response. The information carried in the data processing dialogue response includes: task processing result, which is obtained by performing task analysis on task data using a target fusion model. The target fusion model is generated according to any one of the model fusion methods in the embodiments of this application.

[0198] Display module 1203 is used to display task processing results within a graphical user interface.

[0199] It should be noted that the acquisition module 1201, return module 1202, and display module 1203 correspond to steps S61 to S63 in the above embodiments. The three modules and their corresponding steps implement the same instances and application scenarios, but are not limited to the content disclosed in the above embodiments. It should be noted that the above modules or units can be hardware or software components stored in memory and processed by one or more processors. The above modules can also run as part of the device in the server provided in the above embodiments.

[0200] Figure 13 This is a structural block diagram of another data processing apparatus according to an embodiment of this application, such as... Figure 13 As shown, the device includes:

[0201] The first display module 1301 is used to respond to input commands applied to the operation interface and display task data on the operation interface.

[0202] The second display module 1202 is used to respond to processing instructions applied to the operation interface and display the task processing results on the operation interface; wherein, the task processing results are obtained by performing task analysis on the task data using a target fusion model, and the target fusion model is generated according to any one of the model fusion methods in the embodiments of this application.

[0203] It should be noted that the first display module 1301 and the second display module 1302 mentioned above correspond to steps S71 to S72 in the above embodiments. The two modules and the corresponding steps implement the same instances and application scenarios, but are not limited to the content disclosed in the above embodiments. It should be noted that the above modules or units can be hardware components or software components stored in memory and processed by one or more processors. The above modules can also be part of a device and run in the server provided in the above embodiments.

[0204] It should be noted that the preferred embodiments involved in the above embodiments of this application are the same as the solutions, application scenarios and implementation processes provided in the above embodiments, but are not limited to the solutions provided in the above embodiments.

[0205] Embodiments of this application may provide a data processing system, including: a client for sending task data; a server connected to the client for performing task analysis on the task data using a target fusion model to obtain task processing results; the client is also used to output the task processing results; wherein the target fusion model is generated according to any model fusion method in any of the embodiments of this application.

[0206] Embodiments of this application may provide a computing device. Figure 14 This is a structural block diagram of a computing device according to an embodiment of this application. Figure 14 As shown, the computing device may include one or more (only one is shown in the figure) processors 142, memory 144, memory controller, and peripheral interfaces.

[0207] The aforementioned computing device can be understood as an integrated smart terminal, including but not limited to servers, desktop computers, PCs (Personal Computers), all-in-one model machines, mobile phones, tablets, or other portable smart terminals, and the computing device may have the model described in the above embodiments of this application pre-installed.

[0208] Specifically, this computing device can pre-install various types of models, including but not limited to models in natural language processing, visual processing, speech processing, code processing, and multimodal task processing, thus providing diverse model selection. In different product forms, this computing device can support one or more model usage methods, including but not limited to model training, model invocation, model fine-tuning, model deployment, model inference, and application. In some product forms, this computing device also supports model management, including but not limited to multi-type model management (supporting the management of discriminative, generative, and other model types), model version control (supporting the control of different model versions), and model evaluation (evaluating model performance and effectiveness based on model evaluation tools). In other product forms, this computing device can also create applications based on models, providing API calling capabilities, allowing models to be called into created applications through API interfaces, and providing application management tools for application management and monitoring.

[0209] Furthermore, the computing device may also include data management (supporting the creation and management of model tuning datasets), a training center (providing abundant training resources to help users learn and master AI technology), and basic control capabilities (providing enterprise-level basic control capabilities to ensure the security and efficient operation of the system). Through the above functions, it provides a comprehensive and integrated device for AI development, training, deployment, and application.

[0210] The memory can be used to store software programs and modules, such as the program instructions / modules corresponding to the methods and apparatus in the embodiments of this application. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory, thereby implementing the methods in the above embodiments. The memory may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory may further include memory remotely located relative to the processor, and these remote memories can be connected to the terminal via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0211] The processor can invoke an executable program stored in memory via a transmission device to execute the method described in any of the above embodiments.

[0212] It will be understood by those skilled in the art that the structure shown in the figure is merely illustrative, and the computing device may also be a smartphone (such as an Android phone, an iOS phone, etc.), a tablet computer, a PDA, a mobile internet device (MID), a PAD, or other terminal device. This figure does not limit the structure of the aforementioned computing device. For example, the computing device may include more or fewer components (such as network interfaces, display devices, etc.) than shown in the figure, or may have a different configuration than that shown in the figure.

[0213] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing the hardware related to the terminal device. The program can be stored in a computer-readable storage medium, which may include: flash drive, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0214] Embodiments of this application also provide a computer-readable storage medium. Optionally, in this embodiment, the computer-readable storage medium can be used to store program code executed by the method provided in the above embodiments.

[0215] Optionally, in this embodiment, the storage medium may be located in a computing device.

[0216] Optionally, in this embodiment, the computer-readable storage medium is configured to store an executable program, which, when the executable program is running, controls the device where the computer-readable storage medium is located to execute the method described in any of the above embodiments.

[0217] Embodiments of this application also provide a computer program product. Optionally, in this embodiment, the computer program product may include a computer program that, when executed by a processor, implements the methods provided in the embodiments described above.

[0218] Embodiments of this application also provide a computer program product. Optionally, the computer program product may include a non-volatile computer-readable storage medium, which can be used to store a computer program that, when executed by a processor, implements the method provided in the above embodiments.

[0219] Embodiments of this application also provide a computer program. Optionally, in this embodiment, when the computer program is executed by a processor, it implements the method provided in the above embodiments.

[0220] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0221] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0222] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0223] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0224] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard drive, magnetic disk, or optical disk.

[0225] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A model fusion method, characterized in that, include: According to the target model decomposition granularity, the multiple models to be fused are divided into multiple sub-modules. The target model decomposition granularity is selected from multiple model decomposition granularities, and the multiple model decomposition granularities are used to divide the models to be fused using different model division methods. The plurality of sub-modules are subjected to nonlinear evaluation to obtain evaluation results, wherein the evaluation results are used to measure the degree of nonlinearity of the plurality of sub-modules at the decomposition granularity of the target model; In response to the evaluation result meeting the preset conditions, at the target model decomposition granularity, the multiple sub-modules are fused to generate a target fusion model, wherein the target fusion model is used to perform task analysis on task data to obtain task processing results.

2. The model fusion method according to claim 1, characterized in that, The nonlinear evaluation of the multiple sub-modules yields the following evaluation results: Nonlinear evaluation is performed on the multiple sub-modules to obtain the nonlinear scores of the multiple sub-modules; The evaluation result is determined based on the nonlinear scores of the multiple sub-modules.

3. The model fusion method according to claim 2, characterized in that, The nonlinear evaluation of the multiple sub-modules and the acquisition of their nonlinear scores include: Obtain the first model weight and the second model weight, wherein the first model weight is the weight used by multiple transformation layers of the pre-trained model, and the second model weight is the weight used by multiple transformation layers of different target sub-modules in the multiple sub-modules, and the multiple sub-modules are obtained by fine-tuning the pre-trained model respectively; Based on the feature space distance measurement between the first model weight and the second model weight, the nonlinear scores of the multiple sub-modules are obtained.

4. The model fusion method according to claim 3, characterized in that, The nonlinear scores of the multiple sub-modules are obtained by measuring the distance in the feature space based on the first model weights and the second model weights, including: The task vector of the target submodule is determined using the first model weight and the second model weight, wherein the task vector is used to represent the model weight difference between the first model weight and the second model weight; Based on the training task data, the first model weights, the second model weights, and the task vector, a feature space distance metric is performed to obtain the nonlinear scores of the multiple sub-modules.

5. The model fusion method according to claim 4, characterized in that, Based on the training task data, the first model weights, the second model weights, and the task vector, a feature space distance metric is performed to obtain the nonlinear scores of the multiple sub-modules, including: Based on the training task data, the first model weights, the task vector, and the first preset weight coefficients, a first output feature vector is obtained, and based on the training task data, the first model weights, the task vector, and the second preset weight coefficients, a second output feature vector is obtained. A third output feature vector is obtained based on the training task data and the first model weights, and a fourth output feature vector is obtained based on the training task data and the second model weights; The first output feature vector, the second output feature vector, the third output feature vector, and the fourth output feature vector are used to perform feature space distance measurement to obtain the nonlinear scores of the multiple sub-modules; The first preset weighting coefficient is different from the second preset weighting coefficient.

6. The model fusion method according to claim 5, characterized in that, Using the first output feature vector, the second output feature vector, the third output feature vector, and the fourth output feature vector to perform feature space distance measurement, and obtaining the nonlinear scores of the multiple sub-modules includes: The first output feature vector and the second output feature vector are subjected to feature space distance measurement to obtain a first measurement result, and the third output feature vector and the fourth output feature vector are subjected to feature space distance measurement to obtain a second measurement result; Based on the first metric result and the second metric result, the nonlinear scores of the plurality of sub-modules are obtained.

7. The model fusion method according to claim 1, characterized in that, The model fusion method also includes: The average score is obtained by averaging the nonlinear scores of the multiple sub-modules. The nonlinear average score is compared with a preset threshold to obtain a comparison result. In response to the comparison result indicating that the nonlinear average score is less than the preset threshold, it is determined that the evaluation result meets the preset condition.

8. The model fusion method according to claim 1, characterized in that, At the target model decomposition granularity, model fusion is performed on the multiple sub-modules to generate the target fusion model, including: At the target model decomposition granularity, the target fusion weights corresponding to the multiple sub-modules are obtained; The multiple sub-modules are linearly fused based on the target fusion weights to generate the target fusion model.

9. The model fusion method according to claim 8, characterized in that, At the target model decomposition granularity, obtaining the target fusion weights corresponding to the multiple sub-modules includes: At the target model decomposition granularity, the closed-form solution of the fusion weight is calculated based on the first model weight and the second model weight to obtain the target fusion weight corresponding to the plurality of sub-modules. The first model weight is the weight used by the multiple transformation layers of the pre-trained model, and the second model weight is the weight used by the multiple transformation layers of any target sub-module among the plurality of sub-modules. The plurality of sub-modules are obtained by fine-tuning the pre-trained model.

10. The model fusion method according to claim 9, characterized in that, The closed-form solution calculation of the fusion weights based on the first model weights and the second model weights, to obtain the target fusion weights corresponding to the plurality of sub-modules, includes: The fifth feature vector is calculated based on the training task data and the weights of the first model, and the sixth feature vector is calculated based on the training task data and the weights of the second model. Calculate the feature vector difference between the fifth feature vector and the sixth feature vector; The target fusion weights corresponding to the multiple sub-modules are obtained by calculating the closed-form solution of the fusion weights based on the feature vector differences.

11. A data processing method, characterized in that, include: Obtain task data; The task data is analyzed using a target fusion model to obtain the task processing results; The target fusion model is generated according to the model fusion method described in any one of claims 1 to 10.

12. A data processing method, characterized in that, include: Obtain programming task data; The programming task data is analyzed using a target fusion model to obtain the programming task processing results; The target fusion model is generated according to the model fusion method described in any one of claims 1 to 10.

13. A data processing method, characterized in that, include: A data processing request is obtained through a first application programming interface, wherein the request data carried in the data processing request includes: task data; The data processing response is returned through the second application programming interface, wherein the response data carried in the data processing response includes: task processing result, which is obtained by performing task analysis on the task data using a target fusion model, and the target fusion model is generated according to the model fusion method of any one of claims 1 to 10.

14. A data processing method, characterized in that, include: Obtain the currently input data processing dialogue request, wherein the request data carried in the data processing dialogue request includes: task data; In response to the data processing dialogue request, a data processing dialogue response is returned, wherein the information carried in the data processing dialogue response includes: task processing result, which is obtained by performing task analysis on the task data using a target fusion model, and the target fusion model is generated according to the model fusion method described in any one of claims 1 to 10; The task processing results are displayed within a graphical user interface.

15. A data processing method, characterized in that, include: In response to input commands applied to the user interface, task data is displayed on the user interface. In response to processing instructions applied to the operation interface, the task processing results are displayed on the operation interface; The task processing result is obtained by performing task analysis on the task data using a target fusion model, and the target fusion model is generated according to the model fusion method described in any one of claims 1 to 10.

16. A data processing system, characterized in that, include: The client is used to send task data; The server, connected to the client, is used to perform task analysis on the task data using a target fusion model to obtain task processing results. The client is also used to output the task processing results; The target fusion model is generated according to the model fusion method described in any one of claims 1 to 10.

17. An electronic device, characterized in that, include: Memory, which stores executable programs; A processor for running the program, wherein the program, when running, performs the model fusion method according to any one of claims 1 to 10 or the data processing method according to any one of claims 11 to 15.

18. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored executable program, wherein, when the executable program is executed, it controls the device on which the computer-readable storage medium is located to perform the model fusion method of any one of claims 1 to 10 or the data processing method of any one of claims 11 to 15.

19. A computer program product, characterized in that, It includes a computer program that, when executed by a processor, implements the model fusion method of any one of claims 1 to 10 or the data processing method of any one of claims 11 to 15.

Images