Task-aware based large model fine-tuning method and legal information analysis method

By constructing a target domain knowledge graph and a task adaptive optimization mechanism, the internal representation and reasoning methods of the large model are dynamically adjusted, solving the problem that the large model cannot uniformly adapt to multiple tasks in a professional domain. This achieves deep integration and flexible adaptation of professional domain knowledge, improving the model's generalization ability and cross-domain application value.

CN120973957BActive Publication Date: 2026-07-24BEIJING ELECTRONIC DIGITAL INTELLIGENCE TECHNOLOGY CO LTD
View PDF 2 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN Β· China
Patent Type
Patents(China)
Current Assignee / Owner
BEIJING ELECTRONIC DIGITAL INTELLIGENCE TECHNOLOGY CO LTD
Filing Date
2025-09-04
Publication Date
2026-07-24

AI Technical Summary

Technical Problem

Existing large models cannot uniformly adapt to multiple tasks in professional fields, lack a deep integration and understanding of professional field knowledge, and cannot dynamically adjust internal representation and reasoning methods according to different task characteristics, resulting in poor performance when dealing with various legal tasks.

Method used

Construct a knowledge graph for the target domain, inject the initial large model through a multi-layered knowledge architecture, configure a task-adaptive optimization mechanism, adopt progressive unfreezing and mixed precision training, and dynamically adjust the model's internal representation and reasoning methods to achieve flexible model adaptation.

Benefits of technology

It enhances the model's ability to master and apply professional domain knowledge, improves its generalization ability and cross-domain application value in different vertical domain tasks, and achieves resource saving and performance optimization.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120973957B_ABST
    Figure CN120973957B_ABST
Patent Text Reader

Abstract

The application discloses a task-aware-based large model fine-tuning method and a legal information analysis method. The fine-tuning method comprises the following steps: constructing a target field knowledge graph and injecting an initial large model to obtain a first large model, establishing a mapping relationship between a task type and a knowledge subgraph to obtain a second large model, and configuring a task self-adaptive optimization mechanism to obtain a third large model; selecting and expanding initial data samples to obtain target training samples; finally, the third large model is fine-tuned by using a progressive unfreezing strategy and mixed precision training. The method deeply couples the two, uses a task-knowledge mapping table to drive the unfreezing sequence and dynamically update the precision bit width, converges quickly with the least video memory and parameters in the early stage, gradually releases the capacity and precision bit width in the later stage, realizes the triple balance of gradient stability, video memory saving and performance optimization, and the fine-tuned large model realizes the deep fusion of professional field knowledge and parameters, can dynamically adjust the internal representation and reasoning mode according to different task characteristics, and meets the diversified task analysis demand.
Need to check novelty before this filing date? Find Prior Art