Calibration dataset: VMware Open Instruct, 4096 tokens per sample