PopQA

What it measures

14k entity-centric factual questions stratified by monthly Wikipedia pageviews to measure long-tail knowledge retention.

Reading the measurement

Accuracy is only part of the picture. Inspect unsupported answers, abstentions, language coverage and the evidence available to the model. A confident response can still be wrong, and a broad average can hide a specialist gap.

Published evaluation

Llama-2 7B systems are fine-tuned separately on Natural Questions or SQuAD. Never-retrieve, always-retrieve and learned adaptive retrieval are different configurations. Published exact-match accuracy is measured on the full PopQA test set.

Where this fits in an evaluation suite

Read accuracy together with error and abstention rates. Domain averages can hide language gaps, specialist failures, and poorly calibrated confidence.

Start with a capability hypothesis, then choose a second test with a different failure mode. Keep the checkpoint, reasoning effort, agent tools, sampling policy, and scoring procedure fixed when measuring a training intervention. Record failures and missing outputs in the denominator.

A reproducible comparison

Llama-2 7B systems are fine-tuned separately on Natural Questions or SQuAD. Never-retrieve, always-retrieve and learned adaptive retrieval are different configurations. Published exact-match accuracy is measured on the full PopQA test set.

Publish the exact run configuration, item coverage, score units, and available uncertainty. If a source supplies only an aggregate or a single run, preserve that limit. A new model name or a rounded score alone does not establish a capability improvement.