Workflow large model construction and fine adjustment method and system for automatic process arrangement
By constructing and fine-tuning a large language model, the problem of insufficient generation capability of existing models in complex process orchestration tasks is solved, realizing high-quality process script generation and management, and improving the accuracy and reliability of automated process orchestration.
Patent Information
- Application Number
- CN202510866011.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-26
- Publication Date
- 2025-10-31
AI Technical Summary
Existing large-scale language models struggle to generate complete process scripts that meet business needs in complex process orchestration tasks, resulting in functional omissions or logical errors. The lack of dedicated training datasets makes it difficult for models to learn complete process logic and multi-interface combination call methods.
By collecting real expert-level process examples from multiple workflow sharing platforms, converting them into Python-style code and generating structured samples, using large-scale language models for multi-granular annotation, constructing diverse and complex datasets, expanding task queries, employing dedicated generative models and dual quality checks, training dedicated annotation models and fine-tuning the models, and setting maximum sequence lengths to adapt to complex processes.
It significantly improves the model's generation capability and accuracy in complex process orchestration tasks involving multiple applications and APIs, solves the problems of logical complexity and insufficient data in existing models, and improves the accuracy and reliability of automated process orchestration.
Smart Images
Figure CN120875503A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of artificial intelligence and automated process orchestration technology, specifically a method and system for constructing and fine-tuning a large workflow model for automated process orchestration. Background Technology
[0002] With the growing demand for intelligent automation, workflow orchestration has become an important means to improve work efficiency and business process intelligence. However, workflow orchestration tasks often involve multi-step decision-making and multi-domain API calls, resulting in high logical complexity. Currently, although large language models (such as Deepseek and Qwen) perform exceptionally well in natural language processing, they still suffer from performance limitations in complex workflow orchestration tasks. For example, in scenarios with multiple applications and multiple APIs, existing models often struggle to generate complete workflow scripts that meet business requirements, easily leading to functional omissions or logical errors. Furthermore, the lack of massive training datasets specifically designed for workflow orchestration makes it difficult for models to learn complete workflow logic and multi-interface combination call methods.
[0003] Therefore, how to improve the model's ability to understand and generate in automated process orchestration to support the automated generation and management of complex business processes is a technical problem that urgently needs to be solved. Summary of the Invention
[0004] The technical objective of this invention is to provide a method and system for constructing and fine-tuning a large workflow model for automated process orchestration, in order to address the problem of how to improve the model's understanding and generation capabilities in automated process orchestration, so as to support the automated generation and management of complex business processes.
[0005] The technical objective of this invention is achieved as follows: a method for constructing and fine-tuning a large workflow model for automated process orchestration, the specific method of which is as follows:
[0006] Data Acquisition: Systematically collect real expert-level process examples from multiple workflow sharing platforms, convert the workflow code into Python style, generate structured samples, and generate hierarchical multi-granular annotations for each example through a large language model, ultimately forming a structured workflow dataset that combines diversity and complexity;
[0007] Query expansion: To address the issue of insufficient diversity in original task queries, task queries are expanded using a large-scale language model, and a large-scale workflow sample is constructed using a dedicated generative model and a dual quality check mechanism.
[0008] Workflow generation: The workflow is generated using a workflow generation + quality verification mode, which involves training a dedicated labeled model, automatically generating a complete workflow script corresponding to the extended query, and verifying the quality of the generated results.
[0009] Model fine-tuning and application: Fine-tune the pre-trained model (such as Qwen3-8B) using the dataset to obtain the workflow fine-tuning model.
[0010] As a preferred method, the data collection process is as follows:
[0011] Convert the original workflow script into Python-style code to generate structured samples containing information such as titles, function descriptions, and API documentation, thereby enhancing parameter passing and logic control capabilities;
[0012] Large-scale language models (such as Deepseek and Qwen) are used to generate multi-granular annotations for each sample and corresponding task queries. The learning value of the data is improved through mind chain annotation. The multi-granular annotations include fine-grained code explanations and coarse-grained task planning descriptions, which improves the adaptability to large-scale language models.
[0013] As a preferred option, the extended query is as follows:
[0014] Sample complementary interfaces from the API library and combine them with the system's built-in APIs to construct hybrid call scenarios. Use a large language model to generate new task queries for hybrid call scenarios.
[0015] By using real-world workflow examples as prompt templates, large language models are driven to generate extended queries with similar logic depth to the examples, thereby improving the complexity and usability of the extended queries.
[0016] As a preferred approach, the workflow generation is as follows:
[0017] During the workflow generation phase, a dedicated annotation model is trained based on the collected real workflow data to automatically generate workflow scripts corresponding to extended queries;
[0018] During the quality verification phase, the generated workflow scripts undergo two levels of quality verification: First, large language models (such as Deepseek and Qwen) are used to fix local errors in the scripts; then, a rule-based filtering mechanism is used to remove logically inconsistent abnormal samples.
[0019] More specifically, the model fine-tuning and application are as follows:
[0020] Use multi-round iterative training strategies (e.g., using the AdamW optimizer, linear learning rate scheduling, and appropriate warmup) to update model parameters, allowing the model to progressively learn the logical structure required for workflow generation;
[0021] During training, a maximum sequence length exceeding a set threshold (e.g., 8192 tokens) is set to accommodate the needs of complex workflows.
[0022] A system for building and fine-tuning a large-scale workflow model for automated process orchestration, comprising:
[0023] The data acquisition module is used to acquire and preprocess workflow data from multiple domain platforms, convert it into Python code, and generate hierarchical comments.
[0024] The query extension module is used to generate new task queries based on sampled, diverse APIs and real-world example hints;
[0025] The workflow generation module is used to automatically generate workflow scripts for training and labeling models, and to perform quality verification for language model repair and rule filtering.
[0026] The model fine-tuning module is used to fine-tune the training of a large pre-trained model based on the constructed workflow dataset, and output the final large workflow model.
[0027] Preferably, the data acquisition module converts the original workflow script into Python-style code, generating structured samples containing information such as titles, function descriptions, and API documentation, enhancing parameter passing and logic control capabilities; and utilizes large language models (such as Deepseek, Qwen) to generate multi-granular annotations for each sample, and generates corresponding task queries, improving the learning value of the data through mind chain-style annotation; wherein, the multi-granular annotations include fine-grained code explanations and coarse-grained task planning descriptions, improving adaptability to large language models;
[0028] The query extension module samples complementary interfaces from the API library and combines them with the system's built-in APIs to construct hybrid call scenarios. It uses a large language model to generate new task queries for hybrid call scenarios and uses real workflow examples as prompt templates to drive the large language model to generate extended queries with similar logical depth to the examples, thereby improving the complexity and practicality of the extended queries.
[0029] More preferably, the workflow generation module adopts a workflow generation + quality verification mode, specifically: in the workflow generation stage, a dedicated annotation model is trained based on the collected real workflow data to automatically generate workflow scripts corresponding to the extended queries; in the quality verification stage, the generated workflow scripts are subjected to two levels of quality verification, specifically: first, a large language model (such as Deepseek, Qwen) is used to fix local errors in the script; then, a rule filtering mechanism is used to remove abnormal samples that do not conform to expectations logically.
[0030] The model fine-tuning module uses a multi-round iterative training strategy (e.g., using the AdamW optimizer, linear learning rate scheduling, and appropriate warmup) to update model parameters, enabling the model to gradually learn the logical structure required for workflow generation; and sets a maximum sequence length exceeding a set threshold (e.g., 8192 tokens) during training to adapt to the needs of complex workflows.
[0031] An electronic device includes: a memory and at least one processor;
[0032] The memory stores computer-executed instructions;
[0033] The at least one processor executes the computer execution instructions stored in the memory, causing the at least one processor to execute the workflow large model construction and fine-tuning method for automated process orchestration as described above.
[0034] A computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method for constructing and fine-tuning a large workflow model for automated process orchestration.
[0035] The method and system for constructing and fine-tuning a large workflow model for automated process orchestration of the present invention have the following advantages:
[0036] (I) This invention focuses on data and improves the performance of large-scale language models in complex workflow orchestration tasks by constructing high-quality training datasets and optimizing model fine-tuning strategies in stages. Specifically, it collects real-world workflow examples from multiple domain platforms and converts them into Python style, generating rich training data by combining multi-granularity annotations and mind chain-style prompts. Simultaneously, it expands task queries using a large-scale language model and constructs large-scale workflow samples using a dedicated generative model and a dual quality detection mechanism. After fine-tuning the pre-trained model (e.g., Qwen3-8B) using the obtained dataset, the resulting fine-tuned workflow model performs excellently in complex workflow orchestration tasks involving multiple applications and APIs, significantly outperforming existing baseline models in zero-shot tests.
[0037] (ii) This invention effectively makes up for the shortcomings of existing large models in the field of process orchestration automation, and provides a reliable intelligent solution for automated process orchestration. It pays special attention to building high-quality process orchestration training datasets and optimizing fine-tuning methods for large language models to support the automated generation and management of complex business processes.
[0038] (III) This invention improves the model’s understanding and generation capabilities in automated process orchestration through a systematic data construction and training method;
[0039] (iv) This invention constructs a data-driven workflow large model framework, which significantly improves the ability of large pre-trained models in automated process orchestration tasks through a systematic dataset construction and model training strategy.
[0040] (v) This invention constructs a diverse and large-scale process orchestration benchmark dataset, covering multiple application scenarios and massive API calls, enhancing the model's ability to learn complex process logic;
[0041] (vi) This invention proposes targeted data processing and generation methods, including query expansion and process generation strategies, to solve the problem of insufficient generalization of existing models in multi-API and multi-step tasks;
[0042] (vii) Through the optimized model fine-tuning scheme, the present invention enables the final workflow model to have higher accuracy and reliability in complex process orchestration with multiple steps and applications.
[0043] (viii) This invention constructs a large-scale workflow orchestration benchmark dataset covering a wide range of application scenarios. Query expansion and automatic process generation further enrich the diversity and complexity of the data, effectively enhancing the model's adaptability to different task scenarios;
[0044] (ix) Through phased data processing and refined model fine-tuning strategies, the workflow model obtained by this invention performs excellently in complex process orchestration tasks involving multiple steps and applications. In zero-sample and cross-domain tests, the quality of the process scripts generated by the model is significantly better than existing open-source and closed-source large models, thereby significantly improving the accuracy and reliability of automated process orchestration.
[0045] (x) This invention uses multi-dimensional indicators such as CodeBLEU and language model-based pass rate evaluation to verify the grammatical and logical quality of the generated script. The automatic evaluation and manual review results have a high consistency rate, which further proves the effectiveness of the evaluation method and process of this invention.
[0046] It is evident that this invention provides an important benchmark for research in related fields. The described technical solution offers a complete solution for automated process orchestration, promoting the practical application of large-scale language models in this field. Attached Figure Description
[0047] The invention will be further described below with reference to the accompanying drawings.
[0048] Appendix Figure 1 A schematic diagram of the structure of a workflow large-scale model construction and fine-tuning system for automated process orchestration;
[0049] Appendix Figure 2 Build a flowchart for the workflow dataset;
[0050] Appendix Figure 3 A schematic diagram of the workflow generation and quality verification process;
[0051] Appendix Figure 4 This is a schematic diagram of the model fine-tuning and evaluation process. Detailed Implementation
[0052] The following detailed description of the workflow large-scale model construction and fine-tuning method and system for automated process orchestration of the present invention is provided with reference to the accompanying drawings and specific embodiments.
[0053] Example 1:
[0054] This embodiment provides a method for constructing and fine-tuning a large workflow model for automated process orchestration, as detailed below:
[0055] S1. Data Acquisition: Systematically collect real expert-level process examples from multiple workflow sharing platforms, convert the workflow code into Python style, generate structured samples, and generate hierarchical multi-granular annotations for each example through a large language model, ultimately forming a structured workflow dataset that combines diversity and complexity.
[0056] S2. Query Expansion: To address the issue of insufficient diversity in original task queries, task queries are expanded using a large-scale language model, and a large-scale workflow sample is constructed using a dedicated generative model and a dual quality inspection mechanism.
[0057] S3. Workflow Generation: The workflow is generated using a workflow generation + quality verification mode, which involves training a dedicated annotation model, automatically generating a complete workflow script corresponding to the extended query, and verifying the quality of the generated results.
[0058] S4. Model Fine-tuning and Application: Fine-tune the pre-trained model (such as Qwen3-8B) using the dataset to obtain the workflow fine-tuning model.
[0059] As attached Figure 2 As shown, the data acquisition in step S1 of this embodiment is as follows:
[0060] S101. Convert the original workflow script into Python-style code to generate a structured sample containing information such as titles, function descriptions, and API documentation, thereby enhancing parameter passing and logic control capabilities.
[0061] S102. Utilize large-scale language models (such as Deepseek, Qwen) to generate multi-granularity annotations for each sample and generate corresponding task queries. Improve the learning value of the data through mind chain annotation. Among them, multi-granularity annotations include fine-grained code explanations and coarse-grained task planning descriptions, thereby improving the adaptability to large-scale language models.
[0062] As attached Figure 3 As shown, the extended query in step S2 of this embodiment is as follows:
[0063] S201. Sample complementary interfaces from the API library and combine them with the system's built-in APIs to construct hybrid call scenarios. Use a large language model to generate new task queries for hybrid call scenarios.
[0064] S202. Utilize real workflow examples as prompt templates to drive large language models to generate extended queries with similar logic depth to the examples, thereby improving the complexity and usability of the extended queries.
[0065] In step S2 of this embodiment, the extended query works by combining API innovation and example-driven approaches, which not only significantly improves the diversity of task types but also ensures the challenge and practicality of generating queries.
[0066] As attached Figure 3 As shown, the workflow generation in step S3 of this embodiment is as follows:
[0067] S301. In the workflow generation stage, a dedicated annotation model is trained based on the collected real workflow data to automatically generate workflow scripts corresponding to extended queries.
[0068] S302. In the quality verification phase, the generated workflow scripts undergo two levels of quality verification: First, large language models (such as Deepseek, Qwen) are used to fix local errors in the scripts; then, a rule-based filtering mechanism is used to remove abnormal samples that do not conform to logical expectations.
[0069] Through the workflow generation process in step S3 of this embodiment, the quality of the synthesized data can be effectively controlled while significantly expanding the data scale, ensuring that the final dataset has both scalability and reliability.
[0070] As attached Figure 4 As shown, the model fine-tuning and application in step S4 of this embodiment are as follows:
[0071] S401. Use a multi-round iterative training strategy (e.g., using the AdamW optimizer, linear learning rate scheduling, and appropriate warmup) to update the model parameters, so that the model gradually learns the logical structure required for workflow generation.
[0072] S402. During training, set a maximum sequence length exceeding a set threshold (e.g., 8192 tokens) to accommodate the needs of complex workflows.
[0073] Through model fine-tuning and application in step S4 of this embodiment, the model can master the ability to generate multi-step, multi-API call processes, thereby forming the final large-scale workflow orchestration model. The trained model exhibits excellent performance in different evaluation tasks: on cross-domain tasks (such as the T-Eval task planning dataset), its zero-shot performance is significantly better than existing benchmark models, verifying the effectiveness of the technical solution.
[0074] The above steps can be performed in combination or integrated as modular components within a unified framework. Overall, this embodiment, through the construction of large-scale and diverse datasets and sophisticated training strategies, forms a complete workflow-based solution for building and applying large-scale models. The entire framework enhances the model's understanding and generation capabilities for automated orchestration tasks, effectively addressing the limitations of existing models in multi-application, multi-API workflow tasks.
[0075] Example 2:
[0076] As attached Figure 1 As shown, this embodiment provides a workflow large-scale model construction and fine-tuning system for automated process orchestration. The system includes:
[0077] The data acquisition module is used to acquire and preprocess workflow data from multiple domain platforms, convert it into Python code, and generate hierarchical comments.
[0078] The query extension module is used to generate new task queries based on sampled, diverse APIs and real-world example hints;
[0079] The workflow generation module is used to automatically generate workflow scripts for training and labeling models, and to perform quality verification for language model repair and rule filtering.
[0080] The model fine-tuning module is used to fine-tune the training of a large pre-trained model based on the constructed workflow dataset, and output the final large workflow model.
[0081] In this embodiment, the data acquisition module converts the original workflow script into Python-style code, generating structured samples containing information such as titles, function descriptions, and API documentation, enhancing parameter passing and logic control capabilities. It also utilizes large language models (such as Deepseek and Qwen) to generate multi-granular annotations for each sample and corresponding task queries, improving the learning value of the data through mind chain-style annotation. The multi-granular annotations include fine-grained code explanations and coarse-grained task planning descriptions, improving adaptability to large language models.
[0082] In this embodiment, the query extension module samples complementary interfaces from the API library and combines them with the system's built-in APIs to construct hybrid call scenarios. It then uses a large language model to generate new task queries for these hybrid call scenarios and uses real workflow examples as prompt templates to drive the large language model to generate extended queries with similar logical depth to the examples, thereby improving the complexity and practicality of the extended queries.
[0083] In this embodiment, the workflow generation module adopts a workflow generation + quality verification mode. Specifically, in the workflow generation stage, a dedicated annotation model is trained based on the collected real workflow data to automatically generate workflow scripts corresponding to the extended queries. In the quality verification stage, the generated workflow scripts are subjected to two levels of quality verification. Specifically, firstly, a large language model (such as Deepseek, Qwen) is used to fix local errors in the script; then, a rule filtering mechanism is used to remove abnormal samples that do not conform to expectations logically.
[0084] In this embodiment, the model fine-tuning module uses a multi-round iterative training strategy (e.g., using the AdamW optimizer, linear learning rate scheduling, and appropriate warmup) to update the model parameters, enabling the model to gradually learn the logical structure required for workflow generation; and during the training process, a maximum sequence length exceeding a set threshold (e.g., 8192 tokens) is set to adapt to the needs of complex workflows.
[0085] Example 3:
[0086] This embodiment also provides an electronic device, including: a memory and at least one processor;
[0087] The memory stores computer-executed instructions;
[0088] The at least one processor executes the computer execution instructions stored in the memory, causing the at least one processor to execute the workflow large model construction and fine-tuning method for automated process orchestration according to any one of the present invention.
[0089] The processor can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), off-the-shelf programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The processor can be a microprocessor or any conventional processor.
[0090] Memory is used to store computer programs and / or modules. The processor implements various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory, and by accessing data stored in the memory. Memory can mainly include a program storage area and a data storage area. The program storage area can store the operating system, at least one application program required for a function, etc.; the data storage area can store data created based on the use of the terminal, etc. In addition, memory can also include high-speed random access memory, and can also include non-volatile memory, such as hard disks, RAM, plug-in hard disks, smart memory cards (SMC), secure digital cards (SD cards), flash memory cards, at least one disk storage device, flash memory devices, or other volatile solid-state storage devices.
[0091] Example 4:
[0092] This embodiment also provides a computer-readable storage medium storing multiple instructions, which are loaded by a processor to cause the processor to execute the workflow large-scale model construction and fine-tuning method for automated process orchestration in any embodiment of the present invention. Specifically, a system or apparatus equipped with a storage medium may be provided, on which software program code implementing the functions of any of the above embodiments is stored, and the computer (or CPU or MPU) of the system or apparatus may read and execute the program code stored in the storage medium.
[0093] In this case, the program code read from the storage medium can itself implement the function of any of the above embodiments, and therefore the program code and the storage medium storing the program code constitute part of the present invention.
[0094] Examples of storage media used to provide program code include floppy disks, hard disks, magneto-optical disks, optical disks (such as CD-ROM, CD-R, CD-RW, DVD-ROM, DVD-RAM, DVD-RW, DVD+RW), magnetic tapes, non-volatile memory cards, and ROMs. Alternatively, program code can be downloaded from a server computer via a communication network.
[0095] Furthermore, it should be clear that not only can the program code read by the computer be executed, but also the operating system or other components operating on the computer can be instructed based on the program code to perform some or all of the actual operations, thereby realizing the function of any of the embodiments described above.
[0096] Furthermore, it is understood that the program code read from the storage medium is written to the memory set in the expansion board inserted into the computer or to the memory set in the expansion unit connected to the computer. Then, based on the instructions of the program code, the CPU or other components installed on the expansion board or expansion unit execute some and all of the actual operations, thereby realizing the function of any of the embodiments described above.
[0097] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for constructing and fine-tuning a large-scale workflow model for automated process orchestration, characterized in that, The method is as follows: Data Acquisition: Systematically collect real expert-level process examples from multiple workflow sharing platforms, convert the workflow code into Python style, generate structured samples, and generate hierarchical multi-granular annotations for each example through a large language model, ultimately forming a structured workflow dataset that combines diversity and complexity; Query expansion: The task query is expanded through a large language model, and a large-scale workflow sample is built using a dedicated generative model and a dual quality check mechanism; Workflow generation: The workflow is generated using a workflow generation + quality verification mode, which involves training a dedicated labeled model, automatically generating a complete workflow script corresponding to the extended query, and verifying the quality of the generated results. Model fine-tuning and application: Fine-tuning the pre-trained model using the dataset to obtain a workflow fine-tuning model.
2. The method for constructing and fine-tuning a large workflow model for automated process orchestration according to claim 1, characterized in that, The data collection process is as follows: Convert the original workflow script into Python-style code to generate structured samples containing information such as titles, function descriptions, and API documentation, thereby enhancing parameter passing and logic control capabilities; A large-scale language model is used to generate multi-granular annotations for each sample and corresponding task queries. The learning value of the data is improved through mind chain annotation. The multi-granular annotations include fine-grained code explanations and coarse-grained task planning descriptions.
3. The method for constructing and fine-tuning a large workflow model for automated process orchestration according to claim 1, characterized in that, The extended query is as follows: Sample complementary interfaces from the API library and combine them with the system's built-in APIs to construct hybrid call scenarios. Use a large language model to generate new task queries for hybrid call scenarios. By using real-world workflow examples as prompt templates, a large language model is driven to generate extended queries with a similar depth of logic to the examples.
4. The method for constructing and fine-tuning a large workflow model for automated process orchestration according to claim 1, characterized in that, The workflow generation process is as follows: During the workflow generation phase, a dedicated annotation model is trained based on the collected real workflow data to automatically generate workflow scripts corresponding to extended queries; During the quality verification phase, the generated workflow scripts undergo two levels of quality verification: first, local errors in the scripts are fixed using a large language model; then, rule-based filtering mechanisms are used to remove logically inconsistent abnormal samples.
5. The method for constructing and fine-tuning a large workflow model for automated process orchestration according to any one of claims 1-4, characterized in that, The specific details of model fine-tuning and application are as follows: The model parameters are updated using a multi-round iterative training strategy, enabling the model to gradually learn the logical structure required for workflow generation. During training, a maximum sequence length exceeding a set threshold is set to accommodate the needs of complex workflows.
6. A workflow large-scale model construction and fine-tuning system for automated process orchestration, characterized in that, The system includes: The data acquisition module is used to acquire and preprocess workflow data from multiple domain platforms, convert it into Python code, and generate hierarchical comments. The query extension module is used to generate new task queries based on sampled, diverse APIs and real-world example hints; The workflow generation module is used to automatically generate workflow scripts for training and labeling models, and to perform quality verification for language model repair and rule filtering. The model fine-tuning module is used to fine-tune the training of a large pre-trained model based on the constructed workflow dataset, and output the final large workflow model.
7. The workflow large-scale model construction and fine-tuning system for automated process orchestration according to claim 6, characterized in that, The data acquisition module converts the original workflow script into Python-style code, generating structured samples containing information such as titles, function descriptions, and API documentation, enhancing parameter passing and logic control capabilities. It also uses a large language model to generate multi-granular annotations for each sample and generates corresponding task queries, improving the learning value of the data through mind chain annotation. The multi-granular annotations include fine-grained code explanations and coarse-grained task planning descriptions. The query extension module samples complementary interfaces from the API library and combines them with the system's built-in APIs to construct hybrid call scenarios. It then uses a large language model to generate new task queries for these hybrid call scenarios and uses real workflow examples as prompt templates to drive the large language model to generate extended queries with similar logical depth to the examples.
8. The automated workflow orchestration system for building and fine-tuning a large-scale workflow model according to claim 6 or 7, characterized in that, The workflow generation module adopts a workflow generation + quality verification mode. Specifically, in the workflow generation stage, a dedicated annotation model is trained based on the collected real workflow data to automatically generate workflow scripts corresponding to extended queries. During the quality verification phase, the generated workflow scripts undergo two levels of quality verification: first, local errors in the scripts are fixed using a large language model; then, rule-based filtering mechanisms are used to remove logically inconsistent abnormal samples. The model fine-tuning module uses a multi-round iterative training strategy to update model parameters, enabling the model to gradually learn the logical structure required for workflow generation; and sets a maximum sequence length exceeding a set threshold during training to adapt to the needs of complex workflows.
9. An electronic device, characterized in that, include: Memory and at least one processor; The memory stores computer-executed instructions; The at least one processor executes the computer execution instructions stored in the memory, causing the at least one processor to perform the workflow large model construction and fine-tuning method for automated process orchestration as described in any one of claims 1 to 5.
10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by the processor, implement the workflow large-scale model construction and fine-tuning method for automated process orchestration as described in any one of claims 1 to 5.