Standard Deviation

Variance is measured in squared units, which makes it awkward to interpret — the variance of a latency in milliseconds is in milliseconds squared. Taking the square root gives $\sigma$, back in the original units and directly comparable to the mean.

That comparability is the whole reason it is the number people quote. "Mean 30 minutes, standard deviation 2" is immediately meaningful; "variance 4 minutes squared" is not. For a roughly normal distribution the rule of thumb is that about 68% of values fall within one $\sigma$ of the mean and about 95% within two.

In practice this is what error bars on a benchmark are, and reporting an accuracy without one is an incomplete result — two models at 91% mean accuracy with standard deviations of 0.5 and 8 are not comparable claims. Be careful which convention produced it, since NumPy divides by $n$ and pandas by $n-1$.