RAG question setting and automatic question judging system based on large model

Abstract

The invention provides an RAG question setting and automatic question judging system based on a large model, and belongs to the technical field of intelligent education of natural language processing. The problems that in the prior art, efficiency is low, accuracy is insufficient, and safety and expansibility are limited are solved. An overall framework is established based on RAG, and a large model local deployment and fine adjustment module, a teaching resource vectorization storage module and a large model question setting and judging module are arranged under the overall framework; the large model local deployment and fine tuning module is used for realizing data collection and preprocessing, model selection and loading, task fine tuning, model training and verification and localized deployment and reasoning optimization; the teaching resource vectorization storage module is used for realizing data uploading and storage, data vectorization processing and data auditing and monitoring; the large model question setting and judging module is used for realizing question generation and automatic question judgment; the method is applied to wisdom education.

Description

Technical Field

[0001] The present invention provides a RAG question setting and automatic question judging system based on a large model, belonging to the field of intelligent education technology of natural language processing. Background Art

[0002] With the rapid development of artificial intelligence technology, the field of education is undergoing profound changes. Traditional teaching evaluation methods mainly rely on teachers to manually set questions, mark papers and evaluate, which not only consumes a lot of time and energy, but also fails to meet the needs of personalized teaching and large-scale teaching evaluation. In order to solve these problems, researchers have begun to introduce technologies such as natural language processing and machine learning into the field of education, trying to achieve automation and intelligence of teaching evaluation.

[0003] Early automated test-setting systems were mainly based on template matching and rule generation. These systems generated corresponding test questions by associating preset question templates with knowledge points. However, the limitations of this approach are: 1) Lack of flexibility: The number of templates is limited, making it difficult to cover all possible question types and knowledge point combinations; 2) Generate single content: lack of diversity and creativity, which can easily lead to students getting bored with the test questions; 3) Unable to adapt to personalized needs: Unable to generate suitable test questions based on students’ knowledge level and learning progress.

[0004] In terms of automatic question assessment, traditional methods mostly use keyword matching and rule-based assessment, which works well for objective questions (such as multiple-choice questions and fill-in-the-blank questions), but is difficult to accurately assess subjective questions (such as short-answer questions and essays). The main problems include: 1) Limited comprehension ability: unable to deeply understand the semantics and logic of students’ answers; 2) Single scoring criteria: It is difficult to consider multiple scoring factors, such as content accuracy, logic and expression ability; 3) Insufficient feedback: Unable to provide students with targeted suggestions for improvement.

[0005] In the field of education, researchers have begun to explore the application of large models to question setting and grading. For example, using the GPT model to generate test questions that meet specific requirements, or using the BERT model for automatic grading. However, these attempts still face some challenges: 1) Insufficient model generalization ability: Directly using pre-trained models may not be well adapted to specific teaching scenarios and subject areas; 2) High resource requirements: Large models have high computing and storage requirements, and high deployment and operation costs; 3) Security and privacy issues: The transmission and storage of teaching resources and student data may pose security risks.

[0006] Therefore, the Retrieval Augmentation Generation (RAG) model was proposed and used, but the existing RAG model still has shortcomings when applied in educational scenarios: 1) Lack of real-time performance: With large-scale teaching resources, the retrieval process may take a long time, affecting the response speed of the system; 2) Poor adaptability: The model may not be able to fully utilize the rich information of teaching resources, and the generated content may not fully match the teaching objectives; 3) Limited scalability: The system’s performance may decline when faced with growing data volumes and diverse teaching resource formats.

[0007] In the smart education system, the security and privacy protection of teaching resources and student data are crucial. Existing systems often lack a sound security mechanism, which may lead to: 1) Data leakage risk: Teaching resources may be obtained by unauthorized persons during storage and transmission; 2) Lack of access control: unable to effectively manage and monitor data access and usage; 3) Insufficient data backup: In the event of data damage or loss, it cannot be restored quickly, affecting the continuity of teaching activities.

[0008] With the growth of teaching resources and the number of users, the system needs to have good scalability and efficient performance. However, traditional single-machine deployment and simple database structure are difficult to meet the needs of high concurrency and large data volumes. Summary of the invention

[0009] The purpose of the present invention is to provide a RAG question setting and automatic judgment system based on a large model to solve the problems of low efficiency, insufficient accuracy, limited security and scalability in the prior art. Through innovative system architecture and technical solutions, efficient, safe and accurate question setting and judgment functions are achieved to meet the needs of smart education.

[0010] The technical solution adopted by the present invention is: Compared with the prior art, the present invention has the following beneficial effects: a RAG question-setting and automatic question-judgment system based on a large model, which establishes an overall architecture based on RAG, selects a local Faiss or Milvus vector database, performs block vectorization processing on teaching materials, ensures effective storage and fast access to files in various formats, and is provided with a large model local deployment and fine-tuning module, a teaching resource vectorization storage module, and a large model question-setting and question-judgment module under the overall architecture; The large model local deployment and fine-tuning module is used to implement data collection and preprocessing, model selection and loading, fine-tuning tasks, model training and verification, and localized deployment and reasoning optimization; The teaching resource vector storage module is used to realize data uploading and storage, data vector processing and data auditing and monitoring; The large model question setting and judging module is used to realize question generation and automatic question judging.

[0011] Model selection and loading of large model local deployment and fine-tuning modules selects a suitable pre-trained model according to the specific teaching task, loads the pre-trained weights as the initial parameters, and optimizes the model performance in combination with quantization technology.

[0012] The fine-tuning tasks of the local deployment and fine-tuning modules of large models are based on teaching objectives, set appropriate loss functions and optimization algorithms, and use LoRA technology to fine-tune the low-rank matrix of specific model layers. Early stopping and learning rate scheduling are introduced to prevent overfitting and accelerate model convergence. Indicators for evaluating model performance are also formulated.

[0013] The process of model training and verification of large model local deployment and fine-tuning module is as follows: Model training: During fine-tuning, LoRa technology is applied to specific layers of the model; Cross-validation: Use the K-fold cross-validation method to continuously evaluate and optimize model performance and fine-tune parameters to ensure that the model can maintain stable performance on different data sets; Performance evaluation and tuning: Use multiple different indicators to comprehensively evaluate the model. If the performance does not meet expectations, adjust LoRA-related hyperparameters or resample data for further optimization; Training termination and saving: When the model reaches the set performance indicators, save the model weights and LoRA adaptation layer parameters.

[0014] Local deployment and inference optimization of large model local deployment and fine-tuning modules include the following two parts: Reasoning speed optimization: Deploy the optimized model locally and use model pruning and / or distillation methods to further improve the reasoning speed. Real-time update and adaptive fine-tuning: Quickly fine-tune through LoRA, combined with incremental learning mechanism, and update the model regularly.

[0015] The teaching resource vectorization storage module stores the uploaded teaching materials in the sqlite database, and when the teacher logs into the system, it automatically calls the local embedding model to vectorize the materials in the sqlite database, generates high-dimensional vectors and stores them in the vector database.

[0016] Data auditing and monitoring in the vector storage module of teaching resources is achieved by introducing a data auditing mechanism to regularly detect and monitor data access records to ensure the compliance and security of data use. When abnormal access is found, an alarm is issued in time and corresponding measures are taken to ensure that teachers' teaching resources are not accessed by unauthorized persons.

[0017] In the question-setting stage of the large-model question-setting and question-judgment module, the question type, subject and related knowledge points are input through the interactive interface. The system vectorizes these inputs by calling the local embedding model. The system uses the pre-trained BGE model to vectorize the user input. After that, the system performs a cosine similarity query in the vector database to determine the most relevant teaching materials, thereby ensuring that the generated questions are targeted and adaptable. During this process, the system will also return metadata related to the input content to assist teachers in further modification and optimization.

[0018] The large model question-setting and question-judgment module automatically selects the optimal scoring standard prompts for different subjects and question types in the judging function. Based on chain thinking technology, the system can analyze the rationality of the answers from different dimensions. Then, by combining the expert scoring criteria, the system conducts comprehensive scoring from multiple aspects such as content accuracy, logical rigor and language standardization to generate objective and fair evaluation results.

[0019] Improve efficiency: By optimizing the retrieval algorithm and model performance, the system response speed is significantly improved to meet real-time requirements; Enhanced accuracy: Using fine-tuned large models and rich teaching resources, the generated questions and judgment results are more targeted and accurate, meeting the teaching objectives; Ensure security: adopt data audit mechanism to ensure the security of teaching resources and user data, and prevent data leakage and unauthorized access; Improve scalability: Supports teaching resources in various formats, adopts a distributed vector database architecture, and has stable system performance that can adapt to large-scale data processing needs. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The present invention will be further described below in conjunction with the accompanying drawings: Figure 1 It is a system structure block diagram of the present invention; Figure 2 The present invention is a flowchart of the use of the present invention. DETAILED DESCRIPTION

[0021] like Figure 1 and 2As shown, the present invention provides a RAG question-setting and automatic judgment system based on a large model. The overall architecture of the RAG question-setting and automatic judgment system is first constructed, and the overall architecture includes three modules: a large model local deployment and fine-tuning module, a teaching resource vectorization storage module, and a large model question-setting and judgment module. The overall architecture is based on the traditional RAG solution, and selects the local Faiss or Milvus vector database because it has efficient query performance and excellent scalability when processing high-dimensional vector retrieval, which is particularly suitable for the needs of the education field. The system ensures effective storage and fast access to files in various formats (such as PDF, DOCX, CSV, etc.) by vectorizing the teaching materials in blocks, significantly improving the retrieval speed and accuracy of information.

[0022] In the question-setting phase, users input question types, subjects, and related knowledge points through a friendly interactive interface, and the system vectorizes these inputs by calling the local embedding model. Here, the system uses a pre-trained BGE model to vectorize user input to capture richer semantic information. After that, the system performs a cosine similarity query in the vector database to determine the most relevant teaching materials, thereby ensuring that the generated questions are targeted and adaptable. During this process, the system will also return metadata related to the input content to assist teachers in further modification and optimization.

[0023] In terms of the question-judging function, the system automatically selects the optimal scoring criteria for different subjects and question types, and based on chain thinking (CoT) technology, the system can analyze the rationality of answers from different dimensions, such as the reasoning process, logical structure, and completeness of specific details. By combining the expert scoring criteria, the system comprehensively scores from multiple aspects such as content accuracy, logical rigor, and language standardization to generate objective and fair evaluation results.

[0024] The overall architecture has the following advantages: Use local vector database: Based on high-performance vector databases such as Faiss or Milvus, it supports efficient retrieval and scalability of high-dimensional vectors, and is suitable for large-scale data processing needs in the education field.

[0025] Support for multi-format teaching resources: vectorize teaching materials into blocks, and support storage and quick access to files in multiple formats (such as PDF, DOCX, CSV, etc.).

[0026] Optimized information retrieval: Through vectorized retrieval, the speed and accuracy of information retrieval are improved, ensuring that the generated questions and judgment results are targeted and adaptable.

[0027] The large model local deployment and fine-tuning module includes the local deployment and fine-tuning of the pre-trained large model, aiming to improve the performance and applicability of the model in specific teaching scenarios. The specific steps are as follows: a. Data collection and preprocessing Data collection: The system collects a wide range of teaching data, covering different subjects and knowledge points, to enhance the model's adaptability to diverse problems.

[0028] Data cleaning and labeling: Strict data cleaning and manual labeling are performed on the collected data to ensure the accuracy and quality of the data. In order to ensure that the model can adapt to the latest teaching content, the new data can also be adapted through the incremental learning mechanism to gradually update the model parameters.

[0029] Data enhancement and segmentation: To avoid data shortage or bias, data enhancement techniques such as synonym replacement and sentence conversion are used. At the same time, the data is divided into training set, validation set and test set in a certain proportion to ensure the reliability of the subsequent training process.

[0030] This part enables the system to widely collect teaching data from different subjects and knowledge points, and conduct strict data cleaning and labeling to ensure the accuracy and quality of the data. In order to improve the continuous learning ability of the model, it supports the incremental learning mechanism to ensure that the model can adapt to new teaching content and maintain timeliness and adaptability.

[0031] b. Model selection and loading Model selection: According to the characteristics and needs of the teaching task, select a suitable pre-trained model (such as Qwen2.5, Meta-Llama-3, glm-4, gemma-2, etc.). When selecting a model, comprehensively consider the model's inference speed, memory usage, and performance to ensure that it runs efficiently in the local environment.

[0032] Load pre-trained weights: Load the pre-trained weights of the selected model as the initial parameters to reduce training time and improve the initial performance of the model. To meet the needs of running large models locally, you can combine quantization technology (such as INT8 quantization) to optimize model performance.

[0033] This part selects a suitable pre-trained model according to the specific teaching task, and combines the LoRA (Low-Rank Adaptation) technology to reduce the number of training parameters and memory requirements through low-rank matrix adjustment. This not only improves the efficiency of localized deployment, but also combines factors such as inference speed and memory usage to ensure the efficient operation of the model in the local environment. Load the pre-trained weights as the initial parameters, and combine the LoRA technology to optimize the training time and initial performance.

[0034] c. Definition of fine-tuning tasks Fine-tuning strategy selection: During the fine-tuning process, LoRA technology is introduced to replace the update of the complete weights with the adjustment of the low-rank matrix, thereby reducing the amount of training parameters and significantly reducing memory requirements and computing costs.

[0035] Loss function and optimization algorithm: Set a suitable loss function (such as cross entropy loss) according to the teaching objectives, and select a suitable optimization algorithm (such as AdamW or AdaFactor). In order to deal with the overfitting problem, early stopping method and learning rate scheduling strategy are introduced to accelerate model convergence and improve generalization ability.

[0036] Adaptability evaluation indicators: Multiple indicators such as F1 score, precision and recall are used to comprehensively evaluate the adaptability and accuracy of the model to the teaching task during the fine-tuning process to ensure its stable performance.

[0037] Based on the teaching objectives, this part sets appropriate loss functions and optimization algorithms, and uses LoRA technology to fine-tune the low-rank matrices of specific model layers (such as attention layers and fully connected layers). LoRA technology significantly reduces computing costs and memory requirements while ensuring performance. In addition, formulating indicators to evaluate model performance and introducing early stopping and learning rate scheduling can prevent overfitting and accelerate model convergence.

[0038] d. Training and validation process Model training: During fine-tuning, LoRA technology is applied to specific layers of the model, such as attention layers and fully connected layers, to reduce the amount of training parameters and speed up training.

[0039] Cross-validation: Through the K-fold cross-validation method, the model performance is continuously evaluated and optimized, and the fine-tuning parameters are adjusted to ensure that the model can maintain stable performance on different data sets.

[0040] Performance evaluation and tuning: Use indicators such as F1 score, accuracy, and recall rate to comprehensively evaluate the model. If the performance does not meet expectations, adjust LoRA-related hyperparameters or resample data for further optimization.

[0041] Training termination and saving: When the model reaches the set performance indicators, save the model weights and LoRA adaptation layer parameters for subsequent calls and continuous optimization.

[0042] In this part, LoRA technology is applied to specific model layers during model training to reduce the number of training parameters and speed up training. Model performance is continuously evaluated through cross-validation, and parameters are optimized for the fine-tuning process of the model. The model is fully evaluated using multiple indicators such as F1 score, accuracy, and recall to ensure its accuracy and stability in teaching tasks. If the performance does not meet expectations, it can be optimized by adjusting the relevant hyperparameters of LoRA or resampling.

[0043] e. Localized deployment and inference optimization Reasoning speed optimization: Taking into account the small memory footprint of the LoRA technology model, the optimized model is deployed locally, and model pruning, distillation and other methods are used to further improve the reasoning speed.

[0044] Real-time update and adaptive fine-tuning: The system supports rapid fine-tuning through LoRA, combined with incremental learning mechanism, and regularly updates the model to ensure its timeliness and adaptability in teaching scenarios.

[0045] This part uses the low memory usage characteristics of LoRA technology to deploy the optimized model to the local operating environment, and combines methods such as model pruning and knowledge distillation to further improve the reasoning speed. The system also supports fast fine-tuning based on LoRA, combined with incremental learning mechanisms to ensure the timeliness and adaptability of the model in different teaching scenarios.

[0046] The teaching resource vector storage module is used to strictly vectorize the teacher's teaching resources before storing them in the vector database to ensure efficient resource retrieval and privacy protection. The specific steps are as follows: a. Data upload and storage: Teachers upload teaching materials through the system, and the system stores them in the sqlite database for easy management and query.

[0047] b. Vectorization: When the teacher logs in, the system automatically calls the embedded model to vectorize the data in sqlite, generates high-dimensional vectors and stores them in the vector database. This process not only improves the data access efficiency, but also enhances the subsequent similarity query capabilities.

[0048] c. Data security: To ensure data security, the system regularly performs snapshot backups of the vector database to ensure rapid recovery in the event of data damage or loss.

[0049] d. Data audit and monitoring: In addition, the system introduces a data audit mechanism to regularly detect and monitor data access records to ensure the compliance and security of data use. When abnormal access is found, the system can issue an alarm in time and take corresponding measures to ensure that teachers' teaching resources are not accessed by unauthorized personnel. In order to cope with the storage needs of large-scale teaching resources, the system adopts a distributed vector database architecture, which can efficiently manage and query a large amount of vector data, ensuring the stability of system performance as the amount of data continues to grow.

[0050] The large model question generation and judgment module aims to generate high-quality questions and automatically score through the large model to improve teaching effectiveness and efficiency. The specific steps are as follows: a. Question generation process: Input parsing: User input includes information such as question type, subject, knowledge points, etc. The system parses these inputs through the preprocessing module and converts them into a format suitable for model understanding.

[0051] Context construction: The system retrieves relevant context information from the teaching resource vector storage module, including course content, learning objectives, and students’ knowledge mastery, to ensure that the generated questions meet students’ actual learning needs.

[0052] Large model generation: Using a fine-tuned large model, combined with contextual information and user input, the system generates questions that meet the requirements and their corresponding answers and explanations. During this process, the system uses techniques such as sampling and temperature adjustment to control the diversity and creativity of the generated content.

[0053] Quality assessment and optimization: Generated questions are reviewed through the built-in quality assessment module, which mainly evaluates the clarity, logic and difficulty of the questions. If necessary, they are automatically adjusted and optimized to ensure that the final output questions meet the teaching standards. To ensure the quality of generated questions, the system establishes a feedback mechanism to regularly collect teacher scores and student performance data, and use this feedback information to adjust and optimize the generation strategy of the large model, thereby improving the relevance and effectiveness of the questions.

[0054] b. Question judging process: Answer input and analysis: The system receives answers submitted by students and pre-processes them through natural language processing technology to identify keywords, syntactic structures and logical relationships to ensure that the system can accurately understand students' answers.

[0055] Application of scoring criteria: The system automatically selects appropriate scoring criteria prompts based on the question type and subject. These prompts are derived from the scoring criteria formulated by experts and cover multiple dimensions such as accuracy, completeness and logic of the content.

[0056] Scoring execution: Using chain thinking (CoT) technology, the big model analyzes the rationality of students' answers from multiple aspects, including semantic matching, structural evaluation, and detail verification. The system generates scores for each criterion and calculates the final grade.

[0057] Feedback generation: The system automatically generates feedback based on the scoring results, points out the strengths and weaknesses of students' answers, and provides targeted learning suggestions and improvement measures to help students better understand knowledge points and improve their ability to answer questions. The system will automatically collect students' responses to feedback, including the adoption of improvement suggestions, and continuously adjust teaching content and question-setting strategies by analyzing these feedbacks, thus achieving a closed loop of personalized learning and continuous optimization.

[0058] c. Module integration and collaboration: Integration with teaching resource module: This module is tightly integrated with the teaching resource vector storage module to ensure that the latest teaching materials and student data can be called up in real time during question generation and judgment, thereby improving the adaptability and intelligence of the system.

[0059] Real-time data feedback and learning path adjustment: After judging the questions, the system will feed back the analysis results to the learning path planning module, dynamically adjust the student's learning plan based on the student's performance, and ensure the effective implementation of personalized learning.

[0060] The system of the present invention is divided into two roles: teacher and student. The teacher's use process is as follows: 1. Log in to the system.

[0061] 2. The teacher logs into the system using his / her account and password and enters the teacher work interface.

[0062] 3. Upload teaching resources Step 1: Data preparation Teachers prepare teaching resources that need to be uploaded, supporting multiple formats such as PDF, DOCX, CSV, etc.

[0063] Step 2: Upload your data In the "Teaching Resource Management" module, click the "Upload Resources" button, select the file and fill in the relevant information.

[0064] Step 3: Resource classification and labeling Classify the uploaded materials and add tags (such as subject, knowledge point, grade) to facilitate subsequent retrieval and use.

[0065] Step 4: Complete Upload The system stores the data in the SQLite database and starts vectorization processing.

[0066] 4. Vectorization of teaching resources Step 1: Call the embedding model The system automatically calls the embedding model to vectorize the uploaded data and generate a high-dimensional vector representation.

[0067] Step 2: Store in vector database The generated vectors are stored in the Faiss or Milvus vector database to support efficient similarity retrieval.

[0068] 5. Generate test questions Step 1: Enter the question module Select the "Smart Question Generator" function in the system interface.

[0069] Step 2: Enter the question requirements Fill in information such as question type, subject, knowledge point, difficulty level, etc.

[0070] Step 3: Submit a question request Click the "Generate Questions" button and the system will start generating questions.

[0071] Step 4: System generates questions Parsing input: The system parses the requirements input by the teacher and converts them into a format that the model can understand.

[0072] Retrieve relevant information: Retrieve relevant teaching resources from the vector database and build contextual information.

[0073] Call the big model to generate: Use the fine-tuned big model to generate questions, answers, and analysis that meet the requirements.

[0074] Quality assessment: Perform quality assessment on the generated questions to ensure their accuracy and applicability.

[0075] Step 5: View and edit questions Teachers can view the generated questions and modify and optimize them.

[0076] Step 6: Save the question Save satisfactory questions to the question bank or test paper.

[0077] 6. Manage test papers Create a test paper: Add the generated questions to a new test paper and set the test paper information.

[0078] Edit the test paper: adjust the order of questions and modify the test paper content.

[0079] Release test paper: Release the test paper to the specified students or classes.

[0080] The student usage process is as follows: 1. Log in to the system.

[0081] 2. Students log in to the system using their account and password and enter the student interface.

[0082] 3. Check the test papers to be completed.

[0083] 4. Check your pending exams in "My Assignments" or "Exam Center".

[0084] 5. Start answering questions Step 1: Access the test paper Click the test name to enter the answering interface.

[0085] Step 2: Answer the questions Enter the answer according to the question requirements. It supports multiple forms such as text and pictures.

[0086] Step 3: Submit your answers After completing all questions, click the "Submit" button.

[0087] 6. Check ratings and feedback Step 1: The system automatically judges the questions The system automatically scores objective questions and conducts intelligent review of subjective questions.

[0088] Step 2: Check your results Students can check their scores in "Score Query".

[0089] Step 3: Review Feedback Click on a specific question to view detailed answer analysis and improvement suggestions.

[0090] 7. Personalized learning Adjust study plans based on feedback; According to the system's suggestions, review weak knowledge points and complete recommended exercises.

[0091] The present invention achieves efficient retrieval and secure storage of teaching resources by locally deploying and fine-tuning large models, using vectorized storage and security mechanisms. The system can generate targeted questions based on user input, and automatically judge and provide feedback on students' answers, which improves teaching efficiency and quality and has good application prospects.

[0092] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A RAG question setting and automatic question judging system based on a large model, characterized by: The overall architecture is established based on RAG, which includes a large model local deployment and fine-tuning module, a teaching resource vector storage module, and a large model question setting and judgment module. The large model local deployment and fine-tuning module is used to implement data collection and preprocessing, model selection and loading, fine-tuning tasks, model training and verification, and localized deployment and reasoning optimization; The teaching resource vector storage module is used to realize data uploading and storage, data vector processing and data auditing and monitoring; The large model question setting and judging module is used to realize question generation and automatic question judging.

2. A RAG question setting and automatic question judging system based on a large model according to claim 1, characterized in that: Model selection and loading of large model local deployment and fine-tuning modules selects a suitable pre-trained model according to the specific teaching task, loads the pre-trained weights as the initial parameters, and optimizes the model performance in combination with quantization technology.

3. A RAG question setting and automatic question judging system based on a large model according to claim 2, characterized in that: The fine-tuning tasks of the local deployment and fine-tuning modules of large models are based on teaching objectives, set appropriate loss functions and optimization algorithms, and use LoRA technology to fine-tune the low-rank matrix of specific model layers. Early stopping and learning rate scheduling are introduced to prevent overfitting and accelerate model convergence. Indicators for evaluating model performance are also formulated.

4. A RAG question setting and automatic question judging system based on a large model according to claim 3, characterized in that: The process of model training and verification of large model local deployment and fine-tuning module is as follows: Model training: During fine-tuning, LoRa technology is applied to specific layers of the model; Cross-validation: Use the K-fold cross-validation method to continuously evaluate and optimize model performance and fine-tune parameters to ensure that the model can maintain stable performance on different data sets; Performance evaluation and tuning: Use multiple different indicators to comprehensively evaluate the model. If the performance does not meet expectations, adjust LoRA-related hyperparameters or resample data for further optimization; Training termination and saving: When the model reaches the set performance indicators, save the model weights and LoRA adaptation layer parameters.

5. A RAG question setting and automatic question judging system based on a large model according to claim 4, characterized in that: Local deployment and inference optimization of large model local deployment and fine-tuning modules include the following two parts: Reasoning speed optimization: Deploy the optimized model locally and use model pruning and / or distillation methods to further improve the reasoning speed. Real-time update and adaptive fine-tuning: Quickly fine-tune through LoRA, combined with incremental learning mechanism, and update the model regularly.

6. A RAG question setting and automatic question judging system based on a large model according to claim 1, characterized in that: The teaching resource vectorization storage module stores the uploaded teaching materials in the sqlite database, and when the teacher logs into the system, it automatically calls the local embedding model to vectorize the materials in the sqlite database, generates high-dimensional vectors and stores them in the vector database.

7. A RAG question setting and automatic question judging system based on a large model according to claim 6, characterized in that: Data auditing and monitoring in the vector storage module of teaching resources is achieved by introducing a data auditing mechanism to regularly detect and monitor data access records to ensure the compliance and security of data use. When abnormal access is found, an alarm is issued in time and corresponding measures are taken to ensure that teachers' teaching resources are not accessed by unauthorized persons.

8. The RAG question setting and automatic question judging system based on a large model according to claim 6 is characterized by: In the question-setting stage of the large-model question-setting and question-judgment module, the question type, subject and related knowledge points are input through the interactive interface. The system vectorizes these inputs by calling the local embedding model. The system uses the pre-trained BGE model to vectorize the user input. After that, the system performs a cosine similarity query in the vector database to determine the most relevant teaching materials. During this process, the system will also return metadata related to the input content to assist teachers in further modification and optimization.

9. The RAG question setting and automatic question judging system based on a large model according to claim 6 is characterized by: The large model question-setting and question-judgment module automatically selects the optimal scoring standard prompts for different subjects and question types in the judging function. Based on chain thinking technology, the system can analyze the rationality of the answers from different dimensions. Then, by combining the expert scoring criteria, the system conducts comprehensive scoring from multiple aspects such as content accuracy, logical rigor and language standardization to generate objective and fair evaluation results.

Images