Projects
Pi-SQL
Proposed using pivot programming languages (e.g., Python) as an intermediate representation to provide fine-grained structural guidance for Text-to-SQL generation.
MiLoRA
Harnessing minor singular components for parameter-efficient LLM finetuning. A novel approach that leverages underutilized singular components to improve LoRA-based fine-tuning.
LoRA-Flow
Dynamic LoRA fusion for large language models in generative tasks. Enables automatic weight allocation across multiple LoRA modules during inference.
Delta-CoMe
Training-free delta-compression with mixed-precision for large language models. Achieves efficient model compression without additional training.