Personalized ability improvement guidance method and system based on fine-tuning large language model

Abstract

The application discloses a personalized ability improvement guidance method based on fine-tuning of a large language model, which comprises the following steps: constructing a field structured fine-tuning data set; adopting a strategy combining low-rank adaptation and supervised fine-tuning to fine-tune an open-source large language model, obtaining a field fine-tuning model and encapsulating the field fine-tuning model into an API service; in response to a user consultation request, analyzing and obtaining a user identifier and a career intention; searching and fusing multiple source data in parallel to generate a unified context data package containing a student personal ability archive, a target post ability model and available resources on campus; calling the fine-tuning model API for reasoning, generating a structured personalized improvement scheme and visually displaying the scheme. Through field fine-tuning and multi-source data fusion, the application enables the model to deeply understand the education scene and generate a personalized scheme precisely matched with local resources, thereby solving the technical problems of lack of real-time performance, deep personalization and operability in the traditional guidance mode.

Description

Technical Field

[0001] This application relates to the fields of natural language processing and educational informatization, and more specifically, to a personalized ability enhancement guidance method and system based on fine-tuning a large language model. Background Technology

[0002] The current higher education and student career development support system suffers from the following major technical deficiencies:

[0003] 1. One-way and delayed guidance methods: Existing career planning or academic guidance relies heavily on periodic human consultations, lectures, or static online course libraries. These methods are difficult to respond to students' individualized and random consultation needs in real time, and the feedback cycle is long, failing to provide immediate support when students show interest or confusion.

[0004] 2. Lack of in-depth personalization: The suggestions provided by traditional systems are often based on general templates or simple rules (such as recommending positions based on majors), and cannot be deeply combined with students' specific intended positions, real-time academic data, and personal skill profiles for comprehensive analysis, making it difficult to provide targeted and personalized improvement paths.

[0005] 3. Disconnect between cognition and action: Even if students learn about a job they are interested in through some channel, they face the dilemma of knowing but not knowing how to prepare. The existing system lacks the ability to intelligently link and plan job goals with available learning resources on campus, making it difficult for students' interests to be transformed into effective learning actions.

[0006] 4. Isolated data, lack of basis for guidance: Students' on-campus growth data (grades, projects, competitions) and external industry knowledge (job requirements, skill trends) belong to different systems and are not interconnected. Therefore, any guidance lacks comprehensive data support and relies more on the personal experience of the instructor.

[0007] The limitations of existing large model applications: Directly calling the general large language model (LLM) API has the following problems: (1) Inaccurate understanding of education-specific terminology, curriculum system and ability evaluation standards; (2) Inability to stably output structured suggestions that conform to the education scenario (such as course number and project name); (3) Lack of ability to deeply integrate with localized data (school resources and student personal files); (4) Easy to recommend courses or resources that do not exist. Summary of the Invention

[0008] This application aims to provide a personalized ability enhancement guidance system and method based on fine-tuning a large language model. Through domain-adaptive model fine-tuning technology and system integration architecture, it addresses the lack of deep personalization and operability in existing guidance systems. To achieve the above objectives, this application adopts the following technical solution.

[0009] Firstly, embodiments of this application provide a personalized capability enhancement guidance method based on fine-tuning a large language model, including:

[0010] Acquire and preprocess domain data to construct a structured, fine-tuned dataset that includes one or more of the following: student consultation dialogue records, job demand analysis, skills gap assessment reports, and successful employment path cases.

[0011] Based on the open-source large language model as the base model, the structured fine-tuning dataset is used to fine-tune the model by combining low-rank adaptation and supervised fine-tuning optimization strategies. The preset proportional parameters of the base model are frozen, and only the low-rank adaptation layer is trained to obtain the domain-fine-tuned large language model.

[0012] The domain-specific fine-tuning large language model is encapsulated as an API service and deployed on a server;

[0013] In response to a consultation request initiated by a user terminal, the consultation request is parsed to obtain user identification and career intention information;

[0014] Based on the user identifier, multi-source data is retrieved and merged in parallel from a preset database to generate one or more unified context data packages, including student personal ability profiles, target job ability models, and available resources within the school.

[0015] The consultation request and the unified context data packet are encapsulated into an API request, and the API service of the domain fine-tuning large language model is called to perform inference and generate a structured personalized improvement plan.

[0016] The personalized enhancement plan is analyzed and sent to the user's terminal for visual display.

[0017] Furthermore, multi-source data is retrieved and fused in parallel to generate a unified context data package, specifically including:

[0018] Based on the user identifier, retrieve one or more of the student's academic performance, completed courses, project experience, and skill tags from the student database to construct the student's personal ability profile;

[0019] Based on the aforementioned career intention information, the corresponding target job competency model is retrieved from the industry knowledge base, including one or more of the following: job skill list, competency requirement level, and skill development trend.

[0020] Based on the students' grade and major, retrieve one or more of the following from the school's resource database: a catalog of elective courses, experimental projects, tutor resources, and academic activities, and construct the set of available school resources.

[0021] The student's personal competency profile, the target job competency model, and the set of available resources within the school are vectorized and mapped and associated to generate a unified multidimensional feature vector as the unified context data package.

[0022] Furthermore, an optimization strategy combining low-rank fitting and supervised fine-tuning is employed for model fine-tuning, specifically including:

[0023] Inject a trainable low-rank adaptation matrix into each layer of the Transformer structure of the base model;

[0024] During fine-tuning, all original parameters of the base model are frozen, and the parameters of the low-rank fitness matrix are updated only through the backpropagation algorithm.

[0025] The base model is optimized by designing a combined loss function, which includes: a supervised fine-tuning loss to ensure that the model output is consistent with the labeled answer, and a domain consistency loss to constrain the model output to conform to the professional terminology and logic of the education and vocational fields.

[0026] A dynamic hard example mining strategy is adopted. After each training iteration, the sample with the worst prediction performance is identified and screened, and then added to the training set for the next round of reinforcement training.

[0027] Furthermore, the API service for fine-tuning the large language model in the aforementioned domain is invoked for inference to generate a structured, personalized improvement plan, specifically including:

[0028] The domain fine-tunes the large language model to parse the API request and understand the user's core consultation intent;

[0029] The student's personal competency profile in the unified context data package is compared and analyzed with the target job competency model from multiple dimensions to generate a quantitative skills gap analysis report. The skills gap analysis report includes one or more of the following: a list of missing skills, existing strengths, and a comprehensive gap score.

[0030] Based on the skills gap analysis report and the available resources on campus, a phased and executable learning path plan is generated; each phase of the learning path plan includes specific action items, suggested completion time, and one or more of the following: the relationship between each action item and bridging a specific skills gap.

[0031] Based on the target job competency model, simulated interview questions, corresponding reference answers, and problem-solving approaches are generated.

[0032] Furthermore, generating the learning path plan also includes: configuring an interactive tracking identifier for each action item in the learning path plan;

[0033] The guidance method also includes:

[0034] In response to a user's triggering action on the tracking identifier, the user's adoption status for that action item is recorded;

[0035] Based on the adoption status, the student's personal ability profile is updated, and the domain fine-tuning large language model is triggered to dynamically adjust and optimize the remaining learning paths.

[0036] Furthermore, after the visualization step, the following steps are also included:

[0037] Collect user feedback data on the personalized enhancement plan; the feedback data includes at least one of the user's adoption rate of recommended resources, completion rate, and overall satisfaction rating of the plan;

[0038] The feedback data is used as new fine-tuning data and merged with the constructed structured fine-tuning dataset to generate an updated fine-tuning dataset.

[0039] Periodically, or when the size of the updated fine-tuning dataset reaches a preset threshold, incremental fine-tuning or retraining of the domain fine-tuning large language model is triggered, and the updated model is hot-deployed to the API service to replace the old version model.

[0040] Furthermore, encapsulating the domain fine-tuning large language model into an API service also includes: configuring a multi-layer caching strategy for the API service;

[0041] The caching strategy includes:

[0042] For frequently asked questions that are frequently encountered, their standard response templates are pre-calculated and stored in a first-level local cache;

[0043] For consultation requests that involve specific student files but have a similarity to historical requests exceeding a preset threshold, a two-level distributed cache is used to store and reuse the inference results of historical similar requests.

[0044] When a user request is received, the system queries the first-level local cache and the second-level distributed cache in sequence. Only when the cache is not hit will the domain fine-tuning large language model be invoked for real-time inference.

[0045] Secondly, embodiments of this application provide a guidance system capable of implementing any of the foregoing guidance methods, comprising:

[0046] The data preparation module is used to acquire and preprocess domain data, and build one or more structured fine-tuning datasets, including student consultation dialogue records, job demand analysis, skills gap assessment reports, and successful employment path cases.

[0047] The model fine-tuning module is used to fine-tune the model based on the open-source large language model as the base model, using the structured fine-tuning dataset and an optimization strategy that combines low-rank adaptation and supervised fine-tuning. The preset proportional parameters of the base model are frozen, and only the low-rank adaptation layer is trained to obtain the domain-fine-tuned large language model.

[0048] The service encapsulation and deployment module is used to encapsulate the domain fine-tuning large language model into an API service and deploy it on the server;

[0049] The request receiving and parsing module is used to respond to the consultation request initiated by the user terminal and parse the consultation request to obtain user identification and career intention information;

[0050] The data retrieval and fusion module is used to retrieve and fuse multi-source data in parallel from a preset database based on the user identifier, and generate one or more unified context data packages including student personal ability profiles, target job ability models and available resources within the school.

[0051] The intelligent reasoning module is used to encapsulate the consultation request and the unified context data package into an API request, call the API service of the domain fine-tuning large language model to perform reasoning, and generate a structured personalized improvement plan.

[0052] The results presentation module is used to parse the personalized enhancement plan and send it to the user terminal for visual display.

[0053] Thirdly, embodiments of this application provide an electronic device, including: one or more processors;

[0054] A memory for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors are able to perform the steps in the guidance method described in any of the preceding claims.

[0055] Fourthly, embodiments of this application provide a computer-readable medium storing a computer program that, when executed by a processor, can implement the steps in the guidance method described in any of the foregoing claims.

[0056] This application discloses a personalized ability enhancement guidance method based on fine-tuning a large language model. The method includes: constructing a domain-structured fine-tuning dataset; employing a strategy combining low-rank adaptation and supervised fine-tuning to fine-tune an open-source large language model, obtaining a domain-tuned model, and encapsulating it as an API service; responding to user consultation requests by parsing and obtaining user identifiers and career intentions; parallelly retrieving and fusing multi-source data to generate a unified contextual data package containing student personal ability profiles, target job ability models, and available school resources; and calling the fine-tuning model API for inference to generate a structured personalized enhancement plan and visualize it. This application, through domain fine-tuning and multi-source data fusion, enables the model to deeply understand the educational scenario and generate personalized plans that accurately match local resources, solving the technical problems of traditional guidance methods lacking real-time performance, deep personalization, and operability. Attached Figure Description

[0057] Figure 1 The core flowchart of a personalized ability enhancement guidance method based on fine-tuning a large language model provided in this application embodiment;

[0058] Figure 2 A schematic diagram of the module structure of a personalized ability improvement guidance system based on fine-tuning a large language model, provided for an embodiment of this application;

[0059] Figure 3 This is a structural block diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0060] To enable those skilled in the art to better understand the technical solutions of this application, exemplary embodiments of this application are described below with reference to the accompanying drawings, including various details of the embodiments of this application to aid understanding. These should be considered merely exemplary. Therefore, those skilled in the art should recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. Similarly, for clarity and conciseness, descriptions of well-known functions and structures are omitted in the following description. Unless otherwise specified, the various embodiments of this application and the features within those embodiments can be combined with each other.

[0061] As used herein, the term "and / or" includes any and all combinations of one or more of the associated enumerated entries. The terminology used herein is for describing particular embodiments only and is not intended to limit the application. As used herein, the singular forms "a" and "the" are also intended to include the plural forms, unless the context clearly indicates otherwise. It should also be understood that when the terms "comprising" and / or "made of" are used herein, the presence of the stated feature, integral, step, operation, element, and / or component is specified, but the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or groups thereof is not excluded. Terms such as "connected" or "linked" are not limited to physical or mechanical connections but can include electrical connections, whether direct or indirect.

[0062] Unless otherwise specified, all terms used in this application (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art. It should also be understood that terms such as those defined in commonly used dictionaries should be interpreted as having a meaning consistent with their meaning in the context of the relevant art and this application, and will not be interpreted as having an idealized or overly formal meaning, unless expressly so defined in this application.

[0063] refer to Figure 1 One embodiment of this application proposes a personalized ability enhancement guidance method based on fine-tuning a large language model, which may specifically include the following steps.

[0064] S1. Acquire and preprocess domain data to construct a structured, fine-tuned dataset that includes one or more of the following: student consultation dialogue records, job requirement analysis, skills gap assessment reports, and successful employment path cases.

[0065] Specifically, the study collected 5,000 student career consultation dialogues from a university over the past three years, job descriptions from mainstream recruitment platforms (such as Boss Zhipin and Liepin), skills gap assessment reports, and successful employment cases archived by the university's career guidance center. After cleaning and de-identifying the data, a structured question-and-answer pair and scenario-based prompt templates were constructed, which included "user questions - standard answers - job requirements - skills tags". The data covered student consultation dialogue records, job demand analysis, skills gap assessment reports, and successful employment path cases.

[0066] S2. Based on the open-source large language model as the base model, the structured fine-tuning dataset is used to fine-tune the model using an optimization strategy that combines low-rank adaptation and supervised fine-tuning. The preset proportional parameters of the base model are frozen, and only the low-rank adaptation layer is trained to obtain the domain-fine-tuned large language model.

[0067] Some embodiments of this application use DeepSeek-R1-8B as the base model and employ LoRA (Low-Rank Adaptation) technology, setting the LoRA rank r=8, and only fine-tuning the attention layer weight matrix of the model. A trainable low-rank adaptation matrix is ​​injected into each Transformer layer of the base model (i.e., the basic large language model). During fine-tuning, all original parameters of the basic large language model are frozen, and the parameters of the low-rank adaptation matrix are updated only through backpropagation. A combined loss function is designed to optimize the basic large language model. This combined loss function includes a supervised fine-tuning loss to ensure that the model output is consistent with the labeled answer, and a domain consistency loss to constrain the model output to conform to the professional terminology and logic of the education and vocational fields. Simultaneously, a dynamic hard example mining strategy is adopted. After each training iteration, the sample with the worst prediction performance is identified and selected, and it is added back to the training set for reinforcement training.

[0068] Furthermore, a contrastive learning mechanism was introduced during the fine-tuning process. By constructing positive and negative sample pairs (e.g., anchor point: student competency profile "familiar with Java, Spring"; positive sample: "recommended to learn microservice architecture, target position: Java backend development engineer"; negative sample: "recommended to learn TensorFlow, target position: AI algorithm engineer"), the model's ability to identify the differences in skill requirements for similar but different positions was enhanced. In particular, training data for easily confused position pairs such as "big data development engineer" and "data mining engineer" was added, enabling the model to learn to distinguish the unique skill requirements and learning paths of each from subtle semantic differences.

[0069] During training, the model's performance is evaluated on the validation set, including answer accuracy, structural compliance, and the standardization of domain terminology usage. Based on the evaluation results, the training data, prompt templates, or fine-tuning strategies are adjusted, and the model is retrained until it meets the deployment requirements. The final result is a domain-optimized, fine-tuned large language model that can stably output structured career planning suggestions.

[0070] S3. Encapsulate the domain fine-tuning large language model into an API service and deploy it on the server.

[0071] Specifically, the fine-tuned model is deployed on a private server cluster in a university, providing model inference services through a RESTful API interface. In some embodiments, a multi-layered caching strategy is configured for the API service: for frequently occurring common consultation questions (such as "How should computer science students prepare for postgraduate entrance exams?"), standard response templates are pre-calculated and stored in a first-level local cache (such as Caffeine) to achieve nanosecond-level access; for consultation requests involving specific student profiles but with a similarity to historical requests exceeding a preset threshold, a second-level distributed cache (such as Redis) is used to store and reuse the inference results of similar historical requests. When students with similar backgrounds ask almost identical questions, the cosine similarity of the request feature vectors is calculated. If it exceeds 0.95, the historical planning scheme is directly reused as the base template; when a user request is received, the first-level local cache and the second-level distributed cache are queried sequentially. Only when the cache is not hit is the domain-fine-tuned large language model called for real-time inference. In addition, an asynchronous processing mechanism is configured, using asynchronous task queues for complex analysis requests; a load balancing mechanism is configured to distribute inference requests among multiple GPU servers; and a monitoring and alarm mechanism is configured to monitor API response time, success rate, and resource utilization in real time.

[0072] S4. In response to a consultation request initiated by a user terminal, parse the consultation request to obtain user identification and career intention information.

[0073] Specifically, a third-year computer science student entered "I want to work as an AI algorithm engineer, how should I prepare?" into a WeChat mini-program, and the system backend parsed out the user ID and career intention "AI algorithm engineer".

[0074] S5. Based on the user identifier, retrieve and merge multi-source data in parallel from the preset database to generate one or more unified context data packages, including student personal ability profiles, target job ability models, and available resources within the school.

[0075] Parallel retrieval and fusion of multi-source data to generate a unified context data package specifically includes: retrieving students' academic performance, completed courses, project experience, and skill tags from the student database based on the user identifier, and constructing the student's personal ability profile (e.g., completed courses are "Advanced Mathematics 95 points, Linear Algebra 92 points, Python Programming 88 points", and skill tags are "Python Programming, Mathematical Foundations"); and retrieving corresponding job competency models from the industry knowledge base based on the career intention information, including the required skill list, competency level, and skill development trends (e.g., the skill list for the "AI Algorithm Engineer" position is "Python / C++, TensorFlow / PyTorch, Data Structures and Algorithms, Machine Learning Algorithm Principles", with each skill having a specific importance). The system uses metrics such as an importance score of 0.9 for "TensorFlow / PyTorch". Based on the student's grade and major, it retrieves available course catalogs, experimental projects, tutor resources, and academic activities from the university's resource database to construct a list of available resources (e.g., courses available this semester and next, such as "Deep Learning Theory and Practice (CS401)", "Introduction to Machine Learning (CS305)", and the related lab project "Innovative Training for Image Recognition Based on Deep Learning"). Then, a pre-trained encoder maps student skills, job requirements, and course resources to vectors in the same semantic space. The system vectorizes and associates the student's personal ability profile, target job ability model, and the list of available resources, generating a unified multi-dimensional feature vector as the unified context data package. For example, it calculates the similarity between the student's "Python Programming" skill vector and the course description vector of "Deep Learning Theory and Practice" to identify the course as an effective resource for filling the "TensorFlow / PyTorch" skill gap.

[0076] S6. Encapsulate the consultation request and the unified context data packet into an API request, call the API service of the domain fine-tuning large language model for inference, and generate a structured personalized improvement plan.

[0077] The API service of the domain-specific fine-tuning large language model is invoked for inference to generate a structured, personalized improvement plan. Specifically, this includes: the domain-specific fine-tuning large language model parsing the API request to understand the user's core inquiry intent as "to obtain a learning plan to become an AI algorithm engineer"; performing a multi-dimensional comparative analysis of the student's personal ability profile in the unified context data package with the target job ability model to generate a quantitative skills gap analysis report; the skills gap analysis report includes a list of missing skills, existing strengths, and a comprehensive gap score (e.g., missing_skills: ["deep learning frameworks", "data structures and algorithms"], existing_strengths: ["Python programming basics", "mathematical basics"], gap_score: 0.65); and generating a phased, ... An executable learning path plan is provided; each stage of the learning path plan includes specific action items, suggested completion times, and the correlation between each action item and filling specific skill gaps (e.g., the first stage (1-2 months) recommends the course "Data Structures and Algorithms (CS210)", the second stage (3-4 months) recommends the course "Deep Learning Theory and Practice (CS401)" and participation in the "Deep Learning-Based Image Recognition" innovation project, with a reason attached after each recommendation: "Deep Learning Theory and Practice is recommended because Chapter 3 of this course specifically explains the TensorFlow framework, which can effectively fill your knowledge gaps in deep learning frameworks"); based on the target job competency model, simulated interview questions and corresponding reference answers and problem-solving approaches are generated (e.g., "Please explain in detail how the self-attention mechanism in the Transformer model works?" and provide reference answers and problem-solving approaches containing key knowledge points).

[0078] The learning path planning process also includes: configuring an interactive tracking identifier for each action item in the learning path plan (e.g., generating an "Add to My Schedule" button for the recommended course "Deep Learning Theory and Practice (CS401)"). In response to a user's triggering action on the tracking identifier, the system records the user's adoption status for that action item (after a student clicks the button, the system records that the student has adopted the suggestion and updates their status from "Pending" to "Planned"). Based on the adoption status, the system updates the student's personal ability profile and triggers the domain fine-tuning large language model to dynamically adjust and optimize the remaining learning paths (this adoption event triggers a path optimization request; the system repackages the updated student profile into a context and calls the fine-tuning model API again. Based on the new situation that "the student is about to acquire a foundation in deep learning theory," the model adjusts the originally recommended basic projects to more challenging projects, such as "Optimization of Visual Models Based on Transformer").

[0079] S7. Analyze the personalized enhancement plan and send it to the user terminal for visual display.

[0080] After receiving the response, the front-end renders the skill gap as a radar chart, the learning path as a Gantt chart, and includes a course details card, which is then displayed to the student user.

[0081] After the system has been running for a semester, user feedback data on the personalized improvement plan is collected. This feedback data includes the adoption rate and completion rate of recommended resources, as well as the overall satisfaction rating of the plan. For example, data shows that the adoption rate for the course "Deep Learning Theory and Practice" is as high as 80%, while the adoption rate for another lecture, "Frontiers of Computer Vision," is only 20%. Meanwhile, some students gave the generated plan ratings as either "very useful" or "not quite as expected."

[0082] The feedback data is used as new fine-tuning data and merged with the structured fine-tuning dataset from step S1 to generate an updated fine-tuning dataset. The feedback data containing "adoption rate," "completion rate," and "satisfaction" is combined with the original 5000 training data points to form a new dataset containing 6000 data points. The feedback data is labeled as new question-answer pairs, for example: Question: "I'm not interested in the 'Frontiers of Computer Vision' lecture you recommended before, do you have any other recommendations?" Standard answer: "You could try the more practical 'Deep Learning Theory and Practice' course..."

[0083] Periodically, or when the size of the updated fine-tuning dataset reaches a preset threshold (e.g., the accumulated new feedback data exceeds 1000 entries), incremental fine-tuning or retraining (continuing training on top of the existing LoRA adapter) of the domain-specific large language model is triggered. The updated model is then hot-deployed to the API service via Kubernetes' rolling update strategy, replacing the old model without service interruption. Thereafter, new consultation requests will be handled by the new model that has learned from user feedback.

[0084] refer to Figure 2 An embodiment of this application also proposes a personalized ability improvement guidance system based on fine-tuning a large language model, which may include the following modules.

[0085] (1) Data Preparation Module: Deploy a data cleaning and labeling platform to regularly scrape data from the academic affairs system, career guidance center and mainstream recruitment platforms, perform de-identification processing, construct a fine-tuning dataset, and store it in HDFS. This module is used to acquire and preprocess domain data to construct a structured fine-tuning dataset that includes student consultation dialogue records, job requirement analysis, skills gap assessment reports and successful employment path cases.

[0086] (2) Model Fine-tuning Module: Based on a computing cluster of 8 NVIDIA A100 GPUs, the DeepSpeed ​​framework is used in conjunction with LoRA technology to periodically fine-tune the DeepSeek-R1-8B model. This module is used to fine-tune the model based on an open-source large language model, using a structured fine-tuning dataset and an optimization strategy that combines low-rank adaptation and supervised fine-tuning. Most of the parameters of the base model are frozen, and only the low-rank adaptation layer is trained to obtain a domain-fine-tuned large language model.

[0087] (3) Service encapsulation and deployment module: The fine-tuning model is encapsulated into a high-performance API service using NVIDIA Triton Inference Server, and containerized and managed using Kubernetes to ensure high availability and elastic scaling of the service. This module is used to encapsulate the domain fine-tuning large language model into an API service and deploy it on the server.

[0088] (4) Request Receiving and Parsing Module: Built on Spring Boot Gateway, this module is responsible for receiving user requests from the Web client and WeChat Mini Program, performing authentication and rate limiting. It responds to consultation requests initiated by user terminals, parses the requests to obtain user identification and career intention information.

[0089] (5) Data Retrieval and Fusion Module: This module connects to MySQL (for storing student profiles), Neo4j (for storing job competency graphs), and Elasticsearch (for storing on-campus course resources). When a request arrives, this module concurrently queries these three data sources and integrates the results into a JSON-formatted context package using a data fusion engine. This module is used to retrieve and fuse multi-source data in parallel from a preset database based on the user identifier, generating a unified context data package containing the student's personal competency profile, target job competency model, and available on-campus resources.

[0090] (6) Intelligent Inference Module: This module is responsible for communicating with the backend Triton Inference Server, sending user requests and context packets to the LLM service, and receiving the structured JSON results returned by the service. This module encapsulates the consultation request and the unified context packet into an API request, calls the API service of the domain fine-tuning large language model for inference, and generates a structured personalized improvement plan.

[0091] (7) Results Presentation Module: In the Vue.js front-end framework, the module receives the inference results and uses the ECharts library to render the skill gap as a radar chart, the learning path as a timeline, and the recommended resources as a card list. This module is used to parse the structured personalized improvement plan and send it to the user terminal for visualization.

[0092] (8) Feedback and Iteration Module: Records students' clicks, selections, ratings, and other behaviors related to recommended courses. Periodically, this behavioral data is packaged and sent to the data preparation module for the next iteration of model training. This module is used to collect user feedback data on personalized improvement plans and input it as new fine-tuning data into the data preparation module to trigger the model fine-tuning module to update the domain fine-tuning large language model.

[0093] The modules communicate with each other through well-defined RESTful APIs, forming a complete closed-loop system that goes from data to model, then to application, and finally back to the model.

[0094] Overall, the advantages of this application compared to the prior art include:

[0095] 1. It solves the "domain adaptation gap" problem of general large models in educational scenarios, and significantly improves the professionalism and accuracy of guidance.

[0096] Existing general-purpose large language models (LLMs) suffer from technical shortcomings when directly applied to educational career guidance. These include inaccurate understanding of domain terminology, the potential for "illusion" recommendations that do not exist, and difficulty in consistently outputting structured suggestions that align with educational scenarios. Research indicates that even fine-tuned LLMs may underperform dedicated models on educational tasks, exhibiting poor temporal consistency and high early sequence error rates. This application addresses these issues by constructing a structured fine-tuning dataset containing over 5,000 real-world domain data points, including student consultation dialogues, job requirement analyses, and successful employment cases. A domain-adaptive training strategy combining supervised fine-tuning and LoRA is employed to enable the model to deeply understand university curriculum systems (e.g., course number CS401), competency evaluation standards, and industry terminology. Furthermore, the design of a domain consistency loss function effectively constrains the model's output to align with professional logic within educational scenarios, avoiding the illusion problem caused by general-purpose models "borrowing tokens from conceptually similar instances." This ensures that recommended courses, projects, and resources are genuinely usable within the university, resulting in highly credible professional skill gap analyses and learning path planning.

[0097] 2. By integrating multi-source data and dynamic behavior modeling, the problems of "data isolation" and "lack of personalization" have been solved.

[0098] Traditional career guidance systems often rely on single data sources (such as recommending jobs solely based on majors) or simple rule templates, making it difficult to achieve true deep personalization. Existing research in the recommender system field shows that effectively integrating multi-source heterogeneous data and dynamically capturing changes in user behavior is a key challenge in achieving accurate personalized recommendations. This application constructs a three-dimensional data foundation comprising a student database, an industry knowledge base, and an on-campus resource repository. Through parallel retrieval mechanisms and vectorized mapping technology in the data processing layer, it semantically associates and integrates students' academic performance, completed courses, skill tags, and the competency models, skill requirement lists, and available on-campus course catalogs and experimental projects for target positions, generating a unified multi-dimensional feature vector as contextual input. This multi-source data fusion mechanism prevents student profiles, external job requirements, and available on-campus resources from becoming isolated. The fine-tuned model can reason based on the fused context, generating precise suggestions that deeply integrate personal circumstances and local resources, such as "You lack knowledge of deep learning frameworks, and Chapter 3 of the fall semester's 'Deep Learning Theory and Practice' (CS401) course specifically explains TensorFlow." This effectively solves the problem of the disconnect between cognition and action in existing systems.

[0099] 3. A closed-loop system of "use-feedback-optimization" has been built, enabling continuous evolution and personalized adaptation of guidance capabilities.

[0100] Existing guidance systems mostly provide static, one-off suggestions, unable to dynamically adjust based on students' actual choices and progress, and struggling to continuously learn and evolve from usage data. This application incorporates a feedback and iteration module into its system architecture. By configuring interactive tracking tags for each action item in the learning path, it records students' adoption status, completion status, and satisfaction ratings of recommended resources in real time. This feedback data serves as new fine-tuning data, periodically merged with the original training dataset, triggering incremental fine-tuning or retraining of the domain-specific fine-tuning model. New versions are deployed without service interruption through model hot-update technology. This closed-loop mechanism enables the system to continuously learn from real user interactions—for example, when the adoption rate of a course recommendation remains low, the model adjusts its recommendation strategy in subsequent iterations—forming a virtuous cycle of "becoming smarter with use." Simultaneously, the path dynamic adjustment function triggered by user adoption status allows learning plans to be optimized in real time based on students' actual choices, achieving truly continuous, adaptive guidance.

[0101] The embodiments of the aforementioned guidance method and the embodiments of the aforementioned guidance system are identical or related in technical concept, and can be referenced and learned from each other in terms of technical details and technical effects, which will not be repeated here.

[0102] Based on the same inventive concept, embodiments of this application also provide an electronic device. Figure 3This is a structural block diagram of an electronic device provided in an embodiment of this application. Figure 3 As shown in the embodiments of this application, an electronic device includes: one or more processors 101, a memory 102, and one or more I / O interfaces 103. The memory 102 stores one or more programs, which, when executed by the one or more processors, cause the one or more processors to implement any of the guided methods described in the above embodiments; the one or more I / O interfaces 103 are connected between the processor and the memory, configured to enable information interaction between the processor and the memory.

[0103] The processor 101 is a device with data processing capabilities, including but not limited to a central processing unit (CPU); the memory 102 is a device with data storage capabilities, including but not limited to random access memory (RAM, more specifically SDRAM, DDR, etc.), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), and flash memory (FLASH); the I / O interface (read / write interface) 103 is connected between the processor 101 and the memory 102, and can realize information interaction between the processor 101 and the memory 102, including but not limited to a data bus (Bus).

[0104] In some embodiments, the processor 101, memory 102, and I / O interface 103 are interconnected via bus 104, and thus connected to other components of the computing device.

[0105] In some embodiments, the one or more processors 101 include a field-programmable gate array.

[0106] This application also provides a computer-readable medium. The computer-readable medium stores a computer program, which, when executed by a processor, implements the steps of any of the guided methods described in the above embodiments. The computer-readable storage medium may be volatile or non-volatile.

[0107] This application also provides a computer program product, including computer-readable code, or a non-volatile computer-readable storage medium carrying computer-readable code, wherein when the computer-readable code is run in the processor of an electronic device, the processor in the electronic device executes the above-described guided method.

[0108] Those skilled in the art will understand that all or some of the steps, systems, and apparatuses disclosed above, and their functional modules / units, can be implemented as software, firmware, hardware, or suitable combinations thereof. In hardware implementations, the division between functional modules / units mentioned above does not necessarily correspond to the division of physical components; for example, a physical component may have multiple functions, or a function or step may be performed collaboratively by several physical components. Some or all physical components may be implemented as software executed by a processor, such as a central processing unit, digital signal processor, or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit (ASIC). Such software can be distributed on a computer-readable storage medium, which may include computer storage media (or non-transitory media) and communication media (or transient media).

[0109] As is known to those skilled in the art, the term computer storage medium includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information, such as computer-readable program instructions, data structures, program modules, or other data. Computer storage media includes, but is not limited to, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), static random access memory (SRAM), flash memory or other memory technologies, portable compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical disc storage, magnetic cartridges, magnetic tape, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and is accessible to a computer. Furthermore, it is known to those skilled in the art that communication media typically contain computer-readable program instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms, and may include any information delivery medium.

[0110] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0111] The computer program instructions used to perform the operations of this application may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuits, such as programmable logic circuits, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), are personalized by utilizing the status information of the computer-readable program instructions. These electronic circuits can execute the computer-readable program instructions to implement various aspects of this application.

[0112] The computer program product described herein can be implemented specifically through hardware, software, or a combination thereof. In one alternative embodiment, the computer program product is specifically embodied in a computer storage medium; in another alternative embodiment, the computer program product is specifically embodied in a software product, such as a software development kit (SDK), etc.

[0113] Various aspects of this application are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer-readable program instructions.

[0114] These computer-readable program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that, when executed by the processor of the computer or other programmable data processing apparatus, they create means for implementing the functions / actions specified in one or more blocks of the flowchart and / or block diagram. These computer-readable program instructions can also be stored in a computer-readable storage medium that causes a computer, programmable data processing apparatus, and / or other device to operate in a particular manner; thus, the computer-readable medium storing the instructions comprises an article of manufacture that includes instructions for implementing aspects of the functions / actions specified in one or more blocks of the flowchart and / or block diagram.

[0115] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the instructions executed on the computer, other programmable data processing apparatus, or other device to perform the functions / actions specified in one or more boxes of a flowchart and / or block diagram.

[0116] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of an instruction containing one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0117] Exemplary embodiments have been disclosed in this application, and while specific terminology has been used, it is used only and should be interpreted in a general illustrative sense and is not intended to be limiting. In some embodiments, it will be apparent to those skilled in the art that features, characteristics, and / or elements described in conjunction with particular embodiments may be used alone, or in combination with features, characteristics, and / or elements described in conjunction with other embodiments, unless otherwise expressly indicated. Therefore, those skilled in the art will understand that various changes in form and detail may be made without departing from the scope of this application as set forth by the appended claims.

Claims

1. A personalized ability enhancement guidance method based on fine-tuning a large language model, characterized in that, include: Acquire and preprocess domain data to construct a structured, fine-tuned dataset that includes one or more of the following: student consultation dialogue records, job demand analysis, skills gap assessment reports, and successful employment path cases. Based on the open-source large language model as the base model, the structured fine-tuning dataset is used to fine-tune the model by combining low-rank adaptation and supervised fine-tuning optimization strategies. The preset proportional parameters of the base model are frozen, and only the low-rank adaptation layer is trained to obtain the domain-fine-tuned large language model. The domain-specific fine-tuning large language model is encapsulated as an API service and deployed on a server; In response to a consultation request initiated by a user terminal, the consultation request is parsed to obtain user identification and career intention information; Based on the user identifier, multi-source data is retrieved and merged in parallel from a preset database to generate one or more unified context data packages, including student personal ability profiles, target job ability models, and available resources within the school. The consultation request and the unified context data packet are encapsulated into an API request, and the API service of the domain fine-tuning large language model is called to perform inference and generate a structured personalized improvement plan. The personalized enhancement plan is analyzed and sent to the user's terminal for visual display.

2. The guidance method according to claim 1, characterized in that, Parallel retrieval and fusion of multi-source data to generate a unified context data package, specifically including: Based on the user identifier, retrieve one or more of the student's academic performance, completed courses, project experience, and skill tags from the student database to construct the student's personal ability profile; Based on the aforementioned career intention information, the corresponding target job competency model is retrieved from the industry knowledge base, including one or more of the following: a list of skills required for the job, competency requirement levels, and skill development trends. Based on the students' grade and major, retrieve one or more of the following from the school's resource database: a catalog of elective courses, experimental projects, tutor resources, and academic activities, and construct the set of available school resources. The student's personal competency profile, the target job competency model, and the set of available resources within the school are vectorized and mapped and associated to generate a unified multidimensional feature vector as the unified context data package.

3. The guidance method according to claim 1, characterized in that, An optimization strategy combining low-rank fitting and supervised fine-tuning is employed for model fine-tuning, specifically including: Inject a trainable low-rank adaptation matrix into each layer of the Transformer structure of the base model; During fine-tuning, all original parameters of the base model are frozen, and the parameters of the low-rank fitness matrix are updated only through the backpropagation algorithm. The design incorporates a combined loss function to optimize the base model. The combined loss function includes: a supervised fine-tuning loss to ensure that the model output is consistent with the labeled answer, and a domain consistency loss to constrain the model output to conform to the professional terminology and logic of the education and vocational fields. A dynamic hard example mining strategy is adopted. After each training iteration, the sample with the worst prediction performance is identified and screened, and then added to the training set for the next round of reinforcement training.

4. The guidance method according to claim 1, characterized in that, The API service for fine-tuning the large language model in the aforementioned domain is invoked for inference to generate a structured, personalized improvement plan, specifically including: The domain fine-tunes the large language model to parse the API request and understand the user's core consultation intent; The student's personal competency profile in the unified context data package is compared and analyzed with the target job competency model from multiple dimensions to generate a quantitative skills gap analysis report. The skills gap analysis report includes one or more of the following: a list of missing skills, existing strengths, and a comprehensive gap score. Based on the skills gap analysis report and the available resources on campus, a phased and executable learning path plan is generated; each phase of the learning path plan includes specific action items, suggested completion times, and one or more of the following: the relationship between each action item and bridging a specific skills gap. Based on the target job competency model, simulated interview questions, corresponding reference answers, and problem-solving approaches are generated.

5. The guidance method according to claim 4, characterized in that, Generating a learning path plan also includes: configuring an interactive tracking identifier for each action item in the learning path plan; The guidance method also includes: In response to a user's triggering action on the tracking identifier, the user's adoption status for that action item is recorded; Based on the adoption status, the student's personal ability profile is updated, and the domain fine-tuning large language model is triggered to dynamically adjust and optimize the remaining learning paths.

6. The guidance method according to claim 1, characterized in that, Following the visualization step, the following steps are also included: Collect user feedback data on the personalized enhancement plan; the feedback data includes at least one of the user's adoption rate of recommended resources, completion rate, and overall satisfaction rating of the plan; The feedback data is used as new fine-tuning data and merged with the constructed structured fine-tuning dataset to generate an updated fine-tuning dataset. Periodically, or when the size of the updated fine-tuning dataset reaches a preset threshold, incremental fine-tuning or retraining of the domain fine-tuning large language model is triggered, and the updated model is hot-deployed to the API service to replace the old version model.

7. The guidance method according to claim 1, characterized in that, Encapsulating the domain fine-tuning large language model into an API service also includes: configuring a multi-layer caching strategy for the API service; The caching strategy includes: For frequently asked questions that are frequently encountered, their standard response templates are pre-calculated and stored in a first-level local cache; For consultation requests that involve specific student files but have a similarity to historical requests exceeding a preset threshold, a two-level distributed cache is used to store and reuse the inference results of historical similar requests; When a user request is received, the system queries the first-level local cache and the second-level distributed cache in sequence. Only when the cache is not hit will the domain fine-tuning large language model be invoked for real-time inference.

8. A guidance system capable of implementing the guidance method according to any one of claims 1-7, characterized in that, include: The data preparation module is used to acquire and preprocess domain data, and build one or more structured fine-tuning datasets, including student consultation dialogue records, job demand analysis, skills gap assessment reports, and successful employment path cases. The model fine-tuning module is used to fine-tune the model based on the open-source large language model as the base model, using the structured fine-tuning dataset and an optimization strategy that combines low-rank adaptation and supervised fine-tuning. The preset proportional parameters of the base model are frozen, and only the low-rank adaptation layer is trained to obtain the domain-fine-tuned large language model. The service encapsulation and deployment module is used to encapsulate the domain fine-tuning large language model into an API service and deploy it on the server; The request receiving and parsing module is used to respond to the consultation request initiated by the user terminal and parse the consultation request to obtain user identification and career intention information; The data retrieval and fusion module is used to retrieve and fuse multi-source data in parallel from a preset database based on the user identifier, and generate one or more unified context data packages including student personal ability profiles, target job ability models and available resources within the school. The intelligent reasoning module is used to encapsulate the consultation request and the unified context data package into an API request, call the API service of the domain fine-tuning large language model to perform reasoning, and generate a structured personalized improvement plan. The results presentation module is used to parse the personalized enhancement plan and send it to the user terminal for visual display.

9. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors are enabled to perform the steps in the guided method as described in any one of claims 1 to 7.

10. A computer-readable medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it can perform the steps of the guided method as described in any one of claims 1 to 7.

Images