Asset evaluation method based on natural language large model and related device

Abstract

The application discloses an asset evaluation method based on a natural language large model and related devices, and the method comprises the following steps: acquiring a first multi-dimensional corpus and a preset base model; the preset base model is a pre-trained natural language large model; the first multi-dimensional corpus is a multi-source and multi-type corpus set constructed for an asset evaluation task, and comprises text corpus, structured corpus and semi-structured corpus; data preprocessing is performed on the first multi-dimensional corpus to obtain a second multi-dimensional corpus; the preset base model is trained based on the second multi-dimensional corpus to obtain a target data asset evaluation model; target object asset data to be evaluated is acquired; the target data asset evaluation model is used to process the asset data to be evaluated to obtain a target asset evaluation report. According to the embodiment of the application, the reliability of the asset evaluation result is improved.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to an asset valuation method and related apparatus based on a natural language large model. Background Technology

[0002] With the synergistic development of the digital economy and artificial intelligence technologies, asset forms are becoming increasingly diversified, with textual assets and intangible assets accounting for a continuously rising proportion, placing higher demands on the intelligence and precision of asset valuation. Natural Language Modeling (LLM), with its powerful semantic understanding, information extraction, and logical reasoning capabilities, offers a technological possibility for overcoming the dilemmas of traditional asset valuation and is gradually becoming an important technological support in the field of asset valuation.

[0003] Currently, asset valuation methods rely excessively on professionals' manual interpretation of textual information such as asset ownership documents, industry reports, and market dynamics. This is not only inefficient but also prone to subjective bias due to differences in individual experience, resulting in low reliability of the valuation results.

[0004] Therefore, how to improve the reliability of asset valuation results has become an urgent problem to be solved. Summary of the Invention

[0005] This application provides an asset valuation method and related apparatus based on a large natural language model, which improves the reliability of asset valuation results.

[0006] In a first aspect, embodiments of this application provide an asset valuation method based on a large natural language model, including: Obtain a first multi-dimensional corpus and a preset base model; the preset base model is a pre-trained large natural language model; the first multi-dimensional corpus is a collection of multi-source and multi-type corpora constructed for asset appraisal tasks, including text corpora, structured corpora and semi-structured corpora; The first multi-dimensional corpus is preprocessed to obtain the second multi-dimensional corpus; The preset base model is trained based on the second multi-dimensional corpus to obtain the target data asset evaluation model; Obtain the asset data of the target entity to be evaluated; The target asset valuation model is used to process the data of the asset to be valued to obtain a target asset valuation report.

[0007] Secondly, embodiments of this application provide an asset valuation device based on a large natural language model. The device includes: an acquisition module, a model training module, and an asset valuation module, wherein: The acquisition module is used to acquire a first multi-dimensional corpus and a preset base model; the preset base model is a pre-trained large natural language model; the first multi-dimensional corpus is a collection of multi-source and multi-type corpora constructed for asset appraisal tasks, including text corpora, structured corpora and semi-structured corpora; The model training module is used to preprocess the first multi-dimensional corpus to obtain a second multi-dimensional corpus; and to train the preset base model based on the second multi-dimensional corpus to obtain a target data asset evaluation model. The acquisition module is also used to acquire the asset data to be evaluated of the target object; The asset valuation module is used to process the asset data to be valued using the target data asset valuation model to obtain a target asset valuation report.

[0008] Thirdly, embodiments of this application provide an electronic device, including: a processor, a memory, a communication interface, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the processor, and the programs include instructions for performing the steps in the first aspect of embodiments of this application.

[0009] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program for electronic data interchange, wherein the computer program causes a computer to perform some or all of the steps described in the first aspect of embodiments of this application.

[0010] Fifthly, embodiments of this application provide a computer program product, wherein the computer program product includes a non-transitory computer-readable storage medium storing a computer program operable to cause a computer to perform some or all of the steps described in the first aspect of embodiments of this application. The computer program product may be a software installation package.

[0011] Implementing this application will have the following beneficial effects: As can be seen, the asset appraisal method and related apparatus based on a large natural language model described in this application constructs a first multi-dimensional corpus with multiple sources and types for asset appraisal, performs data preprocessing on it to form a high-quality second multi-dimensional corpus, and then trains a preset base model based on the second multi-dimensional corpus to obtain a target data asset appraisal model. This appraisal model is used to perform intelligent processing of the asset data to be appraised throughout the entire process, which plays a role in standardizing the process, unifying the logic, and accurately appraising. Compared with the existing technology that relies on manual interpretation and subjective judgment, this method has the advantages of a more standardized process, more objective logic, more accurate data mining, and no experience bias, thereby improving the reliability of the asset appraisal results. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of this application or the background art, the accompanying drawings used in the embodiments of this application or the background art will be described below.

[0013] Figure 1 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application; Figure 2 This is a flowchart of an asset valuation method based on a large natural language model provided in an embodiment of this application; Figure 3 This is a schematic diagram illustrating the composition of a second multi-dimensional corpus provided in an embodiment of this application; Figure 4 This is a flowchart of a method for directly generating a dataset provided in an embodiment of this application; Figure 5 This is a flowchart of a method for generating a supervised fine-tuning dataset provided in an embodiment of this application; Figure 6 This is a flowchart of a method for obtaining a vertical domain model provided in an embodiment of this application; Figure 7 This is a functional module block diagram of an asset appraisal device based on a large natural language model provided in an embodiment of this application; Figure 8 This is a schematic diagram of the structure of another electronic device provided in an embodiment of this application. Detailed Implementation

[0014] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present application.

[0015] The terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or apparatuses.

[0016] It should be understood that the term "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this document indicates that the preceding and following related objects are in an "or" relationship. In the embodiments of this application, "multiple" refers to two or more.

[0017] In the embodiments of this application, "at least one item" or its similar expression refers to any combination of these items, including any combination of a single item or a plurality of items. "One or more" means one or more, while "multiple" means two or more. For example, "at least one item" of a, b, or c can represent the following seven cases: a, b, c; a and b; a and c; b and c; a, b, and c. Each of a, b, and c can be an element or a set containing one or more elements.

[0018] In this application, the term "connection" refers to various connection methods, such as direct connection or indirect connection, to achieve communication between devices. This application does not impose any limitations on this.

[0019] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0020] The electronic devices described in this application embodiment may include smartphones (such as Android phones, iOS phones, Windows Phones, etc.), tablet computers, PDAs, laptops, video matrices, monitoring platforms, mobile internet devices (MIDs), or wearable devices, etc. The above are merely examples and not exhaustive, and include but are not limited to the above devices.

[0021] Of course, the aforementioned electronic devices can also be servers, such as cloud servers.

[0022] The following describes the relevant content, concepts, meanings, technical issues, technical solutions, and beneficial effects involved in the embodiments of this application.

[0023] First, let me explain some of the technical terms or phrases used in this application: Asset appraisal: refers to the act and process by which a professional entity, under the premise of compliance and legality, comprehensively uses appraisal methods such as the market approach, income approach, and cost approach to analyze and estimate the current value of assets (including tangible assets, intangible assets, data assets, etc.) and issue a professional appraisal opinion. It is mainly used in scenarios such as asset transaction pricing, financial accounting, ownership confirmation, and asset management. The objectivity and accuracy of the appraisal results directly affect the efficiency of asset circulation and the reliability of decision-making.

[0024] Pre-trained large-scale natural language models refer to large-scale language models that are pre-trained based on massive amounts of general text data through self-supervised learning. They possess powerful general language capabilities such as semantic understanding, information extraction, logical reasoning, and text generation. They can be adapted to multiple application scenarios without retraining for specific tasks, and their professional capabilities in vertical fields (e.g., asset valuation) can be further enhanced through fine-tuning of domain corpora.

[0025] Data distillation: a model compression and knowledge transfer technique that transfers the knowledge, reasoning logic, or prediction distribution learned by a complex "teacher model" to a lightweight "student model," enabling the student model to reduce the number of parameters and computational costs while maintaining its core capabilities. In this application, it can be used to transfer the professional evaluation knowledge of a target data asset evaluation model to a lightweight model deployed on the adapted edge, balancing evaluation accuracy and reasoning efficiency.

[0026] Thinking chain: refers to a reasoning pattern in which the reasoning process is broken down into multiple logical steps and presented explicitly when dealing with complex tasks. It forms a complete reasoning chain by gradually deriving and connecting key information, rather than directly outputting the result.

[0027] Data chunking: refers to the processing method of splitting large-scale, high-capacity corpus data or asset data into several standardized data blocks according to preset rules (e.g., semantic integrity, data type, storage unit size).

[0028] "Full-blooded large model" refers to a large natural language model that has not undergone parameter pruning, quantization compression, knowledge distillation, or structural simplification. It fully retains all network layers, parameter weights, computational accuracy, and full capabilities of the original pre-trained model, and has the ability to perform semantic understanding, multi-step reasoning, and complex task processing without attenuation.

[0029] Easydataset tool: refers to a lightweight, standardized dataset building tool designed for vertical domains (e.g., asset valuation). It supports the rapid access, cleaning, deduplication, format unification, and annotation management of multi-source (public / private) and multi-type (text / structured / semi-structured) data. It can efficiently generate high-quality domain corpora adapted for model training and evaluation, reducing the technical threshold and time cost of data preparation in vertical domains.

[0030] EasyDistill is an open-source toolkit whose core goal is to make knowledge distillation techniques for large language models easier to use, thereby efficiently migrating the capabilities of large, high-performance "teacher models" to smaller, faster "student models".

[0031] BERT (Bidirectional Encoder Representations from Transformers) model: It is a pre-trained natural language processing model with powerful semantic understanding, text representation, and semantic similarity calculation capabilities. It can accurately capture the deep semantic information and contextual relationships of text.

[0032] Please see Figure 1 , Figure 1 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application; it can be seen that the electronic device may include: a corpus construction unit, a model training unit, and an asset evaluation unit, wherein: The corpus construction unit is used to acquire and build a multi-source, multi-type, multi-dimensional corpus for asset assessment tasks. It performs data preprocessing operations such as deduplication, cleaning, format unification, and domain adaptation on the multi-dimensional corpus to generate a high-quality multi-dimensional corpus, providing standardized and highly adaptable domain data support for subsequent model training.

[0033] The model training unit is used to acquire a pre-trained natural language large model as a preset base model. Based on the multi-dimensional corpus output by the corpus construction unit, the preset base model is subjected to domain-specific training and optimization to obtain a target data asset valuation model with professional asset valuation capabilities, realizing the deep integration of general large model and asset valuation domain knowledge.

[0034] The asset appraisal unit is used to acquire the asset data of the target object, call the target data asset appraisal model trained by the model training unit, and perform intelligent processing of the asset data to be appraised throughout the entire process, including information extraction, logical reasoning, value estimation and report generation, and finally output a standardized and objective target asset appraisal report, thus completing the intelligent execution of the entire asset appraisal process.

[0035] Please see Figure 2 , Figure 2 This is a flowchart of an asset valuation method based on a large natural language model, provided in an embodiment of this application; the method includes, but is not limited to, the following steps: S201. Obtain a first multi-dimensional corpus and a preset base model; the preset base model is a pre-trained large natural language model; the first multi-dimensional corpus is a collection of multi-source, multi-type corpora constructed for asset appraisal tasks, including text corpora, structured corpora and semi-structured corpora.

[0036] In this embodiment of the application, the preset base model can be preset in advance or defaulted.

[0037] It should be explained that the asset valuation method based on a large natural language model provided in this application can be applied to... Figure 1 The electronic device shown.

[0038] In a specific embodiment, the scope of corpus collection can be clearly defined, and the corpus can be divided into three major sources: basic fields, cutting-edge dynamics, and real business; three major types: text, structured, and semi-structured; and four major content dimensions: basic theory, industry knowledge, dynamic information, and business data. Secondly, we collect corpora from different sources: we obtain basic corpora such as professional books and evaluation criteria through formal channels, dynamic corpora such as cutting-edge papers and reports through academic platforms and industry institutions, and real business corpora by connecting with enterprise business systems and accessing the enterprise's internal knowledge base. Finally, the coverage and quality of the corpus are verified, missing content is supplemented, and a dynamic update mechanism is established to ultimately form the first multi-dimensional corpus for asset assessment tasks, which is multi-source and multi-type. Then, a preset base model can also be obtained.

[0039] In some embodiments, the preset base model can be a DeepSeek-R1 or a Qwen series model.

[0040] Thus, by constructing the first multi-dimensional corpus, we can ensure that the corpus is comprehensive, reliable, and relevant to business needs, providing data support for model training.

[0041] S202. Perform data preprocessing on the first multi-dimensional corpus to obtain the second multi-dimensional corpus.

[0042] In this embodiment of the application, data preprocessing may include at least one of the following: data cleaning, format standardization, handling of missing and outlier values, corpus regularization and quality verification, etc., which are not limited here.

[0043] In a specific embodiment, data preprocessing can include three operations: data cleaning, format unification, and handling of missing and outlier values. First, data cleaning can be performed to remove duplicate, garbled, irrelevant, and low-quality data from the first multi-dimensional corpus, and to de-identify sensitive data such as enterprise privacy and customer information, filtering invalid content to ensure data compliance and purity. Second, format unification can be performed to convert text-based, structured, and semi-structured data into standardized formats (e.g., text to TXT format, structured data to JSON format), unifying encoding rules, field naming, and storage specifications to eliminate format differences between multiple sources. Third, missing and outlier value handling can be performed to complete missing key evaluation information (e.g., asset parameters, valuation indicators) in the corpus, correcting numerical anomalies and format errors to ensure data integrity and accuracy, adapting to model training requirements. After the above operations, a standardized, clean, and complete second multi-dimensional corpus is obtained, providing high-quality data support for model training.

[0044] It should be noted that data preprocessing may also include other processing operations, which can be flexibly selected according to the actual situation, and are not limited here.

[0045] In some embodiments, please refer to Figure 3 , Figure 3 This is a schematic diagram illustrating the composition of a second multi-dimensional corpus provided in an embodiment of this application. As can be seen, the second multi-dimensional corpus may include: basic theoretical corpus data, industry knowledge corpus data, dynamic information corpus data, business practice corpus data, etc., and is not limited thereto; wherein: Basic theoretical corpus data: This includes preprocessed evaluation methods, valuation logic, and ownership rules. Industry knowledge corpus data: standardized assessment points and industry practices for different asset types (e.g., real estate, digital assets, etc.); Dynamic information corpus data: the latest regulatory rules, market trends, and cutting-edge assessment technologies after cleaning; Business practice data: anonymized and completed real asset information, historical valuation parameters, transaction data, etc.

[0046] Optionally, in some embodiments, step S202, which involves preprocessing the first multi-dimensional corpus to obtain a second multi-dimensional corpus, includes: S21. Determine the corpus file format corresponding to the first multi-dimensional corpus; S22. Determine the target segmentation method according to the corpus file format; S23. Divide the corpus data in the first multi-dimensional corpus into blocks according to the target segmentation method to obtain the first segmented corpus dataset; S24. Perform data cleaning on the first segmented corpus dataset to obtain a second segmented corpus dataset; the second segmented corpus dataset includes e segmented corpus data; e is a positive integer greater than 1; S25. Perform data quality scoring on each of the e segments of the corpus data to obtain e data quality scores; S26. Determine the second multi-dimensional corpus based on the e data quality scores and the second segmented corpus dataset.

[0047] In this embodiment of the application, the target segmentation method may include one of the following: parent-child segmentation, semantic paragraph segmentation, fixed-length segmentation, segmentation by format identifier, field-based segmentation, etc., which are not limited here.

[0048] In a specific embodiment, all corpora in the first multi-dimensional corpus can be traversed. By reading the file extension, file header identifier, and data structure features, the original file format (e.g., TXT, JSON, etc.) of each corpus can be determined one by one, and the format information of each corpus can be recorded. Then, the frequency of all identified file formats can be statistically analyzed, and the formats can be categorized into three types: plain text format, structured data format, and semi-structured mixed format. The proportion of each type of format and the specific format can be calculated. The specific file format with the highest proportion in the statistical results can be selected and determined as the core target file format of the corpus. For example, if a certain type of format (e.g., plain text format) has the highest proportion in the first multi-dimensional corpus, then that type of format can be used as the corpus file format corresponding to the first multi-dimensional corpus.

[0049] Next, the target segmentation method can be determined according to the corpus file format. Specifically, the mapping relationship between preset file formats and segmentation methods can be stored in advance, and the target segmentation method corresponding to the corpus file format can be determined based on the mapping relationship.

[0050] For example, if the corpus file format is plain text, the target segmentation method can be parent-child segmentation or semantic paragraph segmentation. Specifically, for plain text corpora with a clear hierarchical structure (e.g., chapters, clauses, sections) (e.g., evaluation criteria, professional books), parent-child segmentation is used; for plain text corpora without hierarchy but with semantic coherence (e.g., rule interpretation, expert experience), semantic paragraph segmentation is used.

[0051] If the corpus file is in a structured data format, the target segmentation method can be field-based segmentation; segmentation is based on core fields of asset valuation business (such as basic asset information, valuation parameters, transaction data, and valuation conclusions) to retain complete data information for each field.

[0052] If the corpus file format is a semi-structured hybrid format, the target segmentation method can include one or more combinations of segmentation by format identifier, semantic paragraph segmentation, and field segmentation. Specifically, segmentation by format identifier is used for parts containing natural structures such as tables, tags, and chapters, semantic paragraph segmentation is used for plain text parts, and field segmentation is used for structured parts to ensure that the semantics of the corpus are complete and the structure is clear after segmentation.

[0053] Furthermore, the corpus data in the first multi-dimensional corpus can be segmented according to the target segmentation method to obtain the first segmented corpus dataset. For example, assuming the target segmentation method is semantic paragraph segmentation, then all corpus data in the first multi-dimensional corpus is segmented according to sentence boundaries, paragraph separators, and semantic coherence, using semantically complete and independently expressive paragraphs as the basic units, so that each segment contains complete semantic information, thus obtaining the first segmented corpus dataset. Then, the first segmented corpus dataset can be cleaned to obtain the second segmented corpus dataset. Specifically, special symbols, garbled characters, whitespace characters, redundant punctuation, and invalid text content in the segmented corpus can be removed using regular expressions, duplicate segments and invalid segments can be removed, and the cleaned corpus blocks can be formatted and their integrity verified, thus obtaining a clean, standardized, and valid second segmented corpus dataset.

[0054] Then, data quality scores can be assigned to each of the e segments of the corpus data to obtain e data quality scores. Specifically, for each segment of the corpus data, semantic features are extracted using a pre-defined language model (e.g., the BERT model), and quality scoring rules are constructed from the perspectives of semantic completeness, redundancy, and complexity. Based on these quality scoring rules, the segment of the corpus data is quantitatively scored to obtain a data quality score. In this way, e data quality scores can be obtained.

[0055] Optionally, in some embodiments, the data quality scoring dimensions may include: redundancy and complexity, with each dimension having a corresponding score range, and a total score of 100 points. The specific quality scoring rules are as follows: Redundancy (0-50 points): Whether the segmented corpus is free of repetition, nonsense, and redundant expressions, and whether the information density is reasonable; Complexity (0-50 points): The semantic logic of the segmented corpus is clear, the expression is standardized, unambiguous, and unbroken, and it contains effective professional information.

[0056] Ultimately, scores are obtained for each dimension, which is the data quality score.

[0057] It should be noted that the average score of each dimension can also be used as the final data quality score.

[0058] Taking a specific segment of the corpus data as an example, the scoring process is as follows: The first step is to use the BERT model to extract the semantic features of the segment, and judge that the corpus has no repetition, no redundancy, and reasonable information density, with a redundancy score of 45. The second step, based on the semantic structure and professional content, judged that the logic was clear, the expression was standardized and unambiguous, and the complexity score was 43 points. The third step is to use the scores of each dimension as the data quality score, which includes: redundancy score 45 and complexity score 43.

[0059] Optionally, in some embodiments, the data quality scoring dimensions may include: semantic integrity, domain relevance, content standardization, and information validity. Each dimension has a corresponding score range, with a total score of 100 points. The specific quality scoring rules are as follows: Semantic integrity (0~25 points): Whether the semantics of the chunks are complete, without omissions, punctuation breaks, or semantic breaks.

[0060] Relevance to the field (0-30 points): Whether the content is relevant to the field of asset valuation and whether it is professional and valid.

[0061] Content standardization (0-25 points): No garbled characters, no special symbols, no redundant whitespace, and standard format.

[0062] Information validity (0-20 points): Whether the content is true, valuable, and free from repetition and redundancy.

[0063] The final score is obtained for each dimension, which is the data quality score.

[0064] Taking a specific segment of the corpus data as an example, the scoring process is as follows: The first step is to use the BERT model to extract the semantic features of the block and determine whether it is semantically complete and unbroken. The semantic completeness score is 22. The second step is to determine whether the content belongs to asset valuation-related texts based on the characteristics of the domain, and the domain relevance score is 26 points. The third step, after data cleaning, is free of garbled characters and redundant symbols, and the format and content are standardized, scoring 23 points for standardization. The fourth step is that the corpus content is effective, non-repetitive, and has reference value, and the information effectiveness score is 18 points; The fifth step yields the following data quality scores: semantic integrity (22 points), domain relevance (26 points), content standardization (23 points), and information effectiveness (18 points).

[0065] Finally, based on the e data quality scores and the second segmented corpus dataset, a second multi-dimensional corpus can be determined. Specifically, f data quality scores from the e data quality scores that are greater than or equal to a preset data quality score can be selected, where f is an integer less than or equal to e. Then, f segmented corpus data corresponding to these f data quality scores are found from the second segmented corpus dataset, and the second multi-dimensional corpus is composed of these f segmented corpus data. The preset data quality scores can be preset in advance or defaulted.

[0066] In this way, by first matching the corpus format and segmentation method, then performing segmentation cleaning, and finally filtering the corpus based on semantic quality scores, we can ensure the semantic integrity of the segments, the cleaning is accurate and efficient, and at the same time, we can quantitatively filter low-quality data. In the end, we can build a standardized, clean, domain-relevant, and semantically high-quality second multi-dimensional corpus, which can improve the effect of subsequent corpus applications.

[0067] S203. The preset base model is trained based on the second multi-dimensional corpus to obtain the target data asset evaluation model.

[0068] In this embodiment of the application, the corpus data in the second multi-dimensional corpus can be used to train the preset base model, thereby obtaining the target data asset evaluation model.

[0069] Optionally, step S203, training the preset base model based on the second multi-dimensional corpus to obtain the target data asset evaluation model, includes: A1. Obtain the target evaluation domain and target evaluation requirements corresponding to the target object; A2. Determine prompt word data based on the target evaluation area and the target evaluation requirements; A3. Generate a first corpus dataset using the first preset corpus tool based on the prompt word data and the second multi-dimensional corpus; A4. Distill the first corpus dataset using a pre-defined large model to obtain a second corpus dataset; the pre-defined large model is a complete natural language large model that has not been simplified or optimized. A5. The preset base model is trained based on the second corpus dataset to obtain the target data asset evaluation model.

[0070] In this embodiment of the application, the first preset corpus tool can be preset in advance or defaulted; the preset large model is also called the "full-blooded large model".

[0071] In a specific embodiment, the target evaluation domain and target evaluation requirements corresponding to the target object are first obtained. Specifically, the business scenario and task purpose of the target object can be analyzed to determine its professional field (i.e., the target evaluation domain) and the required target evaluation requirements. For example, if the target object is a game company, the corresponding target evaluation domain can be the game data asset domain, and the target evaluation requirement is to determine the commercial value of the game data assets. Alternatively, the asset evaluation instruction of the target object can be received directly, which contains the target evaluation domain and target evaluation requirements.

[0072] Then, based on the target evaluation domain and target evaluation requirements, the prompt word data can be determined. Specifically, domain information can be extracted from the target evaluation domain, including: domain keywords, professional terms, business scenarios, etc. At the same time, requirement constraints can be clarified based on the target evaluation requirements, including: corpus purpose, quality requirements, format requirements, output length, content focus, etc. Then, the domain information and requirement constraints can be combined to form structured and clearly defined prompt statement data. The prompt statement data is then standardized and organized to obtain the prompt word data.

[0073] For example, if the target evaluation area is the game data asset area, the following information is extracted: the domain keywords are game data assets, user behavior data, game item assets, and game revenue data; the professional terms are digital asset ownership confirmation, valuation, income approach, and cost approach; and the business scenarios are game operation data analysis and digital asset commercial pricing.

[0074] If the target assessment requirement is to evaluate the commercial value of game data assets, the following requirements constraints should be defined: the corpus is intended for training the asset assessment model; the quality requirements are semantic integrity, professional standardization, no redundancy, and low complexity; the format requirement is paragraph-style text; the output length is 100-200 words per paragraph; and the content focus is on the value influencing factors and assessment methods of game data assets.

[0075] By combining the aforementioned domain information with demand constraint information, a prompt statement is formed: "Generate professional corpus for the commercial value assessment of game data assets, covering user behavior data, item assets, and revenue data, using assessment terms such as the income approach and cost approach, with complete semantics, professional standardization, and no redundancy, with a single paragraph length of 100-200 words, focusing on value influencing factors and assessment logic." After standardization and organization, prompt word data that can be directly used in corpus tools is obtained.

[0076] Next, the first corpus dataset can be generated using the first preset corpus tool based on the prompt word data and the second multi-dimensional corpus. Specifically, the prompt word data can be input into the first preset corpus tool as the generation instruction, and the second multi-dimensional corpus can be used as the reference corpus and knowledge base. The first preset corpus tool performs semantic understanding, content expansion and standardization generation within the knowledge scope of the second multi-dimensional corpus based on the domain requirements, quality constraints and content direction in the prompt word data, and outputs multiple new corpora, thereby obtaining the first corpus dataset.

[0077] In some embodiments, the first preset corpus tool may be the Easydataset tool.

[0078] Then, the first corpus dataset can be distilled using a pre-set large model to obtain the second corpus dataset. Finally, the pre-set base model can be trained based on the second corpus dataset to obtain the target data asset evaluation model.

[0079] In this way, by generating precise prompt words based on the assessment domain and needs, combining them with a high-quality corpus to generate specialized corpus, then refining effective knowledge through large model data distillation, and finally training the base model with the refined corpus, the professionalism and quality of the corpus can be significantly improved, training noise can be reduced, and the model can focus on the asset assessment task more quickly, ultimately resulting in a lightweight, domain-adaptable, and more accurate specialized model.

[0080] In some embodiments, please refer to Figure 4 , Figure 4 This is a flowchart of a method for directly generating a dataset according to an embodiment of this application, as detailed below: 1. Industry knowledge: As the original input data source, it provides professional texts, knowledge content, industry standards and background information in the field, and is the basic source of materials for building datasets.

[0081] 2. Split text blocks: The input long text of industry knowledge is segmented into paragraphs, semantics, or structure to form shorter text units with smaller granularity and regular structure, which facilitates subsequent extraction and question generation.

[0082] 3. Generating questions: Based on the content of the split text blocks, question-and-answer pairs that match industry knowledge are automatically or semi-automatically generated, and the query statements required for the task are constructed, forming the basis of the question-and-answer structure.

[0083] 4. Construct the dataset: The generated questions are paired with corresponding industry knowledge texts, organized, and standardized in format, and then assembled into a sample set that can be used for model training, thus completing the construction of the data structure.

[0084] 5. Export dataset: The completed dataset is output in the specified format to form a training data file that can be directly loaded and used for subsequent model fine-tuning.

[0085] Figure 4 While the method shown can quickly generate datasets, it only completes basic text segmentation, question-and-answer generation, and data assembly. It cannot effectively guide and control the generation of thought chains, which are key to improving the model's reasoning ability, logical rigor, and interpretability. Therefore, the datasets constructed by this method are of low quality and cannot support high-quality reasoning-based fine-tuning.

[0086] To generate a high-quality dataset, please refer to Figure 5 , Figure 5 This is a flowchart of a method for generating a supervised fine-tuning dataset provided in an embodiment of this application, as detailed below: 1. Preprocessed corpus data: As the initial input to the entire process, it refers to the standardized and effective corpus obtained after cleaning, deduplication, format normalization, noise removal, and missing value completion of the original domain corpus, providing high-quality basic materials for subsequent dataset construction.

[0087] 2. First preset corpus tool: The first pre-set corpus tool performs standardized processing on the pre-processed corpus data, including structured parsing, text segmentation, and semantic extraction. This provides the basic corpus processing and format conversion capabilities for subsequent prompt word construction and dataset generation, and is the first tool processing node in the process.

[0088] 3. Construct prompt words for the dataset based on different fine-tuning areas and business needs: Based on the fine-tuning domain of the preset base model, the actual business scenario and the requirements of the dataset generation task, we designed and wrote prompts to guide the model to generate the corresponding dataset, and clarified the content, format, output form and constraints of the data generation.

[0089] 4. Save the generated prompts and domains each time as a knowledge base: The prompts generated in step 3 are categorized, summarized, and persistently stored according to their business domain and task type, forming a reusable, traceable, and iterative prompt knowledge base, providing historical reference for subsequent prompt optimization and iterative generation.

[0090] 5. Generate new prompts based on the previously constructed prompts: "Building an intelligent agent to generate prompts" specifically means that a dedicated automated prompt generation intelligent agent can be built. The intelligent agent can automatically complete the writing, optimization, verification and generation of prompts, generating new prompts with stronger adaptability and better guidance effect, and realizing the iterative upgrade of prompts; Additionally, new suggestions can be fed back to step 3 for optimization.

[0091] 6. Initial fine-tuning of the dataset: Guided by the prompts output in step 3, “Generate a dataset using a model with thought chain”. Specifically, a large model with thought chain reasoning ability can be used to process the corpus data and automatically generate initial sample data containing reasoning logic and derivation process, so that the initial data has basic thought chain features, thereby obtaining an initial fine-tuned dataset.

[0092] 7. Second pre-set corpus tool: The preliminary fine-tuning dataset from step 6 is processed in multiple dimensions using the second pre-set corpus tool, including redundancy detection, complexity scoring, thought chain integrity verification, and sample quality assessment, to provide data support for subsequent prompt word rewriting and indicator selection.

[0093] 8. Construct or rewrite thought chain prompts based on domain and business needs: Receive the corpus data and prompt words output from steps 7 and 5, and strengthen, correct, and standardize the sample thought chain of the corpus data and prompt words in combination with the current business domain and task objectives, so as to improve the reasoning logic and interpretability of the dataset. In addition, the prompts generated or rewritten in step 8 can be synchronously returned to step 4 and stored in the prompt knowledge base, enabling continuous supplementation and iteration of the prompt system.

[0094] 9. Automatically select indicators based on three preset formulas: Using three preset formulas (i.e., the first, second, and third preset formulas mentioned above), the sample quality, redundancy, complexity, and fit of the preliminary fine-tuning dataset are quantitatively calculated, and core evaluation indicators applicable to the current field are automatically selected, while invalid and low-value indicators are eliminated.

[0095] 10. By automatically testing and combining different metrics based on different fields: To address the unique characteristics of different business areas, the core indicators selected in step 9 are automatically combined, tested, and validated to determine the optimal combination of indicators that is highly compatible with the target evaluation area and evaluation needs, thus ensuring the rationality and applicability of the evaluation system.

[0096] 11. Supervised fine-tuning dataset: After optimization of the thought chain, automatic selection of indicators, and testing and verification of the optimal combination of indicators, a high-quality, highly adaptable, standardized dataset containing a standardized thought chain is finally generated, which is also known as the supervised fine-tuning dataset. This supervised fine-tuning dataset can be directly used for subsequent supervised fine-tuning (SFT) training of large models, providing high-quality samples for model training.

[0097] It should be explained that the first preset corpus tool can be the Easydataset tool; the second preset corpus tool can be the EasyDistill tool.

[0098] In some embodiments, please refer to Figure 6 , Figure 6 The flowchart below shows a method for obtaining a vertical domain model according to an embodiment of this application: 1. Obtain the preset base model: A pre-trained natural language processing large model is obtained and used as a preset base model. This preset base model has general language understanding, text generation and basic reasoning capabilities. It serves as the initial model body for subsequent domain fine-tuning training and provides the basic network structure and parameters for vertical domain-specific optimization.

[0099] 2. Supervised fine-tuning dataset: We use a high-quality dedicated dataset that has been optimized through thought chain, automatic index selection and domain adaptation in the early stage. It contains standardized vertical domain question answering, inference samples and labeled content to provide training sample support for supervised learning fine-tuning.

[0100] 3. Supervised fine-tuning model: Supervised learning fine-tuning is performed on a pre-defined base model using a supervised fine-tuning dataset. Specifically, the pre-defined base model can be optimized for parameters and adapted to the domain through a supervised learning training strategy, so that the model learns the knowledge paradigm, output format and reasoning logic of the target domain, thereby obtaining a supervised fine-tuned model.

[0101] The supervised fine-tuning model is an intermediate model obtained after fine-tuning training through supervised learning. This model has completed the initial domain knowledge alignment and task adaptation, and has basic vertical domain processing capabilities. It serves as the model to be optimized for subsequent reinforcement learning fine-tuning.

[0102] 4. Reward Dataset: The reward dataset is an evaluation and feedback dataset specifically built for the reinforcement learning phase. It is used to evaluate the quality and rationality of the output results of the supervised fine-tuning model and generate corresponding reward signals, providing optimization directions and reward / penalty basis for reinforcement learning fine-tuning.

[0103] 5. Vertical Domain Model: The supervised fine-tuning model is fine-tuned using a reward dataset through reinforcement learning. Specifically, by combining the feedback reward signals output from the reward dataset, the supervised fine-tuning model is further refined and optimized through a reinforcement learning training strategy. This corrects the model's inference logic and output performance, improving the model's accuracy, logic, and practicality in vertical domain tasks, thus obtaining a vertical domain model.

[0104] The vertical domain model (i.e., the target data asset valuation model) is a specialized domain model obtained after two levels of iterative optimization through supervised learning fine-tuning and reinforcement learning fine-tuning. This vertical domain model is deeply adapted to the business needs of the target valuation domain, has standardized thinking chain reasoning ability and stable task execution ability, and can be directly applied to actual asset valuation and other vertical domain business scenarios.

[0105] Optionally, step A4, which involves performing data distillation on the first corpus dataset using a preset large model to obtain a second corpus dataset, includes: B1. The first corpus dataset is distilled using the preset large model to obtain a third corpus dataset; the third corpus dataset includes a corpus data; a is a positive integer greater than 1; B2. Perform data quality scoring on the a corpus data to obtain a data quality scores; each data quality score corresponds to one corpus data. B3. Based on the a data quality scores and the a corpus data, construct a thought chains; each thought chain corresponds to one corpus data. B4. Based on the a thought chains and the a corpus data, determine the second corpus dataset.

[0106] In this embodiment of the application, the first corpus dataset can be input into a preset large model, which will then perform data distillation on the corpus data to filter out noise, redundancy, and ambiguous content, retain high-value and high-confidence information, and finally output a "more professional, more standardized, and cleaner" corpus dataset, namely the third corpus dataset.

[0107] In some embodiments, the EasyDistill tool can also be used to perform data distillation on the first corpus dataset. The specific process is as follows: the first corpus dataset is input into the EasyDistill tool, which performs noise filtering, redundancy removal, semantic enhancement, and confidence screening on the corpus data based on the semantic features and probability distribution of the preset large model, retaining effective corpus data with high value, high standardization, and high domain adaptability, and outputting a purified and optimized third corpus dataset.

[0108] Next, the data quality of the 'a' corpus data is scored, resulting in 'a' data quality scores. Specifically, the method for obtaining the 'a' data quality scores can be the same as the method for obtaining the 'e' data quality scores mentioned above, and will not be repeated here. Then, based on the 'a' data quality scores and the 'a' corpus data, 'a' thought chains can be constructed. Finally, based on the 'a' thought chains and the 'a' corpus data, the second corpus dataset can be determined. Specifically, each corpus data can be associated and integrated with its corresponding thought chain to supplement the reasoning logic, steps, and basis, so that the corpus has both content and reasoning process. Then, the integrated data is standardized, deduplicated, and quality-screened to retain corpora with complete structure, clear reasoning, and high confidence, ultimately forming the second corpus dataset.

[0109] In this way, by distilling and purifying the corpus through a large model, then screening it with quality scores, constructing corresponding thought chains based on the scores, and finally merging the corpus and thought chains into a training set, the professionalism, purity, and reasoning integrity of the corpus can be significantly improved. This allows the model to learn high-quality content and complete evaluation logic, making the final trained model more rigorous in reasoning, more reliable in results, and more adaptable to the domain.

[0110] Optionally, each data quality score includes: redundancy score and complexity score; step B3, constructing a thought chains based on the a data quality scores and the a corpus data, includes: C1. Obtain the first corpus data and its corresponding first data quality score; the first corpus data is any corpus data among the a corpus data; the first data quality score includes: a first redundancy score and a first complexity score; C2. Determine the first comprehensive score based on the first redundancy score and the first complexity score; C3. Determine the first thought chain type corresponding to the first comprehensive score; C4. Determine the thought chain corresponding to the first corpus data based on the first thought chain type and the first corpus data.

[0111] In this embodiment of the application, the first thought chain type includes one of the following: short chain, medium chain, long chain, etc., which are not limited here.

[0112] In a specific embodiment, a first corpus data and its corresponding first data quality score can be obtained; then, a first comprehensive score can be determined based on the first redundancy score and the first complexity score. Specifically, the average value of the first redundancy score and the first complexity score can be calculated and used as the first comprehensive score.

[0113] In some embodiments, a first data length corresponding to the first corpus data can be determined. Specifically, character or word segmentation statistics can be performed on the first corpus data, and after removing meaningless symbols, whitespace, and other irrelevant content, the number of valid characters (or the number of valid words) can be counted, and the statistical result can be used as the first data length. A reference redundancy score and a reference complexity score corresponding to the first data length can be determined. Specifically, a pre-stored mapping relationship between preset data lengths and redundancy and complexity scores can be used to determine the reference redundancy and reference complexity scores corresponding to the first data length. Then, based on the first redundancy score and the reference redundancy score, a first ratio is determined, as follows: First ratio = First redundancy score / Reference redundancy score; The second ratio is determined based on the first complexity score and the reference complexity score, as follows: First ratio = First complexity score / Reference complexity score; Next, based on the first ratio and the second ratio, determine the first weight corresponding to the first redundancy score and the second weight corresponding to the first complexity score; wherein the sum of the first weight and the second weight is 1; specifically, the first ratio and the second ratio can be normalized to obtain the first normalized value and the second normalized value respectively; the first normalized value is used as the first weight, and the second normalized value is used as the second weight. For example, assuming the first ratio is 2 and the second ratio is 0.5, then the first ratio is normalized to obtain the first normalized value = 2 / (2+0.5) = 0.8; the second ratio is normalized to obtain the second normalized value = 0.5 / (2+0.5) = 0.2; then, the first comprehensive score can be obtained by weighting the first redundancy score, the first complexity score, the first weight, and the second weight.

[0114] Then, the first thought chain type corresponding to the first comprehensive score can be determined. Specifically, a pre-stored mapping relationship between comprehensive scores and thought chain types can be used to determine the first thought chain type corresponding to the first comprehensive score. Finally, the thought chain corresponding to the first corpus data can be determined based on the first thought chain type and the first corpus data. Specifically, the first thought chain type (short chain, medium chain, or long chain) can be identified first, and then, based on the evaluation object, core indicators, reasoning logic, and conclusions in the first corpus data, the thought chain can be determined according to the step length, reasoning granularity, and structural norms of the corresponding type. Deconstruction and organization: For short chains, extract core judgments and key evidence to form concise reasoning steps; for medium chains, supplement intermediate logical connections and indicator mapping relationships; for long chains, fully unfold the entire process from condition analysis, indicator calculation, logical deduction to conclusion output, ultimately generating a complete thought chain that matches the type and corresponds to the corpus. For example, if the first thought chain type is a short chain and the first corpus data is "a certain game has 500,000 daily active users and monthly sales of items of 8 million yuan, and the data asset value is relatively high", then the thought chain corresponding to the first corpus data can be: first extract key indicators, then determine the asset value.

[0115] In this way, by calculating a comprehensive score based on the redundancy and complexity of the corpus, and automatically matching the corresponding length of the thought chain type, and then generating an appropriate thought chain, the length of the thought chain can be accurately matched with the quality and complexity of the corpus. This avoids the waste of resources caused by pairing simple corpus with long chains, and also prevents insufficient reasoning caused by pairing complex corpus with short chains. Ultimately, this makes the training data structure more reasonable, the generalization stronger, and the training efficiency and model performance better.

[0116] Optionally, step A5, training the preset base model based on the second corpus dataset to obtain the target data asset evaluation model, includes: D1. Obtain the target loss function and target subset size level corresponding to the preset base model; D2. Divide the second corpus dataset according to the target subset size level to obtain b training subsets and a target test set; b is an integer greater than 1; D3. Based on the first preset formula and the preset indicator set, determine c indicator types; c is a positive integer; D4. Based on the c indicator types, determine b indicator sets corresponding to the b training subsets; each indicator set includes c indicator values, each indicator value corresponds to an indicator type; each indicator set corresponds to a training subset. D5. Determine the weights of c indicators according to the second preset formula; each indicator weight corresponds to an indicator type. D6. Determine the values ​​of b comprehensive indicators based on the third preset formula, the weights of the c indicators, and the set of b indicators; D7. Determine the target training subset based on the b comprehensive index values ​​and the b training subsets; D8. Using the target training subset and the target test set, train the preset base model to obtain the target data asset evaluation model.

[0117] In this embodiment of the application, the preset indicator set can be preset in advance or defaulted; the indicator type can include one of the following: redundancy indicator, complexity indicator, accuracy indicator, information density indicator, diversity indicator, noise indicator, etc., which are not limited here.

[0118] In a specific embodiment, a pre-stored mapping relationship between a preset base model, loss function, and subset size level can be used to determine the target loss function and target subset size level corresponding to the preset base model. Then, the second corpus dataset can be divided according to the target subset size level to obtain b training subsets and a target test set. For example, assuming the target subset size level is 1000 to 3000 samples, the total number of samples in the second corpus dataset is counted. According to the upper and lower limits of the target subset size level, the second corpus dataset is divided into multiple training subsets of equal (or approximate) size, so that the number of samples in each training subset falls within the target subset size level range. In addition, after dividing b training subsets from the second corpus dataset, the remaining samples can be used as the target test set, or data of the target subset size level can be extracted from the remaining samples as the target test set. Finally, b training subsets and one target test set are obtained.

[0119] Next, based on the first preset formula and the preset indicator set, c indicator types can be determined, where the first preset formula is as follows: ; in, This represents the optimal combination of indicators after filtering (i.e., c indicator types). This represents a preset set of indicators; This represents a pre-defined indicator evaluation function; Indicates the selection that makes The first preset formula is to automatically select the subset of indicators that best represents the data quality from all candidate indicators, and eliminate invalid or weakly correlated indicators. Then, based on c indicator types, b indicator sets corresponding to b training subsets can be determined. Specifically, for each training subset, the indicator value corresponding to each indicator type is calculated to obtain c indicator values, which is one indicator set. For example, assuming an indicator type is information density, the total effective character length and total effective information content of all corpus data in the training subset can be counted. Then, the information density indicator value corresponding to the training subset is obtained by dividing the effective information content by the total effective character length. In this way, b indicator sets can be obtained.

[0120] Furthermore, the weights of c indicators can be determined according to the second preset formula, which is as follows:

[0121] in, This represents the optimal weight combination (i.e., the weights of c indicators); This represents the weight combination to be solved; This represents the preset weight optimization objective function; Indicates the selection that makes The weight combination that achieves the minimum value, in other words, the role of the second preset formula is to automatically calculate a set of optimal weight combinations that meet the evaluation target through optimization algorithms, rather than setting them manually based on experience.

[0122] Next, based on the third preset formula, the weights of c indicators, and the set of b indicators, the values ​​of b comprehensive indicators can be determined. The third preset formula is as follows:

[0123] in, Indicator set The corresponding comprehensive index value; Indicator set The j-th index value; This represents the j-th weight of the j-th indicator value in the c-th indicator weights; according to the third preset formula, b calculations are performed to obtain b comprehensive indicator values.

[0124] Then, based on b comprehensive index values ​​and b training subsets, the target training subset can be determined; finally, the target training subset and the target test set can be used to train the preset base model to obtain the target data asset assessment model, as follows: First, supervised fine-tuning is performed: a subset of the target training set is used as the supervised fine-tuning dataset and input into a pre-defined base model to fine-tune the model using supervised learning. During training, the difference between the model output and the labeled results is calculated based on the target loss function, and the model parameters are continuously adjusted according to the loss function value, enabling the model to gradually learn the relevant rules and features of data asset assessment. Simultaneously, the performance of the supervised fine-tuned model is validated using the target test set to evaluate the model's effectiveness in terms of assessment accuracy and stability. Secondly, reinforcement learning fine-tuning is performed: a portion of the data is selected from the supervised fine-tuning dataset, and a reward dataset is constructed using other models (e.g., a pre-set large model). Based on this reward dataset, reinforcement learning fine-tuning is further performed on the supervised fine-tuned model. During multiple rounds of iterative optimization, the model continuously adjusts its strategy and parameters according to the reward signal, thereby continuously improving the evaluation accuracy and generalization ability. Through progressive optimization involving supervised fine-tuning and reinforcement learning fine-tuning, and validation using the target test set, a fully trained target data asset evaluation model with performance meeting requirements is finally obtained.

[0125] In this way, by dividing the dataset according to its size level, automatically selecting the optimal indicators, adaptively learning the weights and aggregating them to obtain comprehensive indicators, the training subset can be accurately selected. Low-quality and low-discrimination data can be eliminated, making the training set more concise, of higher quality and more stable in distribution.

[0126] In addition, by using supervised fine-tuning and reinforcement learning to progressively optimize the model, training efficiency can be improved, noise and overfitting risks can be reduced, and the accuracy, generalization ability and output stability of the model in data asset evaluation tasks can be significantly enhanced, ultimately resulting in a specialized model that is adapted to vertical domains and has better performance.

[0127] Optionally, step D7, determining the target training subset based on the b comprehensive index values ​​and the b training subsets, includes: E1. Determine the comprehensive index value that is greater than or equal to the preset index value among the b comprehensive index values ​​to obtain d comprehensive index values; d is a positive integer less than or equal to b; E2. Determine the target training subset based on the d comprehensive index values ​​and the b training subsets.

[0128] In this embodiment of the application, the preset index value can be preset in advance or defaulted.

[0129] In a specific embodiment, b comprehensive index values ​​can be compared with preset index values ​​to obtain all comprehensive index values ​​greater than or equal to the preset index values, i.e., d comprehensive index values. Then, based on the d comprehensive index values ​​and b training subsets, a target training subset can be determined. Specifically, the maximum value among the d comprehensive index values ​​can be determined, and the training subset corresponding to the maximum value in the b training subsets can be used as the target training subset. Alternatively, the d comprehensive index values ​​can be sorted from largest to smallest to obtain a first index sequence. The k training subsets corresponding to the top k comprehensive index values ​​are selected and merged into a larger and more stable target training subset, where k is a positive integer less than d.

[0130] It should be explained that if all b comprehensive index values ​​are less than the preset index values, it indicates that the quality of the current training subsets does not meet the requirements, and new training subsets can be acquired or generated again to ensure the data quality of subsequent model training.

[0131] In this way, by selecting effective comprehensive index values ​​that are not lower than the preset index values, and then determining the target training subset from them, the training subsets that are substandard or have high noise can be eliminated, ensuring that the training data selected subsequently all meet the basic quality requirements, effectively reducing the interference of low-quality data on model training, improving training stability and final model accuracy, while reducing invalid calculations and improving overall training efficiency.

[0132] S204. Obtain the asset data of the target object to be evaluated.

[0133] In this application embodiment, the target object can be one of the following: an individual, a school, a company, a hospital, etc., without limitation.

[0134] In a specific embodiment, an asset appraisal request initiated by the target object can be received. Based on the asset appraisal request, data on the assets to be appraised can be obtained. For example, basic asset information, financial data, operational indicators, risk characteristics, and related text data matching the target object can be extracted from public databases, business platforms, or third-party data sources to obtain the asset data to be appraised. It should be explained that the asset data to be evaluated can also be manually uploaded by the target entity. For example, the target entity can manually enter and upload local asset documents, table data or relevant supporting materials through the front-end interactive interface of an electronic device. After the system receives and parses the data, it will be cleaned and standardized to obtain the asset data to be evaluated.

[0135] S205. The target data asset valuation model is used to process the data of the asset to be valued to obtain a target asset valuation report.

[0136] In this embodiment of the application, the data of the asset to be evaluated is input into the target data asset valuation model to obtain the target asset valuation report. The target asset valuation report may include a number of valuation-related information, such as basic asset information, key indicator scores, risk level, comprehensive valuation results (e.g., asset cost, asset price, potential income, etc.), valuation basis explanation and suggestions, etc., which are not limited here.

[0137] In summary, the asset valuation method based on a large natural language model described in this application constructs a first multi-dimensional corpus with multiple sources and types for asset valuation, preprocesses it to form a high-quality second multi-dimensional corpus, and then trains a pre-set base model based on the second multi-dimensional corpus to obtain a target data asset valuation model. This valuation model is then used for intelligent processing of the asset data to be valued throughout the entire process, standardizing the process, unifying the logic, and ensuring accurate valuation. Compared to existing technologies that rely on manual interpretation and subjective judgment, this method has the advantages of a more standardized process, more objective logic, more accurate data mining, and no experience bias, thereby improving the reliability of the asset valuation results.

[0138] Please see Figure 7 , Figure 7 This is a functional block diagram of an asset appraisal device based on a natural language large model provided in this application embodiment. The asset appraisal device 700 based on a natural language large model includes: an acquisition module 701, a model training module 702, and an asset appraisal module 703, wherein: The acquisition module 701 is used to acquire a first multi-dimensional corpus and a preset base model; the preset base model is a pre-trained large natural language model; the first multi-dimensional corpus is a collection of multi-source and multi-type corpora constructed for asset appraisal tasks, including text corpora, structured corpora and semi-structured corpora. The model training module 702 is used to preprocess the first multi-dimensional corpus to obtain a second multi-dimensional corpus; and to train the preset base model based on the second multi-dimensional corpus to obtain a target data asset evaluation model. The acquisition module 701 is also used to acquire the asset data to be evaluated of the target object; The asset valuation module 703 is used to process the asset data to be valued using the target data asset valuation model to obtain a target asset valuation report.

[0139] Optionally, in the step of training the preset base model based on the second multi-dimensional corpus to obtain the target data asset assessment model, the model training module 702 is specifically used for: Obtain the target evaluation domain and target evaluation requirements corresponding to the target object; Based on the target evaluation domain and the target evaluation requirements, determine the prompt word data; The first corpus dataset is generated by the first preset corpus tool based on the prompt word data and the second multi-dimensional corpus. The first corpus dataset is distilled using a pre-defined large model to obtain the second corpus dataset; the pre-defined large model is a complete natural language large model that has not been simplified or optimized. The preset base model is trained based on the second corpus dataset to obtain the target data asset evaluation model.

[0140] Optionally, in the step of performing data distillation on the first corpus dataset using a preset large model to obtain the second corpus dataset, the model training module 702 is specifically used for: The first corpus dataset is distilled using the preset large model to obtain a third corpus dataset; the third corpus dataset includes a corpus data points; a is a positive integer greater than 1; Data quality scores are assigned to the a corpus data, resulting in a data quality scores; each data quality score corresponds to one corpus data. Based on the a data quality scores and the a corpus data, construct a thought chains; each thought chain corresponds to one corpus data. Based on the a thought chains and the a corpus data, the second corpus dataset is determined.

[0141] Optionally, each data quality score includes: redundancy score and complexity score; in the aspect of constructing a thought chains based on the a data quality scores and the a corpus data, the model training module 702 is specifically used for: Obtain the first corpus data and its corresponding first data quality score; the first corpus data is any corpus data from the a corpus data; the first data quality score includes: a first redundancy score and a first complexity score; A first comprehensive score is determined based on the first redundancy score and the first complexity score; Determine the first thought chain type corresponding to the first comprehensive score; Based on the first thought chain type and the first corpus data, determine the thought chain corresponding to the first corpus data.

[0142] Optionally, in the step of training the preset base model based on the second corpus dataset to obtain the target data asset evaluation model, the model training module 702 is specifically used for: Obtain the target loss function and target subset size level corresponding to the preset base model; The second corpus dataset is divided according to the target subset size level to obtain b training subsets and a target test set; b is an integer greater than 1. Based on the first preset formula and the preset indicator set, determine c indicator types; c is a positive integer; Based on the c indicator types, determine b indicator sets corresponding to the b training subsets; each indicator set includes c indicator values, each indicator value corresponds to an indicator type; each indicator set corresponds to a training subset. The weights of c indicators are determined according to the second preset formula; each indicator weight corresponds to an indicator type. Based on the third preset formula, the weights of the c indicators, and the set of b indicators, determine the values ​​of the b comprehensive indicators; Based on the b comprehensive index values ​​and the b training subsets, determine the target training subset; The target data asset evaluation model is obtained by training the preset base model using the target training subset and the target test set.

[0143] Optionally, in determining the target training subset based on the b comprehensive index values ​​and the b training subsets, the model training module 702 is specifically used for: Determine the comprehensive index values ​​that are greater than or equal to the preset index values ​​among the b comprehensive index values ​​to obtain d comprehensive index values; d is a positive integer less than or equal to b. The target training subset is determined based on the d comprehensive index values ​​and the b training subsets.

[0144] Optionally, in the process of preprocessing the first multi-dimensional corpus to obtain the second multi-dimensional corpus, the model training module 702 is specifically used for: Determine the corpus file format corresponding to the first multi-dimensional corpus; The target segmentation method is determined based on the format of the corpus file; The corpus data in the first multi-dimensional corpus is divided into blocks according to the target block segmentation method to obtain the first block corpus dataset; The first segmented corpus dataset is cleaned to obtain the second segmented corpus dataset; the second segmented corpus dataset includes e segmented corpus data; e is a positive integer greater than 1; Data quality scores are obtained by scoring each of the e segments of the corpus data; The second multi-dimensional corpus is determined based on the e data quality scores and the second segmented corpus dataset.

[0145] In specific implementations, the asset appraisal device 700 based on a natural language large model described in the embodiments of the present invention can also execute other implementations described in the asset appraisal method based on a natural language large model provided in the embodiments of the present invention, which will not be repeated here.

[0146] Please see Figure 8 , Figure 8 This is a schematic diagram of another electronic device provided in an embodiment of this application. The electronic device may include a processor, a memory, a communication interface, and one or more programs. The processor, memory, and communication interface can be interconnected via a bus. The one or more programs are stored in the memory and configured to be executed by the processor. In this embodiment, the programs include instructions for performing some or all of the steps in the method embodiments described above.

[0147] This application also provides a computer-readable storage medium storing a computer program for electronic data interchange, which causes a computer to perform some or all of the steps of any of the methods described in the above method embodiments, wherein the computer includes an electronic device.

[0148] This application also provides a computer program product, which includes a non-transitory computer-readable storage medium storing a computer program operable to cause a computer to perform some or all of the steps of any of the methods described in the above method embodiments. The computer program product may be a software installation package, and the computer may include an electronic device.

[0149] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0150] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0151] In the several embodiments provided in this application, it should be understood that the disclosed apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical or other forms.

[0152] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as ROM or random access memory (RAM), magnetic disks, or optical disks.

[0153] The steps of the methods or algorithms described in the embodiments of this application can be implemented in hardware or by a processor executing software instructions. The software instructions can consist of corresponding software modules, which can be stored in RAM, flash memory, ROM, EPROM, electrically erasable programmable read-only memory (EEPROM), registers, hard disk, portable hard disk, read-only optical disk (CD-ROM), or any other form of storage medium well known in the art. An exemplary storage medium is coupled to a processor, enabling the processor to read information from and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and storage medium can reside in an ASIC. Furthermore, the ASIC can reside in a terminal device or management device. Alternatively, the processor and storage medium can exist as discrete components in the terminal device or management device.

[0154] Those skilled in the art will recognize that, in one or more of the examples above, the functions described in the embodiments of this application can be implemented, in whole or in part, by software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, in the form of a computer program product. This computer program product includes one or more computer instructions. When these computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated.

[0155] The aforementioned computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media.

[0156] The available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., digital video discs (DVDs)), or semiconductor media (e.g., solid-state disks (SSDs)).

[0157] The modules / units included in the various devices and products described in the above embodiments can be software modules / units, hardware modules / units, or a combination of both. For example, for devices and products applied to or integrated into a chip, all modules / units can be implemented using hardware methods such as circuits, or at least some modules / units can be implemented using software programs that run on a processor integrated within the chip, while the remaining (if any) modules / units can be implemented using hardware methods such as circuits. For devices and products applied to or integrated into a chip module, all modules / units can be implemented using hardware methods such as circuits. Different modules / units can be located in the same component (e.g., chip, circuit module, etc.) or different components of the chip module, or at least some modules / units can be implemented using hardware methods such as circuits. The implementation is achieved through a software program that runs on the processor integrated within the chip module. The remaining modules / units (if any) can be implemented using hardware methods such as circuits. For various devices and products applied to or integrated into terminal equipment, each of their modules / units can be implemented using hardware methods such as circuits. Different modules / units can be located in the same component (e.g., chip, circuit module, etc.) or different components within the terminal equipment. Alternatively, at least some modules / units can be implemented through a software program that runs on the processor integrated within the terminal equipment, while the remaining modules / units (if any) can be implemented using hardware methods such as circuits.

[0158] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the embodiments of this application. It should be understood that the above descriptions are merely specific embodiments of the embodiments of this application and are not intended to limit the protection scope of the embodiments of this application. Any modifications, equivalent substitutions, improvements, etc., made on the basis of the technical solutions of the embodiments of this application should be included within the protection scope of the embodiments of this application.

Claims

1. An asset valuation method based on a large natural language model, characterized in that, include: Obtain the first multi-dimensional corpus and the pre-set base model; The preset base model is a pre-trained large natural language model; The first multi-dimensional corpus is a collection of multi-source, multi-type corpora built for asset valuation tasks, including text corpora, structured corpora, and semi-structured corpora; The first multi-dimensional corpus is preprocessed to obtain the second multi-dimensional corpus; The preset base model is trained based on the second multi-dimensional corpus to obtain the target data asset evaluation model; Obtain the asset data of the target entity to be evaluated; The target asset valuation model is used to process the data of the asset to be valued to obtain a target asset valuation report.

2. The method as described in claim 1, characterized in that, The step of training the preset base model based on the second multi-dimensional corpus to obtain the target data asset evaluation model includes: Obtain the target evaluation domain and target evaluation requirements corresponding to the target object; Based on the target evaluation domain and the target evaluation requirements, determine the prompt word data; The first corpus dataset is generated by the first preset corpus tool based on the prompt word data and the second multi-dimensional corpus. The first corpus dataset is distilled using a pre-defined large model to obtain the second corpus dataset; the pre-defined large model is a complete natural language large model that has not been simplified or optimized. The preset base model is trained based on the second corpus dataset to obtain the target data asset evaluation model.

3. The method as described in claim 2, characterized in that, The process of distilling the first corpus dataset using a pre-set large model to obtain the second corpus dataset includes: The first corpus dataset is distilled using the preset large model to obtain a third corpus dataset; the third corpus dataset includes a corpus data points; a is a positive integer greater than 1; Data quality scores are assigned to the a corpus data, resulting in a data quality scores; each data quality score corresponds to one corpus data. Based on the a data quality scores and the a corpus data, construct a thought chains; each thought chain corresponds to one corpus data. Based on the a thought chains and the a corpus data, the second corpus dataset is determined.

4. The method as described in claim 3, characterized in that, Each data quality score includes: redundancy score and complexity score; The step of constructing a thought chains based on the a data quality scores and the a corpus data includes: Obtain the first corpus data and its corresponding first data quality score; the first corpus data is any corpus data from the a corpus data; the first data quality score includes: a first redundancy score and a first complexity score; A first comprehensive score is determined based on the first redundancy score and the first complexity score; Determine the first thought chain type corresponding to the first comprehensive score; Based on the first thought chain type and the first corpus data, determine the thought chain corresponding to the first corpus data.

5. The method according to any one of claims 2-4, characterized in that, The step of training the preset base model based on the second corpus dataset to obtain the target data asset evaluation model includes: Obtain the target loss function and target subset size level corresponding to the preset base model; The second corpus dataset is divided according to the target subset size level to obtain b training subsets and a target test set; b is an integer greater than 1. Based on the first preset formula and the preset indicator set, determine c indicator types; c is a positive integer; Based on the c indicator types, determine b indicator sets corresponding to the b training subsets; each indicator set includes c indicator values, each indicator value corresponds to an indicator type; each indicator set corresponds to a training subset. The weights of c indicators are determined according to the second preset formula; each indicator weight corresponds to an indicator type. Based on the third preset formula, the weights of the c indicators, and the set of b indicators, determine the values ​​of the b comprehensive indicators; Based on the b comprehensive index values ​​and the b training subsets, determine the target training subset; The target data asset evaluation model is obtained by training the preset base model using the target training subset and the target test set.

6. The method as described in claim 5, characterized in that, The step of determining the target training subset based on the b comprehensive index values ​​and the b training subsets includes: Determine the comprehensive index values ​​that are greater than or equal to the preset index values ​​among the b comprehensive index values ​​to obtain d comprehensive index values; d is a positive integer less than or equal to b. The target training subset is determined based on the d comprehensive index values ​​and the b training subsets.

7. The method according to any one of claims 1-4, characterized in that, The step of preprocessing the first multi-dimensional corpus to obtain the second multi-dimensional corpus includes: Determine the corpus file format corresponding to the first multi-dimensional corpus; The target segmentation method is determined based on the format of the corpus file; The corpus data in the first multi-dimensional corpus is divided into blocks according to the target block segmentation method to obtain the first block corpus dataset; The first segmented corpus dataset is cleaned to obtain the second segmented corpus dataset; the second segmented corpus dataset includes e segmented corpus data; e is a positive integer greater than 1; Data quality scores are obtained by scoring each of the e segments of the corpus data; The second multi-dimensional corpus is determined based on the e data quality scores and the second segmented corpus dataset.

8. An asset appraisal device based on a large natural language model, characterized in that, The device includes: an acquisition module, a model training module, and an asset evaluation module, wherein: The acquisition module is used to acquire a first multi-dimensional corpus and a preset base model; the preset base model is a pre-trained large natural language model; the first multi-dimensional corpus is a collection of multi-source and multi-type corpora constructed for asset appraisal tasks, including text corpora, structured corpora and semi-structured corpora; The model training module is used to preprocess the first multi-dimensional corpus to obtain a second multi-dimensional corpus; and to train the preset base model based on the second multi-dimensional corpus to obtain a target data asset evaluation model. The acquisition module is also used to acquire the asset data to be evaluated of the target object; The asset valuation module is used to process the asset data to be valued using the target data asset valuation model to obtain a target asset valuation report.

9. An electronic device, characterized in that, include: Processor, memory, communication interface, and one or more programs; The one or more programs are stored in the memory and configured to be executed by the processor, the programs including instructions for performing the steps of the method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, A computer program for storing electronic data interchange, wherein the computer program causes a computer to perform the method as described in any one of claims 1-7.

Images