An old-age vertical field large model construction method, system, device and storage medium

Abstract

The application discloses a pension vertical field large model construction method, system and device and a storage medium, relates to the cross technical field of artificial intelligence and pension service, and comprises the following steps: collecting pension field data, processing the pension field data, constructing a pre-training data set, a supervision fine-tuning data set and a preference data set according to the pension field data, taking the pre-training data set, the supervision fine-tuning data set and the preference data set as input, constructing a preliminary pension vertical field large model, screening the pension field data, constructing a pension scene structured knowledge base according to the pension field data, and fusing the pension scene structured knowledge base with the preliminary pension vertical field large model to obtain the pension vertical field large model. The method disclosed by the application achieves better effects in the following aspects: the standardization degree of data set construction, the consistency of professional and safety constraints of pension scene answers, and the scene adaptation ability under the enhancement of knowledge base retrieval.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of artificial intelligence and elderly care services, specifically to a method, system, device, and storage medium for constructing a large model in the vertical field of elderly care. Background Technology

[0002] In their daily lives, the elderly not only face basic needs such as physiological health management and safety protection, but also prominent pain points such as lack of emotional companionship, inadequate access to health knowledge, and insufficient personalized life guidance.

[0003] Artificial intelligence technology has made groundbreaking progress, and its capabilities in general fields such as intelligent interaction, information retrieval, and personalized recommendation have been widely verified, providing technical support for solving service adaptation problems in complex scenarios.

[0004] There is still a significant gap in the application of existing artificial intelligence technology in the field of elderly care: On the one hand, general-purpose models lack deep adaptation to vertical scenarios of elderly care, making it difficult to accurately understand the language habits, physiological and psychological characteristics, and professional elderly care knowledge of the elderly. On the other hand, existing intelligent products in the elderly care field mostly focus on single functions and lack comprehensive intelligent solutions that can cover multiple scenarios such as emotional companionship, health management, life assistance, and knowledge popularization, thus failing to meet the multi-dimensional and personalized elderly care needs of seniors.

[0005] Against this backdrop, developing a dedicated large-scale model for the vertical field of elderly care, by integrating professional elderly care knowledge systems, adapting to the interaction habits of the elderly, and optimizing multi-scenario service logic, can achieve functions such as providing emotional companionship for lonely elderly people, offering customized health and lifestyle advice for the elderly, assisting in solving daily life questions, and popularizing knowledge about health protection for the elderly. This can effectively fill the gap in the integration of artificial intelligence and elderly care services and improve the intelligence and personalization of elderly care services. Summary of the Invention

[0006] In view of the above-mentioned problems, the present invention is proposed.

[0007] Therefore, the technical problem solved by this invention is: existing general large-scale model training and application methods for elderly care scenarios suffer from the following problems: the data sources in the elderly care field are scattered and the formats are heterogeneous, making them difficult to use directly for training; the responses are not professional; and the lack of preference alignment and knowledge base integration leads to unstable responses. The problem is how to conduct step-by-step training on a general basic model with a parameter size greater than 0B and less than 10B, based on a pre-trained dataset, a supervised fine-tuning dataset, and a preference dataset, and integrate it with a structured knowledge base for elderly care scenarios to construct a large-scale model for the vertical field of elderly care.

[0008] To address the aforementioned technical problems, this invention provides the following technical solution: a method for constructing a large model in the vertical field of elderly care, comprising collecting and processing data in the elderly care field, and constructing a pre-training dataset, a supervised fine-tuning dataset, and a preference dataset based on the data in the elderly care field.

[0009] Using pre-trained datasets, supervised fine-tuning datasets, and preference datasets as inputs, we construct a preliminary large-scale model for the elderly care vertical field.

[0010] Data in the elderly care field is screened, and a structured knowledge base for elderly care scenarios is constructed based on this data. This knowledge base is then integrated with a preliminary large-scale model for the elderly care vertical field to obtain a large-scale model for the elderly care vertical field.

[0011] The construction of the supervised fine-tuning dataset and the preference dataset involves taking the pre-trained dataset as input, filtering the supervised fine-tuning data and preference data, and constructing the supervised fine-tuning dataset and the preference dataset.

[0012] The initial construction of a large-scale model for the elderly care vertical field includes selecting a general basic model with more than 0B and less than 10B parameters, and training the general basic model step by step through a pre-training dataset, a supervised fine-tuning dataset, and a preference dataset.

[0013] As a preferred embodiment of the method for constructing a large model in the vertical field of elderly care described in this invention, the step of collecting and processing data in the field of elderly care includes collecting data from academic databases whose keywords are related to the field of elderly care.

[0014] Standardize the format of data in the elderly care field through format unification.

[0015] The data in the elderly care field, after being standardized in format, was converted into Markdown format through data conversion.

[0016] Identify textual semantic breakpoints in elderly care data after format unification and format conversion.

[0017] The text is divided into blocks of 5,000-10,000 characters by splitting the text into blocks.

[0018] Remove redundant information from text blocks and output the pre-trained dataset.

[0019] As a preferred embodiment of the large-scale model construction method for the elderly care vertical field described in this invention, wherein: The construction of the supervised fine-tuning dataset includes selecting questions and standard solutions from the professional exercise set, matching data of basic knowledge of elderly care, contextualized questions and structured solutions containing reasoning processes from the pre-trained dataset.

[0020] The questions and standard answers from the professional exercise set, the matching data of basic knowledge of elderly care, the contextualized questions and the structured answers with reasoning processes are standardized and merged to output a supervised fine-tuning dataset.

[0021] As a preferred embodiment of the method for constructing a large model in the vertical field of elderly care described in this invention, the construction of the preference dataset includes setting the response priority and answer selection priority for elderly care scenarios.

[0022] From the pre-trained dataset, the actual needs of elderly care service scenarios are selected according to the response priority of elderly care scenarios.

[0023] Based on the answer selection priority, multiple answers for each item in the actual needs of elderly care service scenarios are reviewed and filtered, and a preference dataset is output.

[0024] The preference datasets are divided into two categories: reinforcement learning stages for direct preference optimization and reinforcement learning stages for cutting-edge policy optimization.

[0025] As a preferred embodiment of the method for constructing a large model in the vertical field of elderly care described in this invention, the step-by-step training of the general basic model includes introducing a reward and punishment mechanism into the basic model and combining a pre-training dataset, a supervised fine-tuning dataset, and a preference dataset to train the general large model step by step.

[0026] As a preferred embodiment of the large-scale model construction method for the elderly care vertical field described in this invention, wherein: The construction of a structured knowledge base for elderly care scenarios and its integration with a preliminary large-scale model for the elderly care vertical field includes text segmentation and embedding of data in the elderly care field to construct a structured knowledge base for elderly care scenarios.

[0027] Convert the knowledge base text blocks into vectors to build the vector library.

[0028] The structured knowledge base for elderly care scenarios is deeply integrated with the preliminary large-scale model of the elderly care vertical field.

[0029] As a preferred embodiment of the large-scale model construction method for the elderly care vertical field described in this invention, the deep fusion includes setting response logic, converting the question into a vector when a user initiates a query, and performing a similarity search in the vector library.

[0030] Return the first 5 to 10 matching text blocks as contextual prompts for inputting a large model in the elderly care vertical field.

[0031] When a user's query for relevant text content does not exist in the structured knowledge base built for the elderly care scenario, the large model in the elderly care vertical field will respond based on the training architecture.

[0032] Another objective of this invention is to provide a large-scale model construction system for the elderly care vertical domain. This system can pre-train and fine-tune a general basic model by taking a pre-trained dataset, a supervised fine-tuning dataset, and a preference dataset as inputs, combining a reward and punishment model, and employing a step-by-step training strategy with low-rank adaptation. Furthermore, it introduces a direct preference optimization algorithm and an advanced policy optimization algorithm to complete preference alignment training. This solves the problem that current technologies for directly applying general-scale models to elderly care scenarios lack support from dedicated datasets for the elderly care domain and have insufficient preference alignment mechanisms, resulting in difficulties in maintaining stable constraints on the professionalism and security of the output.

[0033] As a preferred embodiment of the large-scale model construction system for the elderly care vertical domain described in this invention, it includes: a module for constructing a pre-training dataset, a supervised fine-tuning dataset, and a preference dataset; a module for step-by-step training to construct a preliminary large-scale model for the elderly care vertical domain; a module for constructing a structured knowledge base for elderly care; and a module for model fusion.

[0034] The module for constructing the pre-trained dataset, supervised fine-tuning dataset, and preference dataset is used to collect and process data in the field of elderly care, and to construct the pre-trained dataset, supervised fine-tuning dataset, and preference dataset based on the data in the field of elderly care.

[0035] The step-by-step training module for building a preliminary large-scale model for the elderly care vertical domain is used to take the pre-trained dataset, the supervised fine-tuning dataset, and the preference dataset as inputs to build a preliminary large-scale model for the elderly care vertical domain.

[0036] The structured knowledge base for elderly care is constructed to filter data in the field of elderly care and to build a structured knowledge base for elderly care scenarios based on the data in the field of elderly care.

[0037] The model fusion module is used to merge the structured knowledge base of elderly care scenarios with the preliminary large model of the elderly care vertical field to obtain the large model of the elderly care vertical field.

[0038] Another object of the present invention is to provide a large-scale model building device for the elderly care vertical field, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the large-scale model building method for the elderly care vertical field.

[0039] Another object of the present invention is to provide a storage medium for constructing a large model in the field of elderly care, wherein a computer program is stored thereon, and when the computer program is executed by a processor, the steps of the method for constructing a large model in the field of elderly care are implemented.

[0040] The beneficial effects of this invention are as follows: The method for constructing a large-scale model in the vertical field of elderly care provided by this invention constructs a pre-training dataset by unifying the format of elderly care data, converting it to Markdown, identifying semantic breakpoints, and segmenting it. Furthermore, the pre-training dataset is filtered to form a supervised fine-tuning dataset and a preference dataset. Then, step-by-step training based on low-rank adaptation and preference alignment training are combined with a structured knowledge base for elderly care scenarios, thereby realizing the construction of a large-scale model in the vertical field of elderly care. This invention achieves better results in terms of the standardization of dataset construction, the consistency of professionalism and security constraints in elderly care scenario responses, and the scenario adaptability under enhanced knowledge base retrieval. Attached Figure Description

[0041] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1 This is an overall flowchart of a method for constructing a large model in the vertical field of elderly care, as provided in Embodiment 1 of the present invention.

[0043] Figure 2 The flowchart below shows the pre-training dataset, supervised fine-tuning dataset, and preference dataset modules of a large model construction method for the elderly care vertical field provided in Embodiment 1 of the present invention.

[0044] Figure 3 This is a flowchart illustrating the preliminary process of a large-scale model module for the elderly care vertical domain, as provided in Embodiment 1 of the present invention.

[0045] Figure 4 The flowchart below shows the structured knowledge base construction module for elderly care, which is part of a method for constructing a large model in the vertical field of elderly care provided in Embodiment 1 of the present invention.

[0046] Figure 5 This is a flowchart illustrating the model fusion module of a method for constructing a large model in the vertical field of elderly care, as provided in Embodiment 1 of the present invention.

[0047] Figure 6 This is an experimental flowchart of a method for constructing a large model in the vertical field of elderly care, as provided in Embodiment 2 of the present invention. Detailed Implementation

[0048] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0049] Example 1, referring to Figures 1-5 As an embodiment of the present invention, a method for constructing a large model in the vertical field of elderly care is provided, comprising: S1: Collect and process data in the field of elderly care, and construct a pre-trained dataset 101, a supervised fine-tuning dataset 102, and a preference dataset 103 based on the data in the field of elderly care.

[0050] like Figure 2 As shown, data with the keyword "elderly care" was collected from 100 academic databases.

[0051] The data in the elderly care field is standardized using Format Unification 1011.

[0052] The data in the elderly care field, after being standardized in format, was converted to Markdown format using Data Converter 1012.

[0053] Identify textual semantic breakpoints in elderly care data after format unification and format conversion.

[0054] The text is divided into blocks of 5000-10000 characters by splitting text block 1013.

[0055] Remove redundant information from the text blocks and output the pre-trained dataset 101.

[0056] The academic database 100 includes open online resources, CNKI, and Wanfang.

[0057] Format standardization includes converting non-TXT format files to PDF files.

[0058] Converting to Markdown format includes using MinerU software to convert the file to Markdown format.

[0059] Furthermore, problems and standard solutions from the professional problem set were selected from the pre-training dataset 101.

[0060] Generate paired data of basic knowledge about elderly care in input and output formats according to the structure of questions and answers.

[0061] We extract contextualized questions and structured solutions containing reasoning processes from real-world elderly care service cases.

[0062] The questions and standard answers in the professional exercise set, the matching data of basic knowledge of elderly care, the contextualized questions and the structured answers with reasoning process were standardized and merged.

[0063] Output supervised fine-tuning dataset 102.

[0064] Furthermore, we set the response priority 1031 and the answer filtering priority 1032 for elderly care scenarios.

[0065] Response priorities include health risk warnings, routine consultations, and emotional support.

[0066] The answer selection priority includes professionalism, security, and practicality.

[0067] From the pre-trained dataset 101, the actual needs of elderly care service scenarios 1033 are selected according to the response priority 1031 of the elderly care scenario.

[0068] Based on the answer selection priority of 1032, multiple answers for each item in the actual needs of elderly care service scenarios are reviewed and selected.

[0069] The review and screening process includes removing ambiguous, redundant, and age-friendly data, filtering multiple answers for each item in the actual needs of elderly care service scenarios according to priority, and outputting a preference dataset 103.

[0070] The preference dataset 103 is divided into two categories: the reinforcement learning stage of direct preference optimization and the reinforcement learning stage of cutting-edge policy optimization.

[0071] Using the preference dataset 103 for the reinforcement learning phase of direct preference optimization includes ranking the answers according to their selection priority.

[0072] Using the preference dataset 103 for the reinforcement learning phase of cutting-edge policy optimization includes ranking the answers according to their selection priority.

[0073] S2: Using the pre-trained dataset 101, the supervised fine-tuning dataset 102, and the preference dataset 103 as inputs, construct a preliminary large-scale model 201 for the elderly care vertical field.

[0074] The specific process of building a preliminary large-scale model for the elderly care vertical industry is as follows: Figure 3 As shown, a reward and punishment mechanism is introduced into the basic model, and the general large model is trained step by step by combining the pre-training dataset 101, the supervised fine-tuning dataset 102, and the preference dataset 103.

[0075] A preferred approach to introducing a reward and punishment mechanism is: A scoring mechanism was set based on preference dataset 103, and a comprehensive score for the trained reward and punishment model's answer was calculated. The scoring mechanism included: Professionalism: 5 points. It fully conforms to the professional logic and service standards of elderly care, with no errors and standardized professional expression.

[0076] A score of 4 is in line with the professional logic and service standards of elderly care, with no errors, only a few expressions are slightly inaccurate.

[0077] A score of 3 indicates that the service basically conforms to the professional logic and service standards of elderly care, has no core errors, and meets the qualification standard.

[0078] A score of 2 indicates a minor, non-core deviation that does not affect the core logic of elderly care services.

[0079] One point indicates a core error that violates the core logic of elderly care services.

[0080] Safety section: 5 points. It fully meets the safety requirements of the elderly care scenario, focuses on core safety points such as health risk warning, and has no non-safety related redundant information.

[0081] A score of 4 indicates that the safety requirements of the elderly care scenario are met, with core safety points presented first and only a few non-core details added at the end, which does not affect the safety assessment.

[0082] A score of 3 indicates that the core safety points are not omitted, but are only mentioned in conjunction with secondary information and are not fully presented in advance, thus meeting the passing standard.

[0083] The 2-point system places core security priorities at the end, with secondary information accounting for a high proportion, which affects the assessment of elderly care security risks.

[0084] 1 point is for failing to mention the core safety points in the elderly care scenario and containing only irrelevant content.

[0085] Practicality: 5 points. It fully covers the core dimensions of elderly care service issues, and the methods and suggestions are in line with the actual needs of elderly care scenarios, with no core information missing.

[0086] The score of 4 points covers the core dimensions of elderly care services, with only a few non-core details missing, which does not affect the actual implementation of services.

[0087] A score of 3 indicates coverage of core operations and recommendations for elderly care services, with only one non-core dimension missing, thus meeting the passing standard.

[0088] The score of 2 only covers some of the core content of elderly care services, omitting key implementation dimensions.

[0089] One point is irrelevant and does not cover any core content of elderly care services.

[0090] A preferred scheme for calculating the comprehensive score of training reward and punishment answers is as follows: , in: This represents the overall score for each answer. Indicating professionalism, Indicates security, To indicate practicality, the overall score is normalized, and the range of the overall score is as follows: .

[0091] The step-by-step training process involves using the pre-training dataset 101 as input and pre-training the base model using a low-rank adaptation method.

[0092] A preferred approach to pre-training is: Set the learning rate to 4e-5 to 5e-5, the number of training epochs to 3-10, and the batch size to 16-32.

[0093] The supervised fine-tuning dataset 102 was used as input to fine-tune the model using a low-rank adaptation method.

[0094] A preferred approach for fine-tuning the model is: Set the learning rate to 4e-5 to 5e-5, the number of training epochs to 3-10, and the batch size to 16-32.

[0095] Using preference dataset 103 as input, the algorithm is trained using both advanced policy optimization and direct preference optimization algorithms.

[0096] A preferred approach for training using both advanced policy optimization algorithms and direct preference optimization algorithms is as follows: Set the rank of the low-rank adaptation to 8-16, the learning rate to 4e-5 to 5e-5, the number of training rounds to 3-10, and the batch size to 16-32.

[0097] S3: Filter data in the elderly care field, construct a structured knowledge base 300 for elderly care scenarios based on the data, and integrate it with the preliminary large-scale model 201 for the elderly care vertical field to obtain the large-scale model for the elderly care vertical field.

[0098] The specific process of building a structured knowledge base for elderly care scenarios is as follows: Figure 4 As shown, data in the elderly care field is processed by text segmentation and embedding to construct a structured knowledge base of 300 for elderly care scenarios.

[0099] Convert the knowledge base text blocks into vectors to build the vector library 301.

[0100] The structured knowledge base 300 for elderly care scenarios is deeply integrated with the preliminary large-scale model 201 for the vertical field of elderly care. The specific process of model integration is as follows: Figure 5 As shown.

[0101] Specifically, the response logic is set so that when a user initiates a query, the question is converted into a vector, and a similarity search is performed in the vector library 301.

[0102] Return the first 5 to 10 matching text blocks as contextual prompts for inputting a large model in the elderly care vertical field.

[0103] When a user's query for relevant text content does not exist in the structured knowledge base 300 for elderly care scenarios, the large model for the elderly care vertical field will respond based on the training architecture.

[0104] Example 2, as Figure 6 As shown, this is an embodiment of the present invention, which provides a method for constructing a large model in the vertical field of elderly care. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0105] First, a total of 5,000 benchmark test questions in the field of elderly care were selected. These 5,000 questions were then divided into four categories: 1,500 questions for Content 1, 1,500 questions for Content 2, 1,000 questions for Content 3, and 1,000 questions for Content 4.

[0106] For Contents 1 and 2, the evaluation metric is accuracy. The model is considered correct if it answers the key points of the answer, and incorrect if it does not or only partially answers them. The accuracy of the model in Contents 1 and 2 of the benchmark test is obtained after statistical analysis of all questions.

[0107] Content 3 uses human scoring (1-5), where professionals in the field of elderly care score the reasoning process of the model's answers. The scoring criteria are as follows: 5 points: The reasoning process is completely consistent with the professional thinking of elderly care services. The logic is closed-loop and progressive. It accurately matches the background and actual needs of elderly care service cases. The reasoning basis complies with the professional norms and service standards of elderly care and has no deviation. 4 points: The reasoning process is in line with the professional thinking of elderly care services, the logic is clear, and the reasoning basis is in line with the professional norms of elderly care. Only a few details are slightly less precise, and there is no core logical deviation. 3 points: The reasoning process is basically in line with the professional thinking of elderly care services, the logic is smooth, there are no core errors, and the basic scenario analysis and service plan derivation can be completed, which meets the qualified standard; 2 points: The reasoning process has minor logical flaws and does not fully fit the core information of the elderly care case. Some of the reasoning basis has non-core deviations from the elderly care service standards, but does not affect the overall service decision-making direction. 1 point: The reasoning process is logically chaotic, contradicts the professional thinking of elderly care services, and contains core errors in the reasoning basis, making it impossible to complete effective scenario analysis and elderly care service decision-making.

[0108] Content 4 uses the BERTScore metric, which calculates the similarity score between candidate text and reference text based on the embedding method. The score ranges from 0 to 1, and is used to measure the semantic fit, contextual coherence, and content relevance of the model's responses in multi-turn dialogue scenarios for the elderly. The experimental data is shown in Table 1.

[0109] Table 1 Experimental Data Benchmarking Evaluation indicators Large Model of the Elderly Care Vertical Sector Qwen3 LLaMA ChatGLM DeepSeek ChatGPT Content 1 Accuracy (%) 95 94 87 94 94 91 Content 2 Accuracy (%) 96 91 83 95 96 92 Content 3 Human rating (1-5) 4.6 4.1 3.8 4.4 4.3 4.5 Content 4 BERTScore 0.80 0.79 0.78 0.8 0.76 0.81

[0110] As shown in Table 1, the large-scale model for the elderly care vertical field constructed in this invention performs excellently in 5,000 benchmark tests covering professional exercises, health care knowledge, real service cases, and multi-turn dialogues: the basic accuracy rate is stable at over 95%, the professional reasoning human score reaches 4.6 points, and the multi-turn dialogue BERTS score reaches 0.80. It comprehensively surpasses most mainstream general-purpose large-scale models in terms of accuracy, professional adaptability, and semantic coherence, and significantly makes up for the lack of professional depth of general-purpose models in the field of elderly care.

[0111] This model can provide accurate, safe, and age-friendly intelligent support for elderly care services, effectively reducing service thresholds and operating costs, improving the quality of elderly care services and their adaptability to various scenarios, and providing core technological support for the intelligent upgrading of the elderly care industry. It has outstanding practical value and significance for promotion.

[0112] Example 3, an embodiment of the present invention, provides a large-scale model construction system for the vertical field of elderly care, including a module E1 for constructing a pre-training dataset, a supervised fine-tuning dataset and a preference dataset, a module E2 for step-by-step training to construct a preliminary large-scale model for the vertical field of elderly care, a module E3 for constructing a structured knowledge base for elderly care, and a model fusion module E4.

[0113] The module E1 for constructing the pre-trained dataset, supervised fine-tuning dataset, and preference dataset is used to collect data in the field of elderly care and process the data in the field of elderly care, and to construct the pre-trained dataset 101, the supervised fine-tuning dataset 102, and the preference dataset 103 based on the data in the field of elderly care.

[0114] The step-by-step training constructs a preliminary large-scale model E2 for the elderly care vertical domain. This module takes the pre-training dataset 101, the supervised fine-tuning dataset 102, and the preference dataset 103 as inputs to construct a preliminary large-scale model 201 for the elderly care vertical domain.

[0115] The structured knowledge base for elderly care, E3, is used to filter data in the field of elderly care, and a structured knowledge base for elderly care scenarios, 300, is constructed based on the data in the field of elderly care.

[0116] The model fusion module E4 is used to fuse the structured knowledge base 300 of elderly care scenarios with the preliminary large model 201 of the elderly care vertical field to obtain the large model of the elderly care vertical field.

[0117] This embodiment also provides a computer device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the method for constructing a large model in the elderly care vertical field as proposed in the above embodiment.

[0118] This embodiment also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the method for constructing a large model in the elderly care vertical field as proposed in the above embodiment.

[0119] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0120] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0121] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0122] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0123] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for constructing a large-scale model in the vertical field of elderly care, characterized in that, include: Collect data in the field of elderly care and process the data. Construct a pre-trained dataset (101), a supervised fine-tuning dataset (102), and a preference dataset (103) based on the data in the field of elderly care. Using the pre-trained dataset (101), the supervised fine-tuning dataset (102), and the preference dataset (103) as inputs, a preliminary large model for the elderly care vertical field (201) is constructed. Data in the elderly care field is screened, and a structured knowledge base for elderly care scenarios (300) is constructed based on the data and integrated with the preliminary large-scale model of the elderly care vertical field (201) to obtain the large-scale model of the elderly care vertical field. The construction of the supervised fine-tuning dataset (102) and the preference dataset (103) includes taking the pre-training dataset (101) as input, filtering the supervised fine-tuning data and preference data, and constructing the supervised fine-tuning dataset (102) and the preference dataset (103). The construction of a preliminary large model for the vertical field of elderly care (201) includes selecting a general basic model (200) with a parameter count greater than 0B and less than 10B, and training the general basic model (200) step by step through a pre-training dataset (101), a supervised fine-tuning dataset (102), and a preference dataset (103).

2. The method for constructing a large-scale model in the elderly care vertical field as described in claim 1, characterized in that: The collection and processing of data in the elderly care field includes, Data on the topic of elderly care were collected from academic databases (100). Data in the elderly care field is formatted using a standardized format (1011). The data in the elderly care field, after being standardized in format, was converted to Markdown format using data conversion (1012); Identify textual semantic breakpoints in elderly care data after format unification and format conversion; The text is divided into blocks of 5000-10000 characters by splitting the text into blocks (1013); Remove redundant information from the text blocks and output the pre-trained dataset (101).

3. The method for constructing a large-scale model in the elderly care vertical field as described in claim 1 or 2, characterized in that: Constructing the supervised fine-tuning dataset (102) includes, Select questions and standard solutions from the professional exercise set, paired data of basic knowledge of elderly care, contextualized questions and structured solutions containing reasoning processes from the pre-trained dataset (101); The questions and standard answers in the professional exercise set, the matching data of basic knowledge of elderly care, the contextualized questions and the structured answers with reasoning process are standardized and merged to output a supervised fine-tuning dataset (102).

4. The method for constructing a large-scale model in the vertical field of elderly care as described in claim 3, characterized in that: Constructing the preference dataset (103) includes, Set the response priority (1031) and answer filtering priority (1032) for elderly care scenarios. From the pre-trained dataset (101), the actual needs of elderly care service scenarios (1033) are selected according to the response priority of elderly care scenarios (1031). According to the answer filtering priority (1032), multiple answers for each item in the actual needs of elderly care service scenarios are reviewed and filtered, and the preference dataset (103) is output. The preference dataset (103) is divided into two categories: a reinforcement learning stage for direct preference optimization and a reinforcement learning stage for cutting-edge policy optimization.

5. The method for constructing a large-scale model in the elderly care vertical field as described in claim 1, 2, or 4, characterized in that: The step-by-step training of the general base model (200) includes, A reward and punishment mechanism is introduced into the basic model, and the general large model is trained step by step by combining the pre-training dataset (101), the supervised fine-tuning dataset (102), and the preference dataset (103).

6. The method for constructing a large-scale model in the vertical field of elderly care as described in claim 5, characterized in that: The construction of a structured knowledge base for elderly care scenarios (300) and its integration with a preliminary large-scale model for the vertical field of elderly care (201) includes, Data in the elderly care field is processed by text segmentation and embedding to construct a structured knowledge base for elderly care scenarios (300). Convert the knowledge base text blocks into vectors to build the vector library (301). The structured knowledge base (300) for elderly care scenarios is deeply integrated with the preliminary large model of the elderly care vertical field (201).

7. The method for constructing a large-scale model in the elderly care vertical field as described in claim 6, characterized in that: The deep fusion includes, Set up response logic so that when a user initiates a query, the question is converted into a vector and a similarity search is performed in the vector library (301); Return the first 5-10 matching text blocks as contextual prompts for inputting a large model in the elderly care vertical field; When a user queries and retrieves relevant text content that does not exist in the structured knowledge base (300) for the elderly care scenario, the large model for the elderly care vertical domain responds based on the training architecture.

8. A large-scale model construction system for the elderly care vertical field, employing the large-scale model construction method for the elderly care vertical field as described in any one of claims 1 to 7, characterized in that: The module includes a pre-training dataset, a supervised fine-tuning dataset and a preference dataset module (E1), a step-by-step training module, a preliminary large model for the elderly care vertical domain (E2), a structured knowledge base for elderly care module (E3), and a model fusion module (E4). The module (E1) for constructing the pre-trained dataset, supervised fine-tuning dataset, and preference dataset is used to collect data in the field of elderly care and process the data in the field of elderly care. Based on the data in the field of elderly care, a pre-trained dataset (101), a supervised fine-tuning dataset (102), and a preference dataset (103) are constructed. The step-by-step training module for constructing a preliminary large model (E2) for the elderly care vertical domain is used to take the pre-trained dataset (101), the supervised fine-tuning dataset (102), and the preference dataset (103) as inputs to construct a preliminary large model (201) for the elderly care vertical domain. The structured knowledge base construction for elderly care (E3) is used to filter data in the field of elderly care and construct a structured knowledge base for elderly care scenarios (300) based on the data. The model fusion module (E4) is used to merge the structured knowledge base (300) of elderly care scenarios with the preliminary large model of the elderly care vertical field (201) to obtain the large model of the elderly care vertical field.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method for constructing a large model in the elderly care vertical field as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the method for constructing a large model in the elderly care vertical field as described in any one of claims 1 to 7.

Images