Training method and device of large model and electronic equipment

Abstract

The present disclosure relates to a large model training method, device and electronic equipment, wherein the method comprises: obtaining a sample data analysis task and a sample subtask list corresponding to the sample data analysis task; determining the priority of each sample subtask in the sample subtask list according to the relevant knowledge data of the sample data analysis task, the sample subtask list and a graph neural network; determining a sample data analysis workflow according to the priority of each sample subtask and the sample subtask list; and obtaining a trained large model for data analysis workflow generation according to the sample data analysis task, the sample subtask list, the priority of each sample subtask and the sample data analysis workflow. The large model trained by the above method can automatically adjust the workflow arrangement logic according to different sample data analysis tasks, thereby adapting to diversified business requirements and further improving the accuracy and efficiency of workflow generation.

Description

Technical Field

[0001] This disclosure relates to the field of data processing technology, and in particular to a method, apparatus and electronic device for training large models. Background Technology

[0002] Currently, data analytics workflows cover the entire process from data collection to results application. Their core objective is to provide a basis for decision-making by uncovering patterns, trends, and correlations within the data.

[0003] In the relevant technologies, there is no large-scale model for generating data analysis workflows. Data analysis workflows mainly rely on manual data analysis methods. However, these data analysis methods have limitations in adaptability and automation, and there is still room for improvement in their performance and accuracy. Summary of the Invention

[0004] This disclosure provides a method, apparatus, and electronic device for training large models.

[0005] According to a first aspect of the present disclosure, a method for training a large model is provided. The method includes: acquiring a sample data analysis task and a list of sample subtasks corresponding to the sample data analysis task; determining the priority of each sample subtask in the sample subtask list based on relevant knowledge data of the sample data analysis task, the list of sample subtasks, and a graph neural network; determining a sample data analysis workflow corresponding to the sample data analysis task based on the priority of each sample subtask and the list of sample subtasks; and training an initial large model based on the sample data analysis task, the list of sample subtasks, the priority of each sample subtask, and the sample data analysis workflow to obtain a trained large model for generating the data analysis workflow.

[0006] In one embodiment of this disclosure, obtaining the sample data analysis task and the sample subtask list corresponding to the sample data analysis task includes: performing intent recognition processing on the sample data analysis task to obtain an intent recognition result; determining a task template matching the sample data analysis task based on the relevant knowledge data; the task template includes multiple subtask templates; and determining the sample subtask list based on the intent recognition result and the multiple subtask templates.

[0007] In one embodiment of this disclosure, determining the priority of each sample subtask in the sample subtask list based on the relevant knowledge data of the sample data analysis task, the sample subtask list, and the graph neural network includes: determining the dependencies and weights between the sample subtasks in the sample subtask list based on the relevant knowledge data; generating a directed acyclic graph (DAG) based on the dependencies and weights between the sample subtasks; training the initial graph neural network based on the DAG to obtain a trained graph neural network; and determining the priority of each sample subtask based on the trained graph neural network.

[0008] In one embodiment of this disclosure, determining the sample data analysis workflow corresponding to the sample data analysis task based on the priority of each sample subtask and the sample subtask list includes: sorting each sample subtask according to its priority to obtain a sorting result; obtaining a workflow component library; the workflow component library includes multiple workflow components and description information of the workflow components; the workflow components embed exception handling logic; sequentially determining at least one candidate workflow component matching each sample subtask in the sorting result based on the description information of the multiple workflow components; and assembling the candidate workflow components corresponding to each sample subtask to obtain the sample data analysis workflow.

[0009] In one embodiment of this disclosure, the step of training an initial large model based on the sample data analysis task, the sample subtask list, the priority of each sample subtask, and the sample data analysis workflow to obtain a trained large model for generating the data analysis workflow includes: determining the sample stage data of the sample data analysis task at each analysis stage based on the sample data analysis task, the sample subtask list, the priority of each sample subtask, and the sample data analysis workflow; the each analysis stage refers to a stage in the data analysis workflow generation process; and training the initial large model based on the sample stage data of the sample data analysis task at each analysis stage to obtain a trained large model for generating the data analysis workflow.

[0010] In one embodiment of this disclosure, the step of training the initial large model based on sample stage data from each analysis stage of the sample data analysis task to obtain a trained large model for generating the data analysis workflow includes: sequentially acquiring the data to be processed by the sample data analysis task in each analysis stage; inputting the data to be processed, the relevant knowledge data, and the stage prompt text of the analysis stage into the initial large model to obtain the prediction stage data output by the initial large model; and adjusting the parameters of the initial large model based on the sample stage data of the analysis stage and the prediction stage data to obtain the trained large model.

[0011] In one embodiment of this disclosure, the analysis phase includes at least one of the following: a task decomposition phase, a task sorting phase, and a task component phase; the sample phase data of the task decomposition phase includes the sample subtask list; the sample phase data of the task sorting phase includes the priority of each sample subtask; and the sample phase data of the task component phase includes the sample data analysis workflow.

[0012] In one embodiment of this disclosure, the method further includes: acquiring a target data analysis task; sequentially determining target stage data of historical analysis stages preceding the analysis stage to be processed for each analysis stage in the various analysis stages; inputting the target stage data and the stage prompt text corresponding to the analysis stage to be processed into the large model to acquire the target stage data of the analysis stage to be processed output by the large model; and determining the target data analysis workflow corresponding to the target data analysis task based on the target stage data of the last analysis stage in the various analysis stages.

[0013] In one embodiment of this disclosure, the method further includes: acquiring a workflow experience base; the workflow experience base includes experience texts from the data analysis workflow generation process; and performing verification and adjustment processing on the target data analysis workflow based on the experience texts in the workflow experience base.

[0014] In one embodiment of this disclosure, the method further includes: displaying target stage data of at least one analysis stage to obtain feedback from the object; determining first target stage data related to the feedback; and adjusting the first target stage data according to the feedback.

[0015] According to a second aspect of the present disclosure, a training apparatus for a large model is also provided. The apparatus includes: a first acquisition module, configured to acquire a sample data analysis task and a list of sample subtasks corresponding to the sample data analysis task; a first determination module, configured to determine the priority of each sample subtask in the sample subtask list based on relevant knowledge data of the sample data analysis task, the list of sample subtasks, and a graph neural network; a second determination module, configured to determine a sample data analysis workflow corresponding to the sample data analysis task based on the priority of each sample subtask and the list of sample subtasks; and a training processing module, configured to train an initial large model based on the sample data analysis task, the list of sample subtasks, the priority of each sample subtask, and the sample data analysis workflow to obtain a trained large model for generating the data analysis workflow.

[0016] According to a third aspect of the present disclosure, an electronic device is also provided, comprising: a processor; a memory for storing processor-executable instructions; wherein the processor is configured to: implement the steps of the training method for a large model as described above.

[0017] According to a fourth aspect of the present disclosure, a non-transitory computer-readable storage medium is also provided, wherein when instructions in the storage medium are executed by a processor, the processor is able to execute the training method for a large model as described above.

[0018] The technical solutions provided by the embodiments of this disclosure have at least the following beneficial effects: The process involves: acquiring sample data analysis tasks and a list of corresponding sample subtasks; determining the priority of each sample subtask in the list based on relevant knowledge data, the sample subtask list, and a graph neural network; determining the sample data analysis workflow based on the priorities of each sample subtask and the sample subtask list; and obtaining a trained large-scale model for generating the data analysis workflow based on the sample data analysis tasks, the sample subtask list, the priorities of each sample subtask, and the sample data analysis workflow. This large-scale model, trained using the above method, can automatically adjust the workflow orchestration logic according to different sample data analysis tasks, thereby adapting to diverse business needs and further improving the accuracy and efficiency of workflow generation.

[0019] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0020] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure, and are not intended to unduly limit this disclosure.

[0021] Figure 1 A flowchart illustrating a training method for a large model according to an embodiment of this disclosure; Figure 2 A flowchart illustrating a training method for a large model according to an embodiment of this disclosure; Figure 3 A flowchart illustrating a training method for a large model according to an embodiment of this disclosure; Figure 4 This is a schematic diagram illustrating the principle of a large model training method according to an embodiment of the present disclosure; Figure 5 This is a schematic diagram of the structure of a training device for a large model according to an embodiment of the present disclosure; Figure 6 This is a structural block diagram of an electronic device according to an exemplary embodiment of the present disclosure. Detailed Implementation

[0022] To enable those skilled in the art to better understand the technical solutions of this disclosure, the technical solutions in the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings.

[0023] It should be noted that the terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this disclosure 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 disclosure described herein can be implemented in orders other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure as detailed in the appended claims.

[0024] Currently, data analytics workflows cover the entire process from data collection to results application. Their core objective is to provide a basis for decision-making by uncovering patterns, trends, and correlations within the data.

[0025] In the relevant technologies, there is no large-scale model for generating data analysis workflows. Data analysis workflows mainly rely on manual data analysis methods. However, these data analysis methods have limitations in adaptability and automation, and there is still room for improvement in their performance and accuracy.

[0026] Figure 1This is a flowchart of a large model training method according to an embodiment of the present disclosure. It should be noted that the large model training method of this embodiment can be applied to a large model training device, which can be configured in an electronic device to enable the electronic device to perform large model training functions.

[0027] The electronic device can be any device with computing capabilities, such as a terminal device or a server. The server can be used in training large models, etc. The following embodiments use a server that trains a large model as the execution entity for illustration.

[0028] like Figure 1 As shown, the method includes the following steps: Step 101: Obtain the sample data analysis task and the list of sample subtasks corresponding to the sample data analysis task.

[0029] It should be noted that sample data analysis tasks can be obtained from workflow task descriptions and examples (or task requirement lists), but are not limited to these. The task requirement list contains real user input samples, covering a wide range of possible task descriptions.

[0030] In this embodiment of the disclosure, in order to efficiently and accurately obtain the sample subtask list corresponding to the sample data analysis task, the sample data analysis task can be processed by intent recognition to obtain intent recognition results; based on relevant knowledge data, a task template matching the sample data analysis task can be determined; the task template includes multiple subtask templates; and based on the intent recognition results and multiple subtask templates, the sample subtask list can be determined.

[0031] The relevant knowledge data includes data from multiple dimensions, including but not limited to descriptions and examples of workflow tasks corresponding to data analysis, lists of workflow components and their functions and usage instructions, workflow platform documentation and tutorials, execution records and logs of workflow tasks, data analysis reports and their corresponding modules, and workflow components corresponding to the data analysis report modules. Specifically, the descriptions and examples of workflow tasks help the large model understand the task flow, terminology, and operational steps within a specific domain; the lists of workflow components and their functions and usage instructions help the large model learn the purpose and parameter configuration of each workflow component, enabling it to rationally select and configure workflow components when optimizing workflows; the documentation and tutorials of the workflow platform help the large model deeply understand the overall architecture, design philosophy, and usage methods of the workflow platform, adhere to standards, and improve stability; the execution records and logs of workflow tasks provide detailed records of the execution status of workflow tasks (including success, failure, exceptions, etc.), which helps in diagnosing and optimizing existing problems, improving the efficiency of the data analysis workflow. Performance; Data analysis reports and corresponding modules refer to the multiple modules included in a data analysis report, such as data acquisition, data preprocessing, data statistics, data analysis, and conclusion output. The workflow components corresponding to the data analysis report module are: data acquisition module (e.g., data dictionary query component, database interface component, etc.), data preprocessing module (e.g., missing value handling component, outlier handling component, standardization component, normalization component, etc.), data statistics module (e.g., summation component, average component, grouping component, comparison component, etc.), and data analysis module (e.g., indicator decomposition analysis component, indicator correlation analysis component, attribution analysis component, strategy correlation component, etc.).

[0032] Step 102: Based on the relevant knowledge data of the sample data analysis task, the sample subtask list, and the graph neural network, determine the priority of each sample subtask in the sample subtask list.

[0033] In this embodiment, dependency analysis is first performed based on relevant knowledge data and a list of sample subtasks for the sample data analysis task. Then, a directed acyclic graph (DAG) is constructed using a graph neural network to clarify the dependencies and execution order among the sample subtasks. After the DAG is constructed, the graph neural network is trained. The trained graph neural network transmits the dependency information in the form of a data structure to a priority allocator. The priority allocator dynamically assigns priorities to each sample subtask based on multiple dimensions such as urgency, resource requirements, and impact on results, ensuring that critical sample subtasks are executed first.

[0034] It should be noted that the DAG is constructed using a graph neural network. This involves using graph theory algorithms to define a node for each data analysis subtask and drawing directed edges in the order of task completion, thus constructing a DAG. Based on the constructed DAG, the various sample subtasks are arranged sequentially, clarifying the dependencies and execution order between them.

[0035] Step 103: Determine the sample data analysis workflow corresponding to the sample data analysis task based on the priority of each sample subtask and the sample subtask list.

[0036] In this embodiment, the sample subtasks are sorted according to their priority to obtain a sorting result. Based on the sorted sample subtasks, corresponding workflow components and their input / output requirements descriptions (prompts, such as functions and usage instructions) are generated. These workflow components include modules for data acquisition, data preprocessing, data statistics, and data analysis. Each module encapsulates its own workflow components and specific functions and algorithms, and the modules are linked through standardized interfaces. Finally, the workflow components corresponding to each sample subtask are assembled to obtain the sample data analysis workflow.

[0037] For the refined and sorted sample subtasks mentioned above, one or more workflow components are designed for each sample subtask. For example, for the "data cleaning" subtask, a "missing value handling" workflow component c1 and an "outlier detection" workflow component c2 need to be designed. Based on the designed workflow components, clear descriptions of input and output requirements are defined. For example, for the "missing value handling" workflow component c1, the input needs to be defined as the original dataset D, and the output as the processed dataset D' without missing values. The system will consider the dependencies between sample subtasks in the design of workflow components. For example, if "data modeling" depends on "data cleaning," then it will ensure that the output of the "data cleaning" workflow component can be used as the input of the "data modeling" workflow component. Finally, the workflow components corresponding to each sample subtask are assembled to obtain the sample data analysis workflow.

[0038] Step 104: Based on the sample data analysis task, the sample sub-task list, the priority of each sample sub-task, and the sample data analysis workflow, train the initial large model to obtain the trained large model used for generating the data analysis workflow.

[0039] In this embodiment of the disclosure, in order to achieve automated generation of data analysis workflow, the sample stage data of the sample data analysis task at each analysis stage can be determined according to the sample data analysis task, the sample sub-task list, the priority of each sample sub-task, and the sample data analysis workflow; each analysis stage is a stage in the data analysis workflow generation process; based on the sample stage data of the sample data analysis task at each analysis stage, the initial large model is trained to obtain the trained large model used for data analysis workflow generation.

[0040] To illustrate the process of training the initial large model, as an example, for each analysis stage, the sample data analysis task's data to be processed in the analysis stage is obtained sequentially; the data to be processed, relevant knowledge data, and stage prompt text of the analysis stage are input into the initial large model to obtain the prediction stage data output by the initial large model; based on the sample stage data of the analysis stage and the prediction stage data, the parameters of the initial large model are adjusted to obtain the trained large model.

[0041] The data to be processed in the analysis phase includes: sample decomposition phase, where the data to be processed is the sample data analysis task; task ranking phase, where the data to be processed is the sample sub-task list; and task component phase, where the data to be processed is the sample sub-task ranking result.

[0042] It should be noted that the analysis phase includes at least one of the following: task decomposition phase, task sorting phase, and task component phase; the sample phase data of the task decomposition phase includes a list of sample subtasks; the sample phase data of the task sorting phase includes the priority of each sample subtask; and the sample phase data of the task component phase includes the sample data analysis workflow.

[0043] Furthermore, in this embodiment of the disclosure, after obtaining the trained large model for generating the data analysis workflow, the method further includes: acquiring the target data analysis task; sequentially determining the target stage data of the historical analysis stages preceding the analysis stage to be processed for each analysis stage in each analysis stage; inputting the target stage data and the stage prompt text corresponding to the analysis stage to be processed into the large model to acquire the target stage data of the analysis stage to be processed output by the large model; and determining the target data analysis workflow corresponding to the target data analysis task based on the target stage data of the last analysis stage in each analysis stage.

[0044] In this embodiment of the disclosure, a workflow experience base can also be obtained; the workflow experience base includes experience texts in the data analysis workflow generation process; based on the experience texts in the workflow experience base, the target data analysis workflow is verified and adjusted.

[0045] As an example, by using a workflow experience base and an embedding model, the experience of sample subtask decomposition, sample subtask execution scheme, and common mistake reminders are vectorized to rigorously verify the rationality of the sample data analysis workflow arrangement. The vectors in the experience base are then used to calculate the matching degree with the current target data analysis task request using cosine similarity, thus optimizing their prompts.

[0046] If an anomaly occurs during the execution of the target data analysis task, the anomaly information and current status are reported to the error handling unit. The error handling unit then passes the anomaly handling results and suggestions to the dynamic adjuster, which optimizes the execution strategy accordingly. The dynamic adjuster synchronizes the latest execution status and optimization results of the target data analysis task to the user interaction module, allowing users to understand and further operate the target data analysis workflow.

[0047] In this embodiment of the disclosure, target stage data of at least one analysis stage may also be displayed to obtain feedback from the object; first target stage data related to the feedback may be determined; and the first target stage data may be adjusted according to the feedback.

[0048] As an example, an interactive assistant interface can be used to demonstrate the comprehensive execution of the target data analysis task. At this stage, a structured and optimized workflow is ready, and users can monitor and precisely intervene in the entire target stage data analysis process in real time through the assistant interface. Then, based on feedback, users can determine whether to adjust the parameters of the first target stage data, rearrange the execution order, or handle unexpected situations.

[0049] In the large model training method of this disclosure embodiment, a sample data analysis task and a list of sample subtasks corresponding to the sample data analysis task are obtained; the priority of each sample subtask in the sample subtask list is determined based on the relevant knowledge data of the sample data analysis task, the sample subtask list, and the graph neural network; the sample data analysis workflow is determined based on the priority of each sample subtask and the sample subtask list; and the trained large model for generating data analysis workflow is obtained based on the sample data analysis task, the sample subtask list, the priority of each sample subtask, and the sample data analysis workflow. The large model trained by the above method can automatically adjust the workflow orchestration logic according to different sample data analysis tasks, thereby adapting to diverse business needs and further improving the accuracy and efficiency of workflow generation.

[0050] To illustrate how the sample data analysis task and the list of sample subtasks corresponding to the sample data analysis task are obtained, Figure 2This is a flowchart of a large model training method according to an embodiment of the present disclosure. It should be noted that the large model training method of this embodiment can be applied to a large model training device, which can be configured in an electronic device to enable the electronic device to perform large model training functions.

[0051] Among them, electronic devices can be any device with computing capabilities, such as terminal devices, servers, etc.

[0052] like Figure 2 As shown, the method includes the following steps: Step 201: Perform intent recognition processing on the sample data analysis task to obtain intent recognition results.

[0053] In this embodiment of the disclosure, a graphical user interface (GUI) is provided for users to describe and query sample data analysis tasks. After receiving the sample data analysis task input by the user, the GUI passes the sample data analysis task to the intent recognition module. Then, based on the received sample data analysis task, the intent recognition module uses part-of-speech tagging, named entity recognition, and dependency parsing to identify keywords and the relationships between keywords in the sentence. Finally, a model (e.g., BERT-Large fine-tuned) is used to train and classify the processed intent (i.e., the relationships between keywords) in order to accurately identify the user's intent and obtain accurate intent recognition results.

[0054] Sequence labeling models (e.g., BERT-based) can be used for part-of-speech tagging and named entity recognition. ; It is a conditional probability distribution, representing the probability of output y given an input x. The softmax function is an activation function that maps a vector to the range (0,1). These are the parameters of the pre-trained BERT model; a Transformer-based dependency parser can be applied for dependency parsing, as follows:

[0055] , It is a scoring function. Transformer is a deep learning model architecture based on self-attention mechanism. `sentence` is the input text, `h` and `m` are different keywords in the text, and `c` is a normalization constant. `MLP` refers to the parameters of a multilayer perceptron used to calculate dependency scores. The principle of BERT-Large fine-tuned for intent classification is as follows: `intent` represents the final predicted intent recognition result, and `argmax` is a function that returns the value of the argument that maximizes the value of the subsequent function. It is a special marker in the input of the BERT model, located at the beginning of the input sequence, and W and b are the linear layer parameters used for fine-tuning.

[0056] Step 202: Based on the relevant knowledge data, determine a task template that matches the sample data analysis task; the task template includes multiple sub-task templates.

[0057] It should be noted that the data analysis reports in the relevant knowledge data and the predefined task templates in each module of the corresponding module contain a series of general sub-tasks.

[0058] In this embodiment of the disclosure, after obtaining the intent recognition result, a task template (including multiple task sub-templates) matching the sample data analysis task is obtained from the task template library based on the relevant knowledge data. Then, the variables of the task sub-templates are filled in based on the intent recognition result, and a specific sub-task list is generated.

[0059] As an example, cosine similarity can be used to match the most similar task templates, and then a pre-trained model (such as T5) can be used to optimize each sample subtask, generating a list of sample subtasks that both conform to a predefined structure and are adapted to specific situations. The parameters of the T5 model are obtained by pre-training the BERT model.

[0060] The principle of cosine similarity is as follows: Where A is the vector representation of the user input, and B is the vector representation of the template. and These are the Euclidean norms of vectors A and B, respectively. The principle of the T5 model is... ,in, These are query, key, and value matrices, respectively. It is the dimension of the key vector.

[0061] Step 203: Determine the sample subtask list based on the intent recognition result and the multiple subtask templates.

[0062] In this embodiment, a sample data analysis task is exemplified by the user input task of "analyzing the reasons for the decline in sales this quarter." This task is a question described in natural language and is input into the system in the form of a query request. After receiving the user input task, the GUI passes the task to the intent recognition module. Based on the received task, the intent recognition module uses part-of-speech tagging and dependency parsing to identify keywords in the sentence and the relationships between keywords. For the task "analyzing the reasons for the decline in sales this quarter," the model will identify "analyze" as the action, "decline in sales" as the object of analysis, and "this quarter" as the time range. The processed intent is trained and classified to accurately identify the user intent recognition result. Subsequently, it will determine whether the detected intent and semantics require multi-turn dialogue to supplement information, that is, to implement a rule-based system to handle the user's supplementary information request. If the recognition result intent is unclear or requires more information, the system will ask the user a specific question through the GUI, such as "Which type of product's sales do you need to analyze?". Finally, based on the intent recognition result, according to the relevant knowledge data, the system matches the corresponding task template from the task template library. For sales trend analysis, the system will match a template that includes steps such as data aggregation, trend calculation, and result visualization. The task template variables are populated with specific information (time range, objects of analysis, etc.) extracted from the user's task description, and a specific list of sample subtasks is generated.

[0063] Step 204: Based on the relevant knowledge data of the sample data analysis task, the sample subtask list, and the graph neural network, determine the priority of each sample subtask in the sample subtask list.

[0064] Step 205: Determine the sample data analysis workflow corresponding to the sample data analysis task based on the priority of each sample subtask and the sample subtask list.

[0065] Step 206: Based on the sample data analysis task, the sample sub-task list, the priority of each sample sub-task, and the sample data analysis workflow, train the initial large model to obtain the trained large model used for generating the data analysis workflow.

[0066] It should be noted that the execution process of steps 204 to 206 can be implemented in any of the embodiments of this disclosure. This disclosure does not limit this and will not elaborate further.

[0067] In the large model training method of this disclosure embodiment, the sample data analysis task is subjected to intent recognition processing to obtain intent recognition results; a task template matching the sample data analysis task is determined based on the relevant knowledge data; the task template includes multiple sub-task templates; a sample sub-task list is determined based on the intent recognition results and the multiple sub-task templates; the priority of each sample sub-task in the sample sub-task list is determined based on the relevant knowledge data of the sample data analysis task, the sample sub-task list, and the graph neural network; a sample data analysis workflow corresponding to the sample data analysis task is determined based on the priority of each sample sub-task and the sample sub-task list; and an initial large model is trained based on the sample data analysis task, the sample sub-task list, the priority of each sample sub-task, and the sample data analysis workflow to obtain a trained large model for generating the data analysis workflow. Specifically, through intent recognition processing, the system enables automatic parsing of complex sample data analysis tasks and generation of sample subtasks. It determines matching task templates based on relevant knowledge data and establishes a sample data analysis workflow based on the priority of each sample subtask and the sample subtask list, thus enabling the trained large model to handle different types of data analysis tasks.

[0068] To explain how the priority of each sample subtask in the sample subtask list is determined based on the relevant knowledge data of the sample data analysis task, the sample subtask list, and the graph neural network, Figure 3 This is a flowchart of a large model training method according to an embodiment of the present disclosure. It should be noted that the large model training method of this embodiment can be applied to a large model training device, which can be configured in an electronic device to enable the electronic device to perform large model training functions.

[0069] Among them, electronic devices can be any device with computing capabilities, such as terminal devices, servers, etc.

[0070] like Figure 3 As shown, the method includes the following steps: Step 301: Obtain the sample data analysis task and the list of sample subtasks corresponding to the sample data analysis task.

[0071] Step 302: Based on relevant knowledge data, determine the dependencies and weights between the various sample subtasks in the sample subtask list.

[0072] It should be noted that the dependencies and weights between the various sample subtasks can be determined by combining the sample data with the execution records and logs of the workflow tasks.

[0073] As an example, graph theory algorithms can be used to define a node for each sample subtask, determine the workflow components of that node, and the dependencies and weights between the sample subtasks in the sample subtask list can be implemented by describing the dependencies and weights between the workflow components.

[0074] For example, directed acyclic graphs (e.g., weighted directed acyclic graphs, WDAGs) can be used to represent the dependencies and weights between workflow components, such as... ,in, It is a vertex set, representing n workflow components, corresponding to the workflow component list; It is an edge set, representing the dependencies between workflow components; It is a weighting function that represents the strength or importance of a dependency, i.e., the weight.

[0075] Step 303: Generate a directed acyclic graph based on the dependencies and weights between the various sample subtasks.

[0076] In this disclosed embodiment, through the above , , It can quickly and accurately construct directed acyclic graphs.

[0077] Step 304: Train the initial graph neural network based on the directed acyclic graph to obtain the trained graph neural network.

[0078] In this embodiment of the disclosure, in order to capture the graph structure information of the data analysis workflow in a directed acyclic graph, a graph neural network (GNN) can be used for node embedding learning, for example, a graph attention network (GAT) can be used: , is used to compute the embedded representation of nodes. Wherein, It is the feature representation of node i of the workflow component at layer l; It is the set of neighbors of node i of the workflow component, which is defined based on the edge E in the directed acyclic graph G; It is the attention coefficient, representing the importance of node j of the neighboring workflow component to node i of the workflow component. It is a learnable weight matrix. It is a non-linear activation function (such as ReLU), and the attention coefficients... Calculated as follows: .

[0079] Where a is a learnable attention vector, and It represents the linear transformation of nodes i and j at layer l, || denotes the vector concatenation operation, and LeakyReLU is an activation function.

[0080] Graph Attention Network (GAT) utilizes the edge information E and weight information W in a directed acyclic graph G to train an initial graph neural network, resulting in a trained graph neural network.

[0081] Step 305: Determine the priority of each sample subtask based on the trained graph neural network.

[0082] In this embodiment of the disclosure, each sample subtask is calculated. Priority: ,in, The final embedded representation of node i is obtained through iterative calculation using the first two formulas. These are the various sample sub-tasks Additional feature vectors, and These are learnable parameters. Each sample subtask The overall loss function for priority consists of three parts: .

[0083] in, For the reconstruction loss of graph embedding,

[0084] , The sequence loss for topological sorting. , The mean squared error of the priority prediction. ; The balancing coefficient is used to adjust the weights of the loss from topology sorting and the mean squared error of priority prediction in the overall loss function. E represents the workflow components that actually have dependencies. These are the embedding vectors of vertex i and vertex j, respectively, and t represents the time step. Workflow components executed at time step t The next workflow component to be executed is The probability of.

[0085] Step 306: Determine the sample data analysis workflow corresponding to the sample data analysis task based on the priority of each sample subtask and the sample subtask list.

[0086] Step 307: Based on the sample data analysis task, the sample sub-task list, the priority of each sample sub-task, and the sample data analysis workflow, train the initial large model to obtain the trained large model used for generating the data analysis workflow.

[0087] It should be noted that the execution processes of steps 301, 306 and 307 can be implemented in any of the embodiments of this disclosure. This disclosure does not limit these processes and will not elaborate further.

[0088] In the large model training method of this disclosure embodiment, a sample data analysis task and a list of sample subtasks corresponding to the sample data analysis task are obtained; based on relevant knowledge data, the dependencies and weights between each sample subtask in the sample subtask list are determined; a directed acyclic graph (DAG) is generated based on the dependencies and weights between each sample subtask; an initial graph neural network is trained based on the DAG to obtain a trained graph neural network; the priority of each sample subtask is determined based on the trained graph neural network; the sample data analysis workflow corresponding to the sample data analysis task is determined based on the priority of each sample subtask and the sample subtask list; the initial large model is trained based on the sample data analysis task, the sample subtask list, the priority of each sample subtask, and the sample data analysis workflow to obtain a trained large model for generating the data analysis workflow; wherein, the graph neural network is able to learn the topological structure and dependency weight information between subtasks, enhancing the training effect of the large model, and the trained large model can adapt to different types of data analysis tasks and business scenarios.

[0089] In addition, to clearly explain the training method for large models, Figure 4This is a schematic diagram illustrating the principle of a large-scale model training method according to an embodiment of the present disclosure. Specifically, the large-scale model training method can be implemented through a task decomposition module, a process orchestration module, and a workflow component generation module. In response to a user input sample data analysis task, intent recognition processing is performed, and it is determined whether the detected intent and semantics require supplementary information through multiple rounds of dialogue. The intent recognition result is then sent to the task decomposition module. The task decomposition module then decomposes the sample sub-tasks into a process flow to obtain a list of sample sub-tasks. Subsequently, the process orchestration module determines the priority of each sample sub-task (e.g., sub-task A, sub-task B, ..., sub-task N), and sorts the sample sub-tasks according to their priorities. The process involves sorting the results (e.g., subtask 1, subtask 2, ..., subtask N), and then assembling the candidate workflow components (e.g., workflow component 1-1, workflow component 1-2, workflow component 1-3, workflow component 2-1, workflow component 2-2, workflow component 3-1, workflow component 3-2, workflow component 3-3, workflow component 3-4) corresponding to each subtask 1, subtask 2, ..., subtask N using the workflow component generation module. This assembly process yields the sample data analysis workflow. After the sample data analysis workflow is verified and adjusted, it is executed to process the target data analysis task, generating an analysis report corresponding to the target data analysis task. During the processing of the target data analysis task, the accuracy of the sample data analysis workflow is verified and iteratively optimized based on the target data analysis workflow. Furthermore, the sample data analysis workflow is also used to verify the rationality of the workflow component generation module and iteratively optimize it.

[0090] Figure 5 This is a schematic diagram of the structure of a training device for a large model according to an embodiment of the present disclosure.

[0091] like Figure 5 As shown, the training device for the large model includes: a first acquisition module 501, a first determination module 502, a second determination module 503, and a training processing module 504.

[0092] The system includes a first acquisition module 501, used to acquire a sample data analysis task and a list of sample subtasks corresponding to the sample data analysis task; a first determination module 502, used to determine the priority of each sample subtask in the sample subtask list based on the relevant knowledge data of the sample data analysis task, the list of sample subtasks, and the graph neural network; a second determination module 503, used to determine the sample data analysis workflow corresponding to the sample data analysis task based on the priority of each sample subtask and the list of sample subtasks; and a training processing module 504, used to train an initial large model based on the sample data analysis task, the list of sample subtasks, the priority of each sample subtask, and the sample data analysis workflow to obtain a trained large model for generating the data analysis workflow.

[0093] In one embodiment of this disclosure, the first acquisition module 501 is specifically configured to: perform intent recognition processing on the sample data analysis task to obtain intent recognition results; determine a task template matching the sample data analysis task based on the relevant knowledge data; the task template includes multiple sub-task templates; and determine the sample sub-task list based on the intent recognition results and the multiple sub-task templates.

[0094] In one embodiment of this disclosure, the first determining module 502 is specifically configured to: determine the dependencies and weights between each sample subtask in the sample subtask list based on the relevant knowledge data; generate a directed acyclic graph based on the dependencies and weights between each sample subtask; train an initial graph neural network based on the directed acyclic graph to obtain a trained graph neural network; and determine the priority of each sample subtask based on the trained graph neural network.

[0095] In one embodiment of this disclosure, the second determining module 503 is specifically configured to: sort the sample subtasks according to their priorities to obtain a sorting result; acquire a workflow component library; the workflow component library includes multiple workflow components and description information of the workflow components; the workflow components embed exception handling logic; sequentially, for each sample subtask in the sorting result, determine at least one candidate workflow component matching the sample subtask according to the description information of the multiple workflow components; assemble the candidate workflow components corresponding to each sample subtask to obtain the sample data analysis workflow.

[0096] In one embodiment of this disclosure, the training processing module 504 includes: a determining unit, configured to determine sample stage data of the sample data analysis task at each analysis stage based on the sample data analysis task, the sample sub-task list, the priority of each sample sub-task, and the sample data analysis workflow; wherein each analysis stage is a stage in the data analysis workflow generation process; and a training unit, configured to train the initial large model based on the sample stage data of the sample data analysis task at each analysis stage to obtain a trained large model for generating the data analysis workflow.

[0097] In one embodiment of this disclosure, the training unit is specifically configured to: sequentially acquire the data to be processed by the sample data analysis task in each of the analysis stages; input the data to be processed, the relevant knowledge data, and the stage prompt text of the analysis stage into the initial large model to acquire the prediction stage data output by the initial large model; and perform parameter adjustment processing on the initial large model based on the sample stage data of the analysis stage and the prediction stage data to obtain the trained large model.

[0098] In one embodiment of this disclosure, the analysis phase includes at least one of the following: a task decomposition phase, a task sorting phase, and a task component phase; the sample phase data of the task decomposition phase includes the sample subtask list; the sample phase data of the task sorting phase includes the priority of each sample subtask; and the sample phase data of the task component phase includes the sample data analysis workflow.

[0099] In one embodiment of this disclosure, the apparatus further includes: a second acquisition module, configured to acquire a target data analysis task; a third determination module, configured to sequentially determine, for each analysis stage to be processed in each of the analysis stages, the target stage data of the target data analysis task in the historical analysis stages prior to the analysis stage to be processed; the third acquisition module, configured to input the target stage data and the stage prompt text corresponding to the analysis stage to be processed into the large model, and acquire the target stage data of the analysis stage to be processed output by the large model; and a fourth determination module, configured to determine the target data analysis workflow corresponding to the target data analysis task based on the target stage data of the last analysis stage in each of the analysis stages.

[0100] In one embodiment of this disclosure, the apparatus further includes: a fourth acquisition module, configured to acquire a workflow experience library; the workflow experience library includes experience texts from the data analysis workflow generation process; and a first processing module, configured to perform verification processing and adjustment processing on the target data analysis workflow based on the experience texts in the workflow experience library.

[0101] In one embodiment of this disclosure, the apparatus further includes: a second processing module for displaying target stage data of at least one analysis stage to obtain feedback from the object; a fifth processing module for determining first target stage data related to the feedback; and a third processing module for adjusting the first target stage data according to the feedback.

[0102] In the large model training apparatus of this embodiment, a sample data analysis task and a list of sample subtasks corresponding to the sample data analysis task are acquired; the priority of each sample subtask in the sample subtask list is determined based on the relevant knowledge data of the sample data analysis task, the sample subtask list, and the graph neural network; the sample data analysis workflow is determined based on the priority of each sample subtask and the sample subtask list; and the trained large model for generating data analysis workflows is obtained based on the sample data analysis task, the sample subtask list, the priority of each sample subtask, and the sample data analysis workflow. The large model trained by the above method can automatically adjust the workflow orchestration logic according to different sample data analysis tasks, thereby adapting to diverse business needs and further improving the accuracy and efficiency of workflow generation.

[0103] According to a third aspect of the present disclosure, an electronic device is also provided, comprising: a processor; and a memory for storing processor-executable instructions, wherein the processor is configured to: implement the training method for a large model as described above.

[0104] To implement the above embodiments, this disclosure also proposes a storage medium.

[0105] When the instructions in the storage medium are executed by the processor, the processor is able to execute the training method for the large model as described above.

[0106] To implement the above embodiments, this disclosure also provides a computer program product.

[0107] When the computer program product is executed by the processor of the electronic device, it enables the electronic device to perform the above-described method.

[0108] Figure 6 This is a structural block diagram of an electronic device according to an exemplary embodiment. Figure 6 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments disclosed herein.

[0109] like Figure 6As shown, the electronic device 600 includes a processor 611, which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 612 or a program loaded from memory 616 into random access memory (RAM) 613. The RAM 613 also stores various programs and data required for the operation of the electronic device 600. The processor 611, ROM 612, and RAM 613 are interconnected via a bus 614. An input / output (I / O) interface 615 is also connected to the bus 614.

[0110] The following components are connected to I / O interface 615: memory 616 including hard disk; and communication section 617 including network interface card such as local area network (LAN) card, modem, etc., communication section 617 performs communication processing via a network such as the Internet; drive 618 is also connected to I / O interface 615 as needed.

[0111] In particular, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 617. When the computer program is executed by processor 611, it performs the functions defined in the methods of this disclosure.

[0112] In an exemplary embodiment, a storage medium including instructions is also provided, such as a memory including instructions, which can be executed by the processor 611 of the electronic device 600 to perform the above-described method. Optionally, the storage medium may be a non-transitory computer-readable storage medium, such as a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical data storage device.

[0113] In this disclosure, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this disclosure, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium, which can transmit, propagate, or transfer a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wireless, wireline, optical fiber, RF, etc., or any suitable combination thereof.

[0114] Furthermore, the term “exemplary” is used herein to mean serving as an example, instance, or illustration. Any aspect or design described herein as “exemplary” is not necessarily to be construed as advantageous compared to other aspects or designs. Rather, the use of the term “exemplary” is intended to present the concept in a concrete manner. As used herein, the term “or” is intended to mean an inclusive “or” rather than an exclusive “or.” That is, unless otherwise specified or clear from the context, “X applies A or B” is intended to mean any of the natural inclusive arrangements. That is, “X applies A or B” satisfies any of the foregoing instances if X applies A; X applies B; or both X applies A and B. Additionally, unless otherwise specified or clear from the context to refer to the singular form, the articles “a” and “an” as used in this application and the appended claims are generally understood to mean “one or more.”

[0115] Similarly, although this disclosure has been shown and described with respect to one or more implementations, equivalent variations and modifications will occur to those skilled in the art upon reading and understanding this specification and the accompanying drawings. This disclosure includes all such modifications and variations and is limited only by the scope of the claims. In particular, with respect to the various functions performed by the components described above (e.g., elements, resources, etc.), unless otherwise indicated, the terminology used to describe such components is intended to correspond to any component (functionally equivalent) that performs the specific function of the described component, even if structurally not equivalent to the disclosed structure. Furthermore, although specific features of this disclosure may have been disclosed with respect to only one of several implementations, such features may be combined with one or more other features of other implementations, as may be desired and advantageous to any given or particular application. Moreover, with regard to the terms “comprising,” “owning,” “having,” “having,” or variations thereof as used in the detailed description or claims, such terms are intended to be inclusive in a manner similar to the term “including.”

[0116] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the following claims.

[0117] It should be understood that this disclosure is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this disclosure is limited only by the appended claims.

Claims

1. A training method for a large model, characterized in that, The method includes: Obtain the sample data analysis task and the corresponding sample subtask list; the sample data analysis task is a question described in natural language and is input in the form of a query request. Based on the relevant knowledge data of the sample data analysis task, the sample subtask list, and the graph neural network, the priority of each sample subtask in the sample subtask list is determined; the relevant knowledge data includes data from multiple dimensions; the dimensions include at least one of the following: description and examples of workflow tasks corresponding to data analysis, a list of workflow components and their functions and usage instructions, documentation and tutorials of the workflow platform, execution records and logs of workflow tasks, data analysis reports and corresponding modules, and workflow components corresponding to the data analysis report module; Based on the priority of each sample subtask and the sample subtask list, the sample data analysis workflow corresponding to the sample data analysis task is determined; the sample data analysis workflow includes candidate workflow components corresponding to each sample subtask in the sample subtask list; Based on the sample data analysis task, the sample sub-task list, the priority of each sample sub-task, and the sample data analysis workflow, the initial large model is trained to obtain the trained large model used for generating the data analysis workflow. The step of determining the priority of each sample subtask in the sample subtask list based on the relevant knowledge data of the sample data analysis task, the sample subtask list, and the graph neural network includes: Based on the relevant knowledge data, determine the dependencies and weights among the sample subtasks in the sample subtask list; Based on the dependencies and weights among the various sample subtasks, a directed acyclic graph is generated; The initial graph neural network is trained based on the directed acyclic graph to obtain the trained graph neural network. The priority of each sample subtask is determined based on the trained graph neural network. The process of training the initial large model based on the sample data analysis task, the sample sub-task list, the priority of each sample sub-task, and the sample data analysis workflow to obtain a trained large model for generating the data analysis workflow includes: Based on the sample data analysis task, the sample subtask list, the priority of each sample subtask, and the sample data analysis workflow, the sample stage data of the sample data analysis task at each analysis stage is determined; each analysis stage is a stage in the data analysis workflow generation process. Based on the sample data of the sample data analysis task at each analysis stage, the initial large model is trained to obtain the trained large model used for generating the data analysis workflow.

2. The method according to claim 1, characterized in that, The task of acquiring and analyzing sample data, and the list of sample subtasks corresponding to the task of acquiring and analyzing sample data, include: The sample data analysis task is subjected to intent recognition processing to obtain intent recognition results; Based on the relevant knowledge data, a task template matching the sample data analysis task is determined; the task template includes multiple sub-task templates. Based on the intent recognition results and the multiple subtask templates, the sample subtask list is determined.

3. The method according to claim 1, characterized in that, The step of determining the sample data analysis workflow corresponding to the sample data analysis task based on the priority of each sample sub-task and the sample sub-task list includes: Based on the priority of each sample subtask, the sample subtasks are sorted to obtain the sorting result; Obtain the workflow component library; the workflow component library includes multiple workflow components and description information of the workflow components; the workflow components embed exception handling logic; For each sample subtask in the sorting results, at least one candidate workflow component matching the sample subtask is determined based on the description information of the plurality of workflow components. The candidate workflow components corresponding to each of the sample sub-tasks are assembled to obtain the sample data analysis workflow.

4. The method according to claim 1, characterized in that, The step of training the initial large model based on sample stage data from each analysis stage of the sample data analysis task to obtain a trained large model for generating the data analysis workflow includes: For each of the analysis stages, the data to be processed by the sample data analysis task in that analysis stage is obtained sequentially. Input the data to be processed, the relevant knowledge data, and the stage prompt text of the analysis stage into the initial large model to obtain the prediction stage data output by the initial large model; Based on the sample stage data from the analysis phase and the prediction stage data, the parameters of the initial large model are adjusted to obtain the trained large model.

5. The method according to claim 1, characterized in that, The analysis phase includes at least one of the following: task decomposition phase, task sorting phase, and task component phase; The sample phase data of the task decomposition phase includes the sample subtask list; The sample phase data of the task sorting phase includes the priority of each sample subtask; The sample phase data of the task component phase includes the sample data analysis workflow.

6. The method according to claim 1, characterized in that, The method further includes: Acquire target data for analysis tasks; For each analysis stage to be processed in each of the analysis stages, the target stage data of the target data analysis task in the historical analysis stages preceding the analysis stage to be processed is determined. Input the target stage data and the stage prompt text corresponding to the analysis stage to be processed into the large model, and obtain the target stage data of the analysis stage to be processed output by the large model; Based on the target stage data of the last analysis stage in each of the analysis stages, determine the target data analysis workflow corresponding to the target data analysis task.

7. The method according to claim 6, characterized in that, The method further includes: Acquire a workflow experience library; the workflow experience library includes experience texts from the data analysis workflow generation process; Based on the experience texts in the workflow experience library, the target data analysis workflow is validated and adjusted.

8. The method according to claim 6, characterized in that, The method further includes: Display and process the target stage data from at least one analysis phase to obtain feedback from the target audience. Determine the first target phase data related to the feedback; The data for the first target stage is adjusted based on the feedback received.

9. A training device for a large model, characterized in that, The device includes: The first acquisition module is used to acquire sample data analysis tasks and a list of sample subtasks corresponding to the sample data analysis tasks; the sample data analysis tasks are questions described in natural language and are input in the form of query requests. The first determining module is used to determine the priority of each sample subtask in the sample subtask list based on the relevant knowledge data of the sample data analysis task, the sample subtask list, and the graph neural network; the relevant knowledge data includes data in multiple dimensions; the dimensions include at least one of the following: description and examples of workflow tasks corresponding to data analysis, list of workflow components and their functions and usage instructions, documentation and tutorials of the workflow platform, execution records and logs of workflow tasks, data analysis reports and corresponding modules, and workflow components corresponding to the data analysis report module; The second determining module is used to determine the sample data analysis workflow corresponding to the sample data analysis task based on the priority of each sample subtask and the sample subtask list; the sample data analysis workflow includes candidate workflow components corresponding to each sample subtask in the sample subtask list. The training processing module is used to train the initial large model according to the sample data analysis task, the sample sub-task list, the priority of each sample sub-task, and the sample data analysis workflow, so as to obtain the trained large model for generation by the data analysis workflow. The first determining module is specifically used to: determine the dependencies and weights between each sample subtask in the sample subtask list based on the relevant knowledge data; generate a directed acyclic graph based on the dependencies and weights between each sample subtask; train an initial graph neural network based on the directed acyclic graph to obtain a trained graph neural network; and determine the priority of each sample subtask based on the trained graph neural network. The training processing module is specifically used to determine the sample stage data of the sample data analysis task at each analysis stage based on the sample data analysis task, the sample sub-task list, the priority of each sample sub-task, and the sample data analysis workflow; each analysis stage is a stage in the data analysis workflow generation process; and to train the initial large model based on the sample stage data of the sample data analysis task at each analysis stage to obtain the trained large model for generating the data analysis workflow.

10. An electronic device, characterized in that, include: processor; Memory used to store the processor's executable instructions; The processor is configured as follows: The steps of implementing the training method for the large model as described in any one of claims 1 to 8.

11. A non-transitory computer-readable storage medium, wherein instructions in the storage medium, when executed by a processor, enable the processor to perform a training method for a large model as described in any one of claims 1 to 8.

Images