Optimizing AI Model Training Costs with Stratified Sampling and Self-Adaptive Testing

The data processing and generation of datasets which are required for training Large Language Models (LLMs) and Generative AI (GenAI) applications have become costly these days. This paper presents an approach to optimize the cost of generating model training datasets by combining a stratified sampling algorithm integrated with a self-adaptive, comprehensive test suite framework. The methodology proposed in this paper will reduce the size of training datasets without impacting model performance and data integrity. This helps to reduce the cost of AI model training. Our empirical analyses demonstrate that our proposed approach reduces the cost of generating datasets up to 60%, without losing model accuracy and performances. It provides a way for efficient use of compute and storage resources for generation of high quality datasets for AI model training.