by Meta
WearableQA is Meta's benchmark of 4,084 ten-option multiple-choice questions about health reasoning over real wearable data. It is meant for researchers testing how well language models reason over sensor time series, blood biomarkers and demographics.
DatasetFree
Step 1: Read the license terms and confirm your use is non-commercial research or benchmarking.
Step 2: Install the Hugging Face datasets library.
Step 3: Load the test split with load_dataset("facebook/WearableQA", split="test"), using streaming=True since prompts are long.
Step 4: Pick a configuration such as row, markdown or structured to choose how sensor data is shown.
Step 5: Send each question to your model and compare its chosen letter with the answer field.
CC BY-NC 4.0 with additional terms. These forbid re-identification and use in clinical, diagnostic or treatment decisions, and limit use to research and benchmarking.
The row config, which is the default. It shows sensor data as one line per day and is the setting the paper reports.
10%, since each question has ten options and answers are balanced across A to J.
Yes. The structured config has no prompt text and gives each record its own sensor window as structured values.