Enterprise credit risk identification method and device and computer readable storage medium

Abstract

The invention provides an enterprise credit risk identification method and device and a computer readable storage medium. According to the scheme, a large language model and a training sample are obtained, the training sample is used for pre-training the large language model, a pre-trained initial large language model is obtained, and the training sample comprises financial data and label information corresponding to the financial data; applying the fine tuning task to the hybrid expert model to perform fine tuning on the initial large language model to obtain a fine-tuned large language model; optimizing the fine-tuned large language model through a preset prediction task to obtain a target large language model; and inputting financial data of the enterprise to be subjected to credit risk identification into the target large language model to obtain a credit risk identification result of the enterprise to be subjected to credit risk identification. According to the scheme, the problem of low accuracy of enterprise credit risk identification in the prior art is solved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology. Specifically, it relates to a method, device, computer-readable storage medium, and electronic device for identifying enterprise credit risks. Background Art

[0002] In the financial field, especially for small and micro enterprises, traditional methods and technologies face significant challenges and limitations in the credit risk assessment of enterprises. The financial information of small and micro enterprises is often opaque and non-standard, and coupled with the high uncertainty of their operations, it makes the credit risk assessment complex and difficult to be precise. Traditional risk assessment methods, such as expert experience judgment, financial ratio analysis, cash flow prediction, etc., although can reveal risks to a certain extent, when dealing with the unique non-financial information of small and micro enterprises and the rapidly changing market environment, these methods usually cannot effectively integrate unstructured information, such as public opinion analysis, industry dynamics, etc., and these information are crucial for a comprehensive assessment of credit risk.

[0003] In recent years, with the development of artificial intelligence and big data technologies, attempts have been made to use machine learning and data mining technologies to improve credit risk assessment. For example, by constructing algorithm models to analyze the historical financial data and credit records of small and micro enterprises, combined with market public opinion, to predict default risks in a more objective and comprehensive manner. However, these models often rely on features and rules manually designed by experts, with limited generalization ability, and are difficult to adapt to the specific situations of different industries and regions. Especially when dealing with non-financial information, their accuracy and timeliness still need to be improved. In addition, although the large model technology based on chatGPT shows powerful natural language processing capabilities, in the application in the financial field, due to the lack of in-depth training for specific fields such as enterprise fund transactions, its professionalism and accuracy in credit risk identification are still in doubt.

[0004] All in all, there is a lack of a solution in the prior art that can both understand complex business environments and accurately predict credit risks, thereby improving the accuracy of enterprise credit risk identification. Summary of the Invention

[0005] The main purpose of this application is to provide a method, device, computer-readable storage medium, and electronic device for identifying enterprise credit risks, so as to at least solve the problem of low accuracy in enterprise credit risk identification in the prior art.

[0006] To achieve the above object, according to one aspect of the present application, a method for identifying enterprise credit risk is provided, including: obtaining a large language model and training samples, and using the training samples to pre-train the large language model to obtain a pre-trained initial large language model, where the training samples include financial and financial data and label information corresponding to the financial and financial data, and wherein the large language model adopts a Decode-only architecture; applying a fine-tuning task to a mixture-of-experts model to fine-tune the initial large language model to obtain a fine-tuned large language model, where the fine-tuning task is determined based on a fine-tuning task dataset; optimizing the fine-tuned large language model through a preset prediction task to obtain a target large language model; and inputting the financial and financial data of the enterprise to be identified for credit risk into the target large language model to obtain the credit risk identification result of the enterprise to be identified for credit risk.

[0007] Optionally, the mixture-of-experts model is a mixture-of-experts model based on Monte Carlo routing. Applying the fine-tuning task dataset to the mixture-of-experts model to fine-tune the initial large language model to obtain a fine-tuned large language model includes: obtaining the fine-tuning task dataset and setting a plurality of fine-tuning tasks according to the fine-tuning task dataset; and based on the plurality of fine-tuning tasks, using the mixture-of-experts model based on Monte Carlo routing to fine-tune the initial large language model to obtain the fine-tuned large language model.

[0008] Optionally, based on the plurality of fine-tuning tasks, using the mixture-of-experts model based on Monte Carlo routing to fine-tune the initial large language model to obtain the fine-tuned large language model includes: a sampling step, where the mixture-of-experts model based on Monte Carlo routing uses Monte Carlo routing technology to dynamically select and activate a target expert model, where the target expert model is an expert model whose relevance to the current fine-tuning task is greater than a preset relevance threshold; a determination step, determining the weights of each of the target expert models through the Monte Carlo routing technology; and repeatedly executing the sampling step and the determination step until a preset stop condition is reached to obtain the fine-tuned large language model.

[0009] Optionally, optimizing the fine-tuned large language model through a preset prediction task to obtain a target large language model includes: a setting step, setting the prediction task, where the output of the prediction task is numerical; a prediction step, inputting the prediction task into the fine-tuned large language model for prediction to obtain a prediction result; an adjustment step, adjusting the parameters of the fine-tuned large language model according to the difference between the prediction result and the actual result; and repeatedly executing the prediction step and the adjustment step until a preset optimization criterion is reached to obtain the target large language model.

[0010] Optionally, after optimizing the fine-tuned large language model through a preset prediction task to obtain a target large language model, the method further includes: setting and storing corresponding Prompt guiding words based on different financial activity scenarios, where the financial activity scenario is a specific operation environment set according to different business stages and requirements including pre-loan management, in-loan management, and post-loan management; parsing credit risk-related questions according to the Prompt guiding words to generate decision texts that meet the requirements of different financial activity scenarios.

[0011] Optionally, obtaining a large language model and training samples, and pre-training the large language model with the training samples to obtain a pre-trained initial large language model, includes: scraping the training samples from multiple data sources, where the data sources include news social media, regulatory announcements and filings, business exception announcements, and financial professional reports; pre-training the large language model with the training samples to obtain the pre-trained initial large language model, so that the initial large language model learns the relevant information of the training samples.

[0012] Optionally, before obtaining the large language model and training samples, the method further includes: evaluating the financial and accounting data using a preset evaluation method to obtain the label information of the financial and accounting data, where the preset evaluation method includes a scoring card model and a GPT-based model.

[0013] According to another aspect of the present application, there is provided an apparatus for identifying enterprise credit risk, including: a pre-training unit configured to obtain a large language model and training samples, and pre-train the large language model with the training samples to obtain a pre-trained initial large language model, where the training samples include financial and accounting data and label information corresponding to the financial and accounting data, and the large language model adopts a Decode-only architecture; a fine-tuning unit configured to apply a fine-tuning task to a mixture of experts model to fine-tune the initial large language model to obtain a fine-tuned large language model, where the fine-tuning task is determined based on a fine-tuning task data set; an optimization unit configured to optimize the fine-tuned large language model through a preset prediction task to obtain a target large language model; and an input unit configured to input the financial and accounting data of an enterprise to be identified for credit risk into the target large language model to obtain the credit risk identification result of the enterprise to be identified for credit risk.

[0014] According to yet another aspect of the present application, there is provided a computer-readable storage medium, where the computer-readable storage medium includes a stored program, and when the program runs, it controls the device where the computer-readable storage medium is located to execute any one of the methods for identifying enterprise credit risk.

[0015] According to another aspect of the present application, there is provided an electronic device, including: one or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the one or more programs include methods for performing any one of the enterprise credit risk identification methods.

[0016] Applying the technical solution of the present application, by using training samples containing financial and financial data and their tags, pre-training and fine-tuning a large language model adopting a Decode-only architecture, wherein the fine-tuning is based on a specific financial scenario task dataset, and then optimizing through a preset prediction task, finally obtaining a target large language model; inputting the financial and financial data of an enterprise into this model, thus solving the problem of low accuracy in enterprise credit risk identification in the prior art. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The accompanying drawings forming a part of this application are used to provide a further understanding of this application. The schematic embodiments of this application and their descriptions are used to explain this application and do not constitute an improper limitation to this application. In the drawings:

[0018] Figure 1 It shows a hardware structure block diagram of a mobile terminal for performing a method of enterprise credit risk identification provided in an embodiment of the present application;

[0019] Figure 2 It shows a schematic flowchart of a method of enterprise credit risk identification provided in an embodiment of the present application;

[0020] Figure 3 It shows a structural diagram of a hybrid expert model based on Monte Carlo routing for a method of enterprise credit risk identification provided in an embodiment of the present application;

[0021] Figure 4 It shows a schematic diagram of the Monte Carlo routing principle for a method of enterprise credit risk identification provided in an embodiment of the present application;

[0022] Figure 5 It shows a structural block diagram of a device for enterprise credit risk identification provided in an embodiment of the present application;

[0023] Figure 6 It shows an overall flowchart of a system for enterprise credit risk identification provided in an embodiment of the present application;

[0024] Figure 7 It shows a schematic diagram of a real-time data engineering pipeline module of a system for enterprise credit risk identification provided in an embodiment of the present application;

[0025] Figure 8 The internal schematic diagram of the credit risk decision-making module of an enterprise credit risk identification system provided according to an embodiment of the present application is shown.

[0026] Among them, the above-mentioned drawings include the following reference numerals:

[0027] 102, processor; 104, memory; 106, transmission device; 108, input / output device. Detailed implementation manners

[0028] It should be noted that, without conflict, the embodiments in the present application and the features in the embodiments may be combined with each other. The present application will be described in detail below with reference to the drawings and in combination with the embodiments.

[0029] In order to enable those skilled in the art to better understand the solution of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application.

[0030] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and do not necessarily need to be used to describe a specific order or sequence. It should be understood that such used data can be interchanged under appropriate circumstances so as to describe the embodiments of the present application here. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device including a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0031] As introduced in the background art, although the existing enterprise credit risk identification solutions based on large models have demonstrated powerful natural language processing capabilities, in the application in the financial field, due to the lack of in-depth training in specific fields such as enterprise fund transactions, the accuracy of credit risk identification is lacking. To solve the problem of low accuracy of enterprise credit risk identification in the existing technology, the embodiments of the present application provide a method, device, computer-readable storage medium and electronic device for enterprise credit risk identification.

[0032] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention.

[0033] The method embodiments provided in the embodiments of the present application can be executed on a mobile terminal, a computer terminal, or a similar computing device. Taking running on a mobile terminal as an example, Figure 1 is a hardware structure block diagram of a mobile terminal for a method of identifying enterprise credit risk according to an embodiment of the present invention. As Figure 1 shown, the mobile terminal may include one or more ( Figure 1 only one is shown in the figure) processors 102 (the processors 102 may include, but are not limited to, processing devices such as a microprocessor MCU or a field programmable gate array FPGA), and a memory 104 for storing data. Among them, the above-mentioned mobile terminal may further include a transmission device 106 for communication functions and an input / output device 108. Those of ordinary skill in the art can understand that Figure 1 the structure shown is only schematic and does not limit the structure of the above-mentioned mobile terminal. For example, the mobile terminal may further include more or fewer components than Figure 1 shown in the figure, or have a different configuration from Figure 1 shown in the figure.

[0034] The memory 104 can be used to store computer programs. For example, software programs and modules of application software, such as the computer program corresponding to the display method of device information in the embodiments of the present invention. The processor 102 executes various functional applications and data processing by running the computer program stored in the memory 104, that is, implements the above-mentioned method. The memory 104 may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memories, or other non-volatile solid-state memories. In some instances, the memory 104 may further include a memory remotely disposed relative to the processor 102, and these remote memories may be connected to the mobile terminal through a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an enterprise intranet, a local area network, a mobile communication network, and their combinations. The transmission device 106 is used to receive or send data via a network. Specific examples of the above-mentioned network may include a wireless network provided by a communication provider of the mobile terminal. In one instance, the transmission device 106 includes a network adapter (Network Interface Controller, abbreviated as NIC), which can be connected to other network devices through a base station and thus communicate with the Internet. In one instance, the transmission device 106 may be a radio frequency (Radio Frequency, abbreviated as RF) module, which is used to communicate with the Internet wirelessly.

[0035] In this embodiment, a method for identifying enterprise credit risks running on a mobile terminal, a computer terminal, or a similar computing device is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. And although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order than here.

[0036] Figure 2 It is a schematic flowchart of the method for identifying enterprise credit risks according to an embodiment of the present application. As Figure 2 shown, the method includes the following steps:

[0037] Step S201, obtain a large language model and training samples, and pre-train the large language model with the above training samples to obtain an initial large language model after pre-training. The above training samples include financial and financial data and label information corresponding to the above financial and financial data. Among them, the above large language model adopts a Decode-only architecture;

[0038] Specifically, a large language model (LLM) refers to a deep learning model with billions to trillions of parameters. By performing unsupervised pre-training on a large amount of text data, it can understand and generate human language. The large language model is based on advanced natural language processing technology. The most significant feature is its large scale, complex deep learning architecture, which can capture the richness and nuances of language, thus showing extraordinary capabilities in various language tasks.

[0039] The large language model usually has billions to trillions of parameters, and these parameters are optimized during training to adapt to the statistical patterns of the input text. An unsupervised or self-supervised training method is adopted to learn the structure and semantics of language by predicting missing words or sentences in the text.

[0040] The large language model often based on the Transformer architecture uses the self-attention mechanism to process sequence data. Compared with the traditional recurrent neural network (RNN), Transformer can process long texts more efficiently and perform parallel computing, thus accelerating the training process. Due to training on a large amount of diverse text data, the large language model has strong generalization ability and can process unseen data to complete various language tasks such as text generation, answering questions, translation, summarization, and sentiment analysis. Although the large language model is trained on general text data, through the fine-tuning technology, it can be further trained on specific fields or tasks to make it more specialized.

[0041] Typical applications of large language models include text generation, question answering systems, translation, content review and filtering, knowledge graph construction, and financial risk control. Among them, text generation can generate articles, stories, poems, code, etc. The question answering system can answer various questions and provide information retrieval and knowledge consultation. Translation can achieve efficient and accurate multilingual translation services. Content review and filtering can identify and filter inappropriate content. Knowledge graph construction can extract entities and relationships from text for building a rich knowledge graph. Financial risk control can analyze financial reports, market news, social media, etc., evaluate credit risks, and predict market trends.

[0042] The specific implementation method of step S201 is as follows. First, a pre-trained large language model is required as the basis. This model is usually trained on a large amount of general text data and has broad language understanding and generation capabilities. In this embodiment, a large language model with a Decoder-only architecture is used, which means that the model is mainly focused on text generation tasks during design. Models such as GPT (Generative Pre-trained Transformer) adopt this architecture. The Decoder-only architecture is a deep learning network design for building generative language models. Different from the traditional Encoder-Decoder architecture, it removes the dependence on the encoder and only retains the decoder part. This architecture was initially adopted in the GPT series of models. Especially in GPT-2, GPT-3, and subsequent variants, the Decoder-only architecture has demonstrated strong capabilities in text generation tasks.

[0043] After determining the large language model used, obtain training samples specific to the financial field. These samples include financial and financial data, such as financial-related content in news and social media, or regulatory announcements, business exception announcements, credit histories of enterprises, etc., as well as label information corresponding to these financial and financial data. The label information can be an analysis report including credit risk levels, repayment probabilities, market sentiment, etc., to guide the large language model on how to accurately understand and predict the credit risks of these data.

[0044] After determining the large language model and obtaining the training samples, use the training samples to pre-train the large language model. Pre-training means that before a specific task, a large amount of data is used to preliminarily train the model so that it can master the knowledge and patterns in the field. In the pre-training stage, the large language model will learn how to parse and understand financial and financial data, and how to make corresponding credit risk assessments based on the label information of the financial and financial data. Through this process, the large language model can better understand the professional terms, rules, and trends in the financial field, laying a foundation for subsequent fine-tuning and prediction tasks.

[0045] Through the above pre-training process, the large language model will contain preliminary knowledge and skills in the financial field, that is, the initial large model after pre-training. Although this initial large model has not been refined through specific credit risk identification tasks, it already has the ability to process and understand financial data, providing the necessary knowledge background for the execution of fine-tuning and prediction tasks.

[0046] In summary, step S201 pre-trains the large language model on data related to the financial field, enabling it to more effectively understand and process financial and accounting data, providing a strong preliminary knowledge framework for subsequent credit risk identification tasks. This initial large language model will serve as the basis for subsequent fine-tuning and optimization, guiding the model to achieve higher accuracy and efficiency in specific credit risk identification tasks.

[0047] In step S202, the fine-tuning task is applied to the Mixture of Experts model to fine-tune the above initial large language model, obtaining a fine-tuned large language model, where the above fine-tuning task is determined based on the fine-tuning task dataset.

[0048] Specifically, fine-tuning refers to further training on a smaller dataset for a specific domain or task based on the initial large language model obtained after pre-training, so that the model can better adapt to this specific domain task. In the scenario of enterprise credit risk identification in this application, the fine-tuning task is designed based on the fine-tuning task dataset in the financial field. These datasets include various types of text data closely related to enterprise credit risk identification, such as financial news, enterprise financial reports, regulatory announcements, social media discussions, etc., as well as the corresponding label information, such as default risk levels, loan risk assessment results, etc. The label information can be in the form of research reports, analysis reports, or business plans, etc.

[0049] The Mixture of Experts (MoE) model is a model that uses multiple sub-models (or expert models) to handle complex tasks. Each expert model focuses on a specific sub-task or specific type of input. Specifically, the Mixture of Experts model is a machine learning architecture that combines the predictions or decisions of multiple expert models to solve complex and diverse task problems. In a traditional single-model architecture, one model attempts to learn the distribution of all data and solve all problems, which may encounter challenges when the data distribution is extensive or the task complexity is high. In contrast, the Mixture of Experts model improves the overall performance and efficiency of the model by decomposing the task and allowing each expert model to focus on solving specific sub-tasks or data subsets.

[0050] A key component of the mixture-of-experts model is the expert model, which is a deep learning model trained for specific subtasks or subsets of data. Each expert model can have different architectures, parameters, and training strategies to meet the needs of processing different types of data or solving specific problems.

[0051] Another key component of the mixture-of-experts model is the gating network, which is responsible for determining which expert models should be activated or participate in the prediction given an input. The gating network is usually a lightweight model, such as a multi-layer perceptron (MLP), that calculates the weights of each expert model based on the input feature vector. These weights represent the contribution of the expert models to the current prediction. The output of a traditional mixture-of-experts model is obtained by weighted summing the outputs of all expert models, which is determined by the weights of the gating network, ensuring that the expert models most relevant to the input have the greatest influence. The working principle is as follows: when an input is fed into the mixture-of-experts model, the gating network first calculates the weights of each expert model based on the input features. Then, the input is sent to all expert models simultaneously, but the output of each expert model will be weighted and summed according to its weight to generate the final output. This embodiment improves the traditional mixture-of-experts model. For specific details, please refer to the corresponding content later.

[0052] By applying a fine-tuning task to the mixture-of-experts model to fine-tune the initial large language model, a fine-tuned large language model is obtained, which has a deeper understanding and more accurate prediction ability in enterprise credit risk identification. The fine-tuned large language model can more effectively process and analyze complex data in the financial field, providing reliable support for credit risk identification.

[0053] In summary, step S202 transforms the initial large language model into a fine-tuned large language model with better performance in the field of credit risk identification by applying a fine-tuning task based on a specific dataset in the mixture-of-experts model, providing an intelligent risk assessment and management tool.

[0054] Step S203: Optimize the above fine-tuned large language model through a preset prediction task to obtain a target large language model;

[0055] In the specific implementation process, the preset prediction task is a task defined in the model design stage for a specific target or business scenario. These tasks are usually quantitative and directly related to the final application scenario of the model. In this embodiment, the preset prediction tasks cover all aspects of enterprise credit risk identification.

[0056] The optimization process is as follows:

[0057] First, clarify the specific details of the preset prediction tasks, including the inputs and expected outputs of the tasks, which are designed to enable the model to make more accurate and quantitative predictions when generating financial products, policies, or assessing risks.

[0058] Input the preset prediction tasks into the fine-tuned large language model. The preset prediction tasks cover multi-dimensional information such as the business operation status and market environment. The fine-tuned large language model generates prediction results based on the input, such as the expected loan default probability, the amount of future loan demand, etc. Then, compare the results predicted by the model with the actual historical data or expert evaluation results to calculate the prediction error. The optimization process uses the prediction error to adjust the parameters of the model so that the model can learn how to complete the prediction task more accurately. This adjustment is usually completed through the backpropagation algorithm to minimize the difference between the prediction error and the actual result. After the optimization training of the preset prediction tasks, the target large language model is obtained. This model not only includes extensive language understanding and generation capabilities but also has a high degree of professionalism and accuracy in the field of enterprise credit risk identification. The target large language model can conduct in-depth analysis of financial and financial data and generate reasonable and quantitative predictions.

[0059] Through optimization with preset prediction tasks, the large language model not only performs excellently in general language capabilities but also demonstrates a high degree of professionalism and accuracy in specific fields such as small and micro enterprise credit risk identification. This optimization process is a necessary step to improve the final application efficiency of the model and ensure that the model plays the greatest value in actual business scenarios.

[0060] Step S204: Input the financial and financial data of the enterprise to be identified for credit risk into the above-mentioned target large language model to obtain the credit risk identification result of the enterprise to be identified for credit risk.

[0061] Specifically, input a series of financial and financial data of the enterprise to be identified into the target large language model. These financial and financial data can include financial statements, credit records, market information, and social media and public opinion analysis. Among them, financial statements include the balance sheet, income statement, and cash flow statement, providing information on the financial condition, profitability, and cash flow of the enterprise; credit records include the credit history of the enterprise and its major shareholders, such as overdue repayments, reflecting the credit level of the enterprise; market information includes industry trends, competitor analysis, market news, etc., helping the enterprise evaluate the impact of the external environment on credit risk; social media and public opinion analysis include social media discussions and news reports related to the enterprise, reflecting the public's attitude towards the financial health and risk perception of the enterprise.

[0062] Before inputting data into the target large language model, preprocessing is required to ensure that the data format is compatible with the model and that the data quality meets the model's requirements. Unstructured or semi-structured data can be converted into a structured format readable by the model, such as extracting key numbers from financial statements and organizing them into text descriptions; standardizing numerical data to eliminate the influence of measurement units so that the model can better understand the data; converting all data into text form because the target large language model is essentially a text processing model that can understand and generate text information; integrating the processed data into a coherent corpus text, and specialized prompt templates need to be written to guide the model to understand the context and intent of the input data.

[0063] Financial and financial data is preprocessed and integrated into a suitable format for input, and then fed into the target large language model for inference. The target large language model will analyze the input financial and financial data based on the knowledge, rules, and patterns learned during its training phase:

[0064] The target large language model first interprets the input text data, understands the meaning of the enterprise's financial situation, analyzes financial indicators, market trends, and other relevant factors, and generates text for predicting and evaluating the enterprise's credit risk, including risk levels, predicted default probabilities, recommended loan conditions, etc. Finally, the text generated by the model is structured to generate the final credit risk identification result.

[0065] The entire process of step S204 demonstrates the application of the target large language model in credit risk identification. It can use deep learning and natural language processing technologies to deeply analyze complex financial and financial data, generate credit risk identification results with high interpretability and practical value, and provide strong risk management support.

[0066] Through this embodiment, a large language model is pre-trained using financial and financial data and corresponding tag information, resulting in an initial large language model after pre-training. The fine-tuning task is applied to the mixture of experts model to fine-tune the initial large language model, obtaining a fine-tuned large language model. The fine-tuned large language model is optimized through a preset prediction task to obtain a target large language model. The financial and financial data of the enterprise to be identified for credit risk is input into the target large language model to obtain the credit risk identification result of the enterprise to be identified for credit risk. This enterprise credit risk identification method effectively improves the professional performance of the large language model in the field of credit risk assessment through three stages: pre-training, fine-tuning, and optimization. First, the pre-training stage enables the model to have a basic understanding of financial domain knowledge. Subsequently, the fine-tuning stage uses the mixture of experts model structure to finely enhance the model's capabilities for specific tasks in financial risk control. Finally, the optimization stage further improves the model's quantitative analysis and accurate prediction capabilities through a preset prediction task. Overall, this solution can analyze the financial and financial data of the enterprise to be identified in real time and accurately, generating highly interpretable credit risk identification results, thereby improving the accuracy of enterprise credit risk identification.

[0067] In the specific implementation process, this embodiment improves the traditional mixture of experts model. The mixture of experts model in this embodiment is a mixture of experts model based on Monte Carlo routing. Applying the fine-tuning task dataset to the mixture of experts model to fine-tune the above initial large language model to obtain a fine-tuned large language model includes: obtaining the above fine-tuning task dataset and setting multiple fine-tuning tasks according to the above fine-tuning task dataset; based on multiple above fine-tuning tasks, using the above mixture of experts model based on Monte Carlo routing to fine-tune the above initial large language model to obtain the above fine-tuned large language model.

[0068] Specifically, the mixture of experts model based on Monte Carlo routing adopts the mechanism of Monte Carlo routing. The main purpose of this mechanism is to dynamically select expert models to handle specific tasks to improve the performance and efficiency of the model.

[0069] In the actual business activities of small and micro enterprises, there are participants from multiple industries, fields, and disciplines. In economic activities, when all parties conduct corresponding risk assessments and make decisions on the development of small and micro enterprises, they need to comprehensively consider various factors. For example, credit risk control needs to consider financial risks, macroeconomic factors, etc. simultaneously. The hybrid expert model based on Monte Carlo routing proposed in this embodiment starts from multiple industries and disciplines, takes the latent variables reflected above as inputs, and activates the corresponding expert model channels through the Monte Carlo sampling scheme. During the learning process of the model, it imitates the process of using multi-field and interdisciplinary professional knowledge in the real world to complete the processes of understanding, thinking, and decision-making. The unactivated expert model channels do not participate in the calculation during the next round of calculation, thus achieving the process of accelerating training.

[0070] First, it is necessary to obtain a fine-tuning task dataset. The fine-tuning task dataset is prepared for specific fine-tuning tasks and contains more in-depth and specific financial and financial data related to the identification of enterprise credit risks. These datasets not only include the original financial data but also cover specific labels or expected outputs related to these data, such as risk levels, default probabilities, etc., which are used to guide the fine-tuning training of the model and make it more professional in the credit risk identification task.

[0071] Based on the fine-tuning task dataset, multiple fine-tuning tasks can be set from different angles, and each task aims to optimize the performance of the model in a specific aspect. The fine-tuning tasks in the embodiments of this application can be formulated based on multiple different roles. The following are examples of several fine-tuning tasks in the context of small and micro enterprises:

[0072] Fine-tuning task 1: Role of a senior enterprise operator: According to the company's financial data, combined with the social environment and market trends, complete business activities such as the company's production planning and marketing decisions; at the same time, estimate the company's profit situation, fulfill the existing IOUs, and decide whether to conduct a new round of capital borrowing from the bank.

[0073] Fine-tuning task 2: In the role of a bank loan manager, generate a risk assessment report on small and micro enterprises for the bank. Focusing on small and micro enterprises in the local and specific regions, output an analysis report on the possibility of loan recovery, such as the solvency and cash flow stability of small and micro enterprises, by professional personnel through financial analysis tools such as financial ratio analysis, cash flow analysis, and 5C (Character, Capacity, Capital, Collateral, Condition), and at the same time use public opinion text analysis to analyze the compliance and risk hazards of the enterprise's loan use.

[0074] Fine-tuning Task 3: Generate a risk assessment report of a securities company for small and micro enterprises in the role of a securities practitioner. Analyze the diversity of the asset pool, default correlation, and predictable cash flow presented when small and micro enterprises serve as the underlying assets of asset securitization (ABS); thus, conduct a risk assessment of the credit rating and investment attractiveness when they are securitized products.

[0075] Fine-tuning Task 4: Generate an analysis report on the development scenario of small and micro enterprises in the role of an administrative staff. Based on the economic development level of the region, industrial structure, and the matching degree of infrastructure construction with the business activities of small and micro enterprises; the market saturation and development prospects of the industries where small and micro enterprises are located; analyze the business operation time, supplier relationships, price control ability over customers, experience of the management team, and rationality of the organizational structure, etc.

[0076] After determining the fine-tuning tasks, a mixture-of-experts model based on Monte Carlo routing is used to fine-tune the initial large language model. The structural diagram of the mixture-of-experts model based on Monte Carlo routing is as Figure 3 shown. The model first receives the input financial and accounting data, processes it through a multi-attention mechanism, extracts and fuses the deep semantic features of these data to form semantic latent variables. These latent variables are the key for the model to understand the input, containing the multi-dimensional meanings and context information of the data.

[0077] Monte Carlo routing is a mechanism that allows the model to dynamically decide which expert models to use to handle different inputs. The schematic diagram of Monte Carlo routing is shown in Figure 4 . Monte Carlo routing constructs an initial solution space set according to the semantic latent variables of the input through random sampling and weight assignment. This set contains all possible solutions or decision paths. Specifically, a set of initial solution spaces for scenario texts is constructed using the semantic vectors extracted by the shallow network of the model, and random sampling is performed by adding Gaussian noise, enabling the model to evaluate various possible decision paths at each sampling point (i.e., the semantic latent variables after adding noise), completing the process of solving combined with sampling. Subsequently, a multi-layer perceptron is used as a transformation network to map these semantic latent variables into the expert activation matrix space. To ensure the stability of the entire solving process, it is necessary to sample N times to reach a stable solution space. The number of sampling times N is a hyperparameter of Monte Carlo routing. The specific steps are as follows:

[0078] Step 1, through the multi-attention mechanism for the above text, obtain the high-dimensional semantic latent variables of the text, add Gaussian noise, and complete the process of constructing and sampling from the initial solution space;

[0079] Step 2, then use a multi-layer perceptron to transform the sampled semantic latent variables into the solution space of the expert activation model;

[0080] Step 3, repeat Steps 1 and 2 N times;

[0081] Step 4: After adding the N expert activation matrices obtained through N cycles and taking the mean, the result of the final expert activation matrix is obtained.

[0082] Among them, the answer space of the expert activation matrix is a matrix of the number of expert models * 2, denoted by the symbol NumOfExp.

[0083] Multiple expert models each focus on different aspects, such as enterprise operation, bank credit, securities analysis, or administrative management. When the fine-tuning task dataset is fed into the mixture-of-experts model, the Monte Carlo routing mechanism determines which expert models should be activated and their relative importance based on the semantic latent variables of the data through random sampling and weight assignment. The fine-tuning task helps the large language model learn and accumulate professional knowledge and skills in the financial field by providing training data for specific scenarios, while the mixture-of-experts model ensures that the model can comprehensively apply these skills to solve practical problems by dynamically selecting and integrating this professional knowledge. For example, when dealing with the credit risk identification of small and micro enterprises, the mixture-of-experts model may simultaneously activate the enterprise operation expert and the bank credit expert, and through their respective perspectives and analysis capabilities, generate a more comprehensive risk assessment.

[0084] During the fine-tuning process, based on multiple above-mentioned fine-tuning tasks, the above-mentioned mixture-of-experts model based on Monte Carlo routing is used to fine-tune the above-mentioned initial large language model to obtain the above-mentioned fine-tuned large language model, including: a sampling step, where the above-mentioned mixture-of-experts model based on Monte Carlo routing uses the Monte Carlo routing technique to dynamically select and activate target expert models, where the above-mentioned target expert models are expert models with a relevance to the current fine-tuning task greater than a preset relevance threshold; a determination step, where the weights of each of the above-mentioned target expert models are determined through the above-mentioned Monte Carlo routing technique; the above-mentioned sampling step and the above-mentioned determination step are repeatedly executed until a preset stop condition is reached to obtain the above-mentioned fine-tuned large language model.

[0085] Specifically, the above steps further refine how to use the mixture-of-experts model based on Monte Carlo routing to fine-tune the initial large language model to obtain the fine-tuned large language model. The following are the key implementation steps:

[0086] Sampling step: The Monte Carlo routing technique dynamically selects and activates the expert models most relevant to the current fine-tuning task. The specific process is as follows:

[0087] Semantic vector extraction: First, the model processes the input fine-tuning task data, extracts the semantic features and latent variables therein through a multi-attention mechanism, and forms a vector representation that can reflect the key information of the data.

[0088] Gaussian noise addition: To improve the exploration and generalization capabilities of the model, Gaussian noise is added to the semantic vectors to construct an initial solution space set for subsequent sampling processes.

[0089] Expert model activation sampling: The multi-layer perceptron is used to transform the noise-added semantic vectors into the solution space of the expert activation matrix, and then the expert models are randomly selected from this space through Monte Carlo sampling techniques. The sampling process ensures that the probability of each expert model being activated is based on a dynamic assessment of its relevance to the current task.

[0090] Evaluating the relevance of expert models: A preset relevance threshold is set, and only those expert models with a relevance higher than the threshold will be considered "target expert models", that is, models that may be activated in the current fine-tuning task.

[0091] Determination step: After the sampling step, it is necessary to determine the weights of each target expert model in the current fine-tuning task, and this weight will determine the influence of the model in the fusion prediction. The process of determining the weights includes:

[0092] Weight calculation: For each sampled target expert model, a weight value is calculated based on its performance on a specific task and its correlation with the input data. The weight value reflects the relative importance and effectiveness of the expert model in handling the current task.

[0093] Weight stabilization: The Monte Carlo routing technique requires multiple sampling processes (N loops) to achieve stable weight distribution. This process helps the model learn the most stable and effective combination of expert models, thereby improving the accuracy of overall prediction and decision-making.

[0094] To meet the weight stabilization condition required by the Monte Carlo routing technique, the steps of sampling and determining weights are repeated. This repeated execution process will continue until a preset stop condition is met, such as the weight change being less than a certain threshold, or reaching a predetermined number of training epochs. After reaching the stop condition, the model will stop fine-tuning and obtain the fine-tuned large language model.

[0095] The key significance of the above process is that it enables the large language model to dynamically select and activate the most suitable expert models according to different input tasks, and at the same time ensures the comprehensiveness and effectiveness of the model's prediction or decision-making through the weight allocation mechanism. In the scenario of enterprise credit risk identification, this means that the model can flexibly call on professional knowledge in different fields such as enterprise operation, bank credit, and securities for analysis according to changes in different enterprise backgrounds and industry characteristics, so as to provide more accurate risk assessment results. In this way, the fine-tuning process of the model not only enhances its understanding ability of specific financial fields, but also ensures that the model can effectively respond to changing business scenarios and provide personalized decision-making support.

[0096] In summary, through the above fine-tuning process of the hybrid expert model based on Monte Carlo routing, the initial large language model is optimized into a fine-tuned large language model. When dealing with the enterprise credit risk identification task, this model not only has a broader knowledge background, but also can dynamically select the most relevant expert model for prediction according to the characteristics of the input data, thus significantly improving the decision-making ability and efficiency of the model in complex scenarios.

[0097] After obtaining the fine-tuned large language model, optimize the above fine-tuned large language model through a preset prediction task to obtain the target large language model, including: a setting step of setting the above prediction task, and the output of the above prediction task being numerical; a prediction step of inputting the above prediction task into the above fine-tuned large language model for prediction to obtain a prediction result; an adjustment step of adjusting the parameters of the above fine-tuned large language model according to the difference between the above prediction result and the actual result; repeatedly executing the above prediction step and adjustment step until a preset optimization standard is reached to obtain the above target large language model.

[0098] It can be understood that the above is an elaboration on how to further optimize the fine-tuned large language model through a preset prediction task. Specifically, first, it is necessary to clarify the nature and goal of the prediction task. The output result of the set prediction task is numerical, so that the model can learn quantitative prediction. Take whether the IOU generated by small and micro enterprises is overdue as the prediction task. IOU performance as a quantitative model judgment task reflects the analysis and judgment of the large model on the operation status of small and micro enterprises. As a model training task, it is of high difficulty and conforms to the design of the downstream tasks of the language model. At the same time, generate training tasks that fit the quantitative attributes of the financial field, enabling the model to generate amounts approaching those in real life, such as the amount of IOUs, the capital scale of securities products, and the amount of capital support. Numerical and quantitative reinforcement learning tasks make the model sensitive to numbers and truly understand the relationship and principle between enterprise capital changes and actual enterprise operation conditions. For the small and micro enterprise credit risk identification scenario, the prediction task can involve the following types:

[0099] Prediction task 1: The generation task of bank loan products. Provide re-lending for small enterprises, inclusive small and micro enterprise loan support tools, intellectual property pledge loans, equity pledge loans to meet the diverse financing needs of enterprises. Increase third-party guarantee institutions to reduce risks and provide larger amounts and more suitable loan products. Around the core enterprises of the industrial chain, provide comprehensive financial service solutions for upstream and downstream small and micro enterprises to help the common development of all links of the industrial chain.

[0100] Prediction task 2: The generation task of securitization products. Use the revolving purchase structure and introduce measures to support liquidity to match the short-term loans of small and micro enterprises with long-term securities products issued by securities companies.

[0101] Prediction Task 3: The task of generating economic policies for the administrative department to boost enterprise development. Through monetary policy tools, differentiated regulatory policies, and fiscal and economic measures, guide financial institutions to increase credit supply to small and micro enterprises. Implement a support plan for selected enterprises in key industrial chains and supply chains, and ensure the supply of production factors for small and micro enterprises that meet key nodes.

[0102] More specifically, the probability of enterprise default can be predicted, and a value between 0 and 1 can be output, representing the likelihood of default; the risk rating can also be predicted, and a quantitative rating can be output, such as a score from A to E, representing the degree of risk; the loan amount can also be predicted, and the output is a specific loan amount value, which is used to guide loan decisions.

[0103] Next, input the set prediction task into the fine-tuned large language model, and let the model make predictions and generate prediction results. First, data preprocessing needs to be carried out, preprocess the data related to the prediction task to ensure that the data format matches the input requirements of the model. Input the preprocessed data and use the fine-tuned large language model to make predictions. The model generates prediction results based on the learned financial knowledge and skills, and the results are specific numerical outputs.

[0104] After obtaining the prediction results of the model, compare them with the actual results, calculate the differences, and adjust the parameters of the fine-tuned large language model based on these differences to optimize its prediction accuracy. By comparing the prediction results of the model with real historical data or expert evaluation results, calculate the prediction error. A loss function can be set to quantify the prediction error, such as mean squared error (MSE) or cross-entropy error, as the goal of model optimization. Through the backpropagation algorithm, according to the gradient of the loss function, adjust the model parameters layer by layer from the output layer to the input layer to minimize the prediction error. Based on the calculated gradient, use an optimization algorithm (such as gradient descent, etc.) to update the model parameters to improve the performance of the model in the prediction task.

[0105] The entire optimization process is an iterative process, and the prediction and adjustment steps need to be repeated until the prediction error of the model reaches the preset optimization standard, or in several consecutive iterations, the change in the prediction error caused by parameter adjustment is lower than a certain threshold, indicating that the model has converged to a better solution. After reaching the preset optimization standard, the target large language model is obtained. The target large language model performs better in specific numerical tasks of predicting enterprise credit risks and can more accurately evaluate and predict quantitative indicators such as risk levels, default probabilities, or loan amounts.

[0106] Through the above steps, the target large language model not only has the ability to deeply understand the complex knowledge in the financial field, but also can optimize the quantitative prediction task, making its decision-making in enterprise credit risk identification more intelligent, accurate and reliable.

[0107] After optimizing the above fine-tuned large language model through preset prediction tasks to obtain the target large language model, the above method further includes: setting and storing corresponding Prompt guiding words based on different financial activity scenarios, where the above financial activity scenarios are specific operating environments set according to different business stages and requirements including pre-loan management, in-loan management and post-loan management; parsing credit risk-related questions according to the above Prompt guiding words to generate decision-making texts that meet the requirements of different above financial activity scenarios.

[0108] Specifically, Prompt guiding words play an important role in natural language processing and can guide the model to focus on specific information or perform specific tasks. In the scenario of enterprise credit risk identification in this embodiment, in order to ensure that the model can understand and process the requirements of different business stages such as pre-loan management, in-loan management and post-loan management, a series of Prompt guiding words related to these scenarios need to be set, such as:

[0109] In the pre-loan management scenario, the Prompt guiding words can include "assessing the credit status of loan applicants", "analyzing the financial statements of applicants", "predicting the probability of loan default", etc. These guiding words will direct the model to focus on the applicant's credit history, financial health status and potential repayment ability.

[0110] In the in-loan management scenario, the Prompt guiding words can include "monitoring the use of loan funds", "evaluating loan amount adjustments", "predicting the impact of interest rate changes on borrowers", etc. These guiding words direct the model to track the borrower's behavior in real time, evaluate the current risk level of the loan, and consider the impact of market changes on existing credit relationships.

[0111] In the post-loan management scenario, the Prompt guiding words can involve "analyzing fluctuations in borrowers' repayment ability", "predicting the possibility of recovering non-performing loans", "formulating strategies for renewing loans", etc. These prompt words will help the model evaluate the repayment trend of borrowers, predict potential loan recovery problems, and formulate strategies for loan renewal or recovery.

[0112] When the target large language model is called to solve problems in specific financial activity scenarios, appropriate Prompt guiding words will be selected according to the current scenario. The Prompt guiding words will be combined with actual loan or borrower information and sent to the target large language model for processing. The target large language model first understands the intention in the Prompt guiding words, identifies and focuses on the characteristic information or tasks related to the current financial scenario. Based on the parsed questions and scenario requirements, the target large language model generates decision texts including analysis results, risk assessments, and recommended measures. These decision texts conform to the business logic of the scenario and provide actionable guiding information.

[0113] The above process ensures that the target large language model can generate credit risk analysis and decision recommendations in specific scenarios according to different business needs, improving the pertinence and effectiveness of decisions. For example, in the pre-loan management scenario, the model can generate a detailed credit assessment report, including an in-depth analysis of the borrower's credit history and a default risk prediction based on their financial situation; in the post-loan management stage, the model can generate updated strategy recommendations for the borrower's repayment ability status to help financial institutions identify and address potential credit risks in advance. By using the Prompt guiding word technology, the target large language model can flexibly adapt to the needs of different financial activity scenarios, providing customized risk identification and management decision support, and improving the intelligent level of credit risk identification and the accuracy of decisions.

[0114] It should be noted that obtaining the large language model and training samples, and using the above training samples to pre-train the above large language model to obtain the pre-trained initial large language model includes: scraping the above training samples from multiple data sources, where the above data sources include news social media, regulatory announcements and filings, business exception announcements, and financial professional reports; inputting the above training samples into the above large language model for pre-training to obtain the pre-trained above initial large language model, so that the above initial large language model learns the relevant information of the above training samples.

[0115] Specifically, the above content describes the process of obtaining a large language model and using specific training samples for pre-training to form an initial large language model after pre-training. First, training samples are scraped from multiple data sources. This stage involves collecting text data related to financial activity scenarios from multiple channels as training samples for the pre-training of the large language model. The data sources include news social media, regulatory announcements and filings, business exception announcements, and financial professional reports. Among them, news social media provides instant information on economic dynamics in various regions or industries, including financial news, market analysis, corporate announcements, etc. Regulatory announcements and filings include announcements, filing documents, and records submitted by enterprises, such as financial statements, audit reports, etc. These data can reflect compliance, financial status, and industry norms. Business exception announcements involve business risks of enterprises, such as equity pledges, restricted consumption, overdue debts, etc., which are important information for evaluating the credit and stability of enterprises. Financial professional reports include analyst reports, industry research reports, investment recommendations, etc., written by experts in the financial field, providing in-depth market analysis and forecasts, and are a source of advanced financial knowledge and complex business logic.

[0116] After collecting the above training samples, the next step is to input these training samples into the large language model for pre-training. Pre-training is a machine learning method aimed at enabling the large language model to learn general language and world knowledge before a specific task. For the large model of credit risk identification, the goal of pre-training is to enable the large language model to understand financial terms and concepts. Through the learning of a large amount of financial text, it can master relevant industry-specific vocabulary and concepts, such as credit, interest rate, risk assessment, etc.; it can also enable the large language model to capture industry dynamics. For example, the training of data sources such as news social media and professional reports allows the large language model to capture economic changes, including industry trends, market fluctuations, etc.; it can also enable the large language model to learn enterprise business risk models. Through the training of business exception announcements and regulatory filing documents, the large language model can identify and understand various risk models in enterprise operations, providing a knowledge basis for subsequent credit risk prediction.

[0117] During the pre-training process, the large language model will repeatedly process these training samples, learn and adjust its parameters to better understand and generate text. Through this process, the initial large language model can learn relevant information in the training samples, laying a solid knowledge foundation for subsequent fine-tuning and reinforcement learning, and ultimately improving its performance in the field of enterprise credit risk identification.

[0118] After the pre-training is completed, the resulting initial large language model will possess the basic ability to process financial texts and be able to understand and generate texts related to credit, market analysis, etc. In short, the above steps ensure that the large language model can learn and understand a wide range of financial concepts, market dynamics, and enterprise risk patterns, providing a powerful and general language processing ability for subsequent scenario-specific training and optimization.

[0119] It should also be noted that before obtaining the large language model and training samples, the above method further includes: evaluating the above financial and accounting data using a preset evaluation method to obtain the above label information of the above financial and accounting data, and the above preset evaluation method includes a scoring card model and a GPT-based model.

[0120] The above describes the process of using a preset evaluation method to preliminarily evaluate the collected financial and accounting data to generate label information. This process is crucial for screening and preparing high-quality training samples, ensuring that the model can learn the most relevant and valuable information during the pre-training stage.

[0121] Before starting the pre-training, evaluate the financial and accounting data scraped from data sources such as news social media, regulatory announcement filings, abnormal operation announcements, and financial professional reports. The purpose of the evaluation is to classify and label the data for easier utilization and management during subsequent training. The preset evaluation method mainly involves two types: a scoring card model and a GPT-based model.

[0122] The scoring card model is a traditional evaluation method used to quantify the quality and diversity of data. By constructing a scoring card model, a series of evaluation indicators can be set, such as the timeliness of the data, the completeness of the information, the reliability of the source, etc., and then each piece of data is scored. Data with high scores will be regarded as high-quality samples in the training set, while data with low scores may be excluded or require further processing.

[0123] The GPT-based model refers to an evaluation model based on the Generative Pre-trained Transformer framework. This model can be used to predict the perplexity of data to evaluate the quality and complexity of the data. The lower the perplexity, the more accurately the model can predict the content of the data, indicating higher data quality. A perplexity threshold can be set, and data above this threshold will be discarded. In addition, the GPT-based model can also be used to calculate the diversity of data. By comparing the data with the existing knowledge of the model, samples containing new information or novel perspectives are identified to increase the diversity of the training set.

[0124] Financial and financial data are evaluated through the above methods to generate corresponding label information. The label information includes quality labels, diversity labels, category labels, risk labels, topic labels, etc. Among them, the quality label reflects the reliability, integrity, and timeliness of the data; the diversity label indicates whether the data contains novel information and the degree to which it is different from other samples; the category label classifies the data into different categories such as news, announcements, reports, etc. according to the source and content of the data; the risk label marks the enterprise risk level involved in the data, such as low risk, medium risk, or high risk; the topic label annotates the specific topics involved in the data, such as credit policies, market trends, enterprise finance, etc.

[0125] These label information not only helps in the management and retrieval of data, but also during the pre-training phase, the model can selectively learn based on the label information, improving the efficiency and effectiveness of training. For example, the model can preferentially learn high-risk and high-quality data to more quickly master the key knowledge required for identifying credit risks.

[0126] When implementing the preset evaluation method, a scoring card model and a GPT-based model are usually used in combination. The scoring card model focuses on the hard quality indicators of the data, such as timeliness and source, while the GPT-based model pays more attention to the soft features of the data, such as the complexity and novelty of the information. By combining these two methods, financial and financial data can be comprehensively evaluated to ensure that the data learned by the model is both high-quality and diverse.

[0127] The label information generated through the preset evaluation method not only improves the quality of financial and financial data but also ensures the diversity of training samples, which is crucial for constructing an initial large language model capable of deeply understanding financial domain knowledge and having cross-domain reasoning capabilities. This process provides high-quality data support for subsequent fine-tuning and reinforcement learning, enabling the model to be more accurate and reliable in identifying enterprise credit risks.

[0128] The embodiments of this application also provide a device for identifying enterprise credit risks. It should be noted that the device for identifying enterprise credit risks in the embodiments of this application can be used to execute the method for identifying enterprise credit risks provided in the embodiments of this application. The device is used to implement the above embodiments and preferred implementation manners, and those that have been described will not be repeated. As used hereinafter, the term "module" can be a combination of software and / or hardware that can achieve a predetermined function. Although the devices described in the following embodiments are preferably implemented in software, implementation in hardware, or a combination of software and hardware is also possible and contemplated.

[0129] The following introduces the device for identifying enterprise credit risks provided in the embodiments of this application.

[0130] Figure 5It is a structural block diagram of a device for enterprise credit risk identification according to an embodiment of the present application. As Figure 5 shown, the device includes a pre-training unit 10, a fine-tuning unit 20, an optimization unit 30, and an input unit 40. Among them, the pre-training unit 10 is used to obtain a large language model and training samples, and use the above training samples to pre-train the above large language model to obtain an initial large language model after pre-training. The above training samples include financial and financial data and label information corresponding to the above financial and financial data. Among them, the above large language model adopts a Decoder-only architecture; the fine-tuning unit 20 is used to apply a fine-tuning task to a mixture-of-experts model to fine-tune the above initial large language model to obtain a fine-tuned large language model. Among them, the above fine-tuning task is determined based on a fine-tuning task dataset; the optimization unit 30 is used to optimize the above fine-tuned large language model through a preset prediction task to obtain a target large language model; the input unit 40 is used to input the financial and financial data of the enterprise to be identified for credit risk into the above target large language model to obtain the credit risk identification result of the enterprise to be identified for credit risk.

[0131] In the device for enterprise credit risk identification of the present application, the above fine-tuning unit includes an acquisition module and a fine-tuning module. Among them, the acquisition module is used to obtain the above fine-tuning task dataset and set a plurality of fine-tuning tasks according to the above fine-tuning task dataset; the fine-tuning module is used to fine-tune the above initial large language model based on a plurality of the above fine-tuning tasks by using the above mixture-of-experts model based on Monte Carlo routing to obtain the above fine-tuned large language model.

[0132] The above fine-tuning module includes a sampling module, a determination module, and a first repetition module. Among them, the sampling module is used to dynamically select and activate a target expert model by using Monte Carlo routing technology. Among them, the above target expert model is an expert model whose relevance to the current fine-tuning task is greater than a preset relevance threshold; the determination module is used to determine the weights of the above target expert models by using the above Monte Carlo routing technology; the first repetition module is used to repeatedly call the above sampling module and the above determination module until a preset stop condition is reached to obtain the above fine-tuned large language model.

[0133] The above optimization unit includes a setting module, a prediction module, an adjustment module, and a second repetition module. Among them, the setting module is used to set the above prediction task, and the output of the above prediction task is numerical; the prediction module is used to input the above prediction task into the above fine-tuned large language model for prediction to obtain a prediction result; the adjustment module is used to adjust the parameters of the above fine-tuned large language model according to the difference between the above prediction result and the actual result; the second repetition module is used to repeatedly call the above prediction module and the adjustment module until a preset optimization criterion is reached to obtain the above target large language model.

[0134] The above device further includes a setting and storage unit and a generation unit. The setting and storage unit is used to set and store corresponding Prompt guiding words based on different financial activity scenarios after optimizing the above fine-tuned large language model through a preset prediction task to obtain a target large language model. Among them, the above financial activity scenarios are specific operation environments set according to different business stages and requirements including pre-loan management, in-loan management, and post-loan management; the generation unit is used to parse credit risk-related questions according to the above Prompt guiding words and generate decision-making texts that meet the requirements of different above financial activity scenarios.

[0135] Further, the above pre-training unit includes a scraping module and a pre-training module. The scraping module is used to scrape the above training samples from multiple data sources, where the above data sources include news social media, regulatory announcements and filings, business exception announcements, and financial professional reports; the pre-training module is used to pre-train the above large language model with the above training samples to obtain the above initially pre-trained large language model, so that the above initially large language model learns the relevant information of the above training samples.

[0136] Further, the above device further includes an evaluation unit. The evaluation unit is used to evaluate the above financial and accounting data using a preset evaluation method before obtaining the large language model and the training samples to obtain the above label information of the above financial and accounting data. The above preset evaluation methods include a scoring card model and a GPT-based model.

[0137] The above device for identifying enterprise credit risk includes a processor and a memory. The above pre-training unit, fine-tuning unit, optimization unit, input unit, etc. are all stored in the memory as program units, and the processor executes the above program units stored in the memory to implement corresponding functions. The above modules are all located in the same processor; or, the above each module is located in different processors in any combination form.

[0138] The memory may include non-permanent memory in a computer-readable medium, in the form of random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash memory (flash RAM), and the memory includes at least one storage chip.

[0139] The embodiment of the present application also provides an enterprise credit risk identification system, and the overall flowchart is as Figure 6As shown in the figure, the enterprise credit risk identification system includes a real-time data engineering pipeline module, a business simulation fine-tuning module, a credit risk prediction reinforcement learning module, and a credit risk decision-making module. Among them, the real-time data engineering pipeline module is used to capture the information text data of real-time business development, and when injecting new text, it avoids the balance of the training text distribution through an intelligent data mixing system; the business simulation fine-tuning module is used to introduce the macro decision-making roles of the entities associated with the enterprise's business activities and perform the fine-tuning task of predicting business role behaviors, helping the large model to better understand the mechanism of business activities; the credit risk prediction reinforcement learning module is used to establish a credit risk prediction task with better task difficulty, refine from the breadth of business operation to risk prediction, and use the reinforcement learning algorithm to focus the attention of the large model on the risk control tasks that meet the focus of the banking industry. Through the gradual knowledge precipitation of the above three modules, a credit risk identification system with the characteristics of the banking field is constructed.

[0140] 1. Real-time data engineering pipeline module

[0141] For the schematic diagram of the real-time data engineering pipeline module, see Figure 7 . Online financial and financial data, including but not limited to: financial news, company filings, credit reporting anomalies announcements, social media discussions, trend types (analysts, experts, financial blogs, forum observations, etc.). Financial news timely and dynamically tracks and reports important information about the economy, specific industries, and individual companies, reflecting the economic state and the mood changes in the industry market; the market sentiment shown by news and media will affect the decisions of financial staff, thus leading to drastic changes in the market. Due to the limited attention of small and micro enterprises by the media and researchers, the filing data and announcement data provided by small and micro enterprises to regulatory agencies are even more important, with better data quality in terms of the granularity of content details, the reliability of information disclosure, and the characteristics of periodicity. Business anomaly announcements refer to text information such as equity pledges, restricted consumption, and overdue promissory notes of other banks by enterprise operators, which significantly mark the abnormal financial status of individual small and micro enterprises.

[0142] 1.1 Corpus scraping component

[0143] News and social media: collected at the daily time granularity; regulatory announcement filings: collected monthly, quarterly, and annually; business anomaly announcements: access multiple channels of credit reporting to capture enterprise abnormal behavior notices in real time; professional reports: adopt the opinions, market forecasts, analysis reports, specific securities, and trends and investment suggestions of financial professionals such as researchers and analysts.

[0144] 1.2 Corpus screening component

[0145] This embodiment proposes a corpus screening component to complete the training data screening step that meets the requirements of sampling diversity and high quality, reducing the training pressure of the large model while ensuring the model's effect. The corpus screening component can be divided into 2 sub-components.

[0146] <1>Automated evaluation and screening sub-component

[0147] Quality screening (quality scoring model): Use a lightweight language model to predict the perplexity of text corpora. The lower the perplexity, the higher the data quality. Set a perplexity threshold, and discard those above the threshold. Diversity screening (diversity scoring model): Use an existing language model to calculate the sentence vectors of texts. After the sentence vectors are reduced in dimension by PCA, K-Means is used for vector clustering. Select the top several data with the highest scores in each cluster, and after removing duplicates, they are used as the training corpora screened in this stage.

[0148] <2>Manual evaluation and screening sub-component

[0149] Construct a fine-grained label system that can explain the corpus, and classify the labels in ways such as the round, role, scene, conversation topic, source, use, language style, and text format of the conversation. In the automated evaluation and screening sub-component, if the classification ratio of the final obtained corpus in the label system is unbalanced, manual judgment is used to complete the adjustment and supplementation of the training corpus.

[0150] 2. Business simulation fine-tuning module

[0151] As an important financial entity, banks need to direct the flow of funds to small, medium and micro enterprises to optimize the financing structure of enterprise operations. The capital flow cycle of small and micro enterprises will continuously have a capital interaction relationship of borrowing, repayment, and borrowing with banks according to their own profit conditions and development needs. Using the historical IOU data and past historical financial data of small and micro enterprises, the large model analyzes to obtain the operating status of the enterprises and analyzes the trend of the next capital transactions initiated by small and micro enterprises with banks. In supervised learning, the large model uses the IOU information generated in the future as the output label for the following text.

[0152] 2.1 Fine-tuning tasks

[0153] Based on different role settings, fine-tuning tasks are generated. For example, based on the role of a senior enterprise operator, a risk assessment report on small and micro enterprises by a bank can be generated as a bank loan manager, or a risk assessment report on small and micro enterprises by a securities company can be generated as a securities practitioner.

[0154] 2.2 Mixture of experts model

[0155] In the actual business activities of small and micro enterprises, there are participants from multiple industries, fields, and disciplines. In social and economic activities, when all parties complete corresponding risk assessments and make decisions on the development of small and micro enterprises, multiple factors need to be considered comprehensively. For example, credit risk control needs to consider financial risks, macroeconomic factors, etc. at the same time.

[0156] The hybrid expert model proposed in this embodiment starts from multiple industries and disciplines, takes the latent variables reflected above as inputs, and activates the corresponding expert model channels through the Monte Carlo sampling scheme. Let the model imitate the process of using multi-field and interdisciplinary professional knowledge to complete understanding, thinking, and decision-making in the real world during the learning process. The unactivated expert model channels do not participate in the calculation during the next round of calculation, so as to achieve the process of accelerating training.

[0157] 3. Credit Risk Prediction Reinforcement Learning Module

[0158] Taking whether the IOU generated by small and micro enterprises is overdue as the prediction task. IOU performance as a quantitative model judgment task reflects the analysis and judgment of the operation status of small and micro enterprises by the large model. As the model training task is highly difficult, it conforms to the design of the downstream tasks of the language model. At the same time, generating training tasks fits the quantitative attributes of the financial field, enabling the model to generate predictions approaching the amounts of IOUs, the capital scales of securities products, etc. in the real world. Numerical and quantitative reinforcement learning tasks make the model sensitive to numbers and truly understand the relationship and principle between enterprise fund changes and actual enterprise operation conditions.

[0159] 4. Credit Risk Decision-making Module

[0160] Based on the training tasks designed by commercial finance simulation, the large language model has the professional ability of cognition, analysis, reasoning, and decision-making in the economic activities of small and micro enterprises under the scenarios of market dynamics, business environment, and financial risks. The large model is integrated as the core intelligent expert engine inside the decision-making system. Through the proprietary Prompt task guiding words, the model generates a reply after understanding and analyzing the interactive questions. The decision-making system then performs adaptive processing on the generated text and displays an intelligent decision-making plan for credit risk activities with rationality and interpretability. For the internal schematic diagram of the credit risk decision-making module, please refer to Figure 8 .

[0161] 4.1 Financial Scenario Management Component

[0162] The financial scenario management component models the actual work scenarios of credit risk identification to form corresponding scenario models. System users select the corresponding financial scenarios to complete the financial tasks that need to be decided. Common financial scenarios are as follows:

[0163] Pre-loan management: credit scoring, customer access, due diligence.

[0164] During-loan management: credit line upgrade, interest rate setting, fund usage warning.

[0165] Post-loan management: renewal loan decision, repayment ability fluctuation; non-performing asset disposal.

[0166] 4.2 Prompt template management component

[0167] Based on different financial activity scenarios, design, collect, and store the corresponding Prompt guiding words for each scenario, so as to guide the model to correctly understand questions and generate answers in different scenarios.

[0168] 4.3 Credit risk decision-making component

[0169] The credit risk decision-making component hosts a large language model that has completed two training stages of fine-tuning and reinforcement learning, and serves as the intelligent decision-making engine of the large model to complete the generation of decisions. First, preprocess the Prompt guiding words suitable for the current decision-making scenario and the materials of bank credit activities into a format that the large model can accept; then send the preprocessed corpus into the large language model to complete the generation of decision texts.

[0170] After the fine-tuning training task, the large language model completes the learning of common sense, knowledge, etc. information in the financial field and has the reserve ability of financial experts. The reinforcement learning task masters the business operations, financial regulation, and economic activity laws of small and micro enterprises from multiple perspectives.

[0171] 4.4 Decision structuring component

[0172] This module presents the decision text generated intelligently through structuring to the user; this module allows the user to provide feedback and uses the user's new instructions to complete functions such as fine-tuning, regenerating, and refining the description of the decision task.

[0173] To more intuitively understand how to implement the enterprise credit risk identification method of this application in real time, the following gives a specific embodiment, which shows the process of handling a small and micro enterprise credit risk identification scenario, from data scraping to model fine-tuning, reinforcement learning, and until the generation of credit risk decisions.

[0174] Scrape the latest information about a small and micro enterprise (hereinafter referred to as the "target enterprise"), including news and social media dynamics in the past week, regulatory announcement filings in the recent three months, and business exception announcements and financial professional reports in the recent year.

[0175] Preprocess the captured data, use a lightweight language model to evaluate the perplexity of the data, set the threshold to 30, and filter out all data above this value. Then, all texts are converted into sentence vectors, after dimensionality reduction by PCA, they are clustered using the K-Means algorithm, and the top 5% of the data with the highest scores in each cluster are selected as the training corpus screened in this stage.

[0176] Next is the design of the fine-tuning task, which is fine-tuned with the roles of senior enterprise operators, bank loan managers, securities practitioners, and administrative staff. Fine-tuning for the role of senior enterprise operator: Input the historical financial data and the latest market analysis report of the target enterprise, and let the model predict the production plan and marketing strategy of the target enterprise in the next year, and estimate its profit situation and future fund transaction intention with the bank. Fine-tuning for the role of bank loan manager: Provide the current financial status and historical borrowing records of the target enterprise, and let the model generate a risk assessment report, including the analysis of debt repayment ability and cash flow stability. Fine-tuning for the role of securities practitioner: Input the industry trend of the target enterprise, and let the model evaluate the risk level and investment attractiveness of the target enterprise as the underlying asset of asset securitization. Fine-tuning for the role of administrative staff: Provide regional economic data, industry market analysis, and key indicators of enterprise operation, and let the model generate an evaluation report covering the comprehensive analysis of regional economic matching, industry prospects, enterprise operation sustainability, supply chain stability, pricing strategy, management experience, and organizational effectiveness.

[0177] After setting the fine-tuning task, use the Mixture of Experts model for fine-tuning. Using the principle of Monte Carlo routing, the model activates the most relevant expert model channels according to the characteristics of the input corpus during training. For example, when processing the market prediction of the target enterprise, the model is more inclined to activate the expert channel of securities practitioners. This step obtains the fine-tuned large language model.

[0178] Optimize the fine-tuned large language model through preset prediction tasks, and optimize it with the two prediction tasks of predicting whether the target enterprise has a default record and quantifying financial products. Among them, predicting whether the target enterprise has a default record: Take the prediction of the operating status of the target enterprise in the next year as a reinforcement learning task, and use whether the enterprise has overdue repayment as the reward signal to train the model to predict its default possibility. Quantifying financial product prediction: Input market dynamics and let the model predict the types and amounts of bank loan products that the target enterprise may apply for, so as to improve the model's quantitative prediction ability.

[0179] After optimizing the fine-tuned large language model, the target large language model is obtained, and then the target large language model is used to identify the credit risk of the target enterprise. For example, a bank credit manager selects the "pre-loan management" scenario and prepares to conduct a credit risk assessment of the target enterprise. Select "Assess loan default risk based on financial statements" as the prompt from the Prompt management component, input the prompt and the specific financial data of the target enterprise into the target large language model, and generate a detailed credit risk analysis report, which includes an assessment of the default risk of the target enterprise, an analysis of potential risk factors, and loan strategy recommendations based on the current market environment.

[0180] The above credit risk analysis report points out that the cash flow of the target enterprise has been slightly unstable recently, but the overall financial condition is acceptable. It is recommended to reduce the loan interest rate by 0.5%, and at the same time require the enterprise to provide quarterly financial reports as the basis for monitoring. Based on this intelligent decision, the bank credit manager conducted pre-loan communication with the target enterprise, and finally determined the loan terms, effectively reducing the credit risk.

[0181] Through the above specific embodiments, it can be seen that by capturing and screening high-quality financial data, using fine-tuning tasks to let the large language model learn the decision-making logic of different financial roles, and introducing preset prediction tasks to enhance the large language model's ability to quantify and predict risks, a targeted credit risk analysis report is finally generated, realizing the intelligent identification and decision support of the credit risk of small and micro enterprises. This process reflects the advantages of the embodiments of the present application in dealing with complex financial scenarios and can provide more accurate and comprehensive credit risk management strategies.

[0182] The embodiment of the present invention provides a computer-readable storage medium, and the above computer-readable storage medium includes a stored program, wherein when the above program runs, it controls the device where the above computer-readable storage medium is located to execute the above method for identifying enterprise credit risk.

[0183] The embodiment of the present invention provides an electronic device, the device includes a processor, a memory, and a program stored on the memory and executable on the processor, and when the processor executes the program, it implements the steps in the above method for identifying enterprise credit risk.

[0184] The present application also provides a computer program product, which is suitable for executing a program initialized with the steps in the method for identifying enterprise credit risk when executed on a data processing device.

[0185] Obviously, those skilled in the art should understand that the various modules or steps of the present invention described above can be implemented by a general-purpose computing device. They can be concentrated on a single computing device or distributed over a network composed of multiple computing devices. They can be implemented by program codes executable by the computing device. Thus, they can be stored in a storage device and executed by the computing device. And in some cases, the steps shown or described can be executed in a different order than here, or they can be separately fabricated into individual integrated circuit modules, or multiple modules or steps among them can be fabricated into a single integrated circuit module for implementation. In this way, the present invention is not limited to any specific combination of hardware and software.

[0186] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0187] The present application is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to the embodiments of the present application. It should be understood that each flow and / or block in the flowchart and / or block diagram, as well as the combination of flows and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing devices generate a device for implementing the functions specified in Figure 1 one or more flows Figure 1 or multiple flows and / or blocks

[0188] These computer program instructions can also be stored in a computer-readable memory capable of guiding the computer or other programmable data processing devices to work in a specific manner, such that the instructions stored in the computer-readable memory generate a manufactured article including an instruction device, and the instruction device implements the functions specified in Figure 1 one or more flows Figure 1 or multiple flows and / or blocks

[0189] These computer program instructions can also be loaded onto the computer or other programmable data processing devices, such that a series of operation steps are executed on the computer or other programmable devices to generate a computer-implemented process. Thus, the instructions executed on the computer or other programmable devices provide for implementing the functions in the flowFigure 1 one or more processes and / or blocks Figure 1 steps of functions specified in one or more blocks

[0190] In a typical configuration, a computing device includes one or more processors (CPUs), an input / output interface, a network interface, and memory.

[0191] The memory may include non-permanent memory in the form of computer-readable media, random access memory (RAM) and / or non-volatile memory such as read-only memory (ROM) or flash memory (flash RAM). The memory is an example of computer-readable media.

[0192] Computer-readable media includes permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology for information storage. The information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette tapes or other magnetic storage devices, or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable media does not include transitory media such as modulated data signals and carrier waves.

[0193] It should also be noted that the term "comprising", "including" or any other variation thereof is intended to cover non-exclusive inclusion, such that a process, method, commodity or device comprising a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, commodity or device. Without further limitation, an element defined by the statement "comprising an..." does not exclude the presence of additional identical elements in the process, method, commodity or device comprising the element.

[0194] The above description is only the preferred embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for identifying corporate credit risk, characterized in that: include: Obtaining a large language model and training samples, and pre-training the large language model using the training samples to obtain an initial large language model after pre-training, wherein the training samples include financial data and label information corresponding to the financial data, wherein the large language model adopts a decoder-only architecture; Applying a fine-tuning task to the hybrid expert model to fine-tune the initial large language model to obtain a fine-tuned large language model, wherein the fine-tuning task is determined based on a fine-tuning task dataset; Optimizing the fine-tuned large language model through a preset prediction task to obtain a target large language model; The financial data of the enterprise to be identified for credit risk is input into the target large language model to obtain the credit risk identification result of the enterprise to be identified for credit risk.

2. The method according to claim 1, characterized in that The hybrid expert model is a hybrid expert model based on Monte Carlo routing. The fine-tuning task dataset is applied to the hybrid expert model to fine-tune the initial large language model to obtain a fine-tuned large language model, including: Acquire the fine-tuning task data set, and set a plurality of fine-tuning tasks according to the fine-tuning task data set; Based on the multiple fine-tuning tasks, the initial large language model is fine-tuned using the Monte Carlo routing-based hybrid expert model to obtain the fine-tuned large language model.

3. The method according to claim 2, characterized in that Based on the plurality of fine-tuning tasks, the initial large language model is fine-tuned using the Monte Carlo routing-based hybrid expert model to obtain the fine-tuned large language model, including: A sampling step, wherein the Monte Carlo routing-based hybrid expert model dynamically selects and activates a target expert model using Monte Carlo routing technology, wherein the target expert model is an expert model whose relevance to the current fine-tuning task is greater than a preset relevance threshold; A determination step, determining the weight of each target expert model by using the Monte Carlo routing technique; The sampling step and the determining step are repeatedly performed until a preset stopping condition is reached, thereby obtaining the fine-tuned large language model.

4. The method according to claim 1, characterized in that: The fine-tuned large language model is optimized by a preset prediction task to obtain a target large language model, including: A setting step, setting the prediction task, wherein the output of the prediction task is a numerical value; A prediction step, inputting the prediction task into the fine-tuned large language model for prediction to obtain a prediction result; An adjustment step, adjusting the parameters of the fine-tuned large language model according to the difference between the predicted result and the actual result; The prediction step and the adjustment step are repeatedly performed until a preset optimization standard is reached to obtain the target large language model.

5. The method according to claim 1, characterized in that After optimizing the fine-tuned large language model through a preset prediction task to obtain a target large language model, the method further includes: Based on different financial activity scenarios, corresponding prompt guide words are set and stored, wherein the financial activity scenarios are specific operating environments set according to different business stages and needs including pre-loan management, loan management and post-loan management; The credit risk-related issues are analyzed according to the prompt guiding words to generate decision texts that meet the needs of different financial activity scenarios.

6. The method according to claim 1, characterized in that Acquiring a large language model and training samples, and pre-training the large language model using the training samples to obtain an initial large language model after pre-training, including: Capturing the training samples from multiple data sources, wherein the data sources include news social media, regulatory announcement filings, abnormal operation announcements, and financial professional reports; The large language model is pre-trained using the training samples to obtain the pre-trained initial large language model, so that the initial large language model learns relevant information of the training samples.

7. The method according to claim 1, characterized in that Before obtaining the large language model and the training sample, the method further includes: The financial data is evaluated using a preset evaluation method to obtain the label information of the financial data, wherein the preset evaluation method includes a scoring card model and a GPT-based model.

8. A device for identifying corporate credit risk, characterized in that: include: A pre-training unit, used to obtain a large language model and training samples, and use the training samples to pre-train the large language model to obtain an initial large language model after pre-training, wherein the training samples include financial data and label information corresponding to the financial data, wherein the large language model adopts a decoder-only architecture; A fine-tuning unit, configured to apply a fine-tuning task to the hybrid expert model to fine-tune the initial large language model to obtain a fine-tuned large language model, wherein the fine-tuning task is determined based on a fine-tuning task data set; An optimization unit, configured to optimize the fine-tuned large language model through a preset prediction task to obtain a target large language model; The input unit is used to input the financial data of the enterprise to be identified for credit risk into the target large language model to obtain the credit risk identification result of the enterprise to be identified for credit risk.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored program, wherein when the program is executed, the device where the computer-readable storage medium is located is controlled to execute the method for identifying corporate credit risk according to any one of claims 1 to 7.

10. An electronic device, characterized in that: include: One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, and the one or more programs include a method for executing the corporate credit risk identification method described in any one of claims 1 to 7.

Images