This is quantized llama 7B is trained on the dataset without chain-of-thought instruction finetune. We use it for ablation study.