=

The passage around this formula

Reframe the object. A benchmark score $\hat{s}$ is an estimator of a population quantity s — the model’s success rate over some distribution of tasks $\mathcal{D}$ that somebody hopes resembles the work you actually have. Every property that makes an estimator trustworthy applies: $\hat{s} = \frac{1}{n}\sum_{i=1}^{n} \mathbf{1}\!\left[\,\text{success on } x_i\,\right], \qquad x_i \sim \mathcal{D}_{\mathrm{bench}}$. The number is only as good as three things: whether $\mathcal{D}_{\mathrm{bench}}$ resembles $\mathcal{D}$ , whether the $x_i$ are genuinely held out, and whether the indicator is measured with enough repetition to characterise its spread. All three fail routinely, and they fail in different directions.

Read this part in the article →