Enterprise private domain large model construction method and corresponding system

Abstract

The invention discloses an enterprise private domain large model construction method and a corresponding system, and aims to solve the problem that an existing large model is difficult to meet accurate adaptation of enterprise businesses. The method comprises the following steps: integrating enterprise knowledge data, and constructing an enterprise knowledge pre-training set through multi-source data textualization conversion, knowledge extraction and desensitization and full text splitting processing; selecting an open source architecture basic large model, performing incremental pre-training based on the pre-training set, and adopting a full-parameter updating strategy and dynamically adjusting a training round and a learning rate to form an initial private domain large model injected with enterprise domain knowledge; constructing a general instruction data set and a business instruction data set, mixing according to a proportion of not less than 5: 1 to carry out instruction fine tuning, and then carrying out two-stage special fine tuning on the model based on business instruction data; by trying on the model and collecting expert feedback data, an enhanced fine tuning data set is constructed, enhanced fine tuning is performed on the model by adopting a preference-based enhanced learning algorithm, and continuous alignment and optimization of model output and business requirements are realized.

Description

Technical Field

[0001] This invention relates to the field of large model technology, and in particular to a method and system for constructing a private domain large model for an enterprise. Background Technology

[0002] In recent years, the rapid development of large-scale pre-trained language models (LLMs) has driven the widespread application of artificial intelligence in general domains. General-purpose large models, such as GPT and LLaMA, have demonstrated powerful natural language understanding and generation capabilities, achieving significant results in scenarios such as text creation, question answering, and code generation. However, these general-purpose models still face significant technical bottlenecks in the actual deployment of enterprise-level professional scenarios: 1. Insufficient depth of domain knowledge; While general-purpose large-scale models have broad training data coverage, their depth is limited, making it difficult to grasp the unique professional knowledge systems within enterprises. For example, in highly specialized fields such as network threat detection, medical device manufacturing, and financial risk assessment, the models lack a deep understanding of detection methods, high-value samples, industry terminology, technical standards, and process specifications, resulting in insufficient professionalism in the generated content and failing to meet the precise decision-making needs of enterprises.

[0003] 2. Lack of enterprise data assets; Core knowledge assets accumulated internally by enterprises, such as patent documents, technical manuals, customer case studies, and operational data, cannot be shared with public models due to trade secrets and data security requirements. This data isolation makes it difficult for general-purpose models to access the enterprise's unique knowledge graph and business logic, resulting in critical information gaps when solving specific enterprise problems.

[0004] 3. Business compatibility issues; Existing technical solutions mainly introduce internal knowledge adaptation to business through prompt engineering or retrieval enhancement (RAG), but neither of them can achieve knowledge internalization and fusion, resulting in insufficient response consistency and reasoning coherence, making it difficult to cover complex business scenarios. Summary of the Invention

[0005] This invention provides a method and system for constructing a large private domain model for enterprises, which solves the problem that existing large models cannot meet the needs of enterprises for accurate decision-making.

[0006] The enterprise private domain large model construction method of this invention includes: Collect enterprise knowledge-related data, and perform multi-source data textification, knowledge extraction and desensitization, and full text splitting on the collected data to construct an enterprise knowledge pre-training set; A basic large model is selected, and incremental pre-training is performed on the selected basic large model based on the enterprise knowledge pre-training set to form an initial enterprise private domain large model; during the incremental pre-training process, a full parameter update strategy is adopted, and the training rounds and learning rate are dynamically adjusted; A two-stage fine-tuning strategy is adopted. First, a general instruction dataset and a business instruction dataset are constructed. Then, a hybrid training set is constructed with a ratio of general instruction data to business instruction data of no less than 5:1 to fine-tune the initial enterprise private domain large model. Finally, the business instruction data is used separately to fine-tune the initial enterprise private domain large model to form the enterprise private domain large model.

[0007] According to some embodiments of the present invention, the method further includes: The enterprise private domain model was deployed and tested in an enterprise environment, and the results were evaluated based on the test results. The evaluation levels were divided into first level and second level. For the first level, the test set and the corresponding test results will be used as business instruction fine-tuning data for subsequent continuous optimization of the enterprise private domain model. For the second level, experts provide test answers for the test set and combine the test results to construct a reinforcement fine-tuning dataset. Using reinforcement learning algorithms, the enterprise private domain large model is optimized.

[0008] According to some embodiments of the present invention, the method further includes: continuously collecting real interaction data after the deployment of the enterprise private domain model, and continuously iterating the enterprise private domain model based on the real interaction data after deployment.

[0009] According to some embodiments of the present invention, the multi-source data textification process includes: Semantic enhancement processing is performed on the unstructured text in the enterprise knowledge entries; Transform the structured data in the enterprise knowledge entries into natural language; The multimedia data in the enterprise knowledge entries is converted into text data.

[0010] According to some embodiments of the present invention, the selected basic large model includes: The model parameter magnitude is determined based on the size of the enterprise knowledge base and the complexity of the application scenario, and a basic large model is selected from a large model library of open source architectures based on the model parameter magnitude.

[0011] According to some embodiments of the present invention, the construction of the general instruction dataset and the business instruction dataset includes: Collect interaction data from community open source and enterprise content to build a general instruction dataset; With the goal of covering the diverse intelligent assistance needs of enterprise users, a business instruction dataset is constructed for different business scenarios and task types. The business instruction dataset follows a standardized data format and includes four key fields: business category, system instruction, user input, and model response.

[0012] According to some embodiments of the present invention, in the process of fine-tuning the enterprise private domain large model using the general instruction dataset and the business instruction dataset, the low-rank adaptive LoRA method is used to adjust some model parameters.

[0013] The enterprise private domain large model construction system of this invention includes: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements the steps of the enterprise private domain large model construction method as described above.

[0014] This invention also proposes a computer-readable storage medium storing an information transmission implementation program, which, when executed by a processor, implements the steps of the enterprise private domain large model construction method described above.

[0015] This invention also proposes a computer program product, including a computer program that, when executed by a processor, implements the steps of the enterprise private domain large model construction method described above.

[0016] This invention can solve the problems of insufficient domain knowledge depth, lack of enterprise data assets, and business adaptability obstacles faced by general large models in enterprise business application scenarios. It can deeply integrate professional knowledge into the model and ensure the security and controllability of enterprise core data.

[0017] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0018] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of the embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. In the drawings: Figure 1 This is a flowchart of the enterprise private domain large model construction method in an embodiment of the present invention; Figure 2 This is a flowchart illustrating the construction process of the enterprise knowledge pre-training set in an embodiment of the present invention; Figure 3This is a flowchart of the business instruction fine-tuning process in an embodiment of the present invention; Figure 4 This is a schematic diagram of the interface of the human feedback data collection system in an embodiment of the present invention. Detailed Implementation

[0019] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the invention and to fully convey the scope of the invention to those skilled in the art. Furthermore, in some instances, well-known methods, structures, and techniques have not been shown in detail so as not to obscure the understanding of this specification.

[0020] The enterprise private domain large model construction method of this invention includes: Collect enterprise knowledge-related data, and perform multi-source data textification, knowledge extraction and desensitization, and full text splitting on the collected data to construct an enterprise knowledge pre-training set.

[0021] First, it needs to be stated that the data collection process complies with relevant laws and regulations, as well as the ethical requirements of the convention.

[0022] There are no restrictions on the format of the collected enterprise knowledge data; it can be text data, image data, video data, etc., including both structured and unstructured data. Therefore, this data needs to be processed, converting all data into text format, de-identifying sensitive fields, filtering noise and duplicate data through feature extraction, and splitting the entire text to ensure that no single data entry exceeds the maximum token value while maintaining semantic integrity.

[0023] A basic large model is selected, and incremental pre-training is performed on the selected basic large model based on the enterprise knowledge pre-training set to form an initial enterprise private domain large model. During the incremental pre-training process, a full parameter update strategy is adopted, and the training rounds and learning rate are dynamically adjusted until the loss function converges and the training loss value fluctuates within the threshold range.

[0024] Construct a general instruction dataset and a business instruction dataset, and build a hybrid training set with a ratio of general instruction data to business instruction data of not less than 5:1 to fine-tune the initial enterprise private domain large model. Then, use business instruction data alone to fine-tune the initial enterprise private domain large model to form the enterprise private domain large model.

[0025] General instruction data is used to enhance the basic interactive capabilities of large models, ensuring targeted responses to common questions. This part of the data needs to pay special attention to data diversity and generalization to guarantee the model's performance in common scenarios. Business instruction data enables large models to output high-value content that meets business needs based on professional instructions.

[0026] This invention leverages the accumulated knowledge and human-computer interaction experience of enterprises over many years. First, it employs incremental pre-training to efficiently inject private domain knowledge into the enterprise. Then, it combines business instruction fine-tuning and reinforcement fine-tuning techniques based on human feedback to significantly enhance the model's professional task processing and intelligent assistance capabilities in enterprise scenarios. This method provides an effective solution for enterprises to quickly build large-scale private domain models with high-value output capabilities, supporting efficient application in scenarios such as intelligent assistance and knowledge management, and possessing significant practical application value. As a result, enterprise employees can achieve intelligent assistance and knowledge management based on the constructed large-scale private domain model. Examples include, but are not limited to, text creation, knowledge question answering, and code generation.

[0027] Based on the above embodiments, further variant embodiments are proposed. It should be noted that, in order to keep the description brief, only the differences from the above embodiments are described in each variant embodiment.

[0028] According to some embodiments of the present invention, the multi-source data textification process includes: Semantic enhancement processing is performed on the unstructured text in the enterprise knowledge entries; Transform the structured data in the enterprise knowledge entries into natural language; The multimedia data in the enterprise knowledge entries is converted into text data.

[0029] According to some embodiments of the present invention, the selected basic large model includes: The magnitude of model parameters is determined based on the size of the enterprise knowledge base and the complexity of the application scenario. A basic large model is then selected from a large open-source model library based on this parameter magnitude. For example, the Qwen and LLaMA series can be used as basic large models.

[0030] According to some embodiments of the present invention, the construction of the general instruction dataset and the business instruction dataset includes: Collect interaction data from community open source and enterprise content to build a general instruction dataset; With the goal of covering the diverse intelligent assistance needs of enterprise users, a business instruction dataset is constructed for different business scenarios and task types. The business instruction dataset follows a standardized data format and includes four key fields: business category, system instruction, user input, and model response.

[0031] According to some embodiments of the present invention, in the process of fine-tuning the enterprise private domain large model using the general instruction dataset and the business instruction dataset, the low-rank adaptive LoRA method is used to adjust some model parameters.

[0032] According to some embodiments of the present invention, the method further includes: The enterprise private domain model was deployed and tested in an enterprise environment, and the results were evaluated based on the test results. The evaluation levels were divided into first level and second level. For the first level, the test set and the corresponding test results are used as business instruction fine-tuning data to optimize the enterprise private domain large model; For the second level, experts provide test answers for the test set and combine the test results to construct a reinforcement fine-tuning dataset. Using reinforcement learning algorithms, the enterprise private domain large model is optimized.

[0033] According to some embodiments of the present invention, the method further includes: continuously collecting real interaction data of the enterprise private domain model after deployment, and continuously iterating the enterprise private domain model based on the real interaction data.

[0034] The method for constructing a large enterprise private domain model according to an embodiment of the present invention will be described in detail below with reference to the accompanying drawings and a specific example. It is to be understood that the following description is merely exemplary and should not be construed as a specific limitation of the present invention.

[0035] This invention proposes a method for constructing a private domain large-scale model for enterprises, addressing the problems of insufficient domain knowledge depth, lack of enterprise data assets, and business adaptability barriers faced by general-purpose large-scale models in enterprise business application scenarios. First, incremental pre-training allows the general-purpose large-scale model to learn enterprise professional knowledge. Then, fine-tuning is performed to adapt it to specific business needs, ultimately constructing a private domain large-scale model specifically for the enterprise. This method deeply integrates professional knowledge into the model while ensuring the security and controllability of core enterprise data.

[0036] The enterprise private domain large-scale model construction method of this invention includes five steps: model selection, large-scale model incremental pre-training, business instruction fine-tuning, human feedback and reinforcement fine-tuning, and model deployment and iteration. Figure 1 As shown, the detailed steps are as follows: Step 1: Model selection; Choose large open-source architecture models as the base (such as Qwen, LLaMA series, etc.), and dynamically determine the model parameter scale based on the size of the enterprise knowledge base and the complexity of the application scenario. In typical scenarios, a scale of hundreds of billions of parameters (e.g., 30B level) can achieve an optimized balance between performance and cost.

[0037] Step 2: Incremental pre-training; Incremental pre-training injects enterprise domain knowledge into general-purpose basic models, which includes two core steps: building an enterprise knowledge pre-training set and model training.

[0038] Step 2.1: Construction of Enterprise Knowledge Pre-training Set; Enterprise internal knowledge is characterized by diverse formats and complex content, making it unsuitable for direct use in large-scale model training. Special processing of data content and format is required. For example, cybersecurity enterprise data includes unstructured data (threat analysis reports, incident response documents, network topology diagrams, traffic heatmaps) and structured data (firewall logs, anomaly alerts, malicious sample characteristics), while medical institution data includes unstructured data (prescription records, medical images, technical documents). The process for constructing an enterprise knowledge pre-training set is as follows: Figure 2 As shown, the detailed data processing flow is as follows: (1) Multi-source data textification processing; For unstructured text, semantic enhancement is implemented to transform conventional natural language descriptions into more information-dense and semantically complete expressions, thereby carrying more high-value domain information. Specifically, by supplementing the original statements with key information such as the basis for judgment, reasoning process, and detailed conclusions, the information content and accuracy of the text are improved, making it easier for the model to more accurately capture domain knowledge. For example, "anomaly SMB protocol request detected" is enhanced to "unconventional SMB protocol requests frequently occur during the active period of the file server (09:00–11:00), and this pattern is highly consistent with the characteristics of ransomware lateral movement attacks." For structured data (such as database records), it is converted into descriptive text conforming to natural language standards; for multimedia data such as images and audio, corresponding text descriptions are generated.

[0039] (2) Knowledge extraction and desensitization; Replace sensitive information (IP address → [IP], username → [User]), retain key features (such as attack pattern signatures, abnormal traffic statistics), and filter out noise and duplicate data.

[0040] (3) Full text splitting processing; Set a maximum token count threshold (e.g., 4096) to split the entire text, ensuring that a single piece of data does not exceed the maximum token count while maintaining semantic integrity.

[0041] Step 2.2: Incremental pre-training of the model; Based on the constructed enterprise knowledge pre-training set, incremental pre-training is performed on the selected large model. This process adopts a full parameter update strategy, continuously optimizing the model weight parameters by dynamically adjusting the training epochs and learning rate until the loss function converges and the training loss value fluctuates within a threshold range.

[0042] Step 3: Fine-tune business instructions; The business instruction fine-tuning phase aims to enable the large model, infused with enterprise knowledge, to simultaneously possess the ability to follow general instructions and execute domain-specific business instructions. It includes two core components: business instruction fine-tuning dataset construction and model fine-tuning.

[0043] Step 3.1: Construction of the business instruction fine-tuning dataset; The business instruction fine-tuning dataset consists of general instruction data and domain-specific business instruction data, with the general instruction data being significantly larger than the business instruction data. (1) General instruction data: High-quality interaction data collected by enterprises and high-quality datasets open-sourced by the community (preferably real human-computer interaction records) are used to improve the basic interaction capabilities of the large model and ensure that targeted responses are generated for general questions. This part of the data should pay special attention to data diversity and generalization to ensure the performance of the model in general scenarios. (2) Business Instruction Data: A small amount of high-quality, representative data is constructed based on the actual business scenarios of the enterprise, enabling the large model to output high-value content that meets business needs according to professional instructions. The following requirements must be met during the construction process: covering various intelligent assistance needs of enterprise users, customizing designs for different business scenarios and task types, and following standardized data formats, including four key fields: business category, system instructions, user input, and model response. Since the quality of the dataset directly affects the performance of the final model, a strict quality control mechanism must be established.

[0044] Step 3.2: Model fine-tuning; Using the constructed business instruction fine-tuning dataset, the large model with completed enterprise knowledge injection is fine-tuned. For the fine-tuning method, a parameter-efficient approach is adopted; the LoRA method is recommended, allowing adjustments to only a small number of model parameters. The business instruction fine-tuning process is as follows: Figure 3 As shown, the specific implementation adopts a two-stage fine-tuning strategy: The first stage adopts a hybrid instruction fine-tuning approach, mixing general instruction data and domain business instruction data in a ratio of no less than 5:1. This allows the large model to maintain excellent general instruction compliance capabilities while initially possessing domain business instruction execution capabilities based on the enterprise's internal knowledge injected during the incremental pre-training stage. The second phase focuses on fine-tuning using domain-specific business instruction data, with a key emphasis on enhancing the large model's ability to accurately execute tasks tailored to specific enterprise needs. Through these two phases of fine-tuning, the performance of the large model in performing specialized tasks based on internal enterprise data can be effectively improved.

[0045] Step 4: Human feedback and reinforcement fine-tuning; After model fine-tuning, it can be deployed and tested in an internal enterprise environment. Performance is evaluated through trials in real-world application scenarios, and human feedback data is collected. This data is then used to further refine the larger, more robust model, thereby improving performance, reducing illusions, and aligning with the usage preferences of internal personnel. This phase comprises two core steps: collecting human feedback data and constructing the refinement dataset, as well as the refinement process itself.

[0046] Step 4.1: Human feedback data collection and reinforcement fine-tuning dataset construction; Experts from various business areas were invited to participate in the trial, collecting human-computer interaction data from real-world scenarios for subsequent enhancement and fine-tuning. To achieve this functionality, a dedicated data collection system was built: (1) Experts pose questions to the fine-tuned large model and obtain the model's response; (2) Experts evaluate the model output in three levels: "good", "average", and "poor"; When the evaluation is "average" or "poor", the expert submits a high-quality answer they wrote themselves in the input box; (3) The collected human-computer interaction data is categorized and processed according to the evaluation results: data rated "good" by humans is classified as domain business instruction fine-tuning data for subsequent instruction fine-tuning; data rated "average" or "poor" is used to construct an enhancement fine-tuning dataset, which contains two comparative answers for each question—a high-quality answer written by humans and an answer to be improved output by the model (i.e., responses rated "average" or "poor"). A schematic diagram of the human feedback data collection system interface is shown below. Figure 4 As shown.

[0047] The data format specifications are as follows: { "messages": [ {"role": "system", "content": "system prompt"}, {"role": "user", "content": "problem"}, {"role": "assistant", "content": "high-quality answer"} ], "rejected_response": "Answer to be improved" } Step 4.2, Strengthen fine-tuning; The large model is reinforced and fine-tuned using a constructed reinforcement and fine-tuning dataset. Given that the dataset contains a preference comparison between high-quality answers and answers requiring improvement, preference-based reinforcement learning algorithms such as DPO or ORPO are employed for optimization. A contrastive learning mechanism guides the model to favor high-quality outputs. The reinforcement and fine-tuning process is designed as a continuous iterative mechanism: as real-world trials progress, new human interaction data is continuously collected, the reinforcement and fine-tuning dataset is updated, and new fine-tuning cycles are implemented, forming a self-optimizing closed-loop process.

[0048] Step 5: Model Deployment and Iteration; After completing internal environment testing and fine-tuning, the enterprise's private domain large-scale model can be officially deployed once its performance meets basic business needs. During operation, a data collection system continuously collects human-computer interaction data from real-world scenarios for subsequent instruction fine-tuning and reinforcement iterations, enabling continuous optimization and upgrading of model performance.

[0049] The efficient construction method for enterprise private domain large-scale models in this invention has a significant advantage in effectively improving the business processing capabilities of large-scale models within enterprise environments. This method fully leverages the knowledge accumulated by enterprises over many years and their human-computer interaction experience. First, it employs incremental pre-training to efficiently inject enterprise private domain knowledge. Then, it combines business instruction fine-tuning and reinforcement fine-tuning techniques based on human feedback to significantly enhance the model's professional task processing and intelligent assistance capabilities in enterprise scenarios. This method provides an effective solution for enterprises to quickly build private domain large-scale models with high-value output capabilities, supporting the efficient application of intelligent assistance, knowledge management, and other scenarios, and has significant practical application value.

[0050] It should be noted that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0051] The enterprise private domain large model construction system of this invention includes: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements the steps of the enterprise private domain large model construction method as described above.

[0052] This invention also proposes a computer-readable storage medium storing an information transmission implementation program, which, when executed by a processor, implements the steps of the enterprise private domain large model construction method described above.

[0053] This invention also proposes a computer program product, including a computer program that, when executed by a processor, implements the steps of the enterprise private domain large model construction method described above.

[0054] It should be noted that any content not described in detail in this specification is common knowledge to those skilled in the art.

Claims

1. A method for constructing a large-scale private domain model for enterprises, characterized in that, include: Collect enterprise knowledge-related data, and perform multi-source data textification, knowledge extraction and desensitization, and full text splitting on the collected data to construct an enterprise knowledge pre-training set; A basic large model is selected, and incremental pre-training is performed on the selected basic large model based on the enterprise knowledge pre-training set to form an initial enterprise private domain large model. The incremental pre-training process employs a full parameter update strategy and dynamically adjusts the training rounds and learning rate. A two-stage fine-tuning strategy is adopted. First, a general instruction dataset and a business instruction dataset are constructed. Then, a hybrid training set is constructed with a ratio of general instruction data to business instruction data of no less than 5:1 to fine-tune the initial enterprise private domain large model. Finally, the business instruction data is used separately to fine-tune the initial enterprise private domain large model to form the enterprise private domain large model.

2. The method as described in claim 1, characterized in that, The method further includes: The enterprise private domain model was deployed and tested in an enterprise environment, and the results were evaluated based on the test results. The evaluation levels were divided into first level and second level. For the first level, the test set and the corresponding test results are used as business instruction fine-tuning data to optimize the enterprise private domain large model. For the second level, experts provide test answers for the test set and combine the test results to construct a reinforcement fine-tuning dataset. Using reinforcement learning algorithms, the enterprise private domain large model is optimized.

3. The method as described in claim 1 or 2, characterized in that, The method further includes: continuously collecting real interaction data after the deployment of the enterprise private domain big model, and continuously iterating the enterprise private domain big model based on the real interaction data.

4. The method as described in claim 1, characterized in that, The multi-source data textification process includes: Semantic enhancement processing is performed on the unstructured text in the enterprise knowledge entries; Transform the structured data in the enterprise knowledge entries into natural language; The multimedia data in the enterprise knowledge entries is converted into text data.

5. The method as described in claim 1, characterized in that, The selected basic large model includes: The model parameter magnitude is determined based on the size of the enterprise knowledge base and the complexity of the application scenario, and a basic large model is selected from a large model library of open source architectures based on the model parameter magnitude.

6. The method as described in claim 1, characterized in that, The construction of the general instruction dataset and the business instruction dataset includes: Collect interaction data from community open source and enterprise content to build a general instruction dataset; With the goal of covering the diverse intelligent assistance needs of enterprise users, a business instruction dataset is constructed for different business scenarios and task types. The business instruction dataset follows a standardized data format and includes four key fields: business category, system instruction, user input, and model response.

7. The method as described in claim 1, characterized in that, In the process of fine-tuning the enterprise private domain large model using the general instruction dataset and the business instruction dataset, the low-rank adaptive LoRA method is used to adjust some model parameters.

8. A system for constructing a large-scale enterprise private domain model, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the enterprise private domain large model construction method as described in any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores an information transmission implementation program, which, when executed by a processor, implements the steps of the enterprise private domain large model construction method as described in any one of claims 1 to 7.

10. A computer program product, characterized in that, It includes a computer program that, when executed by a processor, implements the steps of the enterprise private domain large model construction method as described in any one of claims 1 to 7.

Images