Digital asset assessment method and system based on large language model dynamic optimization

Abstract

The invention relates to the field of data processing, and discloses a digital asset assessment method and system based on large language model dynamic optimization, and the method comprises the following steps: carrying out the dimension label classification of data needing to be analyzed, and retrieving a corresponding LLM prompt template from a prompt library according to the class; secondly, carrying out confidence calculation and deviation correction on the information carried in the analysis process by using a large language model to obtain deviation-corrected analysis information; furthermore, through the collected deviation correction information, the large model is further fed and trained, the analysis performance of the large model is optimized, and the self-iteration optimization of the model is realized. And finally, providing a data processing and analysis service to the outside through the dynamically optimized large model. According to the method, rich background knowledge and reasoning ability of the large language model are utilized, and the purposes of data processing and classification are achieved through continuous dynamic optimization of the large language model.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of data processing, in particular to a digital asset evaluation method and system based on large language model dynamic optimization. BACKGROUND

[0002] Digital assets refer to any non-monetary data resources that exist in binary format, have ownership or usage rights, and can generate value. Among them, digital assets exist in computers, servers or the cloud. Common digital assets include but are not limited to copyrights, software copyrights, patent rights, trademark rights, etc. Traditional analysis methods will face great challenges when analyzing data, i.e., analyzing digital assets. The significance of the data analysis method based on large language model dynamic optimization lies in that it provides a new and intelligent paradigm to solve the above problems. Its advantages include efficient processing of massive information, such as LLM can instantly read and analyze thousands of data corresponding to digital assets, extract key factors affecting value, and greatly improve analysis efficiency; it also has strong reasoning and correlation ability, and can form a positive cycle of getting smarter through the closed loop of "analysis-correction-training-reanalysis". At the same time, it improves analysis consistency and interpretability, and reduces dependence on experts.

[0003] The core significance of using large language model dynamic optimization to analyze data lies in creating an intelligent analysis system that can continuously learn and evolve, solving the unique, dynamic and complex problems faced by traditional methods when analyzing data. Its greatest advantage is to realize the leap from a static, artificial-dependent analysis mode to a dynamic, automated, human-machine collaborative intelligent mode. However, when a general large language model is applied to data assets, due to the wide range of training data and the inherent uncertainty of the generation logic, it is easy to produce illusions, output content deviating from professional common sense, and evaluation results fluctuating and unreliable technical defects. Therefore, it is urgent to solve the problem of improving the output accuracy, consistency and reliability of large language models in processing digital asset evaluation tasks. SUMMARY

[0004] The present application overcomes the shortcomings of the prior art and provides a digital asset evaluation method and system based on large language model dynamic optimization.

[0005] To achieve the above purpose, the technical scheme adopted by the present application is as follows:

[0006] The present application provides a digital asset evaluation method based on large language model dynamic optimization in the first aspect, comprising the following steps:

[0007] Constructing a large language model to be trained, classifying the dimensions of the data to be evaluated and extracting key features, and constructing an LLM prompt template;

[0008] Perform structural transformation on the key characteristic information of the data to be evaluated, and combine the LLM prompt template and data analysis module to conduct preliminary analysis on the data to be evaluated and generate an initial analysis and evaluation report;

[0009] Perform deviation analysis on the initial analysis and evaluation report to generate deviation correction information for the initial analysis and evaluation report. This deviation correction information is then fed into the large language model to be trained for self-iterative optimization, generating a target analysis and evaluation report and optimizing the large language model.

[0010] By optimizing the large language model, all data to be evaluated is analyzed and evaluated.

[0011] Furthermore, in a preferred embodiment of the present invention, the construction of a large language model to be trained, the dimension label classification and key feature extraction of the data to be evaluated, and the construction of an LLM prompt template are specifically as follows:

[0012] Acquire construction software for constructing a large language model, mark it as target construction software, and construct a large language model to be trained based on the target construction software;

[0013] The large language model to be trained includes different model modules, including a data classification module, a prompt template management module, a data analysis module, a fine-tuning training management module, and a correction feedback management module;

[0014] Determine the data that needs to be analyzed, mark it as data to be evaluated, and obtain all dimension labels of the data to be evaluated. The data that needs to be analyzed is digital asset data;

[0015] Introducing a big data network, determining all feature information of different data to be evaluated in the big data network, and analyzing the number of times all feature information of the data to be evaluated is cited in the big data network, marking feature information with a number of big data network citations greater than a preset number as key feature information of the data to be evaluated, and saving the key feature information of the data to be evaluated in a data classification module;

[0016] Retrieving a prompt library in a big data network, and retrieving and designing an LLM prompt template in the prompt library, wherein different LLM prompt blank templates are preset in the prompt library;

[0017] Among them, the LLM prompt template is a structured input text template of a large language model. The method for designing the LLM prompt template is to combine the dimensional labels of all the data to be evaluated, define the key elements of the LLM prompt template, including input format, output requirements and task instructions, and based on the key elements of the LLM prompt template, on the LLM prompt blank template with the highest degree of correlation with the key elements of the LLM prompt template, construct a structured input text template that can generate dimensional features through the key feature information of the data to be evaluated, that is, the LLM prompt template, and save the LLM prompt template in the prompt template management module.

[0018] Furthermore, in a preferred embodiment of the present invention, the key feature information of the data to be evaluated is structured and converted, and combined with the LLM prompt template and data analysis module, a preliminary analysis of the data to be evaluated is performed to generate an initial analysis and evaluation report, specifically:

[0019] In the large language model to be trained, the key feature information of the data to be evaluated in the data module is structurally converted according to the input format of the key elements in the LLM prompt template, so that the format of the key feature information of the data to be evaluated is equal to the input format of the key elements in the LLM prompt template;

[0020] The key feature information of the data to be evaluated after structured conversion is marked as target feature information, and the data classification module and the prompt template management module are connected so that the target feature information can be imported into the LLM prompt template;

[0021] Within the LLM prompt template, field analysis is performed on the target feature information. The analyzed fields include text fields and numeric fields. Field integrity is checked using the LLM prompt template before field analysis. If the field integrity is not within the preset range, a secondary structural transformation is performed on the target feature information until the field integrity remains within the preset range.

[0022] After performing field analysis on the target feature information, based on the output requirements in the LLM prompt template, the key fields of the target feature information are screened and parsed. Based on the parsing results, the dimension labels corresponding to the data to be evaluated are output in the LLM prompt template.

[0023] Combined with the dimension labels corresponding to the data to be evaluated, an initial analysis and evaluation report of the data to be evaluated is generated in the data analysis module of the large language model to be trained.

[0024] Furthermore, in a preferred embodiment of the present invention, correction information of the initial analysis and evaluation report is generated, and the correction information of the initial analysis and evaluation report is fed into the large language model to be trained for model self-iteration optimization, generating a target analysis and evaluation report and optimizing the large language model, specifically:

[0025] In the large language model to be trained, the initial analysis evaluation report is imported into the fine-tuning training management module for storage;

[0026] In the fine-tuning training management module of the large language model to be trained, a rectification trigger mechanism is set, wherein the rectification trigger mechanism can calculate the confidence of the initial analysis evaluation report of the to-be-evaluated data, and determine whether the initial analysis evaluation report of the to-be-evaluated data is biased according to the confidence;

[0027] The rectification trigger mechanism is provided with historical analysis evaluation reports of all data of the same type as the to-be-evaluated data and corresponding feature information, and the data of the same type as the to-be-evaluated data is marked as to-be-analyzed historical data;

[0028] The target feature information of the input to-be-evaluated data is analyzed, and the similarity between the feature information of the to-be-analyzed historical data different from the rectification trigger mechanism is calculated;

[0029] The to-be-analyzed historical data with the highest similarity is selected as a type of historical data, the historical analysis evaluation report of the type of historical data is determined, and the report comparison between the initial analysis evaluation report of the to-be-evaluated data and the historical analysis evaluation report of the type of historical data is performed to output the confidence of the to-be-evaluated data;

[0030] If the coincidence degree between the initial analysis evaluation report of the to-be-evaluated data and the historical analysis evaluation report of the type of historical data is greater than a preset value, the confidence of the to-be-evaluated data is greater than a standard value, and it is determined that the to-be-evaluated data is not biased, otherwise it is determined that the initial analysis evaluation report of the to-be-evaluated data is biased;

[0031] When the initial analysis evaluation report of the to-be-evaluated data is biased, the rectification information of the initial analysis evaluation report of the to-be-evaluated data is calculated, and the rectification information of the initial analysis evaluation report is fed to the large language model to be trained for model self-iterative optimization to generate a target analysis evaluation report and an optimized large language model.

[0032] Further, in a preferred embodiment of the present application, when the initial analysis evaluation report of the to-be-evaluated data is biased, the rectification information of the initial analysis evaluation report of the to-be-evaluated data is calculated, and the rectification information of the initial analysis evaluation report is fed to the large language model to be trained for model self-iterative optimization to generate a target analysis evaluation report and an optimized large language model, specifically:

[0033] The fine-tuning training deviation interval is set, when the initial analysis evaluation report of the to-be-evaluated data has a deviation, and the initial analysis evaluation report of the to-be-evaluated data after the deviation is not in the fine-tuning training deviation interval, the non-coincidence position between the initial analysis evaluation report of the to-be-evaluated data and the historical analysis evaluation report of the first type of historical data is determined, and is marked as preliminary correction information of the analysis evaluation report;

[0034] An interactive page of the to-be-trained large language model is constructed, and based on the interactive page, the preliminary correction information of the analysis evaluation report is provided to the user for secondary correction of the correction information to obtain target correction information of the analysis evaluation report;

[0035] The target correction information of the analysis evaluation report is imported into the fine-tuning training management module of the to-be-trained large language model, and the to-be-trained large language model is fine-tuned to optimize the analysis performance of the to-be-trained large language model, and an optimized large language model is obtained, wherein the fine-tuning training is model self-iteration optimization of the to-be-trained large language model;

[0036] Wherein, the information number of the target correction information of the analysis evaluation report is analyzed in real time, if the information number of the target correction information of the analysis evaluation report is greater than a predetermined value, the model self-iteration optimization is started, and 60% of the target correction information of the analysis evaluation report is used for fine-tuning training, and 40% of the target correction information of the analysis evaluation report is used for verification test;

[0037] In combination with the optimized large language model and the imported target correction information of the analysis evaluation report, the initial analysis evaluation report of the to-be-evaluated data is corrected and optimized to obtain a corrected analysis evaluation report of the to-be-evaluated data, which is marked as a target analysis evaluation report.

[0038] Further, in a preferred embodiment of the present application, the initial analysis evaluation report of the to-be-evaluated data is corrected and optimized to obtain a corrected analysis evaluation report of the to-be-evaluated data, which is marked as a target analysis evaluation report, specifically:

[0039] The imported target correction information of the analysis evaluation report is processed for format unification and data set construction to obtain a to-be-evaluated data set;

[0040] The to-be-evaluated data set is configured in the fine-tuning training management module of the to-be-trained large language model to control the to-be-trained large language model to perform model parameter fine-tuning, wherein the model parameter fine-tuning is configuration update of the model parameters of the to-be-trained large language model, and the model KL divergence is calculated during the update process, and the model parameter fine-tuning is stopped when the model KL divergence is maintained within a preset range to obtain a large language model;

[0041] In a large language model, input test correction information and test analysis and evaluation report, and perform correction test on the test analysis and evaluation report through a large language model, and judge the sensitivity and accuracy of the correction test;

[0042] If the sensitivity and accuracy of the correction test are greater than the preset value, the model parameters of the large language model are frozen, and the large language model is labeled as an optimized large language model;

[0043] If the sensitivity and accuracy of the correction test are not greater than the preset value, the model parameters of the large language model are continuously fine-tuned until the sensitivity and accuracy of the correction test are greater than the preset value.

[0044] Further, in a preferred embodiment of the present application, the analysis and evaluation processing of all to-be-evaluated data by the optimized large language model is specifically:

[0045] In the optimized large language model, all to-be-evaluated data are imported for data analysis to generate analysis and evaluation reports corresponding to different to-be-evaluated data;

[0046] Based on the optimized large language model, the analysis and evaluation reports corresponding to different to-be-evaluated data are pushed to the owners of the to-be-evaluated data through an interactive page.

[0047] In a second aspect of the present application, a digital asset evaluation system based on dynamic optimization of a large language model is provided, characterized in that the digital asset evaluation system integrates a high-performance computing architecture and a data storage module, including a non-volatile memory composed of a DDR4 RDIMM memory module with ECC verification and a NVMe solid-state storage array using 3D NAND flash, and a multi-core processor based on Zen4 microarchitecture; the memory is solidified and deployed with a digital asset evaluation method program having a digital asset evaluation engine, and when the program is executed in parallel by the superscalar pipeline execution unit in the processor, the following steps are implemented:

[0048] A to-be-trained large language model is constructed to perform dimension label classification and key feature extraction on to-be-evaluated data, and an LLM prompt template is constructed;

[0049] The key feature information of the to-be-evaluated data is structured and converted, and the to-be-evaluated data is preliminarily analyzed in combination with the LLM prompt template and the data analysis module to generate an initial analysis and evaluation report;

[0050] The initial analysis and evaluation report is analyzed for deviation, the correction information of the initial analysis and evaluation report is generated, and the correction information of the initial analysis and evaluation report is fed to the to-be-trained large language model for model self-iteration optimization to generate a target analysis and evaluation report and an optimized large language model;

[0051] By optimizing the large language model, all to be evaluated data are analyzed and evaluated.

[0052] The technical defects existing in the background of the present application are solved, and the present application has the following beneficial effects: the data to be analyzed are classified by dimension label, and the corresponding LLM prompt template is retrieved from the prompt library according to the category. Secondly, the information carried in the analysis process is calculated by the large language model Confidence and rectification, and the rectified analysis information is obtained. Thirdly, by collecting the rectified information, the large model is further fed and trained, and the analysis performance of the large model is optimized, realizing the self-iteration optimization of the model. Finally, through the dynamic optimization of the large model, data processing and analysis services are provided. The present application utilizes the rich background knowledge and reasoning ability of the large language model, and realizes the processing and classification purposes of data through the continuous dynamic optimization of the large language model. BRIEF DESCRIPTION OF DRAWINGS

[0053] In order to more clearly illustrate the technical solutions of the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can obtain other drawings of embodiments according to these drawings without creative labor.

[0054] Figure 1 The flowchart of the digital asset evaluation method based on dynamic optimization of large language model is shown;

[0055] Figure 2 The flowchart of the method for generating target analysis and evaluation report and optimizing large language model is shown;

[0056] Figure 3 The system architecture of the digital asset evaluation system based on dynamic optimization of large language model is shown. DETAILED DESCRIPTION

[0057] In order to more clearly illustrate the technical solutions of the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can obtain other drawings of embodiments according to these drawings without creative labor.

[0058] In the following description, many specific details are set forth in order to provide a thorough understanding of the present application, however, the present application can also be implemented in other ways different from those described herein, therefore, the scope of protection of the present application is not limited by the specific embodiments disclosed below.

[0059] Figure 1 The flowchart of the digital asset evaluation method based on dynamic optimization of large language model is shown, including the following steps:

[0060] S102: Constructing a large language model to be trained, dimension label classification and key feature extraction are performed on the to-be-evaluated data, and an LLM prompt template is constructed;

[0061] S104: The key feature information of the to-be-evaluated data is structured and converted, and the to-be-evaluated data is preliminarily analyzed in combination with the LLM prompt template and the data analysis module, to generate an initial analysis evaluation report;

[0062] S106: The initial analysis evaluation report is analyzed for deviation, deviation information of the initial analysis evaluation report is generated, and the deviation information of the initial analysis evaluation report is fed into the large language model to be trained for model self-iteration optimization, to generate a target analysis evaluation report and an optimized large language model;

[0063] S108: All to-be-evaluated data is analyzed and evaluated by the optimized large language model.

[0064] Further, in a preferred embodiment of the present application, the construction of the large language model to be trained, the dimension label classification and the key feature extraction of the to-be-evaluated data, and the construction of the LLM prompt template are specifically as follows:

[0065] Obtain a construction software for constructing a large language model, which is labeled as a target construction software, and based on the target construction software, a large language model to be trained is constructed;

[0066] The large language model to be trained includes different model modules, including a data classification module, a prompt template management module, a data analysis module, a fine-tuning training management module, and a deviation correction feedback management module;

[0067] Determine the data that needs to be analyzed, which is labeled as to-be-evaluated data, and obtain all dimension labels of the to-be-evaluated data, wherein the data that needs to be analyzed is digital asset data;

[0068] Introduce a big data network, determine all feature information of different to-be-evaluated data in the big data network, and analyze the number of big data network references of all feature information of the to-be-evaluated data, label the feature information with a number of big data network references greater than a preset number as key feature information of the to-be-evaluated data, and save the key feature information of the to-be-evaluated data to the data classification module;

[0069] Search the prompt library in the big data network, and search and design an LLM prompt template in the prompt library, wherein different LLM prompt blank templates are preset in the prompt library;

[0070] The LLM prompt template is a structured input text template of a large language model. A method for designing the LLM prompt template is as follows: in combination with dimension labels of all to-be-evaluated data, key elements of the LLM prompt template are defined, including an input format, an output requirement and a task instruction, and according to the key elements of the LLM prompt template, a structured input text template capable of generating dimension features via key feature information of the to-be-evaluated data is constructed on an LLM prompt blank template having the highest correlation degree with the key elements of the LLM prompt template, that is, the LLM prompt template, and the LLM prompt template is saved in a prompt template management module.

[0071] It should be noted that the large language model is a fine-tuned model, such as Deepseek, LoRA, etc., which is responsible for generating initial analysis and analysis process information of a to-be-analyzed object according to a prompt template. The analysis information is an evaluation of data. In this application, the data to be classified is a digital asset, so the data analysis is an evaluation of the digital asset. It includes a data classification module, a prompt template management module, a data analysis module, a fine-tuning training management module and a rectification feedback management module. Different modules have different functions. In this application, the data to be classified is a digital asset, which includes but is not limited to patents, trademarks, artistic works, literary works, software copyrights, papers, etc. The dimension labels include but are not limited to technical value analysis, legal status evaluation, market value prediction, comprehensive evaluation, technical scarcity, enterprise endorsement and market heat, etc. First, the feature information of the data needs to be extracted, the purpose of which is to determine the corresponding dimension label according to the feature information. Secondly, the LLM prompt template is designed. The LLM is a large language model. The LLM prompt template can determine the dimension label to which the data, i.e. the digital asset, belongs by importing the feature information.

[0072] Further, in a preferred embodiment of the present application, the key feature information of the to-be-evaluated data is structured and converted, and in combination with the LLM prompt template and the data analysis module, the to-be-evaluated data is preliminarily analyzed to generate an initial analysis evaluation report, specifically as follows:

[0073] In the to-be-trained large language model, the key feature information of the to-be-evaluated data in the data module is structured and converted according to the input format in the key elements of the LLM prompt template, so that the format of the key feature information of the to-be-evaluated data is equal to the input format in the key elements of the LLM prompt template;

[0074] The key feature information of the to-be-evaluated data after the structured conversion is marked as target feature information, and the data classification module and the prompt template management module are connected, so that the target feature information can be imported into the LLM prompt template;

[0075] In the LLM prompt template, the target feature information is subjected to field analysis, wherein the analyzed fields include text fields and numerical fields, and before field analysis, field integrity is checked through the LLM prompt template, if the field integrity is not within the preset range, the target feature information is subjected to secondary structured conversion until the field integrity is kept within the preset range;

[0076] After the field analysis of the target feature information, based on the output requirements in the LLM prompt template, the key fields of the target feature information are screened and parsed, and according to the parsing result, the dimension label corresponding to the to-be-evaluated data is output in the LLM prompt template;

[0077] In combination with the dimension label corresponding to the to-be-evaluated data, an initial analysis and evaluation report of the to-be-evaluated data is generated in the data analysis module of the to-be-trained large language model.

[0078] It should be noted that the LLM prompt template has a template format, and the input feature information needs to have a corresponding format to be recognized, so the format of the key feature information of the to-be-evaluated data is subjected to structured conversion. The LLM prompt template is analyzed according to the text fields and numerical fields in the key feature information, and thus the dimension label is output, so the field integrity needs to be verified, if the field is not complete, the structured conversion needs to be continued.

[0079] Further, in a preferred embodiment of the present application, the analysis and evaluation processing of all to-be-evaluated data is performed by optimizing the large language model, specifically:

[0080] In the optimized large language model, all to-be-evaluated data is imported for data analysis to generate analysis and evaluation reports corresponding to different to-be-evaluated data;

[0081] Based on the optimized large language model, the analysis and evaluation reports corresponding to different to-be-evaluated data are pushed to the owners of the to-be-evaluated data through an interactive page.

[0082] It should be noted that, since the data represents a digital asset, the generated analysis and evaluation report corresponding to the to-be-evaluated data is a digital asset analysis and evaluation report. The optimized large language model is used to push the analysis and evaluation report to others, that is, to push the digital asset to the owner. The evaluation report of the digital asset is a comprehensive, structured, and professional document with legal and commercial reference value, which aims to clearly define the value of a digital asset at a specific purpose and at a specific point in time, and elaborates on the source of the value. Among them, the evaluation report of the digital asset includes but is not limited to the abstract, which clearly defines the owner of the digital asset, the evaluation purpose, the evaluation benchmark date, the evaluation value, etc. Among them, the evaluation purpose clearly defines what the evaluated asset is, including patents, software copyrights, legal status, etc. At the same time, the value of the digital asset caused by technical factors, market factors, financial factors, strategic factors, etc. is analyzed in depth. The data evaluation report is not only the basis for transaction pricing, but also a powerful tool for enterprises to understand the value of their digital assets, make strategic decisions, and manage risks.

[0083] Figure 2 A method flowchart for generating a target analysis and evaluation report and optimizing a large language model is shown, including the following steps:

[0084] S202: Perform offset analysis on the initial analysis and evaluation report, generate correction information of the initial analysis and evaluation report, and feed the correction information of the initial analysis and evaluation report to the to-be-trained large language model for model self-iteration optimization to generate a target analysis and evaluation report and an optimized large language model;

[0085] S204: When the initial analysis and evaluation report of the to-be-evaluated data has an offset, calculate the correction information of the initial analysis and evaluation report of the to-be-evaluated data, and feed the correction information of the initial analysis and evaluation report to the to-be-trained large language model for model self-iteration optimization to generate a target analysis and evaluation report and an optimized large language model;

[0086] S206: Fine-tune the to-be-trained large language model to optimize the analysis performance of the to-be-trained large language model to obtain an optimized large language model, wherein the fine-tuning is model self-iteration optimization of the to-be-trained large language model.

[0087] Further, in a preferred embodiment of the present application, the offset analysis on the initial analysis and evaluation report, the generation of the correction information of the initial analysis and evaluation report, and the feeding of the correction information of the initial analysis and evaluation report to the to-be-trained large language model for model self-iteration optimization to generate a target analysis and evaluation report and an optimized large language model are specifically:

[0088] In the to-be-trained large language model, the initial analysis and evaluation report is imported into the fine-tuning management module for saving;

[0089] In the fine-tuning training management module of the large language model to be trained, a rectification trigger mechanism is set, wherein the rectification trigger mechanism can calculate the confidence of the initial analysis evaluation report of the to-be-evaluated data, and determine whether the initial analysis evaluation report of the to-be-evaluated data is biased according to the confidence.

[0090] In the rectification trigger mechanism, historical analysis evaluation reports of all data of the same type as the to-be-evaluated data and corresponding feature information are set, and the data of the same type as the to-be-evaluated data are marked as to-be-analyzed historical data.

[0091] The target feature information of the input to-be-evaluated data is analyzed, and the similarity between the feature information of the to-be-analyzed historical data different from the rectification trigger mechanism is calculated.

[0092] The to-be-analyzed historical data with the highest similarity is selected as a type of historical data, the historical analysis evaluation report of the type of historical data is determined, and the report comparison between the initial analysis evaluation report of the to-be-evaluated data and the historical analysis evaluation report of the type of historical data is performed to output the confidence of the to-be-evaluated data.

[0093] If the coincidence degree between the initial analysis evaluation report of the to-be-evaluated data and the historical analysis evaluation report of the type of historical data is greater than a preset value, the confidence of the to-be-evaluated data is greater than a standard value, and it is determined that the to-be-evaluated data does not have bias, otherwise it is determined that the initial analysis evaluation report of the to-be-evaluated data has bias.

[0094] When the initial analysis evaluation report of the to-be-evaluated data has bias, the rectification information of the initial analysis evaluation report of the to-be-evaluated data is calculated, and the rectification information of the initial analysis evaluation report is fed to the large language model to be trained for model self-iteration optimization to generate a target analysis evaluation report and an optimized large language model.

[0095] It should be noted that the fine-tuning training management module is used to fine-tune the model after rectification analysis of the to-be-evaluated data, so that the model can automatically adjust the analysis evaluation report according to the rectification information. First, the initial analysis evaluation report of the to-be-evaluated data is calculated for confidence calculation. High confidence proves that there is no bias or the bias is very small, and no rectification processing is needed, otherwise it is needed. The data is common, and there is the same type of data in the historical data, i.e. historical data. If the similarity between the feature information of the historical data is greater than a preset value, the corresponding historical data can be used as a template to compare the corresponding analysis evaluation report to determine the coincidence degree of the analysis evaluation report to determine whether there is preliminary bias. When there is preliminary bias, the bias amount needs to be further judged and rectification information is generated.

[0096] Further, in a preferred embodiment of the present application, when the initial analysis and evaluation report of the to-be-evaluated data has deviation, the bias correction information of the initial analysis and evaluation report of the to-be-evaluated data is calculated, and the bias correction information of the initial analysis and evaluation report is fed into the to-be-trained large language model for model self-iterative optimization to generate a target analysis and evaluation report and an optimized large language model. Specifically,

[0097] The fine-tuning training deviation interval is set. When the initial analysis and evaluation report of the to-be-evaluated data has deviation, and the initial analysis and evaluation report of the to-be-evaluated data after deviation is not within the fine-tuning training deviation interval, the non-overlapping position between the initial analysis and evaluation report of the to-be-evaluated data and the historical analysis and evaluation report of a type of historical data is determined, and is labeled as analysis and evaluation report preliminary bias correction information.

[0098] An interactive page of the to-be-trained large language model is constructed. Based on the interactive page, the analysis and evaluation report preliminary bias correction information is provided to the user for secondary correction of the bias correction information to obtain analysis and evaluation report target bias correction information.

[0099] The analysis and evaluation report target bias correction information is imported into the fine-tuning training management module of the to-be-trained large language model, and the to-be-trained large language model is fine-tuned to optimize the analysis performance of the to-be-trained large language model to obtain an optimized large language model. The fine-tuning is a model self-iterative optimization of the to-be-trained large language model.

[0100] The information number of the analysis and evaluation report target bias correction information is analyzed in real time. If the information number of the analysis and evaluation report target bias correction information is greater than a predetermined value, the model self-iterative optimization is started, and 60% of the analysis and evaluation report target bias correction information is used for fine-tuning, and 40% of the analysis and evaluation report target bias correction information is used for verification test during the model self-iterative optimization.

[0101] The initial analysis and evaluation report of the to-be-evaluated data is corrected and optimized in combination with the optimized large language model and the imported analysis and evaluation report target bias correction information to obtain a corrected analysis and evaluation report of the to-be-evaluated data, which is labeled as a target analysis and evaluation report.

[0102] It should be noted that the interactive page is used for experts and owners to analyze the analysis evaluation report in real time, further correct and correct the target deviation information of the analysis evaluation report. The non-overlapping position between the initial analysis evaluation report of the to-be-evaluated data and the historical analysis evaluation report of the historical data is the deviation position, and the position needs to be corrected. And there is a deviation interval that can be corrected, for example, when the deviation position is less than 10%, it is a deviation within the allowable range, and no correction processing is required. On the contrary, if the threshold is exceeded, the model needs to be fine-tuned, that is, the model is iteratively optimized. The correction information can be iteratively optimized after reaching a certain data size, such as more than 1000, fine-tuned once, and according to a preset proportion, 60% for fine-tuning and 40% for verification test, to ensure the accuracy of fine-tuning. After optimizing the model, the initial analysis evaluation report of the to-be-evaluated data is corrected and optimized by the model to obtain the analysis evaluation report of the to-be-evaluated data after correction, which is calibrated as the target analysis evaluation report. At this time, the analysis evaluation report has accuracy.

[0103] Further, in a preferred embodiment of the present application, the initial analysis evaluation report of the to-be-evaluated data is corrected and optimized to obtain the analysis evaluation report of the to-be-evaluated data after correction, which is calibrated as the target analysis evaluation report, specifically:

[0104] The imported analysis evaluation report target deviation information is subjected to format unification and data set construction processing to obtain a to-be-evaluated data set;

[0105] The to-be-evaluated data set is configured in the fine-tuning training management module of the to-be-trained large language model, and the to-be-trained large language model is controlled to perform model parameter fine-tuning, wherein the model parameter fine-tuning is to configure and update the model parameters of the to-be-trained large language model, and calculate the model KL divergence during the updating process. When the model KL divergence is maintained within a preset range, the model parameter fine-tuning is stopped, and a large language model of a type is obtained;

[0106] The test analysis evaluation report is input into the large language model of a type, and the test analysis evaluation report is corrected and tested by the large language model of a type to judge the sensitivity and accuracy of the correction and test;

[0107] If the sensitivity and accuracy of the correction and test are greater than a preset value, the model parameters of the large language model of a type are frozen, and the large language model of a type is calibrated as an optimized large language model;

[0108] If the sensitivity and accuracy of the correction and test are not greater than a preset value, the model parameters of the large language model of a type are continuously fine-tuned until the sensitivity and accuracy of the correction and test are greater than a preset value.

[0109] It should be noted that the large model needs to be dynamically optimized and fine-tuned to realize the purpose of automatically analyzing and correcting the data. The training needs to configure the data set in the fine-tuning training management module of the large language model to be trained, and control the large language model to be trained to fine-tune the model parameters. Among them, the KL divergence is a deviation loss function, which is used for training function in training process, when the KL divergence is maintained in the preset range, the model parameter fine-tuning is stopped, and a large language model is obtained, at this time, it is the fine-tuned large language model, which can be used for correction test to judge whether the correction sensitivity and accuracy are qualified, if qualified, it can be directly output as a data analysis model, if not qualified, it continues to train. The data analysis model is a model for evaluating digital assets.

[0110] As shown in Figure 3 The second aspect of the present application also includes a digital asset evaluation system based on dynamic optimization of a large language model, characterized in that the digital asset evaluation system integrates a high-performance computing architecture and a data storage module, including a non-volatile memory composed of an ECC-verified DDR4 RDIMM memory module and an NVMe solid-state storage array using 3D NAND flash, and a multi-core processor based on Zen4 microarchitecture; the memory has a digital asset evaluation method program with a digital asset evaluation engine solidified and deployed therein, when the program is executed in parallel by the superscalar pipeline execution unit in the processor, the following steps are realized:

[0111] A large language model to be trained is constructed, dimension label classification and key feature extraction are performed on the data to be evaluated, and an LLM prompt template is constructed;

[0112] The key feature information of the data to be evaluated is structured and converted, and the data to be evaluated is preliminarily analyzed in combination with the LLM prompt template and the data analysis module to generate an initial analysis evaluation report;

[0113] The initial analysis evaluation report is analyzed for deviation, deviation information of the initial analysis evaluation report is generated, and the deviation information of the initial analysis evaluation report is fed into the large language model to be trained for model self-iteration optimization to generate a target analysis evaluation report and an optimized large language model;

[0114] All the data to be evaluated is analyzed and evaluated by the optimized large language model.

[0115] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A digital asset evaluation method based on dynamic optimization of a large language model, characterized in that: The following steps are involved: Build a large language model to be trained, classify the dimension labels and extract key features of the evaluation data, and build an LLM prompt template; Perform structural transformation on the key characteristic information of the data to be evaluated, and combine the LLM prompt template and data analysis module to conduct preliminary analysis on the data to be evaluated and generate an initial analysis and evaluation report; Perform deviation analysis on the initial analysis and evaluation report to generate deviation correction information for the initial analysis and evaluation report. This deviation correction information is then fed into the large language model to be trained for self-iterative optimization, generating a target analysis and evaluation report and optimizing the large language model. By optimizing the large language model, all data to be evaluated is analyzed and evaluated.

2. The digital asset evaluation method based on dynamic optimization of a large language model according to claim 1 is characterized in that: The construction of the large language model to be trained, the dimension label classification and key feature extraction of the evaluation data, and the construction of the LLM prompt template are as follows: Acquire construction software for constructing a large language model, mark it as target construction software, and construct a large language model to be trained based on the target construction software; The large language model to be trained includes different model modules, including a data classification module, a prompt template management module, a data analysis module, a fine-tuning training management module, and a correction feedback management module; Determine the data that needs to be analyzed, mark it as data to be evaluated, and obtain all dimension labels of the data to be evaluated. The data that needs to be analyzed is digital asset data; Introducing a big data network, determining all feature information of different data to be evaluated in the big data network, and analyzing the number of times all feature information of the data to be evaluated is cited in the big data network, marking feature information with a number of big data network citations greater than a preset number as key feature information of the data to be evaluated, and saving the key feature information of the data to be evaluated in a data classification module; Retrieving a prompt library in a big data network, and retrieving and designing an LLM prompt template in the prompt library, wherein different LLM prompt blank templates are preset in the prompt library; Among them, the LLM prompt template is a structured input text template of a large language model. The method for designing the LLM prompt template is to combine the dimensional labels of all the data to be evaluated, define the key elements of the LLM prompt template, including input format, output requirements and task instructions, and based on the key elements of the LLM prompt template, on the LLM prompt blank template with the highest degree of correlation with the key elements of the LLM prompt template, construct a structured input text template that can generate dimensional features through the key feature information of the data to be evaluated, that is, the LLM prompt template, and save the LLM prompt template in the prompt template management module.

3. The digital asset evaluation method based on dynamic optimization of a large language model according to claim 1 is characterized in that: The key feature information of the data to be evaluated is structured and converted, and combined with the LLM prompt template and data analysis module, a preliminary analysis of the data to be evaluated is performed to generate an initial analysis and evaluation report, specifically: In the large language model to be trained, the key feature information of the data to be evaluated in the data module is structurally converted according to the input format of the key elements in the LLM prompt template, so that the format of the key feature information of the data to be evaluated is equal to the input format of the key elements in the LLM prompt template; The key feature information of the data to be evaluated after structured conversion is marked as target feature information, and the data classification module and the prompt template management module are connected so that the target feature information can be imported into the LLM prompt template; Within the LLM prompt template, field analysis is performed on the target feature information. The analyzed fields include text fields and numeric fields. Field integrity is checked using the LLM prompt template before field analysis. If the field integrity is not within the preset range, a secondary structural transformation is performed on the target feature information until the field integrity remains within the preset range. After performing field analysis on the target feature information, based on the output requirements in the LLM prompt template, the key fields of the target feature information are screened and parsed. Based on the parsing results, the dimension labels corresponding to the data to be evaluated are output in the LLM prompt template. Combined with the dimension labels corresponding to the data to be evaluated, an initial analysis and evaluation report of the data to be evaluated is generated in the data analysis module of the large language model to be trained.

4. The digital asset evaluation method based on dynamic optimization of a large language model according to claim 1 is characterized in that: The deviation analysis is performed on the initial analysis and evaluation report to generate correction information for the initial analysis and evaluation report, and the correction information of the initial analysis and evaluation report is fed into the large language model to be trained for model self-iteration optimization, to generate a target analysis and evaluation report and optimize the large language model, specifically as follows: In the large language model to be trained, import the initial analysis and evaluation report into the fine-tuning training management module and save it; In the fine-tuning training management module of the large language model to be trained, a correction trigger mechanism is set, wherein the correction trigger mechanism can calculate the confidence level of the initial analysis and evaluation report of the data to be evaluated, and determine whether the initial analysis and evaluation report of the data to be evaluated is biased based on the confidence level; The deviation correction trigger mechanism is provided with historical analysis and evaluation reports of all data of the same type as the data to be evaluated and the corresponding feature information, and the data of the same type as the data to be evaluated is marked as the historical data to be analyzed; Analyze the target feature information of the input data to be evaluated, and calculate the similarity between the feature information of the different historical data to be analyzed in the correction trigger mechanism; Select the historical data to be analyzed with the highest similarity, mark it as a category of historical data, determine the historical analysis and evaluation report of the category of historical data, compare it with the initial analysis and evaluation report of the data to be evaluated, and output the confidence level of the data to be evaluated; If the overlap between the initial analysis and evaluation report of the data to be evaluated and the historical analysis and evaluation report of a category of historical data is greater than a preset value, then the confidence level of the data to be evaluated is greater than the standard value, and it is determined that there is no deviation in the data to be evaluated. Otherwise, it is determined that there is a deviation in the initial analysis and evaluation report of the data to be evaluated. When there is an offset in the initial analysis and evaluation report of the data to be evaluated, the correction information of the initial analysis and evaluation report of the data to be evaluated is calculated, and the correction information of the initial analysis and evaluation report is fed into the large language model to be trained for model self-iteration optimization, generating a target analysis and evaluation report and optimizing the large language model.

5. The digital asset evaluation method based on dynamic optimization of a large language model according to claim 4 is characterized in that: When there is a deviation in the initial analysis and evaluation report of the data to be evaluated, the correction information of the initial analysis and evaluation report of the data to be evaluated is calculated, and the correction information of the initial analysis and evaluation report is fed into the large language model to be trained for model self-iteration optimization, generating a target analysis and evaluation report and optimizing the large language model. Specifically: Set a fine-tuning training deviation interval. When the initial analysis and evaluation report of the data to be evaluated has a deviation, and the initial analysis and evaluation report of the data to be evaluated after the deviation is not within the fine-tuning training deviation interval, determine the non-overlapping position between the initial analysis and evaluation report of the data to be evaluated and the historical analysis and evaluation report of a type of historical data, and mark it as the preliminary correction information of the analysis and evaluation report; Construct an interactive page for the large language model to be trained. Based on the interactive page, provide the preliminary correction information of the analysis and evaluation report to the user for secondary correction of the correction information to obtain the target correction information of the analysis and evaluation report; Import the target correction information from the analysis and evaluation report into the fine-tuning training management module of the large language model to be trained, and perform fine-tuning training on the large language model to be trained to optimize the analysis performance of the large language model to be trained, thereby obtaining an optimized large language model. Fine-tuning training involves self-iterative optimization of the large language model to be trained. The target deviation correction information of the analysis and evaluation report is analyzed in real time. If the target deviation correction information of the analysis and evaluation report is greater than a predetermined value, the model self-iteration optimization is started. During the model self-iteration optimization process, 60% of the target deviation correction information of the analysis and evaluation report is used for fine-tuning training, and 40% of the target deviation correction information of the analysis and evaluation report is used for verification testing. Combined with the optimized large language model and the imported analysis and evaluation report target correction information, the initial analysis and evaluation report of the data to be evaluated is corrected and optimized to obtain the analysis and evaluation report of the data to be evaluated after correction, which is marked as the target analysis and evaluation report.

6. The digital asset evaluation method based on dynamic optimization of a large language model according to claim 5 is characterized in that: The initial analysis and evaluation report of the data to be evaluated is corrected and optimized to obtain the analysis and evaluation report of the data to be evaluated after correction, which is calibrated as the target analysis and evaluation report, specifically: The imported analysis and evaluation report target correction information is formatted and the data set is constructed to obtain the data set to be evaluated; Configuring the dataset to be evaluated in a fine-tuning training management module of the large language model to be trained, and controlling the large language model to be trained to perform model parameter fine-tuning, wherein the model parameter fine-tuning is to update the configuration of the model parameters of the large language model to be trained, and calculating the KL divergence of the model during the updating process. When the KL divergence of the model remains within a preset range, the model parameter fine-tuning is stopped, and a class of large language models is obtained; Input the test correction information and the test analysis and evaluation report into a large language model, and perform a correction test on the test analysis and evaluation report using the large language model to determine the sensitivity and accuracy of the correction test; If the sensitivity and accuracy of the correction test are both greater than the preset values, the model parameters of the first-class large language model are frozen, and the first-class large language model is calibrated as the optimized large language model; If the sensitivity and accuracy of the correction test are not greater than the preset values, continue to fine-tune the model parameters of the large language model until the sensitivity and accuracy of the correction test are greater than the preset values.

7. The digital asset evaluation method based on dynamic optimization of a large language model according to claim 1 is characterized in that: By optimizing the large language model, all the data to be evaluated are analyzed and evaluated, specifically: In optimizing the large language model, all the data to be evaluated is imported for data analysis, and analysis and evaluation reports corresponding to different data to be evaluated are generated; Based on the optimized large language model, analysis and evaluation reports corresponding to different data to be evaluated are pushed to the owners of the data to be evaluated through an interactive page.

8. A digital asset evaluation system based on dynamic optimization of a large language model, characterized by: The digital asset evaluation system integrates a high-performance computing architecture and a data storage module, including a non-volatile memory consisting of a DDR4 RDIMM memory module equipped with ECC check and an NVMe solid-state storage array using 3D NAND flash memory, and a multi-core processor based on the Zen4 microarchitecture; a digital asset evaluation method program with a digital asset evaluation engine is fixedly deployed in the memory, and when the program is decoded and executed in parallel by a superscalar pipeline execution unit in the processor, the data processing steps according to any one of claims 1 to 7 are implemented.

Images