kaimeta/wearables_benchmarks

Wearables Benchmarks

This page consolidates the GitHub and HuggingFace links to wearables AI benchmarks from Meta.

Datasets highly relevant to wearable use cases

Benchmark	Paper	GitHub	HuggingFace	TLDR
WearVox	arXiv	facebookresearch/wearvox	WearVox	A benchmark for evaluating voice assistants in realistic wearable scenarios using 3,842 egocentric audio recordings collected via AI glasses across multiple task types like QA, tool calling, and speech translation.
WearVQA	arXiv		WearVQA	A benchmark for evaluating visual question answering on wearable devices using 2,520 image-question pairs that reflect egocentric challenges like occlusion, poor lighting, and blur.
CRAG	arXiv	facebookresearch/CRAG		A comprehensive RAG benchmark for factual question answering spanning five domains, eight question categories, and varied entity popularity and temporal dynamics.
CRAG-MM	arXiv	facebookresearch/CRAG-MM	Single-Turn & Multi-Turn	A multimodal conversational benchmark featuring image-based QA across 13 domains with single- and multi-turn conversations captured via smart glasses and public sources.
MemoryQA	arXiv	facebookresearch/MemoryQA		A benchmark for answering recall questions about visual content retrieved from previously stored multimodal memories.
PLM-VideoBench	arXiv	facebookresearch/perception_models		A comprehensive video understanding benchmark suite covering fine-grained QA, egocentric smart glasses QA, region captioning, temporal localization, and dense video captioning.

Datasets moderately relevant to wearable use cases

Benchmark	Paper	GitHub	TLDR
Head-to-tail	arXiv	facebookresearch/head-to-tail	A benchmark specifically designed to assess how well LLMs incorporate factual knowledge across head, torso, and tail popularity distributions.
VisualLens	arXiv	facebookresearch/visuallens	A recommendation benchmark evaluating systems under task-agnostic visual user histories using data from Google Local and Yelp.
MeetingQA	arXiv	facebookresearch/AssoMem	A benchmark simulating real-world meeting scenarios where multi-turn dialogues form the memory base, paired with diverse QA examples.
SemiBench	arXiv	facebookresearch/SemiBench	A benchmark for evaluating knowledge extraction quality and its impact on question answering from semi-structured webpages.

License: Please refer to each dataset's license terms and conditions before using these datasets. License information can be found in the respective GitHub repositories and HuggingFace dataset pages.