An enterprise-level financial large model platform

Abstract

This application provides an enterprise-level financial big data model platform, relating to the fields of artificial intelligence and fintech. The platform includes: a model layer for collaboratively integrating the capabilities of models with different modalities and parameter quantities, providing model capabilities to the component layer and application scenario layer; a component layer for encapsulating model capabilities into standardized components; and an application scenario layer for calling standardized components corresponding to task requirements based on financial business scenarios, combined with multimodal data, to meet the task requirements of the financial business scenarios. The method of this application improves the development efficiency of financial applications, reduces development costs, and shortens the development cycle by integrating the interfaces of different models through the componentization of big and small model capabilities. Through an automated learning mechanism, it continuously optimizes model capabilities based on actual business feedback, achieving continuous model optimization, enhancing the intelligence of financial applications, and reducing the labor and maintenance costs of commercial banks.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and fintech, and in particular to an enterprise-level financial big data platform. Background Technology

[0002] In the financial industry, with the deepening of digital transformation, banks and other financial institutions face multiple challenges, including massive data processing, intelligent upgrades of complex business scenarios, and stringent compliance regulations. The scale of unstructured and structured data, such as financial transaction data, user behavior data, and market sentiment data, is growing exponentially, and traditional data processing methods are struggling to meet the demands for real-time processing, accuracy, and in-depth analysis. Currently, the financial industry urgently needs an enterprise-level large-scale model platform to achieve a fully intelligent closed loop from data collection, model training, scenario application to continuous optimization.

[0003] Existing platforms do not provide a standardized model component encapsulation mechanism. Model calls rely on manual configuration and only support simple tasks in a single scenario. This results in high model development costs and long development cycles, and makes it impossible to flexibly combine models with different modalities and parameter amounts to adapt to complex financial business needs.

[0004] Therefore, existing platforms suffer from low development efficiency and are unable to support the financial industry's demand for large-scale model platforms. Summary of the Invention

[0005] This application provides an enterprise-level financial large-scale model platform to solve the technical problems of existing platforms having low development efficiency and being unable to support the financial industry's demand for large-scale model platforms.

[0006] Firstly, this application provides an enterprise-level financial big data model platform, including a data layer, a model layer, a component layer, and an application scenario layer:

[0007] The data layer is used to store and manage multimodal data in financial business scenarios;

[0008] The model layer is used to collaboratively integrate the capabilities of models with different modalities and parameter counts, providing the component layer and application scenario layer with the capability of models with different modalities and parameter counts;

[0009] The component layer is used to encapsulate the capabilities of models with different modalities and different parameter quantities into standardized components.

[0010] The application scenario layer is used to call standardized components corresponding to the task requirements of financial business scenarios based on multimodal data, so as to meet the task requirements of financial business scenarios.

[0011] Secondly, this application provides an enterprise-level financial big data model platform management method, applicable to any of the enterprise-level financial big data model platforms described in the first aspect. The platform includes a data layer, a model layer, a component layer, and an application scenario layer. The method includes:

[0012] Through the model layer, models with different modalities and different parameter quantities can be obtained;

[0013] The capabilities of models with different modalities and different parameter values ​​are synergistically integrated;

[0014] Through the component layer, the capabilities of models with different modalities and different parameter quantities are encapsulated into standardized components;

[0015] Based on the task requirements corresponding to financial business scenarios, and combined with the input multimodal data, the application scenario layer calls the standardized components corresponding to the task requirements to meet the task requirements of the financial business scenarios.

[0016] Thirdly, this application provides an electronic device, including: a processor and a memory communicatively connected to the processor;

[0017] The memory stores the instructions that the computer executes;

[0018] The processor executes computer-executable instructions stored in memory to implement any of the methods in the second aspect.

[0019] Fourthly, this application provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the method of any of the second aspects.

[0020] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the method of any of the second aspects.

[0021] This application provides an enterprise-level financial large-scale model platform that integrates interfaces for models from different domains, modalities, and sizes through modularization and fusion of large and small model capabilities. This improves the development efficiency of financial applications, reduces development costs, and shortens development cycles, providing technical support for commercial banks to rapidly build financial application scenarios at scale. Through an automated learning mechanism, the platform continuously optimizes model capabilities based on actual business feedback, achieving continuous optimization of model performance. This enhances the intelligence and compliance of financial applications and significantly reduces the labor and maintenance costs for commercial banks. Attached Figure Description

[0022] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0023] Figure 1 This is a schematic diagram illustrating solutions for the application of artificial intelligence technology in the financial industry.

[0024] Figure 2 A schematic diagram of the structure of an enterprise-level financial big data platform provided in this application embodiment;

[0025] Figure 3 A logical architecture diagram of an enterprise-level financial big data platform provided for embodiments of this application;

[0026] Figure 4 A deployment architecture diagram of an enterprise-level financial big data model platform provided in this application embodiment;

[0027] Figure 5 A security architecture diagram of an enterprise-level financial big data platform provided in this application embodiment;

[0028] Figure 6 A flowchart illustrating the overall interaction process of a large model platform provided in this application embodiment;

[0029] Figure 7 A flowchart illustrating a work order classification and summary scenario provided in this application embodiment;

[0030] Figure 8 A flowchart illustrating an enterprise-level financial big data platform management method provided in this application embodiment;

[0031] Figure 9 A flowchart illustrating the component-based task execution of a size model fusion capability is provided in this application embodiment;

[0032] Figure 10 A flowchart for document parsing and execution is provided as an embodiment of this application;

[0033] Figure 11 A flowchart illustrating an automated learning closed-loop mechanism for a model, as provided in this application embodiment;

[0034] Figure 12 A flowchart for automatic model evaluation provided in this application embodiment;

[0035] Figure 13 A dynamic security governance flowchart provided in this application embodiment;

[0036] Figure 14 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0037] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0038] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0039] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of the relevant data all comply with the relevant laws, regulations, and standards of the relevant countries and regions, have taken necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation access points for users to choose to authorize or refuse.

[0040] Furthermore, the technical solution involved in this application, which involves big data analysis of user information (including but not limited to personal biometrics, identity data, consumption data, asset data, electronic terminal operation data, etc.) and the use of artificial intelligence technology for automated decision-making, and makes decisions that have a significant impact on personal rights based on the results of automated decision-making, provides users with corresponding operation entry points for users to choose to agree to or reject the results of automated decision-making; if the user chooses to reject, the process will proceed to the expert decision-making process.

[0041] It should be noted that the enterprise-level financial big data platform provided in this application can be used in the fields of artificial intelligence and fintech, or in any field other than artificial intelligence and fintech. The application fields of the enterprise-level financial big data platform in this application are not limited.

[0042] This application specifically targets the construction of enterprise-level large-scale model platforms in the banking, securities, and other financial industries. With the accelerated digital transformation of the financial industry, financial institutions face the need for efficient processing and intelligent analysis of massive amounts of financial data. Transaction data, user behavior data, and market dynamics data in financial scenarios are experiencing explosive growth. This data includes not only structured information such as transaction records and customer profiles, but also unstructured data such as voice interactions and image credentials. Traditional financial business systems struggle to effectively mine and utilize the value of this data, thus limiting the level of intelligence in financial services.

[0043] In the current technology landscape, the financial industry mainly relies on traditional AI platforms or single-model solutions for the application of artificial intelligence technology. Figure 1 Solutions for the application of artificial intelligence technology in the financial industry, such as Figure 1 As shown, existing technical solutions typically employ discrete small models or general large models for task processing.

[0044] Therefore, existing solutions have the following limitations: 1. Single architecture: Existing platforms mostly adopt a single-layer model architecture, lacking the ability to collaboratively integrate multimodal and multi-size models, resulting in low processing efficiency in complex financial scenarios (such as multi-turn dialogues and cross-modal information processing). 2. High development and maintenance costs: Traditional platforms rely on manual development and customized model training, lacking standardized components and automated processes, resulting in long model iteration cycles, high R&D costs, and difficulty in quickly adapting to new business needs. 3. Insufficient model reusability: Existing solutions have fragmented model capabilities and lack unified component-based encapsulation. Different business scenarios require repeated development of similar functions, leading to resource waste and system redundancy. 4. Weak security protection capabilities: Traditional platforms mostly adopt static security strategies (such as alignment during training + filtering during inference), lacking real-time monitoring and dynamic repair capabilities for model output, making it difficult to cope with complex compliance risks and data security requirements in financial scenarios. 5. Lack of learning loop: Existing model training relies on manually labeled data and lacks a mechanism for automatically optimizing the model from actual business feedback, making it difficult to continuously improve model performance and adapt to the rapidly changing financial business environment.

[0045] Therefore, this application aims to address the following issues to overcome the shortcomings of existing solutions: 1. Insufficient high availability and stability: Existing platforms lack distributed architecture support and dynamic resource scheduling capabilities, making it difficult to cope with the high concurrency and multi-center deployment requirements of financial scenarios, resulting in insufficient system stability. 2. Poor model capability reusability: In traditional solutions, model capabilities are scattered and lack standardized component encapsulation, leading to repeated development for different business scenarios, increasing R&D costs and reducing efficiency. 3. Low model iteration efficiency: Existing platforms rely on manually labeled data and manual training processes, lacking an automated learning closed-loop mechanism, resulting in long model update cycles and difficulty in quickly responding to business changes. 4. Insufficient security protection capabilities: Traditional security strategies cannot achieve dynamic governance throughout the entire lifecycle, lacking real-time monitoring, rapid repair, and feedback training capabilities for model output, making it difficult to meet the compliance and security requirements of financial-grade systems. 5. Weak multimodal and multi-task collaborative capabilities: Existing technologies struggle to effectively integrate multimodal inputs such as text, voice, and images, and collaboratively process complex financial tasks, resulting in poor scenario adaptability.

[0046] This application provides an enterprise-level financial large-scale model platform, which is achieved through the organic integration of a data layer, a model layer, a component layer, and an application scenario layer. An automated learning closed-loop mechanism enables continuous and automatic optimization of each model, thereby providing robust model capabilities for the component layer. A component-based mechanism that integrates large and small model capabilities allows for the organic combination of the component and model layers, enabling flexible combinations of different components and models to meet diverse application scenarios, aiming to solve the aforementioned technical problems of existing technologies.

[0047] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will be described below with reference to the accompanying drawings.

[0048] Figure 2 This application provides a schematic diagram of the structure of an enterprise-level financial big data platform, as shown in the embodiments. Figure 2 As shown, the enterprise-level financial big data model platform 20 includes a data layer 201, a model layer 202, a component layer 203, and an application scenario layer 204.

[0049] The data layer 201 is used to store and manage multimodal data in financial business scenarios; the model layer 202 is used to coordinate and integrate the capabilities of models with different modalities and parameter counts, providing the component layer 203 and application scenario layer 204 with the capabilities of models with different modalities and parameter counts; the component layer 203 is used to encapsulate the capabilities of models with different modalities and parameter counts into standardized components; the application scenario layer 204 is used to call the standardized components corresponding to the task requirements based on the task requirements corresponding to the financial business scenarios, combined with multimodal data, to meet the task requirements corresponding to the financial business scenarios.

[0050] In one example, the enterprise-level financial big data model platform 20 also includes an infrastructure layer, which provides the operating environment and hardware facilities for the platform. The operations management layer is used to monitor, manage, and optimize the entire lifecycle of the enterprise-level financial big data model platform 20, as well as the continuous optimization and training of the models.

[0051] Specifically, the infrastructure layer, as the cornerstone of the enterprise-level financial big data platform, employs high-performance heterogeneous computing power, RoCE high-performance networks, and distributed high-speed cloud storage. Deployed based on a domestically developed cloud computing architecture, it provides the operating environment and hardware facilities, including network, computing power, and storage resources, for the platform's normal operation. Simultaneously, the infrastructure layer provides physical resources and drivers to the upper layers, serving as the fundamental support for the enterprise-level financial big data platform.

[0052] The data layer forms the data foundation of the enterprise-level financial big data model platform. It is responsible for storing and managing various types of data and providing data support for model training corpora. The data provided by the data layer includes structured data from the data platform and unstructured data from the unified knowledge base, covering both externally public and internally private data. This data is stored in a vectorized format using a unified vector library. Simultaneously, the data layer provides retrieval services.

[0053] For example, Figure 3 A logical architecture diagram of an enterprise-level financial big data platform provided in this application embodiment is shown below. Figure 3 As shown, the overall architecture of the enterprise-level financial big data platform is divided into six layers: infrastructure layer, data layer, model layer, component layer, application scenario layer, and operations management layer. The application scenario layer is used to configure solutions according to different financial business needs, providing end-to-end solutions for various financial business scenarios. The infrastructure layer provides the operating environment and hardware facilities for the normal operation of the platform, consisting of a RoCE (Remote Direct Memory Access Ethernet) high-performance network, high-performance heterogeneous computing power, and distributed high-speed cloud storage, building the basic framework of the enterprise-level financial big data platform based on cloud computing architecture.

[0054] The data layer manages the financial data of the enterprise-level financial big data model platform. It primarily includes a unified vector library for managing vectorized data, a data platform for managing structured text data, and a unified knowledge base for managing text data. The unified vector library mainly stores vectorized data, providing data support for the model layer. The data platform stores structured data generated during platform operation; this structured data can be used for subsequent model fine-tuning.

[0055] The model layer comprises models of different modalities and sizes. It consists of large and small model groups. The large model group comprises multi-domain (question-answering, code, conference, etc.), multi-size (7B, 14B, 32B, etc.), and multi-modal (text, image, audio, etc.) large models. The small model group typically consists of ASR (Automatic Speech Recognition), TTS (Text-to-Speech), and OCR (Optical Character Recognition) models with fewer parameters. Both large and small models include open-source models and fine-tuned models. These models provide interfaces with different capabilities to the component layer, which coordinates and uses them uniformly.

[0056] The component layer abstracts the different capability interfaces in the model layer into unified, reusable service components, providing services to the application scenario layer. These include general basic components, intelligent office components, intelligent R&D components, intelligent customer service components, multimodal components, and intelligent agents, among other service components tailored to different application scenarios. The service components provided by the component layer are not simply single interfaces from the model layer; rather, they combine large and small model capabilities according to different task requirements, thereby serving complex application scenarios in the financial field. By integrating and componentizing the large and small model capabilities of the model layer, complex financial tasks that traditional platforms struggle to accomplish with a single model can be solved. A complex scenario may require different component combinations, significantly improving the AI ​​application capabilities of enterprise-level financial large model platforms.

[0057] The operations management layer is responsible for monitoring, managing, and optimizing the platform throughout its entire lifecycle, as well as the continuous optimization and training of models. Enterprise-level financial large-scale model platforms not only require high security and reliability but also high upgrade continuity to meet the ever-changing needs of financial businesses. Therefore, the operations management layer consists of a security operations module and a training management module. The operations management layer, data layer, and model layer of the enterprise-level financial large-scale model platform form a dynamic security governance and automated learning closed-loop mechanism. Dynamic security governance primarily involves security review of the model's output, ensuring platform security through a closed-loop governance capability of "runtime monitoring -> rapid repair -> dynamic adjustment -> feedback training." The automated learning closed-loop mechanism, designed to meet the ever-changing needs of financial businesses, continuously learns and optimizes the models in the model layer through an automated closed loop of "business feedback -> data update -> model retraining -> automatic evaluation -> component upgrade -> scenario deployment optimization."

[0058] Optionally, the enterprise-level financial big data platform 20 adopts a distributed deployment architecture and uses a dynamic traffic scheduling strategy to allocate task requirements to the corresponding target nodes to ensure the continuity of financial business.

[0059] In one example, the enterprise-level financial big data platform adopts a multi-active, multi-center, off-site disaster recovery deployment architecture, supports load balancing, dual-active, disaster recovery and other mechanisms, and has the characteristics of high availability, high reliability and high scalability to ensure business continuity.

[0060] Specifically, taking the enterprise-level financial big data platform 20, which adopts a "dual-center in the same city + off-site disaster recovery" architecture as an example, Figure 4 This application provides a deployment architecture diagram of an enterprise-level financial big data platform, as shown in the embodiments below. Figure 4As shown, the deployment architecture of the enterprise-level financial big data model platform includes: (1) Load balancing. Gateways and proxy servers are deployed in the DMZ (Demilitarized Zone), and load balancing proxy functions are implemented using SLB server load balancing technologies such as Nginx proxy (high-performance web service and proxy engine). In the event of a disaster at the A center data center, the data storage service will switch to the slave database, the service currently connected to the A center will be interrupted, and the B center will continue to provide services. (2) Disaster recovery in other locations. When the service of the two centers in the same city is abnormal, the request traffic will be switched to the disaster recovery in other locations (C center) to ensure that the overall system can provide services normally. (3) Big data model services. Relying on the existing big data model and training fine-tuning technology, core big data model services such as knowledge Q&A, report generation, meeting transcription, code generation, work order summary, document parsing, data analysis, and script generation are provided. (4) Data storage. MySQL (Relational Database Management System) stores the underlying data, MongoDB (a distributed file-based database) stores vectorized knowledge data, and Redis (Remote Dictionary Server) stores user information and large-scale model session permission information. A dual-center deployment (one master and one slave) ensures high availability through master-slave synchronization. Multimodal and other unstructured data are stored in OSS (Object Storage Service), leveraging its high durability and reliability to meet data usage requirements.

[0061] In another example, a multi-active, multi-center architecture is deployed, distributing the infrastructure resources (such as computing power and storage) of the enterprise-level financial big data platform across multiple data centers, each with independent service capabilities. When a user requests access to the platform, load balancing technology dynamically allocates request traffic based on the real-time load status of each data center (such as CPU utilization and network latency), ensuring resource utilization and response efficiency in high-concurrency scenarios. Simultaneously, master-slave synchronization technology synchronizes critical business data (such as user session information and model parameters) in real time through a data replication mechanism between the master and slave data centers (such as database master-slave replication), ensuring data consistency across data centers. If a data center experiences service disruptions due to hardware failure or network interruption, load balancing technology automatically switches traffic to other functioning data centers, while master-slave synchronization technology ensures rapid data recovery after the disrupted data center recovers, thus maintaining business continuity.

[0062] The multi-active, multi-center architecture is a distributed architecture where multiple data centers simultaneously provide services, each with independent request processing capabilities. For example, a "dual-center within the same city + disaster recovery in a different location" architecture involves two data centers within the same city and one data center in a different location, each with independent service capabilities. Load balancing technologies include dynamically adjusting user requests to different data centers using traffic allocation algorithms (such as Nginx proxy SLB) to optimize resource utilization and system stability. For instance, traffic can be dynamically allocated based on metrics such as CPU (Central Processing Unit) utilization and network latency in each data center to avoid overloading a single center. Master-slave synchronization technology achieves data consistency between the master and slave data centers through data replication mechanisms (such as MySQL master-slave replication). For example, database update operations in the master data center are synchronized to the slave data center in real time to ensure data consistency.

[0063] The dynamic traffic allocation algorithm dynamically adjusts the distribution of user request traffic by collecting real-time load data (such as CPU utilization and network latency) from each data center and combining it with preset load balancing strategies (such as weighted round-robin and minimum connection count). For example, when the CPU utilization of a data center exceeds a set threshold, the algorithm switches some request traffic to a data center with a lower load to prevent service degradation due to overload. Simultaneously, the algorithm dynamically adjusts traffic paths based on network latency, prioritizing the allocation of requests to data centers with lower network latency, thereby optimizing the overall system response efficiency.

[0064] By deploying multiple computing nodes in a distributed architecture, the risk of single points of failure is eliminated, ensuring the continued operation of financial services even in the event of node failure. A high-speed data synchronization mechanism guarantees cross-node data consistency, avoiding business anomalies caused by data inconsistency. A dynamic traffic scheduling strategy allocates requests based on real-time load status, improving the overall resource utilization of the system. Through the synergistic effect of these three elements, the financial system achieves business continuity assurance under extreme failure scenarios, while simultaneously meeting the load balancing requirements of high-concurrency scenarios, significantly improving system stability and disaster recovery capabilities.

[0065] In another example, Figure 5 A security architecture diagram of an enterprise-level financial big data platform is provided for embodiments of this application, such as... Figure 5 As shown, the enterprise-level financial big data model platform realizes a multi-dimensional and comprehensive information security protection system for financial applications, covering four aspects: infrastructure security, data security, model security, and application security.

[0066] Specifically, infrastructure security includes physical security, network security, host security, and monitoring response, protecting infrastructure security from the perspectives of systems, networks, and data. Data security protects data security from four dimensions: basic measures, transmission security, data management, and data processing, including mechanisms such as identity authentication, access control, token verification, keys, data version control, privacy protection, and data anonymization. Model security includes content security and process security. Content security utilizes mechanisms such as illusion mitigation and data detoxification to ensure security, while process security is responsible for the secure management of model inputs and outputs, using ten processes including model registration, model verification, model review, model approval, model monitoring, and model optimization to strictly control and ensure process security. In terms of application security, mechanisms such as identity recognition and authentication, network communication encryption, client sandboxing, virtual private networks, and client hardening ensure application security, achieving full-channel, full-link, and full-lifecycle information security protection to meet the security requirements of financial-grade systems. Through data security and model security, a dynamic security governance method is achieved that spans the entire lifecycle of large-scale model data engineering, model training, and model inference.

[0067] In another example, the operations management layer includes a security operations module and a training management module. The security operations module ensures the platform's stability and security by employing a multi-layered management and monitoring mechanism to promptly detect and handle risks and anomalies during operations. This module further includes an operations analysis unit and a security management unit. The training management module monitors, manages, and optimizes the platform throughout its entire lifecycle, as well as continuously optimizes and trains the model. It includes a unified model training unit and a unified corpus unit.

[0068] Specifically, the operations analysis unit is used to evaluate various models used in the enterprise-level financial big data modeling platform at preset time intervals, ensuring the accuracy and effectiveness of the models. The operations analysis unit also manages plugins uniformly, ensuring their security and compatibility. Furthermore, it monitors the operational status of the enterprise-level financial big data modeling platform in real time; when anomalies are detected, it triggers an alarm mechanism to notify relevant personnel for handling. Finally, at a second preset time interval, the operations analysis unit generates operational data reports, providing decision-makers with a comprehensive analysis of the operational situation.

[0069] Security management is used to establish a sensitive keyword database for sensitive information involved in the enterprise-level financial big data model platform, and to identify and filter multimodal data to prevent the leakage of sensitive information. Security management is also used to establish a rapid response mechanism for faults and problems occurring on the enterprise-level financial big data model platform to ensure that problems are resolved in a timely manner. Security management is also used to review the content published on the enterprise-level financial big data model platform to ensure the legality and compliance of the content and prevent the spread of harmful information.

[0070] The unified model training unit supports various training modes, including model fine-tuning, model retraining, reinforcement learning, model augmentation, and model distillation, to improve the model's generalization ability. The unified model training unit is also used to fine-tune the model for specific tasks or domains to improve its performance in specific scenarios.

[0071] The Unified Corpus Unit has the capability to effectively manage and maintain training data, ensuring the high quality and diversity of multimodal data. The Unified Corpus Unit also utilizes automated tools and algorithms to collect and edit corpus data, improving the efficiency and accuracy of data processing.

[0072] in, Figure 6 A flowchart illustrating the overall interaction process of a large model platform provided in this application embodiment is shown below. Figure 6 As shown, for external calls, the calling channels include user access via PC (personal computer), mobile devices, and SDK (Software Development Kit), as well as scheduled batch tasks. Calls are made to the application component cluster through API (Application Programming Interface) interfaces, and through hardware or software load balancing mechanisms, calls are made to six major application components. The component layer calls the large model, small model, and vector inference interfaces to compute the cluster and return the results to the user. Internally, the enterprise-level financial large model platform collects and stores data in a database cluster. The operation analysis unit and security management unit monitor the model's running status and the cluster's status in real time. Once sufficient training data is collected, automated model training and evaluation are performed, and the data is then deployed to the model cluster to form a closed-loop knowledge feedback mechanism.

[0073] In one implementation scenario, taking the intelligent customer service scenario as an example, Figure 7 A flowchart of a work order classification and summary scenario provided in this application embodiment is shown below. Figure 7 As shown, in the interaction process between the enterprise-level financial big data model platform and external systems, the platform integrates core functions such as work order summarization and classification. Through the work order summary function, the platform automatically extracts and generates concise work order summaries, facilitating archiving and subsequent operations. The work order classification function accurately categorizes work order content, significantly improving processing efficiency. Through the customer service system of the enterprise-level financial big data model platform, the platform's capabilities are utilized for knowledge-based Q&A, element extraction, work order classification, and work order summarization, demonstrating the interaction process between users, agents, the customer service system, and the enterprise-level big data model platform.

[0074] The enterprise-level financial big data model platform provided in this embodiment integrates interfaces for models from different domains, modalities, and sizes through modularization and fusion of large and small model capabilities. This improves the development efficiency of financial applications, reduces development costs, and shortens development cycles, providing technical support for commercial banks to rapidly build financial application scenarios at scale. Through an automated learning mechanism, the platform continuously optimizes model capabilities based on actual business feedback, achieving continuous optimization of model performance. This enhances the intelligence and compliance of financial applications and significantly reduces the labor and operational costs for commercial banks.

[0075] Optionally, models with different modalities and different parameter quantities include large models and small models. Among them, large models have multimodal or multi-domain processing capabilities, while small models are used to solve specific tasks. Component layer 203 is specifically used to encapsulate the capabilities of large models and small models to obtain standardized components corresponding to large models and standardized components corresponding to small models, respectively.

[0076] In one example, standardized encapsulation interfaces modularize the capabilities of large and small models. For instance, the OCR small model is encapsulated as an independent component, while the text summarization capability of the large model is encapsulated as a general-purpose basic component. The large model refers to an AI (artificial intelligence) model with a large number of parameters and multimodal or multi-domain processing capabilities, such as question-answering or code-based models. For example, a large model could include a 32-parameter large model for generating work order summaries. The small model refers to an AI model with a small number of parameters that focuses on specific tasks, such as OCR or ASR models. For example, a small model could include an OCR small model for extracting text from images. Componentized service interfaces are used to encapsulate model capabilities into independently callable modular interfaces, supporting standardized access. For example, component interfaces provided through SDKs or APIs.

[0077] By standardizing and encapsulating interfaces, flexible combinations of different model capabilities are achieved. This not only enhances component reusability but also improves the development efficiency, reduces development costs, and shortens development cycles of financial applications by integrating interfaces of different domains, modalities, and sizes through the fusion of large and small model capabilities into modular components. This lays a technological foundation for commercial banks to rapidly build financial application scenarios at scale.

[0078] Optionally, application scenario layer 204 is specifically used to generate service interfaces by combining different standardized components corresponding to task requirements according to task requirements through a dynamic decision-making routing mechanism.

[0079] In one example, application scenario layer 204 is specifically used to automatically select component combinations and generate service interfaces based on task requirements (such as document parsing tasks) through a dynamic decision-making routing mechanism. For example, a document parsing task needs to call an OCR component to extract text, a multimodal component to recognize image features, and a general basic component to generate a summary, ultimately providing services externally through a standardized interface. Dynamic decision-making routing is a mechanism that automatically selects component combinations based on the characteristics of task requirements (such as input type and complexity) corresponding to the financial business scenario. For example, the document parsing task calls an OCR + multimodal component combination. By decoupling the capabilities of large and small models into independent components, modular encapsulation of model capabilities is achieved. In the financial scenario, user-input multimodal data (such as text, voice, and images) and task requirements (such as work order classification) are parsed by the dynamic routing mechanism, and the system selects the optimal component combination based on task characteristics. For example, a document parsing task needs to call an OCR component to extract text from images, a multimodal component to recognize image features, and a general basic component to generate summary text. All components provide services externally through standardized interfaces, supporting rapid integration and deployment of financial business systems. This method reduces model invocation complexity through component-based design and optimizes component combination through dynamic routing mechanism, thereby meeting the diverse task processing needs in financial scenarios.

[0080] By optimizing component combinations through dynamic routing at the application scenario layer, redundant calls are reduced, thereby further shortening the development cycle of financial scenarios and reducing system coupling. Furthermore, a dynamic decision-making routing mechanism selects the optimal component combination based on task characteristics, reducing redundant calls and improving processing efficiency.

[0081] Optionally, the platform also includes an operations management layer. The operations management layer is specifically used to acquire the output data of models with different modalities and different parameter amounts, as well as the corresponding feedback data. The feedback data is then fed back to the training corpus of the data layer. When the feedback data reaches a preset threshold, the platform triggers an incremental fine-tuning task for models with different modalities and different parameter amounts and generates an optimized model image.

[0082] In one example, the operations management layer is specifically used to collect the model's output data and corresponding feedback data in real time after generating componentized service interfaces, and then feed the feedback data back to the training corpus. For example, in an intelligent customer service scenario, user complaint data is automatically identified as feedback data and used to trigger incremental model fine-tuning. When the feedback data reaches a preset threshold, the operations management layer initiates the fine-tuning process, generating an optimized model image to provide update support for subsequent componentized service interfaces. The training corpus stores the dataset used for model training. For example, it may include a unified vector library containing work order classification and annotation data. Incremental model fine-tuning involves updating the local parameters of an existing model based on new data. For example, a risk assessment model adjusts its risk evaluation logic based on the latest market fluctuation data.

[0083] Continuous model optimization is achieved through an automated learning closed-loop mechanism. Feedback data triggers model fine-tuning tasks to improve model accuracy. Business feedback drives model iteration, reducing manual intervention and further improving model performance in specific scenarios, thus supporting dynamic changes in financial business needs.

[0084] Figure 8 This application provides a flowchart illustrating an enterprise-level financial big data model platform management method, applicable to the aforementioned enterprise-level financial big data model platform. The method includes a data layer, a model layer, a component layer, and an application scenario layer, as shown below. Figure 8 As shown, the method includes:

[0085] S801. Obtain models with different modalities and different parameter quantities through the model layer.

[0086] S802, to synergistically integrate the capabilities of models with different modalities and different parameter quantities.

[0087] S803. Through the component layer, the capabilities of models with different modalities and different parameter quantities are encapsulated into standardized components.

[0088] S804. Based on the task requirements corresponding to the financial business scenario, and combined with the input multimodal data, the standardized components corresponding to the task requirements are called through the application scenario layer to meet the task requirements corresponding to the financial business scenario.

[0089] In one example, Figure 9 A flowchart of a component-based task execution method for large and small model fusion capability is provided in an embodiment of this application, such as... Figure 9 As shown, the capabilities of large and small models are integrated to build a unified and reusable component. The component supports the coordination and cooperation between large and small models, achieving unified scheduling and result fusion.

[0090] After the API request reaches the scene layer, multi-dimensional decision information is obtained, including the content of the request, the function of the components, the effect and performance of the model, and the machine load. Through a dynamic decision routing mechanism, the components corresponding to the task requirements are determined to coordinate and cooperate to complete the request. Different components can call small models, large models, or small models + large models. After the component layer calls the model interface, the results are uniformly processed and returned to the front end. The feedback data is saved to the database for subsequent model and strategy optimization.

[0091] In one implementation scenario, Figure 10 A flowchart for document parsing and execution is provided for embodiments of this application, such as... Figure 10As shown, taking a document parsing scenario as an example, by analyzing the user's request, the routing decision determines which components to call to complete the processing. First, the intelligent office component is called to extract images from the document. Then, the multimodal component is used to recognize and extract text information from the images, and accurately classify, recognize, and extract features of various objects, graphics, and other elements in the images. Next, the intelligent data component is called to obtain data information related to the question. Finally, the general basic component is called to process the complete user query, ultimately achieving a comprehensive understanding and parsing of the user-uploaded file content.

[0092] The enterprise-level financial large model platform management method provided in this embodiment dynamically combines models with different parameters and modalities by integrating the capabilities of large and small models into a component-based approach, thereby forming enterprise-level reusable financial large model application components and effectively solving the problems of insufficient model capability scalability and reusability.

[0093] Optionally, the method also includes: acquiring the output data of models with different modalities and different parameter amounts, as well as the corresponding feedback data, and feeding the feedback data back to the training corpus; when the feedback data reaches a preset threshold, triggering an incremental fine-tuning task for models with different modalities and different parameter amounts and generating an optimized model image.

[0094] In one example, Figure 11 A flowchart illustrating an automated learning closed-loop mechanism for a model, as provided in this application embodiment, is shown below. Figure 11 As shown, an automated closed-loop mechanism of "automated continuous learning" is implemented to achieve "business feedback -> data update -> model retraining -> automatic evaluation -> component upgrade -> scenario deployment optimization". Specifically, this includes: 1. Introducing a "model decision deviation monitoring module" in the operations management layer to monitor the difference between model output and manual correction in real time; 2. Feedback of the discrepancy data to the data layer, stored in the specific business database of the unified vector library; 3. Monitoring the data volume in the model training layer. When the set threshold is reached, the data is pulled from the production environment to the unified corpus in the laboratory environment, and the model incremental fine-tuning task is initiated. After the model training is completed, the model image is exported to the test environment and automatically evaluated using the evaluation process; 4. When the model's effect and performance meet the requirements, the model image is pushed to the image repository in the production environment, and a canary release is performed to the model layer. Simultaneously, model monitoring and self-learning are conducted, and the model capabilities are used in the component layer; 5. Finally, seamless switching and optimization of model capabilities are achieved at the scenario layer, forming a "perception-learning-evolution" closed loop to meet constantly changing business needs and scenarios, and reduce product iteration costs.

[0095] For example, an enterprise-level financial big data model platform provides complete automated model evaluation, including the following three core stages: 1. Evaluation Dimension Determination: Evaluation dimensions are automatically selected based on the target model's architectural characteristics and application domain. Each dimension includes both automated and manual evaluation. Models under evaluation are assessed on specific dimensions due to their respective capabilities and applications in different domains. General capability evaluation and system performance evaluation are performed on basic architecture models, while financial capability evaluation and RAG (Retrieval Enhancement Generation) capability evaluation are performed on financial vertical domain models.

[0096] 2. Assessment Data Configuration: Select the assessment dataset from the defined dimensions. The specific content of the assessment dataset is shown in Table 1 below. Table 1 includes the source, dimensions, and dataset types of the assessment datasets. For example, the general ability dimension includes datasets for language, text understanding, logical reasoning, code generation, mathematical reasoning, and long text. Improve the prompt word project by modifying the prompt words according to the assessment task configuration. Determine the answer extraction project by extracting answers from the model's original responses based on the prompt words, while removing invalid formatted content.

[0097] Table 1

[0098]

[0099] 3. Evaluation Result Output: Figure 12 A flowchart for automatic model evaluation provided in this application embodiment, such as Figure 12 As shown, after the model completes the evaluation inference process, the automatic evaluation selects an evaluator (accuracy, Rouge, EM, etc.) based on the task type, while the manual evaluation is handed over to humans, who score online based on scoring criteria (such as adoption rate and fluency). After the evaluation results are evaluated, an evaluation report is generated, which mainly includes three aspects: scores for each capability dimension and performance on the specific evaluation dataset; a multi-dimensional radar chart visualization report; and scenario adaptation suggestions and business scenario matching scores.

[0100] In another example, Figure 13 A dynamic security governance flowchart is provided for embodiments of this application, such as Figure 13 As shown, this is a dynamic security governance method that spans the entire lifecycle of the large model. The specific process is as follows:

[0101] ① Alignment during training: Continuously fine-tune the model and use reinforcement learning to align values. The reward model is incrementally updated weekly, and trained using continuously new anomalous data.

[0102] ② Quick Problem Repair: Establish a "quick repair channel" for problem response, supporting minute-level policy updates. When monitoring detects that the large model outputs inappropriate content (such as customer privacy information), an alarm is immediately triggered, and the request is automatically truncated, returning the specified answer. When a human receives the alarm, they will configure the correct answer for the specified and similar problems in real time. Subsequent questions will directly use the configured correct answer to prevent the problem from spreading.

[0103] ③ Protection during use: Enables dynamic adjustment and version control of prompt templates. The prompt template can be dynamically adjusted based on the user's role, and multiple prompts can be compared across three dimensions: security, accuracy, and user satisfaction, to select the optimal prompt.

[0104] Through a closed-loop mechanism for automated continuous learning of large models, driven by the performance of large models in real business scenarios, the large models are dynamically and automatically updated and trained to achieve continuous optimization of model performance and improve the intelligence and compliance level of financial applications.

[0105] Figure 14 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 14 As shown, the electronic device 140 may include a memory 1401 and a processor 1402. Optionally, the electronic device may also include a transceiver 1403, wherein the memory 1401 and the processor 1402 communicate; for example, the memory 1401, the processor 1402 and the transceiver 1403 may communicate via a communication bus 1404, the memory 1401 is used to store a computer program, and the processor 1402 executes the computer program to implement the method of the above embodiments.

[0106] Optionally, the aforementioned processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps in the method embodiments disclosed in this application can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0107] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the methods in any of the above method embodiments.

[0108] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the methods in any of the above method embodiments.

[0109] All or part of the steps in the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a readable memory. When the program is executed, it performs the steps of the above method embodiments; and the aforementioned memory (storage medium) includes: read-only memory (ROM), RAM, flash memory, hard disk, solid-state drive, magnetic tape, floppy disk, optical disk, and any combination thereof.

[0110] This application describes embodiments with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processing unit of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processing unit of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0111] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0112] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0113] Obviously, those skilled in the art can make various modifications and variations to the embodiments of this application without departing from the spirit and scope of this application. Therefore, if these modifications and variations to the embodiments of this application fall within the scope of the claims of this application and their equivalents, this application also intends to include these modifications and variations.

[0114] In this application, the term "comprising" and its variations can refer to non-limiting inclusion; the term "or" and its variations can refer to "and / or". The terms "first", "second", etc., in this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. In this application, "multiple" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. The character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0115] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to this application.

[0116] It should be further noted that although the steps in the flowchart are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowchart may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0117] It should be understood that the above-described device embodiments are merely illustrative, and the device of this application can also be implemented in other ways. For example, the division of units / modules in the above embodiments is only a logical functional division, and there may be other division methods in actual implementation. For example, multiple units, modules, or components may be combined, or integrated into another system, or some features may be ignored or not executed.

[0118] Furthermore, unless otherwise specified, the functional units / modules in the various embodiments of this application can be integrated into one unit / module, or each unit / module can exist physically separately, or two or more units / modules can be integrated together. The integrated units / modules described above can be implemented in hardware or as software program modules.

[0119] When integrated units / modules are implemented in hardware, the hardware can be digital circuits, analog circuits, etc. The physical implementation of the hardware structure includes, but is not limited to, transistors, memristors, etc. Unless otherwise specified, the processor can be any suitable hardware processor, such as a CPU, GPU, FPGA, DSP, and ASIC, etc. Unless otherwise specified, the storage unit can be any suitable magnetic or magneto-optical storage medium, such as Resistive Random Access Memory (RRAM), Dynamic Random Access Memory (DRAM), Static Random Access Memory (SRAM), Enhanced Dynamic Random Access Memory (EDRAM), High-Bandwidth Memory (HBM), Hybrid Memory Cube (HMC), etc.

[0120] If the integrated unit / module is implemented as a software program module and sold or used as an independent product, it can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned memory includes various media capable of storing program code, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard drive, magnetic disk, or optical disk.

[0121] In the above embodiments, the descriptions of each embodiment have their own emphasis. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments. The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0122] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the following claims.

[0123] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. An enterprise-level financial large-scale model platform, characterized in that, The platform includes a data layer, a model layer, a component layer, and an application scenario layer. The data layer is used to store and manage multimodal data in financial business scenarios; The model layer is used to collaboratively integrate the capabilities of models with different modalities and different parameter quantities, providing the component layer and the application scenario layer with the capabilities of models with different modalities and different parameter quantities; The component layer is used to encapsulate the capabilities of models with different modalities and different parameter quantities into standardized components; The application scenario layer is used to call standardized components corresponding to the task requirements based on the task requirements corresponding to the financial business scenario, combined with the multimodal data, so as to meet the task requirements corresponding to the financial business scenario.

2. The platform according to claim 1, characterized in that... The models with different modalities and different parameter quantities include large models and small models. The large model has multimodal or multi-domain processing capabilities, and the small model is used to solve specific tasks. The component layer is specifically used to encapsulate the standardized components corresponding to the large model and the standardized components corresponding to the small model based on the capabilities of the large model and the small model, respectively.

3. The platform according to claim 2, characterized in that, The application scenario layer is specifically used to generate service interfaces by combining different standardized components corresponding to the task requirements according to the task requirements through a dynamic decision-making routing mechanism.

4. The platform according to claim 3, characterized in that, The platform also includes an operation management layer, which is specifically used to acquire the output data of the models with different modalities and different parameter quantities, as well as the feedback data corresponding to the output data, and to feed the feedback data back to the training corpus of the data layer. When the feedback data reaches a preset threshold, an incremental fine-tuning task for the model with different modalities and different parameter values ​​is triggered, and an optimized model image is generated.

5. The platform according to claim 1, characterized in that, The platform adopts a distributed deployment architecture and uses a dynamic traffic scheduling strategy to allocate the task requirements to the corresponding target nodes to ensure the continuity of financial business.

6. A management method for an enterprise-level financial large-scale model platform, characterized in that, Applied to the enterprise-level financial big data model platform as described in any one of claims 1 to 5, the platform comprising a data layer, a model layer, a component layer, and an application scenario layer, the method comprises: Through the model layer, models with different modalities and different parameter quantities can be obtained; The capabilities of models with different modalities and different parameter values ​​are synergistically integrated; Through the component layer, the capabilities of models with different modalities and different parameter quantities are encapsulated into standardized components; Based on the task requirements corresponding to the financial business scenario, and combined with the input multimodal data, the standardized components corresponding to the task requirements are called through the application scenario layer to meet the task requirements corresponding to the financial business scenario.

7. The method according to claim 6, characterized in that, The method further includes: Obtain the output data of the models with different modalities and different parameter quantities, as well as the corresponding feedback data, and feed the feedback data back to the training corpus. When the feedback data reaches a preset threshold, an incremental fine-tuning task for the model with different modalities and different parameter values ​​is triggered, and an optimized model image is generated.

8. An electronic device, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the method as described in any one of claims 6 to 7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 6 to 7.

10. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the method of any one of claims 6 to 7.

Images