Fine-tuning method, device and equipment of large model in banking business and storage medium

Abstract

Embodiments of the present application disclose a fine-tuning method, device and equipment for large models in banking business and a storage medium. The method comprises receiving a request for a case in the banking business, wherein the request for the case comprises a data type that needs to be processed in the banking business; selecting a corresponding large model and constructing a fine-tuning auxiliary model according to the data type and the request for the case, and connecting the fine-tuning auxiliary model to the large model; receiving labeled data samples, inputting the large model and the fine-tuning auxiliary model, fixing the first parameters of the large model, and training the fine-tuning auxiliary model; when the large model and the fine-tuning auxiliary model converge, sending the second parameters of the fine-tuning auxiliary model to the large model; and fine-tuning the large model according to the second parameters. The method realizes fine-tuning of the large model with a small amount of data, reduces the training cost of the large model when applied in different fields of the banking industry, and makes the service of the banking business more high-quality, convenient and digital.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to methods, apparatus, devices, and storage media for fine-tuning large models in banking operations. Background Technology

[0002] Large models are a trend in the field of artificial intelligence because their large number of parameters makes them more robust in relevant scenarios. However, training large models is extremely costly, and not all companies have the resources and capabilities to train them. Therefore, fine-tuning methods for large models are active in academia.

[0003] Because bank data has high security requirements, it cannot be provided to external parties for training large models. Although many large models have emerged, adjustments and improvements to the implementation of large models are still needed for different scenarios. Therefore, fine-tuning technology is widely used in the implementation of large models in the banking industry. Summary of the Invention

[0004] This invention proposes a method, apparatus, device, and storage medium for fine-tuning large models in banking operations, which solves the problems of requiring a large amount of training data and high training costs when large models are applied in different areas of the banking industry.

[0005] This invention provides a method for fine-tuning a large model in banking operations, the method comprising:

[0006] Receive a scenario request in banking operations, wherein the scenario request includes the data type that needs to be processed in the banking operations;

[0007] Based on the data type and the occasion request, select the corresponding large model and construct a fine-tuning auxiliary model, and connect the fine-tuning auxiliary model to the large model;

[0008] Receive labeled data samples, input the large model and the fine-tuning auxiliary model, fix the first parameter of the large model, and train the fine-tuning auxiliary model;

[0009] When the large model and the fine-tuned auxiliary model converge, the second parameter of the fine-tuned auxiliary model is sent to the large model.

[0010] The large model is fine-tuned based on the second parameter of the fine-tuning auxiliary model.

[0011] Furthermore, the fine-tuning auxiliary model includes an encoder and a decoder. The encoder is used to extract target feature information from the data type, and the decoder is used to reconstruct target context information from the target feature information.

[0012] Furthermore, adjusting the large model according to the second parameter of the fine-tuning auxiliary model includes:

[0013] Based on the second parameter, the target feature information extracted by the encoder in the fine-tuning auxiliary model is added to the corresponding position in the large model;

[0014] Based on the second parameter, the target situation information output by the decoder in the fine-tuning auxiliary model is added to the corresponding position in the large model.

[0015] Furthermore, the step of adding the target feature information extracted by the encoder in the fine-tuning auxiliary model to the corresponding position in the large model according to the second parameter includes:

[0016] Connect the encoder in the fine-tuning auxiliary model to the corresponding encoder in the large model;

[0017] The target feature information of the encoder in the fine-tuning auxiliary model is superimposed onto the corresponding position of the encoder in the large model.

[0018] Furthermore, the step of superimposing the target feature information of the encoder in the fine-tuning auxiliary model onto the corresponding position of the encoder corresponding to the large model specifically involves:

[0019] The target feature information output by the encoder in the fine-tuning auxiliary model is superimposed onto the target feature information output by the corresponding encoder in the large model.

[0020] Furthermore, the step of superimposing the target feature information output by the encoder in the fine-tuning auxiliary model onto the target feature information output by the corresponding encoder in the large model further includes:

[0021] The superimposed target feature information is then input into the next processor of the corresponding encoder in the large model.

[0022] Furthermore, the step of adding the target situation information output by the decoder in the fine-tuning auxiliary model to the corresponding position in the large model according to the second parameter includes:

[0023] The target context information output by the decoder in the fine-tuning auxiliary model is superimposed onto the target context information output by the large model.

[0024] Furthermore, the loss function of the fine-tuning auxiliary model is the difference between the loss function of the large model and the overfitting prevention loss function, wherein the overfitting prevention loss function is obtained from the encoder layer and decoder layer in the fine-tuning auxiliary model.

[0025] This invention also provides a fine-tuning device for a large model in banking operations, the device comprising:

[0026] A receiving module is used to receive occasion requests in banking operations, wherein the occasion requests include data types that need to be processed in banking operations;

[0027] The access module is used to select the corresponding large model and construct the fine-tuning auxiliary model according to the data type, and to connect the fine-tuning auxiliary model to the large model.

[0028] The training module is used to receive labeled data samples, input the large model and the fine-tuning auxiliary model, fix the first parameters of the large model, and train the fine-tuning auxiliary model.

[0029] The sending module is used to send the second parameters of the fine-tuned auxiliary model to the large model when the model converges after training.

[0030] The fine-tuning module is used to adjust the large model according to the second parameter of the fine-tuning auxiliary model.

[0031] This invention also provides a fine-tuning device for a large model in banking operations, including a memory and a processor. The memory stores a computer program, which, when executed by the processor, causes the processor to perform the following steps:

[0032] Receive a scenario request in banking operations, wherein the scenario request includes the data type that needs to be processed in the banking operations;

[0033] Based on the data type and the occasion request, select the corresponding large model and construct a fine-tuning auxiliary model, and connect the fine-tuning auxiliary model to the large model;

[0034] Receive labeled data samples, input the large model and the fine-tuning auxiliary model, fix the first parameter of the large model, and train the fine-tuning auxiliary model;

[0035] When the large model and the fine-tuned auxiliary model converge, the second parameter of the fine-tuned auxiliary model is sent to the large model.

[0036] The large model is fine-tuned based on the second parameter of the fine-tuning auxiliary model.

[0037] This invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the following steps:

[0038] Receive a scenario request in banking operations, wherein the scenario request includes the data type that needs to be processed in the banking operations;

[0039] Based on the data type and the occasion request, select the corresponding large model and construct a fine-tuning auxiliary model, and connect the fine-tuning auxiliary model to the large model;

[0040] Receive labeled data samples, input the large model and the fine-tuning auxiliary model, fix the first parameter of the large model, and train the fine-tuning auxiliary model;

[0041] When the large model and the fine-tuned auxiliary model converge, the second parameter of the fine-tuned auxiliary model is sent to the large model.

[0042] The large model is fine-tuned based on the second parameter of the fine-tuning auxiliary model.

[0043] The fine-tuning method for large models in banking operations provided in this invention trains a fine-tuning auxiliary model constructed from data types and occasion requests, obtains the second parameter corresponding to the fine-tuning auxiliary model, and fine-tunes the large model corresponding to the data types and occasion requests using the second parameter. This achieves fine-tuning of large models with a small amount of data, reduces the training cost of large models when applied in different areas of the banking industry, promotes the application of large models in the banking industry, and makes banking services more efficient, convenient, and digital. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] in:

[0046] Figure 1 A schematic diagram of the network architecture of a fine-tuning method for a large model in banking operations provided by an embodiment of the present invention;

[0047] Figure 2 This is a schematic diagram of the structure of the large model fine-tuning method in banking operations provided in an embodiment of the present invention;

[0048] Figure 3 A flowchart illustrating a method for fine-tuning a large model in banking operations, provided as an embodiment of the present invention;

[0049] Figure 4 This is a schematic diagram illustrating the process of adjusting a large model using a second parameter, as provided in an embodiment of the present invention.

[0050] Figure 5 This is a schematic diagram of the structure of a fine-tuning device for a large model in banking operations, provided by an embodiment of the present invention.

[0051] Figure 6 This is a schematic diagram of the structure of a fine-tuning device for a large model in banking operations, provided by an embodiment of the present invention.

[0052] Figure 7 This is a schematic diagram of the structure of a computer-readable storage medium provided in an embodiment of the present invention. Detailed Implementation

[0053] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] To facilitate understanding, the relevant terms used in this application will be introduced below.

[0055] (1) Large Scale Model refers to a neural network model consisting of millions or even billions of parameters. They can handle large-scale datasets and have very strong learning and generalization capabilities.

[0056] (2) Banking business includes liability business, asset business, and intermediary business. Among them, liability business is the business that commercial banks use to generate funds, the most important of which is deposit business. Asset business includes loan business and securities investment business. Intermediary business, also known as off-balance-sheet business, refers to business that is not reflected in the balance sheet; it includes: wealth management business, transaction clearing business, payment and settlement business, bank card business, agency business, custody business, electronic banking business, etc.

[0057] (3) LoRa technology is a spread spectrum modulation technique, also known as Chirp modulation, which is a proprietary IP of Semtech. Spread spectrum technology is a technique that trades bandwidth for sensitivity. Wi-Fi, ZigBee, and other technologies use spread spectrum technology, but LoRa modulation is characterized by approaching the limits of Shannon's theorem, maximizing sensitivity. Compared to traditional FSK technology, at the same communication rate, LoRa has 8-12 dBm better sensitivity than FSK.

[0058] (4) Hugging Face Hub. Similar to GitHub, Hugging Face Hub is a hub (community). Hugging Face can be considered the GitHub of the machine learning world. Hugging Face provides users with the following main functions: ① Model Repository: Just as Git repositories allow you to manage code versions and open-source code, model repositories allow you to manage model versions and open-source models. The usage is similar to GitHub. ② Models: Hugging Face provides many pre-trained machine learning models for different machine learning tasks, which are stored in the model repository. ③ Datasets: Hugging Face has many public datasets.

[0059] (5) Open source websites are collaborative websites that rely on community production and peer review to use, modify and share open source projects.

[0060] (6) Image generation (IG) refers to the task of generating new images from an existing dataset. Image generation models include two categories: unconditional generation and conditional generation. Unconditional generation refers to generating samples from the dataset without conditions, i.e., p(y); conditional image generation refers to generating samples from the dataset conditionally according to the labels, i.e., p(y|x).

[0061] like Figure 1 The diagram shown illustrates the network architecture of a fine-tuning method for a large model in banking operations, as provided in an embodiment of the present invention. The network architecture may include a server 200 and a cluster of terminal devices. The cluster of terminal devices may include one or more terminal devices; in this embodiment, the number of terminal devices is not limited. Figure 1 As shown, multiple terminal devices may specifically include terminal device 1, terminal device 2, terminal device 3, ..., terminal device n; as... Figure 1 As shown, terminal device 1, terminal device 2, terminal device 3, ..., terminal device n are all connected to server 200 through network 300, so that each terminal device can interact with server 200 through network 300.

[0062] like Figure 1The server 200 shown can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The terminal devices can be smart terminals such as smartphones, tablets, laptops, desktop computers, and smart TVs. The following description uses the communication between terminal device 1, terminal device 2, and server 200 as an example to illustrate the specific implementation of this application.

[0063] Terminal devices 1, 2, 3, ..., n can be used to collect scenario requests in banking operations and send these requests to server 200. Server 200 receives these scenario requests, which include data types that need to be processed in banking operations. Based on the data type, a corresponding large model and a fine-tuning auxiliary model are selected and constructed. The fine-tuning auxiliary model is then integrated into the large model. Labeled data samples are received and input into the large model and the fine-tuning auxiliary model. The first parameter of the large model is fixed, and the fine-tuning auxiliary model is trained. When the fine-tuning auxiliary model converges, its second parameter is sent to the large model. The large model is then fine-tuned based on the fine-tuning results of the fine-tuning auxiliary model. This fine-tuning method for the large model in banking operations can also be directly applied to terminal devices 1, 2, 3, ..., n. This embodiment of the invention uses this method to run on server 200, and the server 200 transmits data with terminal device 1, terminal device 2, terminal device 3, ..., terminal device n to illustrate the specific implementation process of this method.

[0064] like Figure 2 The diagram shown illustrates the structure of a large model fine-tuning method for banking operations provided in this embodiment of the invention, including a large model, a fine-tuning auxiliary model, and data samples. For a detailed description of the fine-tuning implementation process, please refer to [link / reference]. Figure 3 ,like Figure 3 The diagram shown is a flowchart illustrating a method for fine-tuning a large model in banking operations according to an embodiment of the present invention. The method includes:

[0065] Step S101: Receive a scenario request in banking operations, wherein the scenario request includes the data type that needs to be processed in banking operations;

[0066] Specifically, in this embodiment, the data types include image data, voice data, and text data generated during banking operations.

[0067] Context requests refer to specific processing requests for image data, voice data, and text data generated in banking operations. Examples include image processing, image segmentation, image recognition, image enhancement, image classification, image fusion, image stitching, image denoising, image preprocessing, and image registration for image data; speech recognition, speech synthesis, speech signal processing, speech emotion recognition, and speech enhancement for voice data; and text classification, text sentiment analysis, text analysis, text mining, text clustering, text similarity, and text-to-speech processing for text data.

[0068] Step S102: Select the corresponding large model and construct the fine-tuning auxiliary model according to the data type and the occasion request, and connect the fine-tuning auxiliary model to the large model;

[0069] Specifically, in this implementation, when selecting a large model, the appropriate method is used to select the model based on the data type and context of the request to be processed. For example, in this implementation, once the data type and corresponding context of the request to be processed are determined, a suitable large model for processing the request in that context of that data type can be selected from open-source websites (such as GitHub, Hugging Face, codeKK, Android Arsenal, Android Libraries and Resources, and GrepCode).

[0070] The selected large model installation package, containing the model's weight parameters, is then downloaded to the bank's server. Based on the bank's server configuration, the downloaded large model installation package and its dependencies are installed. For toolkits not available on the bank's intranet, offline import can be used. After installation, the large model is run to confirm its proper functioning on the server.

[0071] It should be noted that when downloading and installing large models, the installation directory corresponding to the large model installation package can be configured on the bank's server to facilitate the management of large models.

[0072] Once the main model is running normally, the corresponding auxiliary model can be built according to the selected data type and context request, and then integrated into the main model. For example, in this implementation, a Transformer model can be used as a fine-tuning auxiliary model and integrated into the main model.

[0073] Step S103: Receive labeled data samples, input the large model and the fine-tuning auxiliary model, fix the first parameters of the large model, and train the fine-tuning auxiliary model;

[0074] Specifically, in this embodiment, image generation is used as an example. A bank image dataset and the final generated images from the dataset are collected and input into an image generation model consisting of a large model and an auxiliary model for training. During the training process, the weights of the large model are fixed and the parameters are not changed.

[0075] Step S104: When the large model and the fine-tuning auxiliary model converge, the second parameter of the fine-tuning auxiliary model is sent to the large model.

[0076] Specifically, in this embodiment, following the image generation in step S103 above, the image generation model, composed of the large model and the auxiliary model, is trained by inputting the bank image dataset and the final generated image from that dataset. Training continues until the model converges (i.e., the range of variation is within an acceptable order of magnitude), at which point training stops. The second parameters of the auxiliary model are then obtained and sent to the large model.

[0077] Step S105: Fine-tune the large model according to the second parameter of the fine-tuning auxiliary model.

[0078] Specifically, in this embodiment, an interface for modifying the model is added inside the large model, that is, a function that can modify the large model is added to the large model.

[0079] The second parameter is received through the interface in the large model, and the model is adjusted according to the second parameter to achieve fine-tuning of the large model.

[0080] The fine-tuning method for large models in banking operations provided in this invention trains a fine-tuning auxiliary model constructed from data types and occasion requests, obtains the second parameter corresponding to the fine-tuning auxiliary model, and fine-tunes the large model corresponding to the data types and occasion requests using the second parameter. This achieves fine-tuning of large models with a small amount of data, reduces the training cost of large models when applied in different areas of the banking industry, promotes the application of large models in the banking industry, and makes banking services more efficient, convenient, and digital.

[0081] In some embodiments, the fine-tuning auxiliary model includes an encoder and a decoder, wherein the encoder is used to extract target feature information from the data type; and the decoder is used to reconstruct target occasion information from the target feature information.

[0082] For details, please see Figure 2The fine-tuning auxiliary model includes an encoder and a decoder. The encoder extracts target feature information from the data samples and then inputs the extracted target feature information into the large model and the decoder. The decoder reconstructs the corresponding target context information from the target feature information output by the encoder. For example, in the image generation described above, the decoder extracts feature information of the generated image from the image samples and then inputs it into the large model and the decoder respectively. The decoder then reconstructs the generated image information from the feature information extracted by the decoder.

[0083] It should be noted that the decoder and encoder can be constructed according to actual needs, and no specific limitations are made in this application. In the above embodiments, the encoder and decoder in image generation can both be constructed using an embedding layer, an attention layer, and a forward propagation layer. The embedding layer consists of a matrix of a certain dimension, with each row representing a corresponding embedding vector. In the attention layer, K and Q are both composed of neural networks and bn layers. After matrix multiplication of the input, the attention-related matrix M is obtained. V is also composed of a neural network and bn layers. After operations on the input, it is multiplied by matrix M to output the attention-adjusted value. The forward network is based on a neural network, consisting of hierarchical normalization and activation layers, mainly extracting information from the target input of the attention layer. These layers are stacked in a certain number (2-4) to form the encoder and decoder.

[0084] It should be noted that target feature information is adjusted based on data type, contextual requirements, the overall model, and fine-tuning assistance. Similarly, target contextual information is also adjusted based on data type, contextual requirements, the overall model, and fine-tuning assistance.

[0085] Preferably, LoRa technology can be used to construct a fine-tuning auxiliary model and integrate it with a large model.

[0086] In some embodiments, such as Figure 4 The diagram shown illustrates the process of adjusting a large model using a second parameter according to an embodiment of the present invention. Step S105 includes:

[0087] Step S1051: Add the target feature information extracted by the encoder in the fine-tuning auxiliary model to the corresponding position in the large model according to the second parameter;

[0088] Step S1052: Add the target situation information output by the decoder in the fine-tuning auxiliary model to the corresponding position in the large model according to the second parameter.

[0089] Specifically, in this embodiment, fine-tuning the large model using the second parameter is divided into two steps. First, the target feature information extracted by the encoder is added to the corresponding position of the large model using the second parameter. Then, the target feature information obtained by the decoder to restore the target feature information is added to the corresponding position of the large model using the second parameter. This changes the intermediate and final results of the large model, thereby realizing the distribution adjustment of the output results of the large model by the fine-tuning auxiliary model.

[0090] It should be noted that in this implementation, the location of adding the target feature information extracted by the encoder to the large model is adjusted according to the data type and the specific requirements. This adjustment can be verified through experiments. Similarly, the location of adding the target feature information output by the decoder to the large model is also adjusted according to the data type and the specific requirements. This adjustment can be verified through experiments.

[0091] In some embodiments, adding the target feature information extracted by the encoder in the fine-tuning auxiliary model to the corresponding position in the large model according to the second parameter includes:

[0092] Connect the encoder in the fine-tuning auxiliary model to the corresponding encoder in the large model;

[0093] The target feature information of the encoder in the fine-tuning auxiliary model is superimposed onto the encoder corresponding to the large model.

[0094] Specifically, in this embodiment, the decoder of the fine-tuning auxiliary model is first connected to the decoder of the large model. Then, the target feature information output by the encoder of the fine-tuning auxiliary model is superimposed on the corresponding position of the encoder in the large model. For example, it can be added to the input data of the large model encoder, the data processed by the large model encoder, or the output data of the large model encoder. The specific details are determined through experiments based on the data type and scenario requirements.

[0095] In some embodiments, the step of superimposing the target feature information of the encoder in the fine-tuning auxiliary model onto the corresponding position of the encoder corresponding to the large model specifically involves:

[0096] The target feature information output by the encoder in the fine-tuning auxiliary model is superimposed onto the target feature information output by the corresponding encoder in the large model.

[0097] Specifically, in this embodiment, by adding the target feature information output by the auxiliary model encoder to the target feature output by the large model encoder, the intermediate results of the large model are fine-tuned, so that the large model has a higher recognition accuracy when applied.

[0098] In some embodiments, the step of superimposing the target feature information output by the encoder in the fine-tuning auxiliary model onto the target feature information output by the corresponding encoder in the large model further includes:

[0099] The superimposed target feature information is then input into the next processor of the corresponding encoder in the large model.

[0100] Specifically, in this embodiment, the target feature information output by the auxiliary model encoder is superimposed with the target feature output by the large model encoder and then input into the next processor of the large model encoder for processing to obtain the target situation information of the large model output that changes the intermediate result.

[0101] In some embodiments, adding the target situation information output by the decoder in the fine-tuning auxiliary model to the corresponding position in the large model according to the second parameter includes:

[0102] The target context information output by the decoder in the fine-tuning auxiliary model is superimposed onto the target context information output by the large model.

[0103] Specifically, in this implementation, the target situation information restored by the fine-tuning auxiliary model decoder is added to the target situation information output by the large model, and the output result of the large model is fine-tuned to make the output result more accurate during application.

[0104] In some embodiments, the loss function of the fine-tuning auxiliary model is the difference between the loss function of the large model and the overfitting prevention loss function, wherein the overfitting prevention loss function is obtained based on the encoder layer and decoder layer in the fine-tuning auxiliary model.

[0105] Specifically, in this implementation, the loss function of the fine-tuning auxiliary model is set as L, which includes two parts: the first part is the loss function L1 originally used for training the large model, and the second part is the loss function L2 to prevent overfitting. L2 is obtained from the encoder and decoder layers in the fine-tuning auxiliary model, and L is obtained by subtracting L1 and L2. That is, L = L1 - L2, where, i represents the i-th layer of the neural network in the encoder or decoder, N represents the total number of layers in the neural network, and ω i Let |ω be the parameter of the i-th layer of the neural network. i | represents the absolute value of the parameters of the i-th layer of the neural network. During training, the learning rate can be set to 3e. -6 ~1e -5 (where e represents an exponent with base 10) The parameters of the auxiliary model are updated using gradient descent.

[0106] like Figure 5The diagram shown is a structural schematic of a fine-tuning device for a large model in banking operations provided by an embodiment of the present invention. The device includes:

[0107] The receiving module 501 is used to receive occasion requests in banking business, wherein the occasion requests include data types that need to be processed in banking business;

[0108] The access module 502 is used to select the corresponding large model and construct the fine-tuning auxiliary model according to the data type, and to connect the fine-tuning auxiliary model to the large model.

[0109] Training module 503 is used to receive labeled data samples, input the large model and the fine-tuning auxiliary model, fix the first parameters of the large model, and train the fine-tuning auxiliary model;

[0110] The sending module 504 is used to send the second parameters of the fine-tuning auxiliary model to the large model when the model converges after training.

[0111] The fine-tuning module 505 is used to adjust the large model according to the second parameter of the fine-tuning auxiliary model.

[0112] In some embodiments, the fine-tuning auxiliary model in the access module 502 includes an encoder and a decoder. The encoder is used to extract target feature information from the data type, and the decoder is used to restore target occasion information from the target feature information.

[0113] In some embodiments, the fine-tuning module 505 includes:

[0114] The first fine-tuning module is used to add the target feature information extracted by the encoder in the fine-tuning auxiliary model to the corresponding position in the large model according to the second parameter.

[0115] The second fine-tuning module is used to add the target situation information output by the decoder in the fine-tuning auxiliary model to the corresponding position in the large model according to the second parameter.

[0116] In some embodiments, the first fine-tuning module is further configured to:

[0117] Connect the encoder in the fine-tuning auxiliary model to the corresponding encoder in the large model;

[0118] The target feature information of the encoder in the fine-tuning auxiliary model is superimposed onto the corresponding position of the encoder in the large model.

[0119] In some embodiments, the first fine-tuning module is further configured to superimpose the target feature information output by the encoder in the fine-tuning auxiliary model onto the target feature information output by the corresponding encoder in the large model.

[0120] In some embodiments, the apparatus further includes:

[0121] The input module is used to input the superimposed target feature information into the next processor of the corresponding encoder in the large model.

[0122] In some embodiments, the second fine-tuning module is further configured to superimpose the target context information output by the decoder in the fine-tuning auxiliary model onto the target context information output by the large model.

[0123] In some embodiments, the loss function of the fine-tuning auxiliary model in the access module 502 is the difference between the loss function of the large model and the overfitting prevention loss function, wherein the overfitting prevention loss function is obtained based on the encoder layer and decoder layer in the fine-tuning auxiliary model.

[0124] For further details regarding the implementation of the above technical solutions by each module in the fine-tuning device for the large model in banking operations, please refer to the description in the fine-tuning method for the large model in banking operations provided above, which will not be repeated here.

[0125] In some embodiments, such as Figure 6 The diagram shown is a structural schematic of a fine-tuning device for a large model in banking operations provided by an embodiment of the present invention. The device includes a memory 601 and a processor 602. The memory 601 stores a computer program. When the computer program is executed by the processor 602, the processor 602 performs the following steps:

[0126] Receive a scenario request in banking operations, wherein the scenario request includes the data type that needs to be processed in the banking operations;

[0127] Based on the data type and the occasion request, select the corresponding large model and construct a fine-tuning auxiliary model, and connect the fine-tuning auxiliary model to the large model;

[0128] Receive labeled data samples, input the large model and the fine-tuning auxiliary model, fix the first parameter of the large model, and train the fine-tuning auxiliary model;

[0129] When the large model and the fine-tuned auxiliary model converge, the second parameter of the fine-tuned auxiliary model is sent to the large model.

[0130] The large model is fine-tuned based on the second parameter of the fine-tuning auxiliary model.

[0131] For further details regarding the implementation of the above technical solution by the processor 601 in the fine-tuning device for large models in banking operations, please refer to the description in the above-mentioned fine-tuning method for large models in banking operations, which will not be repeated here.

[0132] The processor 601 can also be called a CPU (Central Processing Unit). The processor 601 may be an integrated circuit chip with signal processing capabilities. The processor 601 can also be a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The general-purpose processor can be a microprocessor, or the processor 601 can be any conventional processor.

[0133] In some embodiments, such as Figure 7 The diagram illustrates the structure of a computer-readable storage medium according to an embodiment of the present invention. The storage medium stores a readable computer program 701. This computer program 701 can be stored in the storage medium in the form of a software product, and includes several instructions to cause a computer device (which may be a personal computer, a server machine, or a network device, etc.) or a processor to perform the following steps:

[0134] Receive a scenario request in banking operations, wherein the scenario request includes the data type that needs to be processed in the banking operations;

[0135] Based on the data type and the occasion request, select the corresponding large model and construct a fine-tuning auxiliary model, and connect the fine-tuning auxiliary model to the large model;

[0136] Receive labeled data samples, input the large model and the fine-tuning auxiliary model, fix the first parameter of the large model, and train the fine-tuning auxiliary model;

[0137] When the large model and the fine-tuned auxiliary model converge, the second parameter of the fine-tuned auxiliary model is sent to the large model.

[0138] The large model is fine-tuned based on the second parameter of the fine-tuning auxiliary model.

[0139] The aforementioned storage media include: USB flash drives, portable hard drives, magnetic disks or optical disks, ROM (Read-Only Memory), RAM (Random Access Memory), and other media that can store program code, or terminal devices such as computers, servers, mobile phones, and tablets.

[0140] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0141] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0142] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for fine-tuning a large model in banking operations, characterized in that, The method includes: Receive a scenario request in banking operations, wherein the scenario request includes the data type that needs to be processed in the banking operations; Based on the data type and the occasion request, select the corresponding large model and construct a fine-tuning auxiliary model, and connect the fine-tuning auxiliary model to the large model; Receive labeled data samples, input the large model and the fine-tuning auxiliary model, fix the first parameter of the large model, and train the fine-tuning auxiliary model; When the large model and the fine-tuned auxiliary model converge, the second parameter of the fine-tuned auxiliary model is sent to the large model. The large model is fine-tuned according to the second parameter of the fine-tuning auxiliary model; The fine-tuning auxiliary model includes an encoder and a decoder. The encoder is used to extract target feature information from the data type; the decoder is used to reconstruct target context information from the target feature information. The step of fine-tuning the large model according to the second parameter of the fine-tuning auxiliary model includes: Based on the second parameter, the target feature information extracted by the encoder in the fine-tuning auxiliary model is added to the corresponding position in the large model; Based on the second parameter, the target situation information output by the decoder in the fine-tuning auxiliary model is added to the corresponding position in the large model.

2. The method for fine-tuning a large model in banking operations according to claim 1, characterized in that, The step of adding the target feature information extracted by the encoder in the fine-tuning auxiliary model to the corresponding position in the large model according to the second parameter includes: Connect the encoder in the fine-tuning auxiliary model to the corresponding encoder in the large model; The target feature information of the encoder in the fine-tuning auxiliary model is superimposed onto the corresponding position of the encoder in the large model.

3. The method for fine-tuning a large model in banking operations according to claim 2, characterized in that, The step of superimposing the target feature information of the encoder in the fine-tuning auxiliary model onto the corresponding position of the encoder corresponding to the large model specifically involves: The target feature information output by the encoder in the fine-tuning auxiliary model is superimposed onto the target feature information output by the corresponding encoder in the large model.

4. The method for fine-tuning a large model in banking operations according to claim 3, characterized in that, After superimposing the target feature information output by the encoder in the fine-tuning auxiliary model onto the target feature information output by the corresponding encoder in the large model, the method further includes: The superimposed target feature information is then input into the next processor of the corresponding encoder in the large model.

5. The method for fine-tuning a large model in banking operations according to claim 1, characterized in that, The step of adding the target situation information output by the decoder in the fine-tuning auxiliary model to the corresponding position in the large model according to the second parameter includes: The target context information output by the decoder in the fine-tuning auxiliary model is superimposed onto the target context information output by the large model.

6. The method for fine-tuning a large model in banking operations according to any one of claims 1-5, characterized in that, The loss function of the fine-tuning auxiliary model is the difference between the loss function of the large model and the overfitting prevention loss function, which is obtained from the encoder layer and decoder layer in the fine-tuning auxiliary model.

7. A fine-tuning device for a large model in banking operations, applied to the fine-tuning method for a large model in banking operations as described in claim 1, characterized in that, The device includes: A receiving module is used to receive occasion requests in banking operations, wherein the occasion requests include data types that need to be processed in banking operations; The access module is used to select the corresponding large model and construct the fine-tuning auxiliary model according to the data type, and to connect the fine-tuning auxiliary model to the large model. The training module is used to receive labeled data samples, input the large model and the fine-tuning auxiliary model, fix the first parameters of the large model, and train the fine-tuning auxiliary model. The sending module is used to send the second parameters of the fine-tuned auxiliary model to the large model when the fine-tuned auxiliary model has converged during training; The fine-tuning module is used to adjust the large model according to the second parameter of the fine-tuning auxiliary model.

8. A device for fine-tuning a large model in banking operations, comprising a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the steps of the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium storing a computer program that, when executed by a processor, causes the processor to perform the steps of the method as claimed in any one of claims 1 to 6.

Images