Method for incremental training of large model

Abstract

The invention relates to the technical field of incremental training, and particularly discloses a method for large model incremental training, comprising the following steps: S1, obtaining a target data set based on an industry corpus data set, dividing the target data set into sub-data sets, obtaining quality indexes of the sub-data sets, and determining quality scores of the sub-data sets; s2, acquiring a first sequence, and determining a performance score of the universal model from the first sub-data set in the first sequence; calculating a performance score difference value, and determining whether the state is an abnormal state or not according to a performance score difference value threshold value; s3, coordinate points are generated and fitted to obtain a fitted curve, and abnormal sorting positions are obtained according to the fitted curve; and determining a target model according to the abnormal sorting position, and performing incremental training on the target model through the to-be-determined data set. According to the method, the negative influence of low-quality data on the model can be reduced, the adaptability of the model to new domain knowledge is improved, and the problem of performance degradation is avoided.

Description

Technical Field

[0001] The present invention relates to the technical field of incremental training, and in particular to a method for incremental training of large models. Background Art

[0002] Large model incremental training is a training method for further optimizing or expanding the capabilities of an already trained large-scale artificial intelligence model. It is mainly used to add new knowledge or adjust model performance based on the original model without retraining the entire model from scratch.

[0003] By adding a specific parameter layer to the network structure of the pre-trained large model, the general catastrophic forgetting during incremental training is solved, and the ability to understand industry content is improved through incremental training of industry corpus. However, in incremental training, data quality directly affects the model's ability to adapt to new domain knowledge. If the quality of the industry corpus is not high (such as containing noise), it may cause the newly added parameter layer during training to absorb wrong knowledge. Not only will the model find it difficult to accurately understand the new domain content, but it may even experience performance degradation. Summary of the invention

[0004] The purpose of the present invention is to provide a method for large model incremental training to solve the following technical problems: In incremental training, data quality directly affects the model's ability to adapt to new domain knowledge. If the quality of the industry corpus is not high (such as containing noise), it may cause the newly added parameter layer during training to absorb erroneous knowledge. Not only will the model find it difficult to accurately understand the content of the new domain, it may even experience performance degradation.

[0005] The purpose of the present invention can be achieved through the following technical solutions: A method for incremental training of a large model, comprising the following steps: S1: Obtain an industry corpus dataset, preprocess the industry corpus dataset to obtain a target dataset, divide the target dataset into m sub-datasets, where m is a preset number, obtain quality indicators of the sub-datasets, which include data integrity, data consistency, and data variability, and calculate the quality scores of the sub-datasets based on the superior and inferior solution distance method; S2: Sort the sub-datasets in descending order according to the corresponding quality scores to obtain a first sorting, starting from the sub-dataset A at the top of the first sorting, divide the sub-dataset A into a training set and a test set according to a preset ratio, train a preset general model based on the training set, and obtain performance indicators of the trained general model through the test set, wherein the performance indicators include accuracy, AUC and loss value, and determine the performance score of the trained general model based on the superior and inferior solution distance method; Calculate the performance score difference ΔD i=D i -D i-1 , D i represents the performance score corresponding to the ith sub-data set in the first sorting, and sets a performance score difference threshold ΔDys. When the performance score difference ΔD i-1 When ≤ΔDys, it is marked as abnormal state; S3: Generate coordinate points (i, ΔD) in a pre-constructed coordinate system i ), fitting the coordinate points to obtain a fitting curve f(x), where x represents the sorting position of the sub-dataset in the first sorting, and f(x)=0, and solving to obtain the abnormal sorting position xyc; The general model trained by the sub-dataset at the xyc-1th position in the first sort is used as the target model, and the target model is incrementally trained by the pending data set until the number of incremental training reaches a preset value. The pending data set is the sub-dataset whose sorting position in the first sorting belongs to the sorting position interval [xyc, m].

[0006] As a further solution of the present invention: in the step S1, when the quality score of the sub-dataset a is less than a preset quality score threshold, the sub-dataset a is removed.

[0007] As a further solution of the present invention: in step S2, when two or more quality scores are of the same size, the quality scores of the same size are sorted in order of data completeness.

[0008] As a further solution of the present invention: in step S2, when the performance score difference ΔD i When ≤ΔDys and i≤0.5m, an early warning message is sent to report an error.

[0009] As a further solution of the present invention: in step S1, the process of preprocessing the industry corpus dataset specifically includes: Cleaning the industry corpus data set, deleting duplicate data in the industry corpus data set, and filling missing values; Standardize the numerical industry corpus data to eliminate the dimension; Encode non-numeric industry corpus data.

[0010] As a further solution of the present invention: in the step S3, during the incremental training of the target model, the target model is incrementally trained with the pending data set according to the order of the first sorting.

[0011] Beneficial effects of the present invention: In this solution, the industry corpus data set obtained is the original data related to a specific industry. By preprocessing the original data such as cleaning, the data quality is improved, and a target data set that can be used for model training is obtained to ensure the effectiveness of model training; dividing the data set facilitates the evaluation of the impact of different data sets on model performance, and by obtaining the quality index of each sub-data set and calculating the quality score of each sub-data set based on the superior and inferior solution distance method, the quality of each sub-data set is quantified, providing an objective basis for the subsequent selection of the data set that is most conducive to model training; by training the model for each sub-data set, evaluating its impact on model performance, and by calculating the performance score difference, the general model performance improvement is identified The critical point provides a basis for subsequent data set screening, avoids too much invalid or negatively impacted data from entering the model training, enables the general model to achieve optimal performance with less data, enhances the generalization ability of the model, and avoids wasting computing resources on data that does not significantly improve the model performance; wherein, if there is no abnormal state, the subsequent steps are not executed until all sub-training sets are used to train the general model; then, the position where the performance score difference is zero is found by fitting the curve, that is, the position where the model performance is no longer improved, and the model before the abnormal point is selected as the basic model to ensure that the model has the best performance, and the remaining data is used for incremental training, so as to fine-tune the model and improve the adaptability of the model. The present invention can reduce the negative impact of low-quality data on the model, improve the model's adaptability to new domain knowledge, and avoid the problem of performance degradation. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] The present invention will be further described below in conjunction with the accompanying drawings.

[0013] Figure 1 It is a flow chart of a method for incremental training of large models according to the present invention. DETAILED DESCRIPTION

[0014] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0015] See also Figure 1 As shown, the present invention is a method for incremental training of a large model, comprising the following steps: S1: Obtain an industry corpus dataset, preprocess the industry corpus dataset to obtain a target dataset, divide the target dataset into m sub-datasets, where m is a preset number, obtain quality indicators of the sub-datasets, which include data integrity, data consistency, and data variability, and calculate the quality scores of the sub-datasets based on the superior and inferior solution distance method; S2: Sort the sub-datasets in descending order according to the corresponding quality scores to obtain a first sorting, starting from the sub-dataset A at the top of the first sorting, divide the sub-dataset A into a training set and a test set according to a preset ratio, train a preset general model based on the training set, and obtain performance indicators of the trained general model through the test set, wherein the performance indicators include accuracy, AUC and loss value, and determine the performance score of the trained general model based on the superior and inferior solution distance method; Calculate the performance score difference ΔD i =D i -D i-1 , D i represents the performance score corresponding to the ith sub-data set in the first sorting, and sets a performance score difference threshold ΔDys. When the performance score difference ΔD i-1 When ≤ΔDys, it is marked as abnormal state; S3: Generate coordinate points (i, ΔD) in a pre-constructed coordinate system i ), fitting the coordinate points to obtain a fitting curve f(x), where x represents the sorting position of the sub-dataset in the first sorting, and f(x)=0, and solving to obtain the abnormal sorting position xyc; The general model trained by the sub-dataset at the xyc-1th position in the first sort is used as the target model, and the target model is incrementally trained by the pending data set until the number of incremental training reaches a preset value. The pending data set is the sub-dataset whose sorting position in the first sorting belongs to the sorting position interval [xyc, m].

[0016] It should be noted that the industry corpus dataset obtained is the original data related to a specific industry. By preprocessing the original data such as cleaning, the data quality is improved, and a target dataset that can be used for model training is obtained to ensure the effectiveness of model training. Dividing the dataset facilitates the evaluation of the impact of different datasets on model performance. By obtaining the quality index of each sub-dataset and calculating the quality score of each sub-dataset based on the superior and inferior solution distance method (TOPSIS), the quality of each sub-dataset is quantified, providing an objective basis for the subsequent selection of the dataset that is most conducive to model training. The superior and inferior solution distance method is a multi-index comprehensive evaluation method. This method evaluates the relative advantages and disadvantages of each evaluation object by calculating the distance between it and the ideal solution and the negative ideal solution. It belongs to the prior art. We will not elaborate on this in detail. By training the model for each sub-dataset, evaluating its impact on the performance of the general model, and calculating the performance score difference, we can identify the critical point for improving the performance of the general model, provide a basis for subsequent data set screening, and avoid too much invalid or negatively impacted data from entering the model training, so that the general model can achieve optimal performance with less data, enhance the generalization ability of the model, and avoid wasting computing resources on data that does not significantly improve the model performance. The method of obtaining performance indicators belongs to the existing technology and is not described here. The general model is the basic model for incremental training, which is used to learn and adapt the knowledge of industry corpus data. This model can choose the following suitable models according to the specific application scenarios and target tasks. The general model type that is suitable for training is a general model, such as BERT in the field of natural language processing (NLP); if there is no abnormal state, the subsequent steps will not be executed until all sub-training sets are used to train the general model; then, the position where the performance score difference is zero is found by fitting the curve, that is, the position where the model performance is no longer improved, and the model before the abnormal point is selected as the basic model to ensure that the model has the best performance, and the remaining data is used for incremental training to fine-tune the model and improve the adaptability of the model; the remaining sub-datasets are subsequently incrementally trained, but the main purpose is not to improve the performance of the core model, but to adapt to the new data domain knowledge, while trying to avoid performance degradation. When the model performance is saturated, the model weight at this time is locked ( If the parameters are saved as snapshots), subsequent incremental training should be carried out on this basis rather than completely retraining. The incremental training process can adopt a small step learning rate to reduce the model's over-adaptation to low-quality data. In addition, during the incremental training process, the model performance indicators are monitored in real time to determine whether to stop training. If the performance drops significantly after incremental training, it can be rolled back to the target model weights to better utilize the target model at saturation as a benchmark, while avoiding the negative impact of low-quality data on the model, ensuring that incremental training mainly plays a role in knowledge adaptation rather than destroying the foundation of existing performance; the process of obtaining the fitting curve is nonlinear fitting, and the specific reference is to the existing technology. The specific process of incremental training also refers to the existing technology and will not be repeated here.

[0017] In another preferred embodiment of the present invention, in step S1, when the quality score of sub-dataset a is less than a preset quality score threshold, the sub-dataset a is removed.

[0018] It is worth noting that by filtering data sets below the threshold, the interference of low-quality data (such as noise and data with serious missing values) on model performance is reduced, and the computing resource consumption of invalid data sets in sorting and model training is reduced; if low-quality data enters the incremental training stage, it may cause the model to learn wrong knowledge or even performance degradation. By eliminating it in advance, this risk can be effectively prevented.

[0019] In another preferred embodiment of the present invention, in step S2, when two or more quality scores are of the same magnitude, the quality scores of the same magnitude are sorted in order of data completeness.

[0020] In another preferred embodiment of the present invention, in step S2, when the performance score difference ΔD i When ≤ΔDys and i≤0.5m, an early warning message is sent to report an error.

[0021] It should be noted that the current sub-dataset is ranked in the top 50%, which means that the problem may occur in the early stage when the model performance has not been fully trained or saturated. During the training process, the top 50% of the sub-datasets are usually high-quality datasets, which theoretically should make a greater contribution to improving model performance. If the performance score difference is detected to be abnormally small (or negative) at this stage, it indicates that there may be data anomalies (such as inconsistency, noise) or model adaptation problems. Therefore, an error needs to be reported.

[0022] In another preferred embodiment of the present invention, in step S1, the process of preprocessing the industry corpus dataset specifically includes: Cleaning the industry corpus data set, deleting duplicate data in the industry corpus data set, and filling missing values; Standardize the numerical industry corpus data to eliminate the dimension; Encode non-numeric industry corpus data.

[0023] It should be noted that in the cleaning stage, completely identical or highly similar records in the industry corpus dataset are detected and deleted to avoid model deviation due to duplicate samples. For nearly duplicate data (such as similar content but different formats), detection can be based on string similarity (such as Jaccard similarity coefficient) or hashing method; data with missing values ​​are reasonably filled to avoid model training interruption or deviation due to missing values. For numerical data, mean filling, median filling, etc. can be used. For non-numeric data, such as categorical data, the most frequently occurring value is used for filling; for standardization and encoding, the specific method can be freely selected in actual situations, and there is no restriction here.

[0024] In another preferred embodiment of the present invention, in step S3, during the incremental training of the target model, the target model is incrementally trained using the pending data set in the order of the first sorting.

[0025] It can be understood that the first sorting is based on the descending order of the quality scores of the sub-datasets, which means that high-quality data is at the front of the sorting. Incremental training of the target model step by step in this order can ensure that the model learns high-quality data first and gradually adapts to data of different quality levels. The specific process of incremental training is: Load the target model: The target model weights locked at the performance score saturation point are used as the starting point for incremental training; Iterative training: Load the sub-datasets in the order of the first sort (from high to low), load one sub-dataset at a time, and divide it into a training set and a validation set for training and performance evaluation of the target model; Dynamically monitor performance: After each incremental training of a sub-dataset, record the performance indicators of the model on the validation set. If the model performance drops significantly, interrupt the incremental training and roll back to the model weights of the previous round. Among them, in order to avoid overfitting of the model or over-learning of low-quality data, the learning rate can be gradually reduced during incremental training, and a learning rate decay strategy (such as exponential decay or piecewise decay) can be used to further reduce the amplitude of parameter updates when processing low-quality data.

[0026] The above is a detailed description of an embodiment of the present invention, but the content is only a preferred embodiment of the present invention and cannot be considered to limit the scope of implementation of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the scope of the patent coverage of the present invention.

Claims

1. A method for incremental training of a large model, characterized in that: The following steps are involved: S1: Obtain an industry corpus dataset, preprocess the industry corpus dataset to obtain a target dataset, divide the target dataset into m sub-datasets, where m is a preset number, obtain quality indicators of the sub-datasets, which include data integrity, data consistency, and data variability, and calculate the quality scores of the sub-datasets based on the superior and inferior solution distance method; S2: Sort the sub-datasets in descending order according to the corresponding quality scores to obtain a first sorting, starting from the sub-dataset A at the top of the first sorting, divide the sub-dataset A into a training set and a test set according to a preset ratio, train a preset general model based on the training set, and obtain performance indicators of the trained general model through the test set, wherein the performance indicators include accuracy, AUC and loss value, and determine the performance score of the trained general model based on the superior and inferior solution distance method; Calculate the performance score difference ΔD i =D i -D i-1 , D i represents the performance score corresponding to the ith sub-data set in the first sorting, and sets a performance score difference threshold ΔDys. When the performance score difference ΔD i When ≤ΔDys, it is marked as abnormal state; S3: Generate coordinate points (i, ΔD) in a pre-constructed coordinate system i ), fitting the coordinate points to obtain a fitting curve f(x), where x represents the sorting position of the sub-dataset in the first sorting, and f(x)=0, and solving to obtain the abnormal sorting position xyc; The general model trained by the sub-dataset at the xyc-1th position in the first sort is used as the target model, and the target model is incrementally trained by the pending data set until the number of incremental training reaches a preset value. The pending data set is the sub-dataset whose sorting position in the first sorting belongs to the sorting position interval [xyc, m].

2. A method for large model incremental training according to claim 1, characterized in that: In the step S1, when the quality score of the sub-dataset a is less than a preset quality score threshold, the sub-dataset a is removed.

3. A method for large model incremental training according to claim 1, characterized in that: In step S2, when two or more quality scores are of the same magnitude, the quality scores of the same magnitude are sorted in order of data completeness.

4. A method for large model incremental training according to claim 1, characterized in that: In step S2, when the performance score difference ΔD i When ≤ΔDys and i≤0.5m, an early warning message is sent to report an error.

5. A method for large model incremental training according to claim 1, characterized in that: In step S1, the process of preprocessing the industry corpus dataset specifically includes: Cleaning the industry corpus data set, deleting duplicate data in the industry corpus data set, and filling missing values; Standardize numerical industry corpus data to eliminate dimensions; Encode non-numeric industry corpus data.

6. A method for large model incremental training according to claim 1, characterized in that: In the step S3, during the incremental training of the target model, the target model is incrementally trained with the pending data set according to the first sorting order.

Images