Adaptive Multi-Fidelity Hyperparameter Optimization in Large Language Models

Fine-tuning of any machine learning models especially large language models is a difficult task because of the large hyper parameter space and high training cost. This paper tries to overcome this by proposing a technique called Adaptive Multi-Fidelity Hyper parameter Optimization (AMF-HPO) which combines multiple advanced techniques like multi-fidelity evaluations, reinforcement learning, Bayesian optimization, and transfer learning. This technique first uses low-fidelity evaluations to eliminate unnecessary hyper parameter space. It then uses a RL agent to identify the next configuration. It then uses Bayesian optimization and Transfer learning to select the optimal hyper parameter configuration.