KAIST Cuts AI Fine-Tuning Time to 3 Hours From 10

Fine-tuning is a critical process for adapting large language models (LLMs) to specialized applications like coding or medicine, involving additional training on relevant material. This process typically uses graphics processing units (GPUs) to divide and parallel-process training data. However, randomly distributing data can lead to inefficiencies, as GPUs might repeatedly process information with similar training effects, akin to redundant learning.

KAIST's CluSTER technique addresses this by analyzing how each piece of data impacts training and then grouping similar data together. From these curated groups, data is selectively chosen and distributed evenly across multiple GPUs. This ensures that each GPU handles a diverse, yet non-redundant, set of training information, optimizing the learning process. Parameters, which are internal values an AI model adjusts as it learns, are key to its acquired knowledge.

Experiments with LLMs having 7 billion and 13 billion parameters, using coding and medical datasets, demonstrated significant efficiency gains. Model performance was largely maintained even when only about 31% of the total data was used for training. This resulted in a reduction of training time by up to 69.6%, effectively cutting a 10-hour process to around three hours. Professor Lee Jae-gil emphasized the importance of 'smart choices' in data selection for AI training efficiency.

KAIST expects this technology to help companies and research institutions reduce both the time and GPU costs associated with additional training for specialized AI applications. While the reported results are based on specific model sizes, fields, and a four-GPU environment, the technique points to a method for more efficient use of existing computational resources rather than solely scaling up hardware.