Meta Releases 'WearableQA' Health Reasoning Benchmark Based on Real Wearable Data

Meta Research has released WearableQA, a health reasoning benchmark based on wearable device data from real users. Unlike existing synthetic data, it utilizes up to 500 days of wearable time-series data and blood biomarkers collected from 200 real users, reflecting real device noise, data missingness, and individual differences.

Benchmark Composition and Classification System WearableQA consists of a total of 4,084 10-choice multiple-choice questions, with correct answers uniformly distributed from A to J, resulting in a random baseline of 10%. The questions are classified into the following 16 types:

• Reasoning group: Data-based (Data, 2,724) vs. Health-based (Health, 1,360)
• Signal complexity: Single signal (Single, 1,682) vs. Cross-signal (Cross, 2,402)
• Grounding: Population-based (Population, 3,154) vs. Literature-based (Literature, 930)

Data Structure and Prompts The provided data is divided into WearableQA.jsonl (for evaluation) and WearableQA_raw.json (structured source). Prompts include user profiles (age, gender, BMI, etc.), up to 500 days of sensor data (16 types including step count, heart rate, sleep patterns, etc.), a panel of 17 blood biomarkers, and cohort reference statistics (p10~p90). The median prompt length is approximately 86,000 characters, reaching up to 124,000 characters, and exhibits sparse characteristics where days lacking specific metrics are omitted from rows.

Usage and Evaluation The original data can be converted into various formats such as CSV and Markdown via the render_raw.py script, and model performance may vary by several percentage points depending on the serialization method. During evaluation, the character selected by the model must exactly match the correct answer, and unparseable responses are treated as incorrect. This benchmark is available for research and benchmarking purposes only (CC BY-NC 4.0), and attempts at re-identification or use for clinical diagnostic purposes are strictly prohibited.