Method, device and equipment for financial data analysis based on model parameter adjustment and medium

Abstract

The application discloses a financial data analysis method and device based on model parameter adjustment, equipment and medium, relates to the computer technical field, including: identifying the task description text corresponding to the model parameter generation task of the financial data analysis model, and generating the target task representation vector based on the task description text, and generating the target parameter weight corresponding to the target task representation vector; the target task representation vector is decomposed into a corresponding low-rank matrix based on the target parameter weight, and a parameter weight file corresponding to the to-be-converted large language model is generated based on the low-rank matrix; the model parameters of the to-be-converted large language model are adjusted based on the parameter weight file, so as to convert the to-be-converted large language model into a target financial data analysis model; and the to-be-analyzed financial data is analyzed by the target financial data analysis model. Therefore, by fine-tuning the parameters and combining the task requirements, the model in other fields can be applied to the financial field, and the training cost of the model is effectively saved.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to a financial data analysis method, apparatus, device, and medium based on model parameter adjustment. Background Technology

[0002] With the rapid adoption of large language models (LLMs) across various industries, the demand for rapid model adaptation to tasks is increasing, especially in enterprise deployments, multi-tenant services, and real-time inference scenarios, where higher requirements are placed on the efficiency, economy, and flexibility of model adaptation. Traditional model fine-tuning methods mainly rely on efficient parameter fine-tuning techniques such as LoRA (Low-Rank Adaptation). Although these methods can reduce the number of training parameters, they still present significant challenges in terms of adaptation time and computational resources.

[0003] To adapt to the financial domain, traditional LoRA fine-tuning methods require complete gradient descent training on each new task. Even parameter-efficient methods require hours to days of training time, making it difficult to meet the demands for rapid deployment and real-time adaptation. Furthermore, current methods require significant GPU (Graphics Processing Unit) computing resources during training, including forward propagation, backpropagation, and gradient updates, increasing deployment costs and energy consumption. Moreover, current methods are typically optimized for specific task types, lacking cross-task adaptability and struggling to handle complex applications involving multiple task modes.

[0004] In the field of financial data analysis, models need to be frequently adapted to different task types, such as risk assessment, fraud detection, and investment analysis. Therefore, the efficiency bottleneck of traditional fine-tuning methods is particularly prominent. Summary of the Invention

[0005] In view of this, the purpose of this invention is to provide a financial data analysis method, apparatus, device, and medium based on model parameter adjustment. By simply fine-tuning the parameters and combining them with task requirements, models from other fields can be applied to the financial field, effectively saving model training costs. The specific solution is as follows:

[0006] Firstly, this application discloses a financial data analysis method based on model parameter adjustment, comprising:

[0007] The model parameters of the financial data analysis model are identified to generate task description text corresponding to the task, and several embedding vectors corresponding to several preset dimensions are generated based on the task description text.

[0008] A target task representation vector is generated based on the aforementioned embedding vectors, and target parameter weights corresponding to the target task representation vector are generated through a preset parameter function.

[0009] Based on the target parameter weights, the target task representation vector is decomposed into a low-rank matrix corresponding to the large language model to be converted, and a parameter weight file corresponding to the large language model to be converted is generated based on the low-rank matrix; the low-rank matrix is ​​a matrix whose rank satisfies the preset low-rank determination condition;

[0010] The model parameters of the large language model to be converted are adjusted based on the parameter weight file to convert the large language model to be converted into a target financial data analysis model corresponding to the model parameter generation task.

[0011] The target financial data analysis model is used to perform data analysis on the financial data to be analyzed.

[0012] Optionally, the step of generating task description text corresponding to the model parameters of the financial data analysis model and generating several embedding vectors corresponding to several preset dimensions based on the task description text includes:

[0013] If a model parameter generation task for a financial data analysis model is received, the task description text corresponding to the task parameter generation task is identified, and several target task descriptions corresponding to several preset dimensions in the task description text are determined.

[0014] The target task descriptions are mapped to a number of embedding vectors corresponding to the preset dimensions.

[0015] Optionally, before generating the target task representation vector based on the plurality of embedding vectors and generating the target parameter weights corresponding to the target task representation vector through a preset parameter function, the method further includes:

[0016] Collect historical model parameters to generate historical task representation vectors and historical parameter weights corresponding to the tasks;

[0017] Determine the correspondence between each vector in the historical task representation vector and each weight in the historical parameter weights;

[0018] Based on the correspondence, the corresponding historical task representation vector and the corresponding historical parameter weight are determined from the historical task representation vector and the historical parameter weight, and several historical pairing data are constructed according to the historical task representation vector and the corresponding historical parameter weight.

[0019] The parameter function to be trained is trained using the historical paired data to obtain the preset parameter function.

[0020] Optionally, the step of generating a target task representation vector based on the plurality of embedding vectors, and generating target parameter weights corresponding to the target task representation vector through a preset parameter function, includes:

[0021] The embedding vectors are fused based on a preset encoder to obtain the target task representation vector;

[0022] The target parameter weights corresponding to the target task vector are predicted by a preset parameter function.

[0023] Optionally, the step of decomposing the target task representation vector into a low-rank matrix corresponding to the large language model to be converted based on the target parameter weights, and generating a parameter weight file corresponding to the large language model to be converted based on the low-rank matrix, includes:

[0024] The model structure of the large language model to be transformed is determined, and the target task representation vector is transformed by several preset generation heads based on preset constraint optimization strategies and the target parameter weights to obtain several low-rank matrices corresponding to the model structure.

[0025] Based on each layer of the model structure, the several low-rank matrices are divided to obtain several target low-rank matrices corresponding to each layer, and several parameter weight files are generated based on the several target low-rank matrices.

[0026] Optionally, the preset constraint optimization strategy is:

[0027] ;

[0028] in, Let A and B be the target parameter weights, A and B be the low-rank matrices to be optimized, and λ be the preset regularization strength coefficient. denoted by Frobenius norm, representing the square root of the sum of squares of the matrix elements.

[0029] Optionally, adjusting the model parameters of the large language model to be converted based on the parameter weight file to convert the large language model to be converted into a target financial data analysis model corresponding to the model parameter generation task includes:

[0030] The parameter weight file is applied to the large language model to be converted, so as to adjust the model parameters of the large language model to be converted through the parameter weight file, so as to obtain the financial data analysis model to be determined.

[0031] The target validation dataset corresponding to the model parameter generation task is determined, and the model performance of the financial data analysis model to be determined is validated using the target validation dataset.

[0032] If the performance of the financial data analysis model to be determined is verified, then the financial data analysis model to be determined will be used as the target financial data analysis model for completing the type conversion.

[0033] Secondly, this application discloses a financial data analysis device based on model parameter adjustment, comprising:

[0034] The vector generation module is used to identify the model parameters of the financial data analysis model to generate task description text corresponding to the task, and generate several embedded vectors corresponding to several preset dimensions based on the task description text.

[0035] The weight generation module is used to generate a target task representation vector based on the plurality of embedded vectors, and to generate target parameter weights corresponding to the target task representation vector through a preset parameter function.

[0036] The weight transformation module is used to decompose the target task representation vector into a low-rank matrix corresponding to the large language model to be transformed based on the target parameter weights, and to generate a parameter weight file corresponding to the large language model to be transformed based on the low-rank matrix; the low-rank matrix is ​​a matrix whose rank satisfies a preset low-rank determination condition;

[0037] The model conversion module is used to adjust the model parameters of the large language model to be converted based on the parameter weight file, so as to convert the large language model to be converted into a target financial data analysis model corresponding to the model parameter generation task.

[0038] The financial data analysis module is used to perform data analysis on the financial data to be analyzed using the target financial data analysis model.

[0039] Thirdly, this application discloses an electronic device, including:

[0040] Memory, used to store computer programs;

[0041] A processor is used to execute the computer program to implement the financial data analysis method based on model parameter adjustment as described above.

[0042] Fourthly, this application discloses a computer-readable storage medium for storing a computer program, wherein the computer program, when executed by a processor, implements the aforementioned financial data analysis method based on model parameter adjustment.

[0043] In this application, the task description text corresponding to the model parameter generation task of the financial data analysis model can be identified, and several embedding vectors corresponding to several preset dimensions can be generated based on the task description text; a target task representation vector can be generated based on the several embedding vectors, and target parameter weights corresponding to the target task representation vector can be generated through a preset parameter function; the target parameter weights can be decomposed into a low-rank matrix corresponding to the initial financial data analysis model, and a parameter weight file corresponding to the initial financial data analysis model can be generated based on the low-rank matrix; the model parameters of the initial financial data analysis model can be adjusted based on the parameter weight file to convert the initial financial data analysis model into a target financial data analysis model of the model type corresponding to the model parameter generation task; finally, data analysis can be performed on the financial data to be analyzed using the target financial data analysis model.

[0044] Therefore, the method of this application can identify the task description text of the model parameter generation task corresponding to the financial data analysis model, convert the task description text into several embedding vectors corresponding to preset dimensions, generate a target task representation vector based on the embedding vectors, and then convert the target task representation vector into target parameter weights through preset weights. Furthermore, the generated target task weights need to be decomposed into a low-rank matrix corresponding to the structure of the large language model to be converted, and a parameter weight file for the model is generated based on the decomposed low-rank matrix. Finally, the parameter weight file is applied to the large language model to be converted to adjust its parameters, transforming the large language model to be converted into a target financial data analysis model suitable for the financial field, and then using the target financial data analysis model to perform data analysis on the financial data to be analyzed. In this way, through the organic combination of deep modeling of task features and parameter generation technology, a direct mapping from task description to network parameters can be achieved, significantly improving the efficiency and practicality of model adaptation. This solves the shortcomings of existing model adaptation technologies in terms of training efficiency, resource consumption, and deployment complexity. Only through parameter fine-tuning, combined with task requirements, models from other fields can be applied to the financial field, effectively saving model training costs. Attached Figure Description

[0045] 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 embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0046] Figure 1 This is a flowchart of a financial data analysis method based on model parameter adjustment disclosed in this application;

[0047] Figure 2 This is a schematic diagram of a financial data analysis device based on model parameter adjustment disclosed in this application;

[0048] Figure 3 This is a structural diagram of an electronic device disclosed in this application. Detailed Implementation

[0049] 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.

[0050] Traditional LoRA fine-tuning methods require complete gradient descent training on each new task, even for parameter-efficient methods, which can take hours to days, resulting in high time costs. Furthermore, current methods require significant GPU computing resources during training, including forward propagation, backpropagation, and gradient updates, increasing deployment costs and energy consumption. Moreover, current methods are typically optimized for specific task types, lacking cross-task adaptability and struggling to handle complex applications involving multiple task modes.

[0051] To overcome the aforementioned technical problems, this application discloses a financial data analysis method, apparatus, device, and medium based on model parameter adjustment. It can directly generate neural network parameters based on task representation (Task-to-Weights, T2W). Through the organic combination of deep modeling of task features and parameter generation technology, it realizes a direct mapping from task description to network parameters. Therefore, by fine-tuning the parameters and combining them with task requirements, models from other fields can be applied to the financial field, effectively saving the training cost of the model.

[0052] See Figure 1 As shown in the figure, this invention discloses a financial data analysis method based on model parameter adjustment, including:

[0053] Step S11: Identify the model parameters of the financial data analysis model to generate the task description text corresponding to the task, and generate several embedding vectors corresponding to several preset dimensions based on the task description text.

[0054] In this embodiment, if a model parameter generation task for a financial data analysis model is received, it is necessary to identify the task description text of the model parameter generation task and generate a corresponding embedding vector based on the identified task description text. Specifically, if a model parameter generation task for a financial data analysis model is received, the task description text corresponding to the task parameter generation task is identified, and several target task descriptions corresponding to several preset dimensions in the task description text are determined. It should be noted that the task description dimensions are divided into task prototype dimensions. Domain context dimension Performance-oriented dimension and constraint dimensions The task prototype dimension defines the basic pattern of the task, such as classification, generation, and extraction, and is the skeleton structure of the task. The domain context dimension represents the professional domain characteristics to which the task belongs, such as sentiment analysis, medical diagnosis, and financial risk control, and determines the semantic environment of the task. The performance-oriented dimension clarifies the optimization goals of the task, such as accuracy priority, speed priority, and robustness priority, and guides the direction of parameter generation. The constraint dimension specifies the specific restrictions and requirements of the task, such as output format, length limit, and style preference, and ensures the applicability of the generated parameters. It should be noted that the task types in this application are not limited to financial data analysis, but are also applicable to other fields, such as sentiment analysis and medical diagnosis.

[0055] Taking the generation of "LoRA parameters for a financial data analysis task" as an example, the input task description is as follows:

[0056] ;

[0057] By recognizing the task text of the above financial data analysis task, it can be determined that the task prototype is "CLASSIFICATION", the domain context is "FINANCE", the performance orientation is "ACCURACY_FIRST", and the constraints are "output_classes":3, "style":"formal", "robustness":"high".

[0058] Then, several target task descriptions need to be mapped to several embedding vectors corresponding to several preset dimensions, where the embedding vector corresponding to the task prototype dimension is e. P The embedding vector corresponding to the domain context dimension is e. D The embedding vector corresponding to the performance-oriented dimension is e. O， The embedding vector corresponding to the constraint dimension is e C .

[0059] Step S12: Generate a target task representation vector based on the aforementioned embedding vectors, and generate the target parameter weights corresponding to the target task representation vector through a preset parameter function.

[0060] In this embodiment, several embedding vectors need to be processed by a preset parameter function to generate a target task representation vector based on these embedding vectors. However, before generating the target task vector based on the embedding vectors, a suitable preset parameter function needs to be constructed to establish a mapping relationship from the abstract task description to the specific network parameters. Specifically, it is necessary to collect historical model parameters to generate the historical task representation vector h corresponding to the task. i and historical parameter weights W i Then, the correspondence between each vector in the historical task representation vector and each weight in the historical parameter weights is determined. For example, historical model parameter generation task A corresponds to historical task representation vector A and historical parameter weights A. Therefore, it is necessary to construct the correspondence between historical task representation vector A and historical parameter weights A. By training the preset parameter function through this correspondence, the parameter function can learn the relationship between the representation vectors and parameter weights of different types of tasks, making the final generated parameter weights more accurate. Then, based on the obtained correspondence, the corresponding historical task representation vectors and parameter weights to be paired need to be determined from the historical task representation vectors and historical parameter weights, respectively. And based on the historical task representation vectors and parameter weights to be paired, several historical pairing data are constructed. Ultimately, the parameter function to be trained can be obtained by training historical paired data. ,in These are the learnable parameters obtained through training. It should be noted that the parameter function to be trained can be obtained by using the correspondence between the historical task representation vectors of historical figures and the weights of each historical parameter. Furthermore, the parameter function design employs a multi-layer neural network architecture, incorporating techniques such as attention mechanisms and residual connections. This allows the parameter function to effectively capture the complex nonlinear relationship between task features and parameter configurations, improving the accuracy and generalization ability of the mapping. In this way, the trained preset parameter function possesses a stronger ability to convert semantic information into corresponding weight parameters.

[0061] Furthermore, the aforementioned embedding vectors need to be fused using a preset encoder to obtain the target task representation vector h. task And it is represented as: , where e P e is the embedding vector corresponding to the task prototype dimension. D e is the embedding vector corresponding to the domain context dimension. Oe is the embedding vector corresponding to the performance-oriented dimension. C The embedding vector corresponds to the constraint dimension, and the encoder is the preset encoder.

[0062] After the task representation vector is constructed, the parameter weights corresponding to the target task vector can be predicted according to a preset parameter function to obtain the target parameter weights corresponding to the target task vector, and the prediction of the weights can be expressed as: ,and These are the predicted LoRA weights. In this way, because the parameter function design employs a multi-layer neural network architecture and incorporates techniques such as attention mechanisms and residual connections, the accuracy and generalization ability of the mapping can be improved, thus making the generated parameter weights corresponding to the target task representation vector more accurate.

[0063] Step S13: Based on the target parameter weights, decompose the target task representation vector into a low-rank matrix corresponding to the large language model to be converted, and generate a parameter weight file corresponding to the large language model to be converted based on the low-rank matrix; the low-rank matrix is ​​a matrix whose rank satisfies the preset low-rank determination condition.

[0064] In this embodiment, the target task representation vector needs to be decomposed into low-rank matrices corresponding to the structure of the large language model to be converted. Then, a parameter weight file corresponding to the large language model to be converted is generated based on the obtained low-rank matrices. Specifically, the model structure of the large language model to be converted needs to be determined, and the target task representation vector is converted using several preset generation heads based on preset constraint optimization strategies and the target parameter weights to obtain several low-rank matrices corresponding to the model structure. It should be noted that a low-rank matrix is ​​a matrix whose rank satisfies a preset low-rank criterion, where the rank of the matrix is ​​much smaller than the number of rows and columns. The target parameter weights... Given a weight matrix and a pre-defined constrained optimization strategy, it can be expressed as:

[0065] ;

[0066] in, The target parameter weights are intended to be decomposed into the product of A and B, where A and B are the low-rank matrices to be optimized, and λ is a preset regularization strength coefficient. Let represent the Frobenius norm, and let represent the square root of the sum of squares of the matrix elements. It should be noted that in the above-mentioned pre-defined constraint optimization strategy, The purpose is to approximate the error as closely as possible to the product of the low-rank matrices AB after decomposition. , Its function is to limit the parameter size by penalizing the Frobenius norm of matrices A and B, thereby preventing overfitting and improving numerical stability.

[0067] It should be noted that the preset generator is a dedicated generator responsible for generating parameters of a specific type. The target task representation vector is input into different generators, which then transform the vector based on a preset constraint optimization strategy and the target parameter weights to obtain several low-rank matrices corresponding to the model structure. For example:

[0068] ;

[0069] Furthermore, it is necessary to partition the generated low-rank matrices based on each layer of the model structure to obtain several target low-rank matrices corresponding to each layer, and then generate several parameter weight files based on these target low-rank matrices. For example, if the large language model to be transformed has a 32-layer Transformer structure, each layer corresponds to a low-rank matrix. Therefore, it is necessary to partition the target low-rank matrix corresponding to each layer of the 32-layer Transformer structure, and then convert the corresponding target low-rank matrix into a weight file.

[0070] Step S14: Adjust the model parameters of the large language model to be converted based on the parameter weight file, so as to convert the large language model to be converted into the target financial data analysis model corresponding to the model parameter generation task.

[0071] In this embodiment, the model parameters of the large language model to be converted need to be adjusted using the generated weight file to achieve type conversion of the large language model, transforming it into a target financial data analysis model corresponding to the model parameter generation task. Specifically, the parameter weight file needs to be applied to the large language model to be converted to adjust its model parameters, resulting in a financial data analysis model to be determined. Then, the performance of the financial data analysis model to be determined needs to be verified using a validation dataset in the financial field. If the model performance of the financial data analysis model to be determined passes the verification, then the financial data analysis model to be determined is used as the target financial data analysis model for the type conversion.

[0072] Step S15: Perform data analysis on the financial data to be analyzed using the target financial data analysis model.

[0073] Therefore, in this embodiment, the task description text of the model parameter generation task corresponding to the financial data analysis model can be identified. Then, the task description text is converted into several embedding vectors corresponding to preset dimensions. Based on these embedding vectors, a target task representation vector is generated. Then, the target task representation vector is converted into target parameter weights through preset weights. Based on the target parameter weights, the target task representation vector is decomposed into a low-rank matrix corresponding to the large language model to be converted. Based on the decomposed low-rank matrix, a parameter weight file of the model is generated. Finally, the parameter weight file is applied to the large language model to be converted to adjust its parameters, converting the large language model to be converted into a target financial data analysis model suitable for the financial field. The target financial data analysis model is then used to perform data analysis on the financial data to be analyzed. In this way, by organically combining deep modeling of task features with parameter generation technology, a direct mapping from task description to network parameters can be achieved, significantly improving the efficiency and practicality of model adaptation. This solves the shortcomings of existing model adaptation technologies in terms of training efficiency, resource consumption, and deployment complexity. By simply fine-tuning the parameters and combining them with task requirements, models from other fields can be applied to the financial field. Due to the reduction in resource consumption, model conversion can be achieved with a single GPU, effectively saving model training costs.

[0074] See Figure 2 As shown in the figure, an embodiment of the present invention discloses a financial data analysis device based on model parameter adjustment, comprising:

[0075] The vector generation module 11 is used to identify the model parameters of the financial data analysis model to generate task description text corresponding to the task, and generate several embedded vectors corresponding to several preset dimensions based on the task description text.

[0076] The weight generation module 12 is used to generate a target task representation vector based on the plurality of embedded vectors, and to generate target parameter weights corresponding to the target task representation vector through a preset parameter function;

[0077] The weight conversion module 13 is used to decompose the target task representation vector into a low-rank matrix corresponding to the large language model to be converted based on the target parameter weights, and generate a parameter weight file corresponding to the large language model to be converted based on the low-rank matrix; the low-rank matrix is ​​a matrix whose rank satisfies a preset low-rank determination condition;

[0078] The model conversion module 14 is used to adjust the model parameters of the large language model to be converted based on the parameter weight file, so as to convert the large language model to be converted into a target financial data analysis model corresponding to the model parameter generation task.

[0079] The financial data analysis module 15 is used to perform data analysis on the financial data to be analyzed through the target financial data analysis model.

[0080] Therefore, this embodiment can identify the task description text of the model parameter generation task corresponding to the financial data analysis model, convert the task description text into several embedding vectors corresponding to preset dimensions, generate a target task representation vector based on these embedding vectors, convert the target task representation vector into target parameter weights using preset weights, decompose the target task representation vector into a low-rank matrix corresponding to the large language model to be converted based on the target parameter weights, and generate a parameter weight file for the model based on the decomposed low-rank matrix. Finally, the parameter weight file is applied to the large language model to be converted to adjust its parameters, transforming the large language model to be converted into a target financial data analysis model suitable for the financial field, and then performing data analysis on the financial data to be analyzed using the target financial data analysis model. In this way, through the organic combination of deep modeling of task features and parameter generation technology, a direct mapping from task description to network parameters can be achieved, significantly improving the efficiency and practicality of model adaptation. This solves the shortcomings of existing model adaptation technologies in terms of training efficiency, resource consumption, and deployment complexity. By only fine-tuning the parameters and combining them with task requirements, models from other fields can be applied to the financial field, effectively saving model training costs.

[0081] In some embodiments, the vector generation module 11 may specifically include:

[0082] The task description determination unit is used to identify the task description text corresponding to the task parameter generation task if a model parameter generation task of a financial data analysis model is received, and to determine several target task descriptions in the task description text that correspond to several preset dimensions.

[0083] The data mapping unit is used to map the several target task descriptions into several embedding vectors corresponding to the several preset dimensions.

[0084] In some embodiments, the financial data analysis device based on model parameter adjustment may further include:

[0085] The historical parameter collection unit is used to collect the historical task representation vector and historical parameter weights corresponding to the historical model parameter generation task.

[0086] The correspondence determination unit is used to determine the correspondence between each vector in the historical task representation vector and each weight in the historical parameter weight;

[0087] A pairing data construction unit is used to determine the corresponding historical task representation vector to be paired and the corresponding historical parameter weights from the historical task representation vector and the historical parameter weights respectively based on the correspondence, and to construct several historical pairing data according to the historical task representation vector to be paired and the corresponding historical parameter weights.

[0088] The function training unit is used to train the parameter function to be trained using the historical paired data to obtain the preset parameter function.

[0089] In some embodiments, the weight generation module 12 may specifically include:

[0090] A vector fusion unit is used to fuse the plurality of embedded vectors based on a preset encoder to obtain a target task representation vector;

[0091] The weight generation unit is used to predict the parameter weights corresponding to the target task vector through a preset parameter function, so as to obtain the target parameter weights corresponding to the target task vector.

[0092] In some embodiments, the weight transformation module 13 may specifically include:

[0093] The weight transformation unit is used to determine the model structure of the large language model to be transformed, and transform the target task representation vector based on a preset constraint optimization strategy and the target parameter weights through several preset generation heads to obtain several low-rank matrices corresponding to the model structure.

[0094] The weight file generation unit is used to divide the plurality of low-rank matrices based on each layer of the model structure to obtain a plurality of target low-rank matrices corresponding to each layer, and to generate a plurality of parameter weight files based on the plurality of target low-rank matrices.

[0095] The preset constraint optimization strategy is as follows:

[0096] ;

[0097] in, Let A and B be the target parameter weights, A and B be the low-rank matrices to be optimized, and λ be the preset regularization strength coefficient. denoted by Frobenius norm, representing the square root of the sum of squares of the matrix elements.

[0098] In some embodiments, the model conversion module 14 may specifically include:

[0099] The parameter adjustment unit is used to apply the parameter weight file to the large language model to be converted, so as to adjust the model parameters of the large language model to be converted through the parameter weight file to obtain the financial data analysis model to be determined.

[0100] The performance verification unit is used to determine the target verification dataset corresponding to the model parameter generation task, and to verify the model performance of the financial data analysis model to be determined through the target verification dataset.

[0101] The model determination unit is used to select the financial data analysis model to be determined as the target financial data analysis model for type conversion if the model performance of the financial data analysis model to be determined passes the verification.

[0102] Furthermore, embodiments of this application also disclose an electronic device, Figure 3 This is a structural diagram of an electronic device 20 according to an exemplary embodiment. The content of the diagram should not be construed as limiting the scope of this application.

[0103] Figure 3 This is a schematic diagram of the structure of an electronic device 20 provided in an embodiment of this application. Specifically, the electronic device 20 may include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 stores a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the financial data analysis method based on model parameter adjustment disclosed in any of the foregoing embodiments. Alternatively, the electronic device 20 in this embodiment may specifically be an electronic computer.

[0104] In this embodiment, the power supply 23 is used to provide operating voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and external devices, and the communication protocol it follows can be any communication protocol applicable to the technical solution of this application, and is not specifically limited here; the input / output interface 25 is used to acquire external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs, and is not specifically limited here.

[0105] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk or optical disk, etc. The resources stored thereon can include operating system 221, computer program 222, etc., and the storage method can be temporary storage or permanent storage.

[0106] The operating system 221 is used to manage and control the various hardware devices on the electronic device 20 and the computer program 222, which may be Windows Server, Netware, Unix, Linux, etc. In addition to including a computer program capable of performing the financial data analysis method based on model parameter adjustment executed by the electronic device 20 as disclosed in any of the foregoing embodiments, the computer program 222 may further include computer programs capable of performing other specific tasks.

[0107] Furthermore, this application also discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the aforementioned disclosed financial data analysis method based on model parameter adjustment. Specific steps of this method can be found in the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.

[0108] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0109] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0110] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0111] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0112] The technical solutions provided in this application have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A financial data analysis method based on model parameter adjustment, characterized in that, include: The model parameters of the financial data analysis model are identified to generate task description text corresponding to the task, and several embedding vectors corresponding to several preset dimensions are generated based on the task description text. A target task representation vector is generated based on the aforementioned embedding vectors, and target parameter weights corresponding to the target task representation vector are generated through a preset parameter function. Based on the target parameter weights, the target task representation vector is decomposed into a low-rank matrix corresponding to the large language model to be converted, and a parameter weight file corresponding to the large language model to be converted is generated based on the low-rank matrix; the low-rank matrix is ​​a matrix whose rank satisfies the preset low-rank determination condition; The model parameters of the large language model to be converted are adjusted based on the parameter weight file to convert the large language model to be converted into a target financial data analysis model corresponding to the model parameter generation task. The target financial data analysis model is used to perform data analysis on the financial data to be analyzed. The process includes generating task description text corresponding to the model parameter generation task of identifying the financial data analysis model, and generating several embedding vectors corresponding to several preset dimensions based on the task description text, including: If a model parameter generation task for a financial data analysis model is received, the task description text corresponding to the model parameter generation task is identified, and several target task descriptions corresponding to several preset dimensions are determined in the task description text. The preset dimensions include a task prototype dimension, a domain context dimension, a performance-oriented dimension, and a constraint dimension. The task prototype dimension defines the basic pattern of the task and is the skeleton structure of the task. The domain context dimension represents the professional domain characteristics to which the task belongs and determines the semantic environment of the task. The performance-oriented dimension clarifies the optimization goal of the task and guides the direction of parameter generation. The constraint dimension specifies the specific restrictions and requirements of the task and ensures the applicability of the generated parameters. The target task descriptions are mapped to a number of embedding vectors corresponding to the preset dimensions; The step of decomposing the target task representation vector into a low-rank matrix corresponding to the large language model to be converted based on the target parameter weights, and generating a parameter weight file corresponding to the large language model to be converted based on the low-rank matrix, includes: The model structure of the large language model to be transformed is determined, and the target task representation vector is transformed by several preset generation heads based on preset constraint optimization strategies and the target parameter weights to obtain several low-rank matrices corresponding to the model structure. Based on each layer of the model structure, the several low-rank matrices are divided to obtain several target low-rank matrices corresponding to each layer, and several parameter weight files are generated based on the several target low-rank matrices. The preset constraint optimization strategy is as follows: ； in, Let A and B be the target parameter weights, A and B be the low-rank matrices to be optimized, and λ be the preset regularization strength coefficient. denoted by Frobenius norm, representing the square root of the sum of squares of the matrix elements.

2. The financial data analysis method based on model parameter adjustment according to claim 1, characterized in that, Before generating the target task representation vector based on the plurality of embedded vectors and generating the target parameter weights corresponding to the target task representation vector through a preset parameter function, the method further includes: Collect historical model parameters to generate historical task representation vectors and historical parameter weights corresponding to the tasks; Determine the correspondence between each vector in the historical task representation vector and each weight in the historical parameter weights; Based on the correspondence, the corresponding historical task representation vector and the corresponding historical parameter weight are determined from the historical task representation vector and the historical parameter weight, and several historical pairing data are constructed according to the historical task representation vector and the corresponding historical parameter weight. The parameter function to be trained is trained using the historical paired data to obtain the preset parameter function.

3. The financial data analysis method based on model parameter adjustment according to claim 1, characterized in that, The step of generating a target task representation vector based on the plurality of embedded vectors, and generating target parameter weights corresponding to the target task representation vector through a preset parameter function, includes: The embedding vectors are fused based on a preset encoder to obtain the target task representation vector; The target parameter weights corresponding to the target task representation vector are predicted by a preset parameter function to obtain the target parameter weights corresponding to the target task representation vector.

4. The financial data analysis method based on model parameter adjustment according to any one of claims 1 to 3, characterized in that, The step of adjusting the model parameters of the large language model to be converted based on the parameter weight file, so as to convert the large language model to be converted into a target financial data analysis model corresponding to the model parameter generation task, includes: The parameter weight file is applied to the large language model to be converted, so as to adjust the model parameters of the large language model to be converted through the parameter weight file, so as to obtain the financial data analysis model to be determined. The target validation dataset corresponding to the model parameter generation task is determined, and the model performance of the financial data analysis model to be determined is validated using the target validation dataset. If the performance of the financial data analysis model to be determined is verified, then the financial data analysis model to be determined will be used as the target financial data analysis model for completing the type conversion.

5. A financial data analysis device based on model parameter adjustment, characterized in that, include: The vector generation module is used to identify the model parameters of the financial data analysis model to generate task description text corresponding to the task, and generate several embedded vectors corresponding to several preset dimensions based on the task description text. The weight generation module is used to generate a target task representation vector based on the plurality of embedded vectors, and to generate target parameter weights corresponding to the target task representation vector through a preset parameter function. The weight transformation module is used to decompose the target task representation vector into a low-rank matrix corresponding to the large language model to be transformed based on the target parameter weights, and to generate a parameter weight file corresponding to the large language model to be transformed based on the low-rank matrix; the low-rank matrix is ​​a matrix whose rank satisfies a preset low-rank determination condition; The model conversion module is used to adjust the model parameters of the large language model to be converted based on the parameter weight file, so as to convert the large language model to be converted into a target financial data analysis model corresponding to the model parameter generation task. The financial data analysis module is used to perform data analysis on the financial data to be analyzed using the target financial data analysis model. The vector generation module includes: The task description determination unit is used to, upon receiving a model parameter generation task for a financial data analysis model, identify the task description text corresponding to the model parameter generation task and determine several target task descriptions in the task description text that correspond to several preset dimensions. These preset dimensions include a task prototype dimension, a domain context dimension, a performance-oriented dimension, and a constraint dimension. The task prototype dimension defines the basic pattern of the task and is its skeletal structure. The domain context dimension represents the professional domain characteristics to which the task belongs and determines the semantic environment of the task. The performance-oriented dimension clarifies the optimization objective of the task and guides the direction of parameter generation. The constraint dimension specifies the specific limitations and requirements of the task, ensuring the applicability of the generated parameters. A data mapping unit is used to map the plurality of target task descriptions into a plurality of embedding vectors corresponding to the plurality of preset dimensions; The weight transformation module includes: The weight transformation unit is used to determine the model structure of the large language model to be transformed, and transform the target task representation vector based on a preset constraint optimization strategy and the target parameter weights through several preset generation heads to obtain several low-rank matrices corresponding to the model structure. The weight file generation unit is used to divide the plurality of low-rank matrices based on each layer of the model structure to obtain a plurality of target low-rank matrices corresponding to each layer, and to generate a plurality of parameter weight files based on the plurality of target low-rank matrices. The preset constraint optimization strategy is as follows: ； in, Let A and B be the target parameter weights, A and B be the low-rank matrices to be optimized, and λ be the preset regularization strength coefficient. denoted by Frobenius norm, representing the square root of the sum of squares of the matrix elements.

6. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the financial data analysis method based on model parameter adjustment as described in any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, Used to store a computer program, wherein the computer program, when executed by a processor, implements the financial data analysis method based on model parameter adjustment as described in any one of claims 1 to 4.

Images