自适应合并的 LoRA 回收的吸引力与现实

The widespread availability of fine-tuned LoRA modules for open pre-trained models has sparked interest in methods for adaptively merging LoRAs to enhance performance. These methods typically involve selecting LoRAs from a pool and tuning merging coefficients based on a task-specific dataset. While adaptive merging techniques have shown improvements in certain settings, no prior work has attempted to recycle LoRAs found 'in the wild' from model repositories like the Hugging Face Hub. To address this gap, the study investigates recycling from a pool of nearly 1,000 user-contributed LoRAs trained from the Llama 3.1 8B-Instruct language model. The empirical study includes a variety of adaptive and non-adaptive merging methods to evaluate their effectiveness in practical applications.