Parameter Efficient Supervised Fine-Tuning
Introduction
Development status: ongoing
Provide an open-source library which implements multiple PEFT (parameter efficient fine tuning) methods, including gradient based SMT, saliency map, and LORA, with HuggingFace’ TRL library.
- First to propose a saliency map–based PEFT approach, enabling interpretable parameter-efficient tuning.
- Ported the SMT (sparse matrix matrices) implementation from DeepSpeed to Hugging Face’s TRL library for improved integration and maint
- He, Haoze, Li, Juncheng Billy, Jiang, Xuan, and Miller, Heather. SMT: Fine-Tuning Large Language Models with Sparse Matrices. ICLR 2025.
- Hu, Edward J., et al. LoRA: Low-Rank Adaptation of Large Language Models. ICLR 2022
- Simonyan, Karen, Andrea Vedaldi, and Andrew Zisserman. Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps. arXiv preprint arXiv:1312.6034 (2013).
