Search

CN-121997044-A - Proxy model training method, device, server and storage medium

CN121997044ACN 121997044 ACN121997044 ACN 121997044ACN-121997044-A

Abstract

The application provides a proxy model training method, a device, a server and a storage medium, wherein the method comprises the steps of acquiring mode data corresponding to a plurality of different types of modes from a mode data information dictionary when a preset proxy model is required to be trained; the method comprises the steps of encoding data of each mode to generate a feature dictionary of the corresponding mode, determining the use frequency of feature atoms in each feature dictionary, quantifying the feature contribution degree of the corresponding mode according to the use frequency of the feature atoms, screening a plurality of target feature atoms from the feature contribution degree of each mode, constructing a composite feature dictionary according to each target feature atom, and training a proxy model according to the composite feature dictionary to complete training of the proxy model so as to improve the training effect of the proxy model.

Inventors

  • LIN BEI

Assignees

  • 中国联合网络通信集团有限公司
  • 中国联通(香港)运营有限公司
  • 中国联通国际有限公司

Dates

Publication Date
20260508
Application Date
20251230

Claims (10)

  1. 1. A proxy model training method, applied to a server, comprising: When a preset proxy model is required to be trained, acquiring mode data corresponding to a plurality of different types of modes from a mode data information dictionary; encoding the data of each mode to generate a feature dictionary of the corresponding mode, wherein each feature dictionary is used for representing the feature representation of the corresponding mode; Determining the use frequency of each feature atom in each feature dictionary; quantifying the characteristic contribution degree of the corresponding mode according to the use frequency of each characteristic atom; Screening a plurality of target feature atoms from the feature contribution of each mode; Constructing a composite feature dictionary according to each target feature atom; and training the proxy model according to the composite feature dictionary to complete training of the proxy model.
  2. 2. The method of claim 1, wherein determining the frequency of use of feature atoms in each feature dictionary comprises: determining a plurality of feature atoms contained in each feature dictionary; determining the use times of each feature atom in the corresponding feature dictionary; determining the total use times of all the feature atoms in the corresponding feature dictionary according to the use times of each feature atom; And determining the use frequency of each characteristic atom according to each use frequency and the total use frequency.
  3. 3. The method according to claim 1, wherein quantifying the feature contribution of the corresponding modality according to the frequency of use of each feature atom comprises: Setting a time attenuation factor and determining the service time of each characteristic atom; setting weights corresponding to the use frequencies of the characteristic atoms according to the use time and the time attenuation factor; And quantifying the characteristic contribution degree of the corresponding mode according to each weight and the use frequency of each characteristic atom.
  4. 4. The method of claim 1, wherein the screening a plurality of target feature atoms from feature contributions of each modality comprises: setting a screening rule of target characteristic atoms; Adjusting the screening rule to finish the adjustment of the screening rule; and after the screening rule is adjusted, screening a plurality of target characteristic atoms from the characteristic contribution degree of each mode.
  5. 5. The method of claim 1, further comprising, after the training of the proxy model according to the composite feature dictionary to complete the training of the proxy model: after the proxy model training is completed, testing the proxy model, and executing the following steps: acquiring a plurality of test mode data, wherein each test mode data is configured with a standard question-answer file; According to each test mode data and the standard question-answering file, question-answering test is conducted on the agent model so as to obtain the correct number of question-answering results; determining the characteristic contribution degree corresponding to each test mode data; determining the question-answering accuracy of the agent model according to the correct quantity and characteristic contribution degree of the question-answering results corresponding to the test modal data; and updating the proxy model according to the question-answer accuracy rate to finish updating the proxy model.
  6. 6. The method of claim 5, wherein testing the proxy model based on each test modality data and the standard question-answer file to obtain the correct number of test results comprises: Inputting each test mode data into the agent model for question-answering processing so as to obtain a plurality of question-answering results; comparing each question and answer result with the standard question and answer file to obtain each comparison result; And determining the correct quantity corresponding to the question-answer results according to the comparison results.
  7. 7. The method according to any one of claims 1 to 6, further comprising: according to a preset framework distributed algorithm and data of each mode, constructing a local feature dictionary and a global feature dictionary of each mode; integrating each local feature dictionary and each global feature dictionary into a feature dictionary corresponding to each mode; and aggregating the feature dictionaries into a composite feature dictionary through a preset aggregation algorithm.
  8. 8. A proxy model training device, applied to a server, comprising: The first acquisition module is used for acquiring the modal data corresponding to a plurality of different types of modalities from the modal data information dictionary when the preset proxy model is required to be trained; the encoding module is used for encoding the data of each mode to generate a feature dictionary of the corresponding mode, wherein each feature dictionary is used for representing the feature representation of the corresponding mode; the first determining module is used for determining the use frequency of each feature atom in each feature dictionary; the quantization module is used for quantizing the characteristic contribution degree of the corresponding mode according to the use frequency of each characteristic atom; the screening module is used for screening a plurality of target feature atoms from the feature contribution of each mode; the first construction module is used for constructing a composite feature dictionary according to each target feature atom; And the training module is used for training the proxy model according to the composite characteristic dictionary so as to complete the training of the proxy model.
  9. 9. A server is characterized by comprising at least one processor and a memory; The memory stores computer-executable instructions; The at least one processor executing computer-executable instructions stored in the memory cause the at least one processor to perform the proxy model training method of any one of claims 1 to 7.
  10. 10. A computer storage medium having stored therein computer executable instructions which, when executed by a processor, implement the proxy model training method of any of claims 1 to 7.

Description

Proxy model training method, device, server and storage medium Technical Field The present application relates to the field of data processing technologies, and in particular, to a proxy model training method, a proxy model training device, a server, and a storage medium. Background Agent model training, i.e., AI Agent training, is the process of server culture of intelligent agents with autonomy and target guidance. The core of the AI Agent is to be given the ability to actively complete tasks, and can autonomously perceive the environment, determine decisions and execute the decisions to achieve the goal. In the prior art, a traditional proxy model training method trains a proxy model by using mode data of a single type mode, combining a hierarchical feature fusion method, a dynamic attention mechanism method and the like, so as to complete training of the proxy model. However, in the prior art, the training mode of training the proxy model through the mode data of a single type mode has single data type, so that the training effect of the proxy model is poor, and the method is not suitable for a complex application environment. Disclosure of Invention The application provides a proxy model training method, a device, a server and a storage medium, which are used for solving the problems that the training mode for training a proxy model through modal data of a single type of modality has single data type, so that the training effect of the proxy model is poor and the method is not suitable for complex application environments. In a first aspect, the present application provides a proxy model training method, applied to a server, including: When a preset proxy model is required to be trained, acquiring mode data corresponding to a plurality of different types of modes from a mode data information dictionary; encoding the data of each mode to generate a feature dictionary of the corresponding mode, wherein each feature dictionary is used for representing the feature representation of the corresponding mode; Determining the use frequency of each feature atom in each feature dictionary; quantifying the characteristic contribution degree of the corresponding mode according to the use frequency of each characteristic atom; Screening a plurality of target feature atoms from the feature contribution of each mode; Constructing a composite feature dictionary according to each target feature atom; and training the proxy model according to the composite feature dictionary to complete training of the proxy model. In one possible design, the determining the use frequency of each feature atom in each feature dictionary includes determining a plurality of feature atoms contained in each feature dictionary, determining the use times of each feature atom in the corresponding feature dictionary, determining the total use times of all feature atoms in the corresponding feature dictionary according to the use times of each feature atom, and determining the use frequency of each feature atom according to the use times and the total use times. In one possible design, the step of quantifying the characteristic contribution of the corresponding mode according to the use frequency of each characteristic atom includes setting a time attenuation factor, determining the use time of each characteristic atom, setting a weight corresponding to the use frequency of each characteristic atom according to each use time and the time attenuation factor, and quantifying the characteristic contribution of the corresponding mode according to each weight and the use frequency of each characteristic atom. In one possible design, the screening of the plurality of target feature atoms from the feature contribution of each mode includes setting a screening rule of the target feature atoms, adjusting the screening rule to complete adjustment of the screening rule, and screening the plurality of target feature atoms from the feature contribution of each mode after the adjustment of the screening rule is completed. In one possible design, after the training of the proxy model according to the composite feature dictionary is completed, the method further comprises the steps of testing the proxy model after the training of the proxy model is completed, obtaining a plurality of test mode data, wherein each test mode data is configured with a standard question-answer file, conducting question-answer tests on the proxy model according to each test mode data and the standard question-answer file to obtain the correct number of question-answer results, determining the feature contribution degree corresponding to each test mode data, determining the question-answer accuracy of the proxy model according to the correct number of question-answer results and the feature contribution degree corresponding to each test mode data, and updating the proxy model according to the question-answer accuracy to complete updating of the proxy model. In one possible design, the p