The Appeal and Reality of Recycling LoRAs with Adaptive Merging
4 hours ago
- #Adaptive Merging
- #Machine Learning
- #LoRA
- The paper explores adaptive merging of LoRA modules to improve performance in machine learning tasks.
- It investigates recycling user-contributed LoRAs from repositories like Hugging Face Hub, using nearly 1,000 LoRAs trained from Llama 3.1 8B-Instruct.
- Findings suggest adaptive merging improves performance over the base model but offers limited benefits compared to training a new LoRA on the same data.
- The study reveals that the choice of LoRAs to merge has little impact, and even randomly initialized LoRAs yield similar performance.
- This suggests adaptive merging may work via regularization rather than positive cross-task transfer.
- Positive transfer is confirmed when highly relevant LoRAs are available in the pool.
- The paper releases model checkpoints and code for further research.