基于大型语言模型微调技术的反诈骗信息识别方法研究
Study on Anti-fraud Information Identification Method Based on Fine-tuning Techniques of Large Language Models
彭成智 1谢园园 1吕光旭1
作者信息
- 1. 中讯邮电咨询设计院有限公司,北京 100048
- 折叠
摘要
针对反诈骗信息识别,对大型语言模型(LLMs)的微调技术进行了深入的实验研究.选取了3种不同规模的LLMs基础模型,并采用了LoRA和p-tuning v2 2种先进的微调技术,以适应特定的反诈骗信息识别任务.通过多个维度的实验评估,微调策略不仅能够显著提升模型在反诈骗信息识别上的性能,还能够在一定程度上保持模型的通用性.此外,探讨了LLMs在少样本情况下的学习能力,并分析了不同微调策略下的资源消耗情况.
Abstract
Aiming at the anti-fraud information identification,it conductes in-depth experimental research on fine-tuning techniques of large language models(LLMs).It selectes three LLMs base models of different scales and employes two advanced fine-tuning technologies,LoRA and p-tuning v2,to adapt to specific anti-fraud information identification tasks.Through experimental evaluations across multiple dimensions,fine-tuning strategies not only significantly enhances the models'performance in anti-fraud information identification,but also maintains the universality of the model to a certain extent.Additionally,it explores the learning capabilities of LLMs under low-sample conditions and analyzes the resource consumption under different fine-tuning strategies.
关键词
大型语言模型/微调技术/反诈骗信息识别/LoRA/p-tuning/v2/少样本学习Key words
LLMs/Fine-tuning techniques/Anti-fraud information identification/LoRA/p-tuning v2/Few-shot learning引用本文复制引用
出版年
2024
NETL
NSTL
万方数据