LoRA-LiteE: A Computationally Efficient Framework for Chatbot Preference-Tuning
Yahe Yang, Chunliang Tao, Xiaojing Fan
TL;DR
The paper tackles the high computational cost of RLHF for chatbot preference tuning by introducing LoRA-LiteE, a framework that merges parameter-efficient Supervised Fine-Tuning with Low-Rank Adaptation and ensemble prediction to fuse lightweight models. Using the Chatbot Arena benchmark, LoRA-LiteE achieves competitive performance relative to un-finetuned GPT-4 and outperforms single larger models under resource constraints, while delivering faster convergence. The approach demonstrates that efficient, scalable preference tuning is feasible for resource-limited deployments, broadening accessibility to advanced, preference-aligned chatbots. The work highlights a practical path toward democratizing high-quality chatbot alignment through lightweight ensembles and targeted fine-tuning.
Abstract
Effective preference tuning is pivotal in aligning chatbot responses with human expectations, enhancing user satisfaction and engagement. Traditional approaches, notably Reinforcement Learning from Human Feedback (RLHF) as employed in advanced models like GPT-4, have demonstrated considerable success in this domain. However, RLHF methods are often computationally intensive and resource-demanding, limiting their scalability and accessibility for broader applications. To address these challenges, this study introduces LoRA-Lite Ensemble (LoRA-LiteE), an innovative framework that combines Supervised Fine-tuning (SFT) with Low-Rank Adaptation (LoRA) and Ensemble Learning techniques to effectively aggregate predictions of lightweight models, which aim to achieve a balance between the performance and computational cost. Utilizing the Chatbot Arena benchmark dataset, we conduct a comprehensive comparative analysis among our LoRA-LiteE model, corresponding base models at different scales, and GPT-4 trained with RLHF. Our empirical results demonstrate that the proposed LoRA-LiteE model achieves comparable performance to un-finetuned GPT-4 and outperforms the single larger-scale models under limited resource constraints. These findings highlight that our LoRA-LiteE provides a feasible and efficient methodology for human preference prediction in chatbot systems, enhancing scalability and accessibility, and thereby broadening the applicability of preference-tuned chatbots in resource-constrained environments.