Beware of the Batch Size: Hyperparameter Bias in Evaluating LoRA
This reconciles inconsistencies in evaluating LoRA methods for fine-tuning large language models, impacting researchers and practitioners.
The paper demonstrates that batch size is a critical factor causing conflicting performance reports among LoRA variants, and shows that properly tuned vanilla LoRA often matches complex variants, with a proposed cost-efficient tuning strategy.
Low-rank adaptation (LoRA) is a standard approach for fine-tuning large language models, yet its many variants report conflicting empirical gains, often on the same benchmarks. We show that these contradictions arise from a single overlooked factor: the batch size. When properly tuned, vanilla LoRA often matches the performance of more complex variants. We further propose a proxy-based, cost-efficient strategy for batch size tuning, revealing the impact of rank, dataset size, and model capacity on the optimal batch size. Our findings elevate batch size from a minor implementation detail to a first-order design parameter, reconciling prior inconsistencies and enabling more reliable evaluations of LoRA variants.