Parameter-efficient fine-tuning of LLMa-2 using quantization low-rank adaptation

Abstract

Llama-2, an advanced neural network with huge potential in text generation, sentiment analysis and language understanding. This report focuses on the fine-tuning process for build chatbot on custom datasets, specification methods, hyperparameters and training strategies. Experimental results on Guanchu datasets show excellent adaptability of the model, outperforming the baseline model in human evaluation and achieving significant BERT scores for help and safety. The analysis includes an in-depth examination of LAMA-2’s architecture, outlining strengths and suggesting areas for improvement. Parameter-efficient fine-tuning and quantization also investigate the transformative potential of LLMA-2 through low-rank adaptation. The objective is to strike a balance between model complexity and efficiency, addressing challenges in resource-constrained environments.