ETL: Ensemble–at–Training for Low-cost Open-source Large Language Models

ETL: Ensemble–at–Training for Low-cost Open-source Large Language Models

연구 분야: Databases

학회: 2025 15th International Conference on Electrical Engineering (ICEENG)

Text summarization with Large Language Models (LLMs) has proven superior to traditional LSTM models, even surpassing human annotators’ preferences over original reference summaries in popular datasets. However, small open-source LLMs are often less capable than proprietary models like GPT-4. To address this, a method of combining multiple LLM outputs via an aggregator LLM has been proposed, though this increases computational requirements during inference. We propose a novel fine-tuning approach that eliminates the need for ensemble models during inference by shifting the ensemble mechanism to the training phase. In our method, a single fine-tuned model is deployed at inference, reducing computational overhead and complexity. Unlike traditional methods relying on static datasets, our approach leverages outputs from multiple pre-trained LLMs during training, enabling the model to generate high-quality summaries without requiring multiple models at inference. This ensemble-at-training paradigm improves summarization performance while ensuring deployment efficiency. Experimental results show that our approach outperforms baseline models and traditional fine-tuning techniques, offering a scalable and practical solution for efficient LLM fine-tuning.