Academic affairs big model system and data cleaning method

Through the design of the academic affairs big model system, using data cleaning and large language model fine-tuning technology, dynamically selecting the pre- or post-model for answering, solving the high-cost problems of manually answering questions in the existing academic affairs system, realizing automatic answering and accurate answering Answer, reducing labor costs.

CN118377877BActive Publication Date: 2024-11-08GUANGDONG WOTU EDUCATION TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410543363.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-05-05
Publication Date
2024-11-08
Estimated Expiration
2044-05-05

AI Technical Summary

Technical Problem

In the existing academic affairs system, when students or teachers encounter uncommon academic affairs problems, it is difficult to find corresponding administrative notices or regulations, resulting in manual answers and high labor costs.

Method used

A large-scale model system for academic affairs was designed, including a data cleaning subsystem, a pre-scale model system and a post-scale model system. The general large-language model is fine-tuned through the LoRA method, combined with the preset offset matrix and feature extraction method. Dynamically select the right model to answer, reducing the need for manual answers.

Benefits of technology

It realizes automatic answering of academic affairs questions, reduces labor costs, improves the efficiency and accuracy of answers, and avoids wasting computing resources caused by the need to supplement identity and campus information every time it enters.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118377877B_ABST
    Figure CN118377877B_ABST
Patent Text Reader

Abstract

The present application relates to the field of big model technology, and in particular to a big model for teaching affairs and a data cleaning method. The big model system for teaching affairs includes: a data cleaning subsystem, which cleans the teaching affairs data according to preset data cleaning rules to obtain first cleaned data and second cleaned data; a front big model system; a rear big model system; a big model retrieval system, which calculates the cosine similarity of the feature vector of the question raised by the user and the offset matrix of the front big model system; if the cosine similarity of the feature vector of the question raised by the user and the offset matrix of the front big model system is greater than a preset first threshold, the front big model system is used to answer the question raised by the user. The big model system for teaching affairs provided by the present application can automatically answer the teaching affairs questions raised by users, reducing labor costs.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of large model technology, and in particular to a large academic affairs model and a data cleaning method. Background Art

[0002] At present, the delivery of teaching activities and various teaching administrative documents is through the website announcement and through the counselor to the class group where the students are. When students or teachers encounter some uncommon teaching problems, it is usually difficult to find the corresponding administrative notices or regulations.

[0003] And because everyone's identity and grade are different, the answers to the same question asked by different people may correspond to the answers in different chapters of the administrative notice or regulations.

[0004] Therefore, the existing solutions to teaching affairs problems need to be done manually, which results in high labor costs. Summary of the invention

[0005] In order to solve the above technical problems or at least partially solve the above technical problems, the present application provides a large academic affairs model system and a data cleaning method, which can reduce labor costs.

[0006] In a first aspect, the present application provides a teaching affairs model system, the teaching affairs model system comprising:

[0007] A data cleaning subsystem, which cleans the teaching data according to a preset data cleaning rule to obtain first cleaned data and second cleaned data, wherein the first cleaned data does not include teaching data related to identity or campus, and the second cleaned data is teaching data related to identity or campus;

[0008] A front-end large model system is used to fine-tune the general large language model through the LoRA method using the first cleaned data to obtain a front-end large model system;

[0009] The post-large model system uses the second cleaning data to fine-tune the general large language model through the LoRA method to obtain the post-large model system;

[0010] The large model retrieval system obtains the offset matrix of the front large model system through a preset offset matrix extraction method, and obtains the feature vector of the question raised by the user through a preset feature extraction method; calculates the cosine similarity between the feature vector of the question raised by the user and the offset matrix of the front large model system; if the cosine similarity between the feature vector of the question raised by the user and the offset matrix of the front large model system is greater than a preset first threshold, the front large model system is used to answer the question raised by the user.

[0011] Optionally, the large model retrieval system further includes, if the cosine similarity between the feature vector of the question raised by the user and the offset matrix of the front large model system is less than a preset first threshold:

[0012] Generate a feature identification question about the user, the feature identification question including the question asked by the user and the campus and identity of the user;

[0013] The offset matrix of the post-large model system is obtained by a preset offset matrix extraction method, and the feature vector of the user's feature identification problem is obtained by a preset feature extraction method;

[0014] Calculate the cosine similarity between the feature vector of the user's feature identification problem and the offset matrix of the post-large model system;

[0015] If the cosine similarity between the feature vector of the user's feature identification question and the offset matrix of the post-large model system is greater than a preset second threshold, the post-large model system is used to answer the question raised by the user.

[0016] Optionally, the large model retrieval system also includes, if the cosine similarity between the feature vector of the user's feature identification question and the offset matrix of the post-large model system is less than a preset second threshold, determining that the academic affairs large model system cannot answer the user's question.

[0017] Optionally, the preset offset matrix extraction method includes the following steps:

[0018] Obtain all ΔW matrices used in weight updates in the general large model system through the LoRA method;

[0019] All ΔW matrices are accumulated to obtain the offset matrix of the large model trained by the LoRA method based on the general large model system.

[0020] Optionally, the preset feature extraction method includes the following steps:

[0021] Use the tokenizer of the front-end large model to convert the question entered by the user into the first token;

[0022] Input the first token into the front-end large model, and obtain the vector representation of the question raised by the user in the word embedding layer through the word embedding layer of the front-end large model;

[0023] Obtain the layer node into which the ΔW matrix has been inserted in the front large model as the first adjustment layer node;

[0024] After the vector representation of the word embedding layer of the question raised by the user is forwarded through a preset number of first adjustment layer nodes in the front large model, the output of the first adjustment layer node is used as the feature vector of the question input by the user.

[0025] Optionally, the preset feature extraction method includes the following steps:

[0026] Use the tokenizer of the post-large model to convert the feature identification problem into the second token;

[0027] The second token is input into the post-large model, and the vector representation of the feature identification problem in the word embedding layer is obtained through the word embedding layer of the post-large model;

[0028] Obtain the layer node of the ΔW matrix inserted in the post-large model as the second adjustment layer node;

[0029] After the vector representation of the feature identification problem at the word embedding layer is forwardly passed through a preset number of second adjustment layer nodes in the post-large model, the output of the second adjustment layer node is used as the feature vector of the feature identification problem.

[0030] Optionally, the preset data cleaning rules include:

[0031] According to the types of academic documents, the academic documents are divided into school calendar documents, lecture notices, activity notices, commendation documents, service guidelines and rules and regulations;

[0032] Screen out the school calendar documents, commendation documents, and rules and regulations as part of the first cleansing data;

[0033] According to the selected preset campus and identity keyword list, the academic documents related to the campus or identity are screened out from the lecture notices, activity notices and service guidelines as the second cleansing data.

[0034] Optionally, the preset campus and identity keyword list includes: campus, grade and class.

[0035] In a second aspect, the present application provides a data cleaning method, which is performed by the data cleaning subsystem in the teaching affairs big model system described in the first aspect, and the data cleaning method includes:

[0036] Obtain the school's academic documents;

[0037] According to the types of academic documents, the academic documents are divided into school calendar documents, lecture notices, activity notices, commendation documents, service guidelines and rules and regulations;

[0038] Screen out the school calendar documents, commendation documents, and rules and regulations as part of the first cleansing data;

[0039] According to the selected preset campus and identity keyword list, the academic documents related to the campus or identity are screened out from the lecture notices, activity notices and service guidelines as the second cleansing data.

[0040] Compared with the prior art, the technical solution provided by this application has the following advantages:

[0041] The teaching affairs big model system provided in this application is based on the characteristics of teaching affairs files. The front-end big model system is trained for teaching affairs files that are not related to identity and campus, and the rear-end big model system is trained for teaching affairs files related to identity and campus. The appropriate model can be dynamically selected for answering according to the specific situation. For questions that are not related to identity and campus, the front-end big model system can be used directly for quick answers, saving time and resources. For complex questions or questions that need to consider identity and campus factors, the user's identity and campus information can be added, and the rear-end big model system can be selected for deeper processing and answers, providing more accurate and personalized answers.

[0042] In addition, the large model retrieval system provided by this application can automatically select the front or back large model system to answer according to the questions and identity information raised by the user. This mechanism avoids the problem of increasing the length of the token due to the need to supplement identity and campus information every time input, and selects the appropriate model according to different situations, reducing the waste of computing resources and improving the efficiency and performance of the system.

[0043] Therefore, the teaching affairs model system provided by this application can automatically answer teaching affairs questions raised by users, reducing labor costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 A schematic diagram of the structure of the teaching affairs model system provided in the embodiment of the present application;

[0045] Figure 2 A schematic diagram of the process of executing the teaching affairs model system provided in the embodiment of the present application. DETAILED DESCRIPTION

[0046] The technical solution in this application will be described below in conjunction with the accompanying drawings.

[0047] In the following description, many specific details are set forth to facilitate a full understanding of the present application, but the present application can also be implemented in other ways different from those described herein; obviously, the embodiments in the specification are only part of the embodiments of the present application, not all of the embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0048] The teaching affairs model system provided in the embodiment of the present application is deployed in a server within a school or in the cloud to provide teaching affairs consulting services to user groups such as students and faculty members, thereby reducing the number of times teaching affairs questions are answered manually and saving labor costs.

[0049] In addition, in order to build a large model for teaching affairs that reduces labor costs, it is necessary to accurately answer questions raised by users, so the teaching affairs documents cannot be directly submitted to the general large language model for learning. This is because, for teaching affairs issues, usually due to different identities and campuses of each person, the same question may have different correct answers. For example, in different campuses, the process and location for handling the same business may be different, students of different grades may have different course schedules, and people with student and teacher identities may also have different administrative processes.

[0050] The large language model performs well only when there are accurate answers to questions. In the case of multiple answers, the large model is prone to give vague answers to users' questions. Therefore, the large teaching model obtained by directly submitting the teaching documents to the general large language model for learning does not perform well.

[0051] Furthermore, if the user's identity information and campus information are automatically supplemented every time the user asks a question, the length of the token input each time will become longer, causing the model to consume more computing resources.

[0052] The present application embodiment provides a teaching affairs model system, and the teaching affairs model system includes:

[0053] S202: Acquire the educational affairs file, and clean it to obtain first cleaned data and second cleaned data.

[0054] The data cleaning subsystem cleans the teaching data according to preset data cleaning rules to obtain first cleaned data and second cleaned data, wherein the first cleaned data does not include teaching data related to identity or campus, and the second cleaned data is teaching data related to identity or campus.

[0055] Specifically, the preset data cleaning rules include:

[0056] According to the types of academic documents, the academic documents are divided into school calendar documents, lecture notices, activity notices, commendation documents, service guidelines and rules and regulations;

[0057] Screen out the school calendar documents, commendation documents, and rules and regulations as part of the first cleansing data;

[0058] According to the selected preset campus and identity keyword list, the academic documents related to the campus or identity are screened out from the lecture notices, activity notices and service guidelines as the second cleansing data.

[0059] Specifically, the preset campus or identity keyword list includes: campus, grade and class.

[0060] Specifically, in the embodiment of the present application, the specific steps of data cleaning are that the academic affairs staff first selects some academic affairs documents that can be used for consultation on academic affairs issues. The types of these academic affairs documents generally include school calendar files, lecture notices, event notices, commendation documents, service guidelines and rules and regulations type files.

[0061] Then, regular expressions are used to filter out lecture notices, activity notices, or service instructions files with "campus", "grade" or "class" from files of the lecture notice, activity notice, and service instruction types. Subsequently, it is manually determined whether these files are all related to the identity or campus. After excluding files containing keywords but not related to the identity or campus, the second cleaned data is obtained.

[0062] The remaining files of the academic affairs file are used as the first cleaned data. That is, the school calendar files, commendation files and rules and regulations are directly screened out as part of the first cleaned data. Then, the files containing keywords in the above steps but not related to identity or campus are used as another part of the first cleaned data.

[0063] S204: Train the front large model system and the rear large model system according to the first cleaning data and the second cleaning data respectively.

[0064] The front-end large model system uses the first cleaning data to fine-tune the general large language model through the LoRA method to obtain the front-end large model system.

[0065] The post-large model system uses the second cleaning data to fine-tune the general large language model through the LoRA method to obtain the post-large model system.

[0066] In the embodiments of the present application, the LoRA (Low-Rank Adaptation) method used is an existing technology for fine-tuning a general large language model by inserting a low-rank matrix into the Transformer layer of the general large language model. Its specific working principle will not be described in detail in this application document.

[0067] In this embodiment, according to the types of files in the first cleaning data and the second cleaning data, it is selected to directly feed the files into the general large language model or to train the general large language model in the form of manually constructing question-answer pairs. In the embodiment of the present application, the school calendar files and commendation files can be directly fed into the general large language model system for training. For the rules and regulations, due to the excessive content of the files, it is possible to construct question-answer pairs and then feed them into the general large language model for training. For example:

[0068] Q1: What are the types of disciplinary sanctions?

[0069] A1: School Rule A. According to Article A of the school rule, the types of disciplinary sanctions include warning, demerit, probation and expulsion.

[0070] Q2: What is the punishment for cheating in an exam?

[0071] A2: School Rule B. According to Article B of the school rule, cheating in exams will result in a demerit and the parents will be notified.

[0072] In the embodiment of the present application, for files of lecture notices, event notices, and service instructions, since the size of each file is not large, it is chosen to directly feed them into the general large language model system for training.

[0073] In another embodiment, the teaching knowledge of the general large language model can be supplemented by directly feeding the first cleaning data and the second cleaning data into the general large language model respectively, so as to use the LoRA method to complete the fine-tuning of the large language model, and then obtain the front large model system and the post large model system respectively.

[0074] S206: Calculate the cosine similarity between the feature vector of the question raised by the user and the offset matrix of the front large model system. If it is greater than a preset first threshold, use the front large model system to answer the question raised by the user.

[0075] The large model retrieval system obtains the offset matrix of the front large model system through a preset offset matrix extraction method, and obtains the feature vector of the question raised by the user through a preset feature extraction method; calculates the cosine similarity between the feature vector of the question raised by the user and the offset matrix of the front large model system; if the cosine similarity between the feature vector of the question raised by the user and the offset matrix of the front large model system is greater than a preset first threshold, the front large model system is used to answer the question raised by the user.

[0076] Specifically, the preset offset matrix extraction method includes the following steps:

[0077] Obtain all ΔW matrices used in weight updates in the general large model system through the LoRA method;

[0078] All ΔW matrices are accumulated to obtain the offset matrix of the large model trained by the LoRA method based on the general large model system.

[0079] In the LoRA method, the ΔW matrix is ​​actually a low-rank matrix inserted in the Transformer layer in the general large language model by the LoRA method, and the updating process of the weight matrix is ​​that the updated weight matrix is ​​equal to the sum of the original weight matrix and the low-rank matrix ΔW.

[0080] The low-rank matrix ΔW in LoRA describes the transformation of the model parameters from the original state to the state after training for a specific domain. This transformation reflects the adjustment of the model to adapt to the data in the academic affairs domain.

[0081] Based on this, figuratively speaking, if the initial capability of the original general large language model is regarded as a circle, then the capability of the fine-tuned front large model or the back large model can be regarded as ellipses of different shapes whose capabilities are shifted toward the academic affairs field. Then the offset matrix obtained by accumulating all ΔW matrices can be used to describe the direction of the capability shift of the front large model or the back large model.

[0082] Specifically, the preset feature extraction method includes the following steps:

[0083] Use the tokenizer of the front-end large model to convert the question entered by the user into the first token;

[0084] Input the first token into the front-end large model, and obtain the vector representation of the question raised by the user in the word embedding layer through the word embedding layer of the front-end large model;

[0085] Obtain the layer node into which the ΔW matrix has been inserted in the front large model as the first adjustment layer node;

[0086] After the vector representation of the word embedding layer of the question raised by the user is forwarded through a preset number of first adjustment layer nodes in the front large model, the output of the first adjustment layer node is used as the feature vector of the question input by the user.

[0087] Specifically, in the embodiments of the present application, the preset number is 5 or 6.

[0088] In the embodiment of the present application, the first threshold is obtained by the following method: artificially select a part of the questions that can be answered and cannot be answered by the front large model. According to the method provided in the above embodiment, the cosine similarity of the feature vector and the offset matrix of these questions is obtained respectively. The calculated similarity is divided into two groups, one group is the similarity of the questions that can be answered by the model, and the other group is the similarity of the questions that cannot be answered by the model. A value that can distinguish the two groups of similarities is selected as the first threshold.

[0089] Its specific working principle is that the ΔW matrix is ​​the change of the model parameters of the layer nodes inserted into the ΔW matrix, and the ΔW matrix is ​​finally adjusted by loss regression in the LoRA method. In this case, the sum of all ΔW matrices, namely the offset matrix, represents the linear change of the model parameters. When the feature vectors of the questions raised by the user are transmitted through the layer nodes of the front-end large model, the front-end large model will perform linear transformation on these feature vectors to generate output. Therefore, the feature vectors and offset matrices of the questions raised by the user after transmission belong to the same linear transformation space.

[0090] At the same time, in an embodiment of the present application, the output of the first adjustment layer node through a preset number is selected as the feature vector of the question raised by the user, and the degree of influence of the feature vector by the domain-specific adjustment of the model is used to help distinguish whether the front-end large model can answer the question raised by the user. Because for problems in the field of academic affairs, if the front-end large model has learned effective representations for problems in the field of academic affairs through its domain-specific training, its output in the first few layers can better map these problems in the field of academic affairs to a similar feature space. For problems outside the field of academic affairs, it is impossible to map them to a feature space close to the field of academic affairs. This improves the accuracy of judging whether the front-end large model can answer the problem based on cosine similarity.

[0091] S208: If the cosine similarity is less than the preset first threshold, a feature identification question is generated to determine whether the cosine similarity between the feature vector of the feature identification question and the post-large model system is greater than a second threshold. If so, the post-large model is selected to answer the question raised by the user.

[0092] Specifically, the large model retrieval system further includes, if the cosine similarity between the feature vector of the question raised by the user and the offset matrix of the front large model system is less than a preset first threshold:

[0093] Generate a feature identification question about the user, the feature identification question including the question asked by the user and the campus and identity of the user;

[0094] The offset matrix of the post-large model system is obtained by a preset offset matrix extraction method, and the feature vector of the user's feature identification problem is obtained by a preset feature extraction method;

[0095] Calculate the cosine similarity between the feature vector of the user's feature identification problem and the offset matrix of the post-large model system;

[0096] If the cosine similarity between the feature vector of the user's feature identification question and the offset matrix of the post-large model system is greater than a preset second threshold, the post-large model system is used to answer the question raised by the user.

[0097] In an embodiment of the present application, a feature identification question is obtained by directly adding the user's identity and text information of the campus where the user is located after the user asks the question.

[0098] Specifically, the preset feature extraction method includes the following steps:

[0099] Use the tokenizer of the post-large model to convert the feature identification problem into the second token;

[0100] The second token is input into the post-large model, and the vector representation of the feature identification problem in the word embedding layer is obtained through the word embedding layer of the post-large model;

[0101] Obtain the layer node of the ΔW matrix inserted in the post-large model as the second adjustment layer node;

[0102] After the vector representation of the feature identification problem at the word embedding layer is forwardly passed through a preset number of second adjustment layer nodes in the post-large model, the output of the second adjustment layer node is used as the feature vector of the feature identification problem.

[0103] Specifically, the feature identification question is passed to the post-large model system so that the post-large model system can answer the user's question.

[0104] Among them, the preset number, the characteristic vector of the feature identification problem and the acquisition process and effect of the offset matrix of the post-large model system are consistent with those in the above embodiment and will not be described in detail here.

[0105] S210: Otherwise, it is determined that the academic affairs model system cannot answer the user's question.

[0106] Specifically, the large model retrieval system further includes, if the cosine similarity between the feature vector of the user's feature identification question and the offset matrix of the post-large model system is less than a preset second threshold, determining that the teaching affairs large model system cannot answer the user's question. The subsequent approach of this step can be switched to manually answering questions that the teaching affairs large model system cannot answer.

[0107] In summary, the beneficial effects of the teaching affairs model provided by this application are:

[0108] The teaching affairs big model system provided in this application is based on the characteristics of teaching affairs files. The front-end big model system is trained for teaching affairs files that are not related to identity and campus, and the rear-end big model system is trained for teaching affairs files related to identity and campus. For questions that are not related to identity and campus, the front-end big model system can be used directly for quick answers, saving time and resources. For complex questions or questions that need to consider identity and campus factors, the user's identity and campus information can be added, and the rear-end big model system can be selected for deeper processing and answers to provide more accurate and personalized answers.

[0109] In addition, the large model retrieval system provided in the embodiment of the present application can automatically select the front or back large model system to answer according to the questions and identity information raised by the user. This mechanism avoids the problem of increasing the length of the token due to the need to supplement identity and campus information every time input, and selects the appropriate model according to different situations, reducing the waste of computing resources and improving the efficiency and performance of the system.

[0110] Therefore, the teaching affairs model system provided in the embodiment of the present application can automatically answer the teaching affairs questions raised by users, thereby reducing labor costs.

[0111] In a second aspect, an embodiment of the present application provides a data cleaning method, which is executed by the data cleaning subsystem in the teaching affairs big model system described in the above embodiment, and the data cleaning method includes:

[0112] Obtain the school's academic documents;

[0113] According to the types of academic documents, the academic documents are divided into school calendar documents, lecture notices, activity notices, commendation documents, service guidelines and rules and regulations;

[0114] Screen out the school calendar documents, commendation documents, and rules and regulations as part of the first cleansing data;

[0115] According to the selected preset campus and identity keyword list, the academic documents related to the campus or identity are screened out from the lecture notices, activity notices and service guidelines as the second cleansing data.

[0116] It should be noted that, in this article, relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. In addition, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "including one..." do not exclude the existence of other identical elements in the process, method, article or device including the elements. Moreover, in the description of the embodiments of the present application, unless otherwise specified, " / " means or, for example, A / B can represent A or B; "and / or" in this article is only a kind of association relationship describing the associated objects, indicating that there can be three relationships, for example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. Furthermore, in the description of the embodiments of the present application, “plurality” refers to two or more than two.

[0117] The above description is only a specific implementation of the present application, so that those skilled in the art can understand or implement the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments described herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. Academic affairs model system, It is characterized in that The teaching affairs model system includes: A data cleaning subsystem, which cleans the teaching data according to a preset data cleaning rule to obtain first cleaned data and second cleaned data, wherein the first cleaned data does not include teaching data related to identity or campus, and the second cleaned data is teaching data related to identity or campus; A front-end large model system is used to fine-tune the general large language model through the LoRA method using the first cleaned data to obtain a front-end large model system; The post-large model system uses the second cleaning data to fine-tune the general large language model through the LoRA method to obtain the post-large model system; The large model retrieval system obtains the offset matrix of the front large model system through a preset offset matrix extraction method, and obtains the feature vector of the question raised by the user through a preset feature extraction method; calculates the cosine similarity between the feature vector of the question raised by the user and the offset matrix of the front large model system; if the cosine similarity between the feature vector of the question raised by the user and the offset matrix of the front large model system is greater than a preset first threshold, the front large model system is used to answer the question raised by the user; If the cosine similarity between the feature vector of the question raised by the user and the offset matrix of the front-end large model system is less than the preset first threshold: Generate a feature identification question about the user, the feature identification question including the question asked by the user and the campus and identity of the user; The offset matrix of the post-large model system is obtained by a preset offset matrix extraction method, and the feature vector of the user's feature identification problem is obtained by a preset feature extraction method; Calculate the cosine similarity between the feature vector of the user's feature identification problem and the offset matrix of the post-large model system; If the cosine similarity between the feature vector of the user's feature identification question and the offset matrix of the post-large model system is greater than a preset second threshold, the post-large model system is used to answer the question raised by the user.

2. The teaching affairs model system according to claim 1, It is characterized in that The large model retrieval system also includes, if the cosine similarity between the feature vector of the user's feature identification problem and the offset matrix of the post-large model system is less than a preset second threshold, determining that the academic affairs large model system cannot answer the user's question.

3. The teaching affairs model system according to claim 1, It is characterized in that The preset offset matrix extraction method comprises the following steps: Obtain all ΔW matrices used in weight updates in the general large model system through the LoRA method; All ΔW matrices are accumulated to obtain the offset matrix of the large model trained by the LoRA method based on the general large model system.

4. The teaching affairs model system according to claim 1, It is characterized in that The preset feature extraction method comprises the following steps: Use the tokenizer of the front-end large model to convert the question entered by the user into the first token; Input the first token into the front-end large model, and obtain the vector representation of the question raised by the user in the word embedding layer through the word embedding layer of the front-end large model; Obtain the layer node into which the ΔW matrix has been inserted in the front large model as the first adjustment layer node; After the vector representation of the word embedding layer of the question raised by the user is forwarded through a preset number of first adjustment layer nodes in the front large model, the output of the first adjustment layer node is used as the feature vector of the question input by the user.

5. The teaching affairs model system according to claim 1, It is characterized in that The preset feature extraction method comprises the following steps: Use the tokenizer of the post-large model to convert the feature identification problem into the second token; The second token is input into the post-large model, and the vector representation of the feature identification problem in the word embedding layer is obtained through the word embedding layer of the post-large model; Obtain the layer node of the ΔW matrix inserted in the post-large model as the second adjustment layer node; After the vector representation of the feature identification problem at the word embedding layer is forwardly passed through a preset number of second adjustment layer nodes in the post-large model, the output of the second adjustment layer node is used as the feature vector of the feature identification problem.

6. The teaching affairs model system according to claim 1, It is characterized in that The preset data cleaning rules include: According to the types of academic documents, the academic documents are divided into school calendar documents, lecture notices, activity notices, commendation documents, service guidelines and rules and regulations; Screen out the school calendar documents, commendation documents, and rules and regulations as part of the first cleansing data; According to the selected preset campus and identity keyword list, the academic documents related to the campus or identity are screened out from the lecture notices, activity notices and service guidelines as the second cleansing data.

7. The teaching affairs model system according to claim 6, It is characterized in that The preset campus and identity keyword list includes: campus, grade and class.

8. Data cleaning methods, It is characterized in that The data cleaning method is performed by the data cleaning subsystem in the teaching affairs big model system according to any one of claims 1 to 5, and the data cleaning method includes: Obtain the school's academic documents; According to the types of academic documents, the academic documents are divided into school calendar documents, lecture notices, activity notices, commendation documents, service guidelines and rules and regulations; Screen out the school calendar documents, commendation documents, and rules and regulations as part of the first cleansing data; According to the selected preset campus and identity keyword list, the academic documents related to the campus or identity are screened out from the lecture notices, activity notices and service guidelines as the second cleansing data.