Recover-LoRA reclaims accuracy in 2-bit LLMs via LoRA and knowledge distillation

Arxiv CS.LG (machine learning)

Paper proposes Recover-LoRA, a method that uses low-rank adaptation and knowledge distillation on synthetic data to recover accuracy in 2-bit quantized language models. It targets severe degradation from aggressive quantization for edge deployment.

llmlora

Arxiv CS.LG (machine learning)·Devleena DasRajeev PatwariElliott DelayeAshish Sirasao

7 days ago