Meta’s research group has published WearableQA, a benchmark of 4,084 multiple-choice questions built entirely from the recorded health data of 200 real people. Each user contributes up to roughly 500 days of daily wearable measurements, a blood biomarker panel, and basic demographics. The stakes are unusual for a machine learning benchmark: this is a model being scored on its ability to reason about somebody’s actual body.
Most health-adjacent AI benchmarks work from synthetic or idealized signals, built to be clean. WearableQA’s authors chose the opposite path. The dataset keeps device noise, missing days, and the variability between individuals intact, on the premise that a model which only performs well on smoothed data has learned something different from a model that performs well on the messy kind.
The questions split into 16 types along two axes. One separates “data” reasoning, computing correlations, excursion counts, recovery times, and trend shapes directly from the numbers, from “health” reasoning, which asks a model to interpret those numbers in bodily terms: predicting a prognosis, weighing a differential diagnosis, or assessing risk. The other separates single-signal questions from cross-signal ones that require integrating multiple metrics at once. Every question offers ten answer options, lettered A through J, with a random-guess baseline of 10 percent.
The scale of what’s packed into each prompt is worth noting on its own. A rendered question can include up to 16 daily metrics per user (steps, resting heart rate, HRV, VO2 max, sleep stages, stress levels, exercise load) plus up to 17 blood biomarkers and population percentiles for context. Meta’s own documentation puts the median prompt at roughly 86,000 characters, with a max near 124,000. Evaluating this benchmark seriously means budgeting real context-window space, not treating it as a lightweight quiz.
On privacy, the repository is precise about restrictions and silent about process. The data ships under a Creative Commons Attribution-Non Commercial license with explicit added terms: no re-identification attempts, no use in clinical, diagnostic, or treatment decisions, and no use that is discriminatory or misleading about health outcomes. What the documentation does not say is how the 200 participants were recruited, what they were told, or how the data was de-identified before release. None of that appears anywhere in the repository as published.
That gap matters more than the benchmark’s headline number. A dataset built from 200 real people’s blood work and daily physiology is a meaningfully different object than a synthetic one, and the license’s ban on clinical use is itself an acknowledgment that these questions are not validated for anything resembling medical decision-making. Meta is not claiming otherwise: this is presented as a reasoning benchmark, not a diagnostic tool, and any accuracy figure a lab reports against it should be read as a capability claim, not a clinical one.
For teams building health-adjacent AI features, WearableQA is a useful stress test precisely because it will not stay clean. A model that handles missing sensor days and noisy biomarkers on this benchmark is being tested against something closer to what a real deployment would see than most existing health datasets offer. But the absence of stated consent and de-identification methodology is worth flagging before treating any published score as evidence the underlying data pipeline is one you’d want to replicate.
Based on the WearableQA repository published by Meta’s research organization (Facebook Research) on GitHub.
