Controllable optimization method, system and equipment based on large model data recovery and medium

By introducing controllable optimization methods of data recycling and constraint merging in large models, the problem of large models lack controllability in supervised fine-tuning data sets is solved, achieving higher usability, accuracy and cost-effectiveness.

CN119721289APending Publication Date: 2025-03-28SHANDONG LANGCHAO YUNTOU INFORMATION TECH CO LTD
0 Cites 0 Cited by

Patent Information

Application Number
CN202411668172.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-11-21
Publication Date
2025-03-28

AI Technical Summary

Technical Problem

Large models lack controllability in supervising fine-tuning data sets, making it difficult to automatically edit fine-tuning data, making it difficult to directly put free-form responses into use in practical applications.

Method used

Through a controllable optimization method based on big model data recycling, multiple constraints are merged into the recovered original data samples according to predefined rules, and the recovered data is added to the expanded constraints to create a new training task to consolidate the controllability of the big model. Specific steps include setting the token length, expanding the data set based on reconstruction instructions and responses, combining it into sequence decisions using an online model selection algorithm, and expanding the computational volume during the training process.

Benefits of technology

The high availability, controllable generation, high accuracy and low cost of large models in supervised fine-tuning datasets is realized, reducing dependence on field experts and improving the multi-constraint controllability of large models.

✦ Generated by Eureka AI based on patent content.
Patent Text Reader

Abstract

The invention discloses a controllable optimization method, system and equipment based on large model data recovery and a medium, belongs to the technical field of large models, and aims to solve the technical problem of how to automatically edit existing fine adjustment data, improve the controllability of the large models and improve the reliability of the large models. According to the technical scheme, the method comprises the steps that multiple constraints are combined into a recycled original data sample according to a predefined rule, expanded constraints are added through recycled data, and a new training task is created to consolidate the controllability of a large model; the method specifically comprises the following steps: setting a token length according to a sample number and a supervision fine tuning data set; augmenting the existing supervised fine tuning dataset based on the reconstruction instruction and the response; an online model selection algorithm is utilized to effectively combine a large model agent into a sequence decision; in the training process, the calculation amount is enlarged, and the available range of a large model is widened.
Need to check novelty before this filing date? Find Prior Art