A Dynamic Adaptive Data Filtering and Sampling Framework for Optimized Fine-Tuning of Large Language Models

Fine-tuning large language models (LLMs), such as Baichuan2-7B, faces challenges in data quality and training efficiency. This paper introduces DADFS (Dynamic Adaptive Data Filtering and Sampling), a new framework designed to optimize LLM fine-tuning. DADFS includes three components: Dynamic Adaptive Filtering (DAF), Adaptive Data Sampling (ADS), and Data Mixing and Proportional Weighting. The framework filters training data based on quality scores to ensure high-quality and diverse samples. It also uses adaptive sampling to select the most informative data points for training and employs data mixing to improve generalization and reduce overfitting. Experimental results show that DADFS outperforms traditional fine-tuning methods across key metrics. DADFS also integrates LoRA (Low-Rank Adaptation) within the PEFT (Parameter-Efficient Fine-Tuning) framework, which helps maintain computational efficiency while improving model adaptability and stability. This research offers an efficient and scalable solution for LLM fine-tuning, showing the importance of data optimization for model performance.