We introduce Humanline as a simple yet surprisingly effective design to close the online-offline alignment gap across multiple model families (Llama-3, Gemma-2) and model sizes (1.5B-27B) on both instruction-following and mathematical reasoning.