WiFi-TTA-Bench (NeurIPS 2026 — Evaluations & Datasets Track, double-blind)

Harm-aware benchmark for test-time adaptation (TTA) on WiFi Channel State Information (CSI) under physics-structured domain shift. This dataset card accompanies the anonymous double-blind submission and mirrors the benchmark's Croissant metadata; raw CSI recordings are not redistributed on HuggingFace (they remain under upstream academic licences).

What This Card Contains

• Split manifests, environment-id vectors, preparation scripts (links in outputs/croissant/wifi_tta_bench_metadata.json).
• Canonical per-method evaluation summaries as JSON cr:FileObject entries.
• A single cr:RecordSet defining the per-method summary schema (method, mean_gain, ci_lower, ci_upper, cohens_d_vs_noadapt, negative_adaptation_rate, n_observations).
• Core + RAI Croissant fields (data collection, biases, PII, maintenance, social impact).

Intended Use

• Rigorous evaluation of published TTA methods (TENT, SHOT/IM, T3A, SAR, CoTTA, LAME, physics-informed variants) on WiFi CSI cross-environment shifts.
• Harm-aware TTA development: negative-adaptation rate and source-drop tracking alongside mean gain.
• Comparative study between controlled synthetic (OFDM channel model) and real cross-environment shifts.

Out of Scope

• WiFi localisation, vital-signs estimation, through-wall imaging.
• Real-time deployment latency claims.
• Cross-modality generalisation claims (acoustic, radar, satellite).

Split Summary

Dataset	Samples	Classes	Environments	Shift type
Widar_BVP	8,964	6 gestures	3 rooms	Cross-room (verified)
NTU-Fi HAR	1,200	6 activities	3 sessions	Cross-session (temporal)
SignFi-10	200	10 gestures	3 splits	Cross-split (temporal)
Synthetic (generator)	configurable	3	2 domains	Controlled physics

Licensing

• Benchmark code, split manifests, evaluation artifacts, and this card: CC-BY-4.0.
• Raw CSI recordings: retain original upstream academic licences — obtain from the respective dataset authors.

Reference

Double-blind; citation to appear after the NeurIPS ED 2026 review period.