CN-122022192-A - Large model optimization method and device for education

CN122022192ACN 122022192 ACN122022192 ACN 122022192ACN-122022192-A

Abstract

The invention belongs to the technical field of intelligent education and relates to a large model optimization method and a large model optimization device for education, wherein the method comprises the steps of generating synthetic data based on Gaussian model fitting real data, evaluating the consistency of the synthetic data and the real data through distribution similarity measurement, obtaining the synthetic data meeting the consistency requirement, and forming a training set; and model parameters of the large model are updated by superposing Gaussian noise on a random gradient through a Lang-hundred thousand dynamics mechanism, and the large model is trained by combining a training set. The invention can reduce the probability of the illusion problem of the large model for education.

Inventors

MA YADONG
SHI BINBIN
Xiang Linggang
YOU JUNJIE
JIA JINGYUAN

Assignees

杭州海亮数字科技有限公司

Dates

Publication Date: 20260512
Application Date: 20260410

Claims (10)

1. A method for large model optimization for education, comprising: The synthetic data generating step is to generate synthetic data based on Gaussian model fitting real data, evaluate the consistency of the synthetic data and the real data through distribution similarity measurement, obtain synthetic data meeting the consistency requirement and form a training set; and a large model training step, namely, model parameters of the large model are updated by overlapping Gaussian noise on random gradients through a Lang-hundred-thousand dynamics mechanism, and the large model is trained by combining a training set.
2. The large model optimizing method for education of claim 1, wherein the synthetic data generating step includes: initializing a first parameter, wherein the first parameter comprises the number of clusters, a cluster center, cluster weights and a covariance matrix; Generating synthetic data by combining the real data according to the probability density function of the Gaussian mixture model based on the initialized first parameter; evaluating the distribution similarity of the synthesized data and the real data through posterior distribution KL divergence; And adding the synthesized data with the distribution similarity greater than the similarity threshold value into the training set.
3. The large model optimization method for education according to claim 2, wherein the gaussian mixture model probability density function is constructed by: Wherein, the Is the first The weight of the individual clusters is determined, ; Is the total number of clusters; And Respectively the first Mean vector and covariance matrix of each cluster; is data of Probability density of (c); or/and, the posterior distribution KL divergence is constructed by the following formula: Wherein, the Differences between the true data distribution and the composite data distribution; is true data belonging to the first The proportion of the number of clusters is, , For the amount of real data, Is true data Belonging to the first Each cluster Posterior probability of (2); To be in the synthesized data The proportion of the number of clusters is, , In order to synthesize the amount of data, To synthesize data Belonging to the first Each cluster Is a posterior probability of (c).
4. The large model optimization method for education of claim 1, wherein the large model training step comprises: a parameter updating step of superposing Gaussian noise on a random gradient through a Lang-Wan dynamics mechanism to update model parameters of a large model; The first evaluation step, namely randomly selecting one or more synthesized data from the training set as a verification set to input an updated large model, judging whether a parameter updating convergence condition is reached or not through a parameter updating evaluation model based on the output of the large model, returning to the parameter updating step if the parameter updating convergence condition is not reached, and executing the training step if the parameter updating convergence condition is reached; Training the large model reaching the parameter updating convergence condition through a training set; The third evaluation step is to judge whether the trained large model reaches the training convergence condition based on the training evaluation model, return to the parameter updating step if the training convergence condition is not reached, and output the trained large model if the training convergence condition is reached; or/and, the method further comprises an iterative optimization step, wherein the iterative optimization step comprises the steps of repeatedly carrying out the synthetic data generation step and the large model training step until convergence conditions are reached; or/and, the method further comprises a monitoring step of monitoring the stability of the output of the large model, and returning to the synthetic data generation step if the stability difference reaches a difference threshold; or/and, the method further comprises a monitoring step, wherein the monitoring step comprises the following steps: monitoring a large model; Constructing an evaluation model; judging whether a second evaluation convergence condition is reached; If the second evaluation convergence condition is not reached, returning to the parameter updating step or returning to the synthetic data generating step; Or/and, further comprising: Collecting feedback information of a user terminal; And adjusting model parameters of the large model according to the feedback information.
5. The large model optimization method for education of claim 4 wherein the large model training step further comprises: the construction step of the multitask generation evaluation model comprises the steps of constructing the multitask generation evaluation model; The second evaluation step is to judge whether the output trained in the large model training step reaches the first evaluation convergence condition or not through the multitask generation evaluation model, if not, return to the parameter updating step, and if so, take the synthesized data in the training set reaching the first evaluation convergence condition as the training set of the next iteration; Or/and, the parameter updating step comprises the following steps: initializing second parameters, wherein the second parameters comprise model parameters, learning rate and iteration times of the large model; randomly taking a plurality of synthesized data from a training set, inputting the initialized large model, and obtaining gradient estimation of model parameters of the large model; and (3) superposing Gaussian noise on the gradient of the model parameters of the large model, and introducing a discretized Lang's ten-thousand dynamics equation to update the model parameters of the large model.
6. The method for optimizing a large model for education according to claim 5, wherein the step of superimposing gaussian noise on the gradient of the model parameters of the large model, introducing discretized langevin dynamics equations to update the model parameters of the large model, comprises the steps of: Updating model parameters of the large model by: Wherein, the Index for iteration times; is the first A momentum term of the second iteration; is the first A second parameter for the second iteration; is a momentum attenuation coefficient; is the learning rate; At the current second parameter as a loss function L A gradient below; is the first The gaussian noise introduced by the multiple iterations, , Is the noise intensity; Or/and, the parameter updating step further comprises: Gradient clipping is carried out on the gradient of the second parameter; Or/and, the parameter updating step further comprises: And (5) introducing cosine annealing to adaptively adjust the learning rate.
7. The method for optimizing large models for education according to claim 6 wherein the step of introducing cosine annealing to adaptively adjust the learning rate comprises: The learning rate is adaptively adjusted by: Wherein, the Is the first Learning rate of the second iteration; Is the minimum learning rate; Is the maximum learning rate; Is the total iteration number; is a learning parameter.
8. A large model optimizing device for education, comprising a synthetic data generating module and a training module: the synthetic data generating module is configured to generate synthetic data based on Gaussian model fitting real data, evaluate the consistency of the synthetic data and the real data through distribution similarity measurement, obtain synthetic data meeting the consistency requirement, and form a training set; The training module is configured to update model parameters of the large model by superimposing Gaussian noise on a random gradient through a Lang-Wen dynamics mechanism, and train the large model in combination with the training set generated by the synthetic data generation module.
9. The large model optimizing apparatus for education of claim 8, further comprising a monitoring module or/and a feedback module: The monitoring module is configured to monitor the stability of the trained large model, and when the stability difference reaches a difference threshold, the monitoring module sends an instruction for generating the synthesized data to the synthesized data generation module; The feedback module is configured to receive feedback information of the user terminal and transmit the feedback information to the synthetic data generation module or/and the training module.
10. A large model optimization apparatus for education comprising a first cluster, a feature extraction module, and a second cluster, the first cluster comprising a plurality of data processing nodes, the second cluster comprising a plurality of computing nodes and at least one training node, wherein: The data processing node is configured to collect real data; The feature extraction module is configured to extract a plurality of features of real data output by the data processing node and convert the features into a learnable feature vector; The computing node is configured to generate synthetic data based on Gaussian model fitting of feature vectors of the real data extracted by the feature extraction module, evaluate consistency of the synthetic data and the real data through distribution similarity measurement, obtain synthetic data meeting consistency requirements, and form a training set; The training nodes are configured to update model parameters of the large model by superimposing Gaussian noise on random gradients through a Lang-Wen dynamics mechanism, and train the large model in combination with a training set constructed by the computing nodes.

Description

Large model optimization method and device for education Technical Field The invention belongs to the technical field of intelligent education, and relates to a large model optimization method and device for education. Background Along with the rapid development of the AI large model technology, AI large model products also continuously grow, and the mass life is continuously facilitated. The AI products continuously enter various aspects of education, thereby facilitating various aspects of teaching, learning, examination, evaluation, management and the like in the education industry, facilitating the work development of staff such as teaching staff, students and the like, and improving the management level of campuses and teachers. However, AI large models present serious illusion problems for the following reasons: Multi-dimensions of educational data has multiple dimensions, e.g., disciplines, practices, diets, mental states, physical attributes, etc., resulting in the AI large model requiring the processing of complex and diverse knowledge systems, knowledge confusion or fiction arises in generating content, such as miscombining different knowledge points to generate plausible but actually wrong answers, creating illusion problems. The intelligent development difference of the education institutions is that the data quality, the technical architecture and the resource investment of the education institutions are different, the data of the education institutions are not communicated with each other, the education scene adaptation and the real-time interaction correction are lacked, the complex teaching requirements are difficult to process, and finally the illusion generation is aggravated. The field limitation is that the knowledge in the professional fields such as education is updated quickly, and the large model is difficult to synchronize the latest progress. The accuracy of the AI large model depends on model training, the quality of the training data directly determines the accuracy of the output of the AI large model, but the training data may contain outdated, wrong or contradictory or even fictional information, and for example, the training data has too few samples and insufficient coverage. Model mechanism problem AI large models employ regression generation frameworks, the model goal is to maximize the probability of outputting sequences, rather than guarantee the fact accuracy, the model will output practically erroneous but statistically more reasonable text. Autoregressive generation preferentially selects high probability words over facts, lacking real-time verification capabilities. Model training bias-AI loss function of large model only measures similarity of generated text and training data, but not true accuracy, which results in model losing reality for pursuing fluency and strengthening error output. For another example, the large model training uses hint words to induce the model to generate error content. The illusion of the large model enables the large model output to generate reasonable and actual wrong contents, which not only can provide wrong information for teachers and students and mislead teaching directions to influence the enthusiasm of the students, but also can provide wrong decision schemes to seriously influence school management. Therefore, reducing the probability of creating hallucination problems in large models for education is an important topic that is currently in need of solution in the industry. Disclosure of Invention In order to reduce the probability of the occurrence of illusion problems in a large model for education, the invention provides a large model optimization method and device for education. According to a first aspect of the present invention, there is provided a large model optimization method for education, comprising: The synthetic data generating step is to generate synthetic data based on Gaussian model fitting real data, evaluate the consistency of the synthetic data and the real data through distribution similarity measurement, obtain synthetic data meeting the consistency requirement and form a training set; and a large model training step, namely, model parameters of the large model are updated by overlapping Gaussian noise on random gradients through a Lang-hundred-thousand dynamics mechanism, and the large model is trained by combining a training set. In one possible implementation, the large model optimization method for education further includes an iterative optimization step of repeating the synthetic data generation step and the large model training step until convergence conditions are reached. In one possible implementation, the synthetic data generating step includes: initializing a first parameter, wherein the first parameter comprises the number of clusters, a cluster center, cluster weights and a covariance matrix; Generating synthetic data by combining the real data according to the probability density function of th