NVIDIA has introduced a new fine-tuning method called DoRA, which improves upon the existing LoRA method without adding extra inference overhead. DoRA has shown significant performance enhancements in various language and vision models, outperforming LoRA in tasks like common-sense reasoning and multi-turn benchmarks. The method has been accepted at ICML 2024, indicating its potential impact in machine learning.
DoRA decomposes pretrained weights into magnitude and directional components, fine-tuning both efficiently. Visualizations show that DoRA makes substantial directional adjustments while maintaining magnitude, resembling full fine-tuning patterns. Across different models, DoRA consistently outperforms LoRA in tasks like commonsense reasoning, image-text understanding, and visual instruction tuning.
DoRA can be integrated into the QLoRA framework for low-bit pretrained models, showing superior accuracy compared to FT and QLoRA. In text-to-image generation applications like DreamBooth, DoRA produces better results than LoRA in challenging datasets.
The method is expected to become a standard choice for fine-tuning AI models, compatible with existing methods and suitable for applications like NVIDIA Metropolis, NeMo, NIM, and TensorRT. For more information, refer to the NVIDIA Technical Blog.
Source link
Source link: https://blockchain.news/news/nvidia-unveils-dora-superior-fine-tuning-method-ai-models