Did anyone compare Alpaca responses between LLaMA fp16 vs 4-bit quantized?