Task scheduling method and device, electronic equipment and storage medium

Abstract

The invention provides a task scheduling method and device, electronic equipment and a storage medium, and the method comprises the steps: carrying out the distillation of a target large model based on a preset loss function, and obtaining a target student model; performing step-by-step fine adjustment on the target student model by using the special data set to obtain a task scheduling model; based on the chain type task processing architecture, calling a task scheduling model to analyze and manage a target task created by a user, and based on an analysis result of the target task, generating a plurality of sub-tasks; task scheduling information corresponding to the target task is acquired, and the task scheduling model performs execution scheduling on the plurality of sub-tasks according to the task scheduling information to generate a task execution plan; and according to the task execution plan and the employee feature data, task scheduling is performed on the plurality of sub-tasks so as to allocate the plurality of sub-tasks to the target employees.

Description

Technical Field

[0001] The present disclosure relates to the technical field of task scheduling, and in particular to a task scheduling method and device, electronic device, and storage medium. Background Art

[0002] In today's digital age, artificial intelligence (AI) technology is developing rapidly. The emergence of large models has brought about a qualitative leap in natural language understanding, creating new development opportunities for businesses and individuals and significantly improving production efficiency. Against this backdrop, traditional enterprise task scheduling models are facing severe challenges.

[0003] Traditional enterprise task scheduling, encompassing diverse tasks such as resource adjustment and policy execution, has long relied on manual operations. This approach is not only inefficient and unable to meet rapidly changing business needs, but also suffers from significant deficiencies in overall scheduling accuracy and timeliness, severely hindering enterprise development and making effective IT governance difficult to achieve.

[0004] Natural language processing and natural language generation, as key areas of natural language understanding, have some potential for assisting in scheduling task generation. However, general machine learning technologies, relying on crude probabilistic data processing, are insufficient for critical tasks such as scheduling task generation and allocation. Using large models for these tasks, while theoretically more effective, comes with the high cost of implementing them. Training and inference of large models require significant computing power and data support, a burden that is prohibitive for most enterprises. Even using APIs for leading large models cannot meet the precise scheduling needs of enterprises. Existing open large models are primarily designed for general-purpose scenarios and have numerous limitations when handling enterprise-specific, precision tasks, making them inapplicable. Therefore, leveraging the advantages of AI and overcoming the shortcomings of existing technologies to achieve efficient, precise, and intelligent task scheduling is a pressing issue. Summary of the Invention

[0005] This disclosure provides a task scheduling method and apparatus, electronic device, and storage medium. Its primary purpose is to address the high training and inference costs of large language models, insufficient domain output accuracy, and the lack of automatic execution capabilities for multi-step instructions in enterprise scheduling scenarios.

[0006] According to a first aspect of the present disclosure, a task scheduling method is provided, comprising:

[0007] Based on the preset loss function, the target large model is distilled to obtain the target student model;

[0008] The target student model is gradually fine-tuned using a dedicated dataset to obtain a task scheduling model;

[0009] Based on the chain task processing architecture, the task scheduling model is called to parse and manage the target tasks created by the user, and multiple subtasks are generated based on the parsed results of the target tasks;

[0010] Obtain the task scheduling information corresponding to the target task. The task scheduling model schedules the execution of multiple subtasks based on the task scheduling information and generates a task execution plan.

[0011] According to the task execution plan and employee characteristic data, multiple subtasks are scheduled to assign the multiple subtasks to target employees.

[0012] Optionally, based on a preset loss function, the target large model is distilled to obtain a target student model, including:

[0013] Perform distillation training on the target large model and the student model using the target dataset;

[0014] The preset loss function is used to calculate the output distribution difference between the target large model and the student model. According to the output distribution difference, it is determined whether the distillation training of the student model is completed. The preset loss function is constructed based on the adaptive weight coefficient.

[0015] Optionally, before fine-tuning the target student model layer by layer using a dedicated dataset to obtain a task scheduling model, the method further includes:

[0016] The target large model is called, and directional domain data is generated for the target dataset according to the preset prompt template to obtain a dedicated dataset.

[0017] Optionally, the target student model is gradually fine-tuned using a dedicated dataset to obtain a task scheduling model, including:

[0018] Starting from the output layer of the target student model, gradually unfreeze the model parameters of the target student model according to the preset number of layers;

[0019] After performing an unfreezing process, the dedicated data set is input into the unfrozen target student model for model fine-tuning until all model parameters are unfrozen to obtain a task scheduling model.

[0020] Optionally, input a dedicated dataset into the unfrozen target student model for model fine-tuning, including:

[0021] Calculate the target learning rate for each model layer in the target student model based on the number of target student model layers, the decay factor, and the base learning rate;

[0022] According to the target learning rate of each model layer, the dedicated dataset is input into the unfrozen target student model for model fine-tuning.

[0023] Optionally, before scheduling the multiple subtasks according to the task execution plan and the employee characteristic data to assign the multiple subtasks to the target employees, the method further includes:

[0024] Based on the task contents of multiple subtasks, execution guidance information corresponding to each subtask is generated.

[0025] According to a second aspect of the present disclosure, a task scheduling apparatus is provided, comprising:

[0026] The distillation unit is used to distill the target large model based on a preset loss function to obtain a target student model;

[0027] A fine-tuning unit, used to gradually fine-tune the target student model using a dedicated dataset to obtain a task scheduling model;

[0028] The creation unit is used to call the task scheduling model to parse and manage the target task created by the user based on the chain task processing architecture, and generate multiple subtasks based on the parsing results of the target task;

[0029] The scheduling unit is used to obtain the task scheduling information corresponding to the target task. The task scheduling model schedules the execution of multiple subtasks based on the task scheduling information and generates a task execution plan;

[0030] The allocation unit is used to schedule multiple subtasks according to the task execution plan and employee characteristic data, so as to allocate the multiple subtasks to target employees.

[0031] Optionally, the distillation unit comprises:

[0032] The training module is used to perform distillation training on the target large model and the student model using the target dataset;

[0033] A determination module is used to calculate the output distribution difference between the target large model and the student model using a preset loss function, and determine whether the distillation training of the student model is completed based on the output distribution difference, wherein the preset loss function is constructed based on an adaptive weight coefficient.

[0034] Optionally, the device further comprises:

[0035] The first generation unit is used to call the target large model and generate directional domain data for the target dataset according to a preset prompt template before fine-tuning the target student model layer by layer using the dedicated dataset to obtain the task scheduling model to obtain the dedicated dataset.

[0036] Optionally, the fine-tuning unit includes:

[0037] An unfreezing module is used to gradually unfreeze the model parameters of the target student model from the output layer of the target student model according to a preset number of layers;

[0038] The fine-tuning module is used to input the dedicated data set into the target student model after the unfreezing process after performing one unfreezing process to perform model fine-tuning until all model parameters are unfrozen to obtain the task scheduling model.

[0039] Optionally, the fine-tuning module is also used to:

[0040] Calculate the target learning rate for each model layer in the target student model based on the number of target student model layers, the decay factor, and the base learning rate;

[0041] According to the target learning rate of each model layer, the dedicated dataset is input into the unfrozen target student model for model fine-tuning.

[0042] Optionally, the device further comprises:

[0043] The second generating unit is used to generate execution guidance information corresponding to each subtask based on the task content of the multiple subtasks before scheduling the multiple subtasks according to the task execution plan and employee characteristic data to assign the multiple subtasks to target employees.

[0044] According to a third aspect of the present disclosure, there is provided an electronic device, including:

[0045] at least one processor; and

[0046] a memory communicatively connected to the at least one processor; wherein,

[0047] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the task scheduling method described in the first aspect.

[0048] According to a fourth aspect of the present disclosure, a non-transitory computer-readable storage medium storing computer instructions is provided, wherein the computer instructions are used to enable the computer to execute the task scheduling method described in the first aspect.

[0049] According to a fifth aspect of the present disclosure, a computer program product is provided, comprising a computer program, wherein when the computer program is executed by a processor, the computer program implements the task scheduling method as described in the first aspect above.

[0050] The present disclosure provides a task scheduling method and apparatus, electronic device, and storage medium, relating to the field of task scheduling technology. The present disclosure distills a target large model based on a preset loss function to obtain a target student model, significantly compressing the model size and reducing the demand for resources such as computing power, enabling the model to operate in resource-limited environments such as CPU terminals. This can reduce the cost of using large models for enterprises and expand the model's application scope in enterprise task scheduling scenarios. A task scheduling model is obtained by gradually fine-tuning the target student model using a dedicated dataset. The dedicated dataset is generated based on a carefully designed prompt template and contains multiple key elements, providing rich and targeted data for specific business scenarios. This enables the task scheduling model to have higher accuracy and greater adaptability when handling complex and diverse enterprise tasks, reduces deviations in model output, and improves the quality of task scheduling. Based on a chained task processing architecture, the task scheduling model is invoked to parse and manage user-created target tasks and generate multiple subtasks. Task scheduling information is then retrieved for execution scheduling and generation of a task execution plan. This process automates and streamlines task processing, significantly improving task processing efficiency, reducing manual intervention, lowering the probability of error, and making task scheduling more scientific and reasonable. Based on the task execution plan and employee characteristic data (such as historical performance, skills, and workload), multiple subtasks are scheduled and assigned to target employees. This allocation method fully considers employees' actual capabilities and work status, effectively matching people to positions, improving employee efficiency, and avoiding problems such as backlogs or excessive employee burdens caused by irrational task allocation, thereby improving the operational efficiency and management level of the entire enterprise.

[0051] It should be understood that the contents described in this section are not intended to identify the key or important features of the embodiments of the present disclosure, nor are they intended to limit the scope of the present disclosure. Other features of the present disclosure will become readily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] The accompanying drawings are provided to facilitate a better understanding of the present invention and do not constitute a limitation of the present disclosure.

[0053] Figure 1 A flowchart of a task scheduling method provided by an embodiment of the present disclosure;

[0054] Figure 2 A flowchart of another task scheduling method provided by an embodiment of the present disclosure;

[0055] Figure 3 A schematic diagram of the structure of a task scheduling device provided in an embodiment of the present disclosure;

[0056] Figure 4 A schematic diagram of the structure of another task scheduling device provided by an embodiment of the present disclosure;

[0057] Figure 5 A schematic block diagram of an exemplary electronic device provided for an embodiment of the present disclosure. DETAILED DESCRIPTION

[0058] The following description of exemplary embodiments of the present disclosure is made in conjunction with the accompanying drawings, including various details of the embodiments of the present disclosure to facilitate understanding. These details should be considered as merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications may be made to the embodiments described herein without departing from the scope and spirit of the present disclosure. Similarly, for the sake of clarity and conciseness, descriptions of well-known functions and structures are omitted in the following description.

[0059] The following describes the task scheduling method and apparatus, electronic device, and storage medium according to embodiments of the present disclosure with reference to the accompanying drawings.

[0060] Figure 1 A flowchart of a task scheduling method provided by an embodiment of the present disclosure.

[0061] like Figure 1 As shown, the method comprises the following steps:

[0062] Step 101: Based on a preset loss function, the target large model is distilled to obtain a target student model.

[0063] In the embodiments of the present disclosure, model distillation aims to effectively transfer the knowledge contained in a large, complex and high-performing target model to a relatively small model, so as to achieve model compression and improve the inference speed. The preset loss function is a key factor in guiding the model distillation process and ensuring the accurate transfer of knowledge. The target student model obtained by model distillation has fewer parameters and lower computational complexity while retaining key knowledge, and can adapt to resource-constrained operating environments. For example, in enterprise task scheduling scenarios, it can run efficiently on ordinary CPU terminals. The preset loss function is used to measure the difference between the target large model (teacher model) and the target student model to be generated. The design of the preset loss function should be based on a deep understanding of the model output characteristics and a precise grasp of the task requirements.

[0064] After completing the definition of the loss function and setting the relevant parameters, an open source dataset with rich data samples, such as the IMDB dataset, is selected to train the target student model. When processing the original dataset, it is converted into a tensor form that the model can handle through a mapping method and input into the target student model for training. During the training process, based on the feedback from the loss function, the parameters of the target student model are continuously adjusted so that it gradually learns the knowledge and characteristics of the target large model. After multiple rounds of training, the target student model is finally obtained. While retaining the core capabilities of the target large model, this target student model reduces the number of layers and parameters of the model, compresses the model size, and has the ability to run in a small resource environment.

[0065] Step 102: Use the dedicated data set to gradually fine-tune the target student model to obtain a task scheduling model.

[0066] In the embodiments of the present disclosure, the target student model obtained through model distillation has been effectively compressed in terms of model size and computational complexity, and can run in a limited resource environment. However, in order to more accurately meet the business needs of specific fields (such as enterprise task scheduling), targeted fine-tuning is still required. This is where a dedicated dataset is introduced. This dataset is generated based on a specific prompt template, which can include, for example, but not limited to, five key elements: role, instruction, format, task, and input.

[0067] At the beginning of fine-tuning, comprehensively adjusting all parameters can cause the model to forget previously learned useful information, resulting in "catastrophic forgetting." To address this issue, layer-by-layer fine-tuning can be employed, but is not limited to. During layer-by-layer fine-tuning, fine-tuning can be performed by freezing each layer. By fine-tuning only parameters near the output layer, the model can quickly adapt to the data characteristics of the new task while retaining existing knowledge, initially adjusting the model output to match the task requirements in the dedicated dataset. As training progresses, the model can gradually unfreeze more layers.

[0068] Throughout the gradual fine-tuning process, the target student model is continuously trained using the dedicated dataset as input. This process is an iterative optimization process, with the model continuously adjusting its parameters based on feedback from each training session. After multiple rounds of training and adjustment, the target student model gradually adapts to the characteristics and requirements of the specific task domain represented by the dedicated dataset, ultimately resulting in a task scheduling model. This task scheduling model can more accurately handle enterprise task scheduling tasks and provide reliable support for subsequent task parsing, scheduling, and allocation.

[0069] Step 103 : Based on the chain task processing architecture, the task scheduling model is called to parse and manage the target task created by the user, and multiple subtasks are generated based on the parsing result of the target task.

[0070] In the embodiments of the present disclosure, the chain task processing architecture is a task processing framework based on process and logical association, which can organize and execute a series of interrelated task operations in an orderly manner. In this step, the architecture provides a stable operating environment and efficient processing flow for task parsing and subtask generation. The task scheduling model is a model specifically used to handle task scheduling related work after being optimized in the previous steps (such as model distillation, fine-tuning, etc.). It has the ability to understand, analyze and process various tasks, and is the core tool for achieving target task parsing and subtask generation.

[0071] Once a user creates a target task, the system, based on a chained task processing architecture, calls upon the task scheduling model to parse and manage the target task. Taking the natural language description of "Increase user activity this quarter" as an example, the task scheduling model first utilizes natural language processing techniques to perform lexical analysis, syntactic analysis, and semantic understanding of the task. Lexical analysis breaks the task description down into individual lexical units; syntactic analysis determines the grammatical structure relationships between these words; and semantic understanding further uncovers the actual intent and key information behind the task description. During the parsing process, the task scheduling model, drawing on its own knowledge and experience, extracts key information from the target task description, such as the task objective ("Increase user activity"), key metrics (such as specific quantitative indicators such as user login frequency and number of interactions), and the timeframe ("This quarter"). Accurately extracting this key information provides an important basis for subsequent task processing.

[0072] Based on the analysis results of the target task, the task scheduling model begins to generate multiple subtasks. This process also relies on the model's intelligent processing capabilities and preset rules and logic. The task scheduling model will break down the large goal into a series of specific, executable subtasks based on the task objectives and key indicators. For example, to achieve the goal of "increasing user activity this quarter," the following subtasks may be generated: "Design a return visit program for existing users," "Plan a new user registration reward mechanism," and "Optimize the product interface to improve user convenience." Each subtask has clear task content and measurable goals, and can be independently executed and managed.

[0073] When generating subtasks, the task scheduling model also considers the logical relationships and dependency order between subtasks. For subtasks that have a sequential order, the model clarifies their execution order; for parallel subtasks, the model rationally arranges resources to ensure they can proceed simultaneously. For example, "designing a return visit program for existing users" may need to be carried out after the list of existing users is determined, which creates dependencies between subtasks. By organizing and planning these relationships, conflicts and confusion during task execution can be avoided, improving the overall efficiency of task processing.

[0074] Step 104: Obtain task scheduling information corresponding to the target task. The task scheduling model schedules execution of multiple subtasks based on the task scheduling information to generate a task execution plan.

[0075] In the embodiments of the present disclosure, task scheduling information covers many key aspects, including but not limited to organizational structure data, time constraints, resource allocation rules, and task priority settings. This information comes from a wide range of sources. Organizational structure data is usually stored in a database, which records in detail the structural relationships between departments, branches, and grassroots employees at all levels of the company, providing clear personnel and department directions for task allocation; time constraints may be set by the user when creating the target task, or determined according to business rules, such as the start time, deadline, and time nodes of each stage of the task, which play a key role in controlling the rhythm of task execution; resource allocation rules stipulate the allocation methods and quantity limits of various resources required to complete the task, such as human, material, financial, and data resources; task priority settings determine the priority execution order of each task based on factors such as the importance and urgency of the task.

[0076] After acquiring comprehensive task scheduling information, the task scheduling model conducts in-depth analysis and processing based on capabilities accumulated through previous training and optimization. Based on organizational structure information and the nature and requirements of each subtask, the task scheduling model determines the most appropriate execution entity for each subtask. For example, subtasks requiring specialized technical knowledge are assigned to employees or departments with the appropriate skills. Subtasks requiring cross-departmental collaboration are appropriately allocated to the multiple departments involved, with each department's responsibilities clearly defined.

[0077] At the same time, the task scheduling model uses a time-planning algorithm to schedule specific execution times for each subtask based on time constraints. It considers the order and dependencies between tasks to avoid time conflicts and ensure consistent and efficient task execution. For example, if the completion of subtask A is a prerequisite for the start of subtask B, the task scheduling model will appropriately schedule the start time of subtask B after subtask A completes.

[0078] In terms of resource allocation, the task scheduling model uses resource allocation rules to scientifically and rationally allocate limited resources. It calculates the resource requirements for each subtask and allocates them based on actual resource availability, ensuring efficient resource utilization and avoiding waste or over-concentration. Based on task priority, the task scheduling model prioritizes high-priority subtasks, ensuring that important and urgent tasks are promptly advanced. When processing multiple subtasks, the task scheduling model systematically schedules their execution by comprehensively considering the above factors. After analyzing and processing task scheduling information, the task scheduling model generates a detailed task execution plan. This plan is presented in a structured format, clearly listing key information such as the executor, execution time, required resources, and execution order for each subtask.

[0079] Step 105 : performing task scheduling on the multiple subtasks according to the task execution plan and the employee characteristic data, so as to assign the multiple subtasks to the target employees.

[0080] In an embodiment of the present disclosure, a task execution plan is generated in step 104 by a task scheduling model after executing and scheduling multiple subtasks based on the scheduling information of the target task. Employee characteristic data is a quantification and description of various aspects of employees, covering multiple dimensions such as historical performance, skill level, and workload. Historical performance reflects the results and quality of employees' past tasks, and is obtained by statistics and analysis of data such as the completion of historical tasks and the indicators achieved; skill level reflects the professional knowledge, technical capabilities, and ability to solve specific problems possessed by employees, which can be determined through skill assessments, training records, etc.; workload records the amount of tasks currently undertaken by employees and the work pressure, which is obtained by comprehensively considering factors such as the number of tasks currently being performed by employees, the complexity of tasks, and the expected completion time. These data comprehensively and objectively reflect the work ability and status of employees.

[0081] When scheduling and assigning tasks, the system first extracts subtask information from the task execution plan. For each subtask, the system screens employee profile data for potentially suitable employees based on attributes such as task type, required skills, and timeframe. For example, if a subtask involves complex data analysis, the system will identify employees with data analysis skills and a proven track record as candidates. The system then considers the employee's current workload. Employees with excessive workloads, even if they possess the skills to complete the subtask, may not be prioritized for new tasks. This prevents excessive workloads from impacting work quality and efficiency. Conversely, employees with relatively low workloads and appropriate skills are given higher priority. After identifying eligible employees, the system further comprehensively evaluates their historical performance. Employees with strong historical performance tend to have higher success rates and better quality assurance when completing similar tasks, giving them an advantage in task assignments. Through this multi-dimensional evaluation and screening, the system identifies target employees for each subtask. Once the target employee is identified, the system assigns the subtask to the appropriate employee and records the assignment information in the task management system. The task management system will promptly update employees' task lists, allowing them to view details of their assigned subtasks, including task content, requirements, and execution time. The system will also track and monitor task assignments to ensure accuracy and timeliness.

[0082] The present disclosure provides a method for task scheduling, which distills a target large model based on a preset loss function to obtain a target student model, greatly compresses the model size, reduces the demand for resources such as computing power, and enables the model to run in a small resource environment such as a CPU terminal. This can reduce the cost of enterprises using large models and expand the scope of application of the model in enterprise task scheduling scenarios. A task scheduling model is obtained by gradually fine-tuning the target student model using a dedicated data set. The dedicated data set is generated based on a carefully designed prompt template and contains a variety of key elements. It can provide rich and targeted data for specific business scenarios, so that the task scheduling model has higher accuracy and stronger adaptability when facing complex and diverse enterprise tasks, reduces the deviation of the model output, and improves the quality of task scheduling. Based on the chain task processing architecture, the task scheduling model is called to parse and manage the target tasks created by the user and generate multiple subtasks, and at the same time obtains task scheduling information to perform execution scheduling and generate a task execution plan. This process realizes the automation and process of task processing, greatly improves task processing efficiency, reduces manual intervention, reduces the probability of error, and makes task scheduling more scientific and reasonable. Based on the task execution plan and employee characteristic data (such as historical performance, skills, and workload), multiple subtasks are scheduled and assigned to target employees. This allocation method fully considers employees' actual capabilities and work status, effectively matching people to positions, improving employee efficiency, and avoiding problems such as backlogs or excessive employee burdens caused by irrational task allocation, thereby improving the operational efficiency and management level of the entire enterprise.

[0083] In order to clearly illustrate the embodiment of the present disclosure, this embodiment provides a flowchart of another task scheduling method.

[0084] like Figure 2 As shown, the method comprises the following steps:

[0085] Step 201: Perform distillation training on the target large model and the student model using the target dataset.

[0086] Specifically in step 201, model distillation aims to transfer the knowledge acquired by the target large model from a large amount of training data and complex network structure to a student model with fewer parameters and lower computational complexity, thereby achieving the purpose of model compression and accelerated reasoning. The embodiments of the present disclosure do not limit the specific type and model of the target large model. The target dataset used here can be, for example, the open source dataset IMDB with rich data samples, which is not limited by the present disclosure. The IMDB dataset contains text and classification information, where the categories cover three types: positive, neutral emotions, and negative. Its rich and diverse data features can provide sufficient information for model distillation training.

[0087] In step 202, a preset loss function is used to calculate the output distribution difference between the target large model and the student model, and based on the output distribution difference, it is determined whether the distillation training of the student model is completed, wherein the preset loss function is constructed based on an adaptive weight coefficient.

[0088] Specifically, in step 202, the preset loss function plays a crucial role in the entire model distillation training process. In this embodiment, the expression of the preset loss function is:

[0089]

[0090] Among them, a i is the adaptive weight coefficient, T i represents the i-th output of the target large model (teacher model), S i represents the i-th output of the student model, σ is the Softmax activation function, and KL represents the Kullback-Leibler divergence.

[0091] Adaptive weight coefficient a i The calculation of depends on the uncertainty of the sample, which is specifically expressed by the entropy output by the teacher model (i.e., the target large model), and the formula is:

[0092]

[0093] Among them, p j represents the predicted probability of the jth class of sample i by the teacher model.

[0094] During the training process, the target model and the student model simultaneously process the input samples and generate output. i It consists of normalized soft labels containing probabilistic attributes and does not contain manually annotated hard labels. This design can obtain richer information within the target large model.

[0095] In each training iteration, the above preset loss function is used to calculate the output distribution difference between the target large model and the student model. In the early stage of training, since the student model has not yet fully learned the knowledge of the target large model, the output distribution difference between the two is large, and L distill As the training progresses, the student model continuously adjusts its own parameters through the back propagation algorithm, gradually learning the key knowledge and features of the target large model, and its output distribution gradually approaches the output distribution of the target large model. distill The value will gradually decrease.

[0096] To judge whether the student model distillation training is completed, it is necessary to follow the pre-set criteria. Generally speaking, when L distill The value decreases to a certain extent, for example, less than a preset threshold, or in multiple consecutive iterations, Ldistill When the change in value is very small (less than a certain set amount of change), it can be considered that the student model has fully learned the knowledge of the target large model, and the difference between its output distribution and the output distribution of the target large model has reached an acceptable range. At this time, the student model distillation training is determined to be complete. This method of calculating the output distribution difference based on a preset loss function to determine whether training is complete can effectively control the model distillation training process, ensuring that the student model retains the performance of the target large model as much as possible while compressing the model size, meeting the application requirements of specific business scenarios (such as enterprise task scheduling).

[0097] Step 203: call the target large model, generate directional domain data for the target dataset according to the preset prompt template, and obtain a dedicated dataset.

[0098] Specifically, in step 203, the target large model has broad application potential in fields such as natural language processing. By invoking the target large model, its generative capabilities are leveraged to construct datasets that meet the needs of specific domains. Preset prompt templates are the core tool for generating targeted domain data. They contain five key elements: roles, instructions, formats, tasks, and inputs. By properly configuring these elements, the target large model can be guided to generate targeted data.

[0099] For role settings, the format "<<You are an excellent {x} person>>" is used, where x can be replaced based on actual needs, such as setting it to "Data Annotation." Before invoking the large model, LangChain (a chained task processing architecture) uses instruction templates to flexibly replace the x value by modifying the configuration file. This allows the model to simulate the working methods and thinking of the corresponding role during task execution, accurately setting work deployment.

[0100] The command parameter is set to "<<Include large model {y} information in the output sample>>", where y is taken to be "confidence, prediction probability", etc. The purpose of this setting is to confirm the accuracy of the sample and include high, medium, and low confidence samples in the generated dataset to enhance the adaptability of the subsequent model fine-tuning process.

[0101] The task type setting is implemented through the format of "<<Please use {z} type to execute this task>>". z can be set to a specific task type such as "execute once every 5 minutes, for a total of 10 times" to clarify the method and frequency of task execution.

[0102] The task format is set to "<<Please use {k} format for output>>", where k is selected as "json", which specifies the output format of the data to facilitate data storage, transmission, and subsequent processing.

[0103] For the specific tasks of the model input, the format of "<<label 1000 data samples about {q}, and set the training set and test set in a ratio of 8:2>>" is adopted. q can be set according to the specific field. This topic is set as "What are the common problems in the scheduling process of resource tasks" to clarify the data content and task requirements that the model needs to process.

[0104] During execution, based on the target dataset, the parameters in the preset prompt template are configured and then input into the target macromodel. Based on this prompt information, the target macromodel utilizes its own generation capabilities to generate data. This generation process fully integrates the characteristics of the target dataset and the needs of specific domains (such as resource and task scheduling) to generate data containing various key information.

[0105] After processing, the generated data becomes a specialized dataset. This dataset, stored in a database in JSON format, is not only trainable and testable, but also focused on specific business domains, meeting the needs of subsequent model training and optimization. Furthermore, manually annotated data can be added to further enhance the data's professionalism, providing strong support for training more accurate models that are better suited to specific domain tasks.

[0106] Step 204 : starting from the output layer of the target student model, the model parameters of the target student model are gradually unfrozen according to a preset number of layers as a step size.

[0107] Specifically, in step 204, the key operation of adjusting the parameters of the target student model is to gradually activate the model parameters for training in a reasonable manner, thereby improving the model's adaptability to specific tasks. This process mainly revolves around starting from the output layer of the target student model and gradually unfreezing the model parameters of the target student model in steps of a preset number of layers.

[0108] The target student model is obtained through the model distillation step. During the model distillation process, to prevent catastrophic forgetting during fine-tuning, a strategy is often adopted: freezing most layers. This means that only a small number of parameters near the output layer are open for fine-tuning. This is because during the model distillation phase, the model has already learned some common features and patterns. Comprehensively adjusting all parameters may cause the model to lose this learned useful information. Setting the preset number of layers (the step size) is a key key to this operation. The preset number of layers is determined based on factors such as the model structure, task complexity, and training experience. To perform this operation, first determine the output layer of the target student model. Starting from this output layer, the model parameters of the M layers immediately adjacent to the output layer are unfrozen at a preset step size M. This allows these parameters to be updated during subsequent training based on feedback from the loss function. After unfreezing these M layers, the model begins a new round of training. During training, these newly unfrozen parameters adjust their values based on the input data and model output using the backpropagation algorithm to better adapt to the training data and task requirements.

[0109] As training progresses, when the model achieves certain training metrics (e.g., loss converges to a certain range, accuracy improves to an expected level) under the currently unfrozen parameters, or after a predetermined number of training rounds, the model parameters are unfrozen again, starting from the next unfrozen layer, using the preset step size M. This repeated training process allows the model parameters to participate in training gradually and orderly, allowing the model to learn features at different levels at different stages, thereby continuously improving its ability to handle specific tasks.

[0110] Step 205: After performing one unfreezing process, the dedicated data set is input into the unfrozen target student model to perform model fine-tuning until all model parameters are unfrozen to obtain a task scheduling model.

[0111] As a possible implementation method of this embodiment, a dedicated dataset is input into the unfrozen target student model for model fine-tuning, which can be done by, but not limited to, the following methods: based on the number of target student model layers, the attenuation factor, and the basic learning rate, the target learning rate of each model layer in the target student model is calculated; according to the target learning rate of each model layer, the dedicated dataset is input into the unfrozen target student model for model fine-tuning.

[0112] Specifically, in step 205, after one thawing process is executed, some parameters of the target student model are unfrozen. At this time, the dedicated dataset is input into the unfrozen target student model for model fine-tuning. The dedicated dataset is obtained by calling the target large model and generating domain-specific data for the target dataset according to a preset prompt template that includes five important elements: role, instruction, format, task, and input. It closely fits the requirements of a specific business domain (such as the resource task scheduling domain) and has strong pertinence. When performing fine-tuning, a method of adjusting the learning rate based on the characteristics and related parameter settings of the target student model itself is adopted to improve the fine-tuning effect. First, based on the number of layers, decay factor, and base learning rate of the target student model, the target learning rate of each model layer in the target student model is calculated. Assume that the total number of layers of the target student model is L, the decay factor is d (0 < d < 1), and the base learning rate is l base . The model layers closer to the input layer usually learn more basic and general features. In the fine-tuning process, a larger learning rate is required to quickly adapt to the new task data in order to better adjust these basic features to fit the specific task; while the model layers closer to the output layer have already learned some high-level features related to the task during the model distillation stage and relatively do not require too large learning rate adjustment to avoid destroying the effective information that has been learned.

[0113] After calculating the target learning rate of each model layer, according to these target learning rates, the dedicated dataset is input into the unfrozen target student model for model fine-tuning. During the fine-tuning process, the model will process the input dedicated dataset according to the respective target learning rates of each model layer. For the parameter updates of different layers, based on the backpropagation algorithm, the parameter values are adjusted according to the difference between the model output and the expected output (measured by the loss function). Since different layers have different target learning rates, the update amplitude and speed of each layer's parameters are also different, enabling the model to more effectively learn the specific domain knowledge contained in the dedicated dataset while retaining the existing knowledge, and continuously optimizing the performance of the model. This process will continue. After each thawing process is executed, the above fine-tuning operation is repeated until all model parameters are unfrozen. After multiple rounds of such thawing and fine-tuning processes, the target student model gradually adapts to the characteristics and requirements of the specific task domain represented by the dedicated dataset, and the parameters of the model are finely adjusted, finally obtaining the task scheduling model. This task scheduling model can accurately handle the work related to enterprise task scheduling, play a core role in subsequent task parsing, scheduling, and allocation, etc., provide efficient and accurate support for the entire task processing process, and meet the strict requirements for task scheduling in complex business scenarios.

[0114] Step 206 : Based on the chained task processing architecture, the task scheduling model is called to parse and manage the target task created by the user, and multiple subtasks are generated based on the parsing result of the target task.

[0115] For the description of step 206, please refer to the explanation of the above embodiment, which will not be described in detail in this embodiment.

[0116] Step 207: Generate execution guidance information corresponding to each subtask based on the task contents of the multiple subtasks.

[0117] Specifically, in step 207, the task scheduling model is invoked based on the chained task processing architecture to parse the target task created by the user and generate multiple subtasks based on the parsing results. These subtasks have clear and specific content, such as "designing a return visit program for old users," "planning a new user registration reward mechanism," and "optimizing the product interface to improve user convenience."

[0118] When generating the execution guidance for each subtask, the task content of each subtask is first analyzed in depth. For example, for the subtask "Designing a revisit program for returning users," key information will be extracted during the analysis process. For example, if the task objective is to revisit returning users and design a program, the execution guidance may include criteria for defining the returning user base. This could be determined by using database data such as user registration date, last login time, and historical spending history. It also includes process guidance for program design, such as conducting market research to understand the needs and preferences of returning users and then formulating the program theme, rules, and prizes based on the research findings. For the subtask "Planning a new user registration incentive program," the execution guidance may include a clear definition of a new user, typically referring to first-time users registering on the platform; considerations for the incentive program design, such as cost budget, expected number of new users, and the choice of incentive format (cash rewards, coupons, virtual items, etc.). It may also include methods for evaluating the effectiveness of the incentive program, such as monitoring new user registrations and initial post-registration activity rates.

[0119] When generating execution guidance information, reference is made to relevant industry standards, internal company practices, and successful past experiences with similar tasks. If the company has previously conducted successful recurring user engagement campaigns, the execution guidance information for the "Design a recurring program for recurring users" subtask can draw on these experiences in terms of process flow, communication techniques, and prize settings, and make adjustments based on the specific circumstances of the current task. The generated execution guidance information is presented in a structured format, making it easier for executors to quickly understand and follow it. For example, a checklist can be used to outline each key execution point, or a flowchart can be used to illustrate the sequence and key milestones of task execution. This execution guidance information provides clear operational guidance for the individuals or systems responsible for executing subtasks, reducing uncertainty and errors during execution, improving the efficiency and quality of subtask execution, and ultimately ensuring the smooth progress of the entire target task and achieving the intended goal.

[0120] Step 208: Obtain task scheduling information corresponding to the target task. The task scheduling model schedules execution of multiple subtasks based on the task scheduling information to generate a task execution plan.

[0121] Step 209 : performing task scheduling on the multiple subtasks according to the task execution plan and the employee characteristic data, so as to assign the multiple subtasks to the target employees.

[0122] For the description of steps 208 - 209 , please refer to the explanation of the above embodiment, which will not be described in detail in this embodiment.

[0123] It should be noted that the embodiments of the present disclosure may include multiple steps. For the convenience of description, these steps are numbered, but these numbers do not limit the execution time slots or execution order between the steps; these steps can be implemented in any order, and the embodiments of the present disclosure do not limit this.

[0124] Corresponding to the above-mentioned task scheduling method, the present disclosure also provides a task scheduling device. Since the device embodiment of the present disclosure corresponds to the above-mentioned method embodiment, details not disclosed in the device embodiment can be referred to the above-mentioned method embodiment and will not be repeated in this disclosure.

[0125] Figure 3 A schematic diagram of a task scheduling device provided by an embodiment of the present disclosure is shown in FIG. Figure 3 As shown, including:

[0126] A distillation unit 31 is configured to distill the target large model to obtain a target student model based on a preset loss function;

[0127] a fine-tuning unit 32 for gradually fine-tuning the target student model using a dedicated data set to obtain a task scheduling model;

[0128] The creation unit 33 is used to call the task scheduling model to parse and manage the target task created by the user based on the chain task processing architecture, and generate multiple subtasks based on the parsing result of the target task;

[0129] The scheduling unit 34 is used to obtain task scheduling information corresponding to the target task. The task scheduling model schedules the execution of multiple subtasks based on the task scheduling information and generates a task execution plan.

[0130] The allocating unit 35 is configured to schedule the multiple subtasks according to the task execution plan and the employee characteristic data, so as to allocate the multiple subtasks to target employees.

[0131] The present disclosure provides a task scheduling device, which distills a target large model based on a preset loss function to obtain a target student model, greatly compresses the model size, reduces the demand for resources such as computing power, and enables the model to run in a small resource environment such as a CPU terminal. This can reduce the cost of enterprises using large models and expand the scope of application of the model in enterprise task scheduling scenarios. A task scheduling model is obtained by gradually fine-tuning the target student model using a dedicated data set. The dedicated data set is generated based on a carefully designed prompt template and contains a variety of key elements. It can provide rich and targeted data for specific business scenarios, so that the task scheduling model has higher accuracy and stronger adaptability when facing complex and diverse enterprise tasks, reduces the deviation of the model output, and improves the quality of task scheduling. Based on the chain task processing architecture, the task scheduling model is called to parse and manage the target tasks created by the user and generate multiple subtasks. At the same time, task scheduling information is obtained to perform execution scheduling and generate a task execution plan. This process realizes the automation and process of task processing, greatly improves task processing efficiency, reduces manual intervention, reduces the probability of error, and makes task scheduling more scientific and reasonable. Based on the task execution plan and employee characteristic data (such as historical performance, skills, and workload), multiple subtasks are scheduled and assigned to target employees. This allocation method fully considers employees' actual capabilities and work status, effectively matching people to positions, improving employee efficiency, and avoiding problems such as backlogs or excessive employee burdens caused by irrational task allocation, thereby improving the operational efficiency and management level of the entire enterprise.

[0132] Furthermore, in a possible implementation of this embodiment, as Figure 4 As shown, the distillation unit 31 comprises:

[0133] A training module 311 is used to perform distillation training on a target large model and a student model using a target dataset;

[0134] The determination module 312 is used to calculate the output distribution difference between the target large model and the student model using a preset loss function, and determine whether the distillation training of the student model is completed based on the output distribution difference, wherein the preset loss function is constructed based on an adaptive weight coefficient.

[0135] Furthermore, in a possible implementation of this embodiment, as Figure 4 As shown, the device also includes:

[0136] The first generating unit 36 is used to call the target large model and generate directional domain data for the target dataset according to a preset prompt template to obtain a dedicated dataset before fine-tuning the target student model layer by layer using the dedicated dataset to obtain the task scheduling model.

[0137] Furthermore, in a possible implementation of this embodiment, as Figure 4 As shown, the fine-tuning unit 32 includes:

[0138] The unfreezing module 321 is used to gradually unfreeze the model parameters of the target student model from the output layer of the target student model according to a preset number of layers as a step size;

[0139] The fine-tuning module 322 is used to input the dedicated data set into the target student model after the unfreezing process is performed once to perform model fine-tuning until all model parameters are unfrozen to obtain the task scheduling model.

[0140] Furthermore, in a possible implementation of this embodiment, the fine-tuning module 322 is further configured to:

[0141] Calculate the target learning rate for each model layer in the target student model based on the number of target student model layers, the decay factor, and the base learning rate;

[0142] According to the target learning rate of each model layer, the dedicated dataset is input into the unfrozen target student model for model fine-tuning.

[0143] Furthermore, in a possible implementation of this embodiment, as Figure 4 As shown, the device also includes:

[0144] The second generating unit 37 is configured to generate execution guidance information corresponding to each subtask based on the task contents of the subtasks before scheduling the subtasks according to the task execution plan and the employee characteristic data to assign the subtasks to the target employees.

[0145] It should be noted that the above explanation of the method embodiment is also applicable to the device of this embodiment, and the principles are the same, which is not limited in this embodiment.

[0146] According to an embodiment of the present disclosure, the present disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0147] Figure 5 A schematic block diagram of an example electronic device 400 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are provided as examples only and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0148] like Figure 5 As shown, the electronic device 400 includes a computing unit 401, which can perform various appropriate actions and processes according to a computer program stored in a ROM (Read-Only Memory) 402 or a computer program loaded from a storage unit 408 into a RAM (Random Access Memory) 403. In the RAM 403, various programs and data required for the operation of the electronic device 400 can also be stored. The computing unit 401, the ROM 402, and the RAM 403 are connected to each other via a bus 404. An I / O (Input / Output) interface 405 is also connected to the bus 404.

[0149] Multiple components in the electronic device 400 are connected to the I / O interface 405, including an input unit 406, such as a keyboard, a mouse, etc.; an output unit 407, such as various types of displays, speakers, etc.; a storage unit 408, such as a magnetic disk, an optical disk, etc.; and a communication unit 409, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 409 allows the electronic device 400 to exchange information / data with other devices via a computer network such as the Internet and / or various telecommunication networks.

[0150] The computing unit 401 can be a variety of general-purpose and / or specialized processing components with processing and computing capabilities. Some examples of the computing unit 401 include, but are not limited to, a CPU (Central Processing Unit), a GPU (Graphic Processing Unit), various specialized AI (Artificial Intelligence) computing chips, various computing units that run machine learning model algorithms, a DSP (Digital Signal Processor), and any appropriate processor, controller, microcontroller, etc. The computing unit 401 performs the various methods and processes described above, such as the task scheduling method. For example, in some embodiments, the task scheduling method can be implemented as a computer software program that is tangibly contained in a machine-readable medium, such as the storage unit 408. In some embodiments, part or all of the computer program can be loaded and / or installed on the electronic device 400 via the ROM 402 and / or the communication unit 409. When the computer program is loaded into the RAM 403 and executed by the computing unit 401, one or more steps of the method described above can be performed. Alternatively, in other embodiments, the computing unit 401 may be configured to execute the aforementioned task scheduling method in any other appropriate manner (for example, by means of firmware).

[0151] Various embodiments of the systems and techniques described herein can be implemented in digital electronic circuit systems, integrated circuit systems, FPGAs (Field Programmable Gate Arrays), ASICs (Application-Specific Integrated Circuits), ASSPs (Application-Specific Standard Products), SOCs (System on Chips), CPLDs (Complex Programmable Logic Devices), computer hardware, firmware, software, and / or combinations thereof. These various embodiments can include being implemented in one or more computer programs that are executable and / or interpreted on a programmable system that includes at least one programmable processor, which can be a special-purpose or general-purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0152] The program code for implementing the method of the present disclosure can be written in any combination of one or more programming languages. These program codes can be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device so that when the program code is executed by the processor or controller, the functions / operations specified in the flow chart and / or block diagram are implemented. The program code can be executed entirely on the machine, partially on the machine, as a stand-alone software package, partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0153] In the context of the present disclosure, a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, device, or apparatus. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or apparatus, or any suitable combination of the foregoing. More specific examples of machine-readable storage media may include an electrical connection based on one or more wires, a portable computer disk, a hard disk, RAM, ROM, EPROM (Electrically Programmable Read-Only-Memory) or flash memory, optical fiber, CD-ROM (Compact Disc Read-Only Memory), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0154] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device (e.g., a CRT (Cathode-Ray Tube) or LCD (Liquid Crystal Display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user can provide input to the computer. Other types of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic input, voice input, or tactile input).

[0155] The systems and techniques described herein can be implemented in a computing system that includes backend components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such backend components, middleware components, or frontend components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: LAN (Local Area Network), WAN (Wide Area Network), the Internet, and blockchain networks.

[0156] A computer system may include a client and a server. The client and server are generally remote from each other and typically interact via a communication network. This client-server relationship is established by computer programs running on the respective computers, establishing a client-server relationship. The server may be a cloud server, also known as a cloud computing server or cloud host, a host product within the cloud computing service ecosystem that addresses the management difficulties and limited scalability of traditional physical hosts and VPS services ("Virtual Private Servers" or simply "VPS"). The server may also be a server in a distributed system or a server integrated with blockchain.

[0157] It's important to note that artificial intelligence (AI) is the study of how computers can simulate certain human thought processes and intelligent behaviors (such as learning, reasoning, thinking, and planning). This encompasses both hardware and software technologies. AI hardware technologies generally include sensors, specialized AI chips, cloud computing, distributed storage, and big data processing. AI software technologies primarily encompass computer vision, speech recognition, natural language processing, machine learning / deep learning, big data processing, and knowledge graphs.

[0158] The various numerical numbers such as first and second involved in the present disclosure are only for the convenience of description and are not used to limit the scope of the embodiments of the present disclosure, and also indicate the order of precedence.

[0159] The at least one in the present disclosure can also be described as one or more, and the multiple can be two, three, four or more, which is not limited in the present disclosure. In the embodiments of the present disclosure, for a technical feature, the technical features in the technical feature are distinguished by "first", "second", "third", "A", "B", "C" and "D", and there is no order of precedence or size between the technical features described by "first", "second", "third", "A", "B", "C" and "D".

[0160] It should be understood that the various forms of the processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this disclosure can be performed in parallel, sequentially, or in a different order, as long as the desired results of the technical solutions disclosed in this disclosure can be achieved. This is not limited herein.

[0161] The above specific embodiments do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art will appreciate that various modifications, combinations, sub-combinations, and substitutions may be made based on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure shall be included within the scope of protection of this disclosure.

Claims

1. A task scheduling method, characterized in that: include: Based on the preset loss function, the target large model is distilled to obtain the target student model; Using a dedicated data set, the target student model is gradually fine-tuned to obtain a task scheduling model; Based on the chain task processing architecture, the task scheduling model is called to parse and manage the target task created by the user, and multiple subtasks are generated based on the parsing results of the target task; Obtaining task scheduling information corresponding to the target task, wherein the task scheduling model performs execution scheduling on the multiple subtasks according to the task scheduling information to generate a task execution plan; The plurality of subtasks are scheduled according to the task execution plan and the employee characteristic data, so as to allocate the plurality of subtasks to target employees.

2. The task scheduling method according to claim 1, characterized in that: The target large model is distilled based on the preset loss function to obtain the target student model, including: Performing distillation training on the target large model and the student model using a target data set; The preset loss function is used to calculate the output distribution difference between the target large model and the student model, and according to the output distribution difference, it is determined whether the distillation training of the student model is completed, wherein the preset loss function is constructed based on an adaptive weight coefficient.

3. The task scheduling method according to claim 1, wherein: Before fine-tuning the target student model layer by layer using a dedicated data set to obtain a task scheduling model, the method further includes: The target large model is called, and directional domain data is generated for the target data set according to a preset prompt template to obtain the dedicated data set.

4. The task scheduling method according to claim 1, characterized in that: The target student model is gradually fine-tuned using a dedicated data set to obtain a task scheduling model, including: Starting from the output layer of the target student model, gradually unfreeze the model parameters of the target student model according to a preset number of layers as a step size; After performing an unfreezing process, the dedicated data set is input into the unfrozen target student model to perform model fine-tuning until all model parameters are unfrozen to obtain the task scheduling model.

5. The task scheduling method according to claim 4, characterized in that: Inputting the dedicated dataset into the unfrozen target student model for model fine-tuning includes: Calculate the target learning rate of each model layer in the target student model based on the number of target student model layers, the attenuation factor, and the basic learning rate; According to the target learning rate of each model layer, the dedicated dataset is input into the unfrozen target student model for model fine-tuning.

6. The task scheduling method according to claim 1, characterized in that: Before performing task scheduling on the plurality of subtasks according to the task execution plan and the employee characteristic data to assign the plurality of subtasks to target employees, the method further includes: Based on the task contents of the multiple subtasks, execution guidance information corresponding to each subtask is generated.

7. A task scheduling device, characterized in that: include: The distillation unit is used to distill the target large model based on a preset loss function to obtain a target student model; a fine-tuning unit, configured to gradually fine-tune the target student model using a dedicated data set to obtain a task scheduling model; A creation unit, configured to call the task scheduling model to parse and manage a target task created by a user based on a chained task processing architecture, and generate a plurality of subtasks based on the parsing result of the target task; The scheduling unit is configured to obtain task scheduling information corresponding to the target task, and the task scheduling model performs execution scheduling on the multiple subtasks according to the task scheduling information to generate a task execution plan; The allocating unit is configured to schedule the plurality of subtasks according to the task execution plan and the employee characteristic data, so as to allocate the plurality of subtasks to target employees.

8. An electronic device, characterized in that: include: at least one processor; as well as a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the task scheduling method according to any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium storing computer instructions, characterized in that: The computer instructions are used to enable the computer to execute the task scheduling method according to any one of claims 1 to 6.

10. A computer program product, characterized in that The invention comprises a computer program, which implements the task scheduling method according to any one of claims 1 to 6 when being executed by a processor.

Images