EMMA-REAL-OOD-v0.1¶

v0.1 · proposed-contract

Purpose. The real-capture out-of-distribution (OOD) subset that anchors EMMA’s sim-to-real gap. It is fidelity only: it never enters the OOD-avg aggregate.

Where it comes from. Real over-the-air captures from public testbeds (Colosseum, POWDER, COSMOS), not rfgen. A consumer does not regenerate it; EMMA ships it as a fixed validation set.

Input¶

Real complex multi-antenna I/Q (4 to 8 elements), shape (num_rx, 2, N) float32, with the testbed’s per-scene metadata.

Labels¶

The same task labels as the synthetic datasets, where the testbed capture supports them, so a model’s synthetic score and real-capture score are directly comparable.

Splits¶

dev only (fidelity). About 8,000 captures. There is no train split: this set exists to measure how much of a synthetic score survives on real signals.

How it is used¶

For every model on the leaderboard, EMMA reports the gap between the synthetic score and the real-capture score in percentage points. A large gap means the model fit the generator rather than learned signal physics. See Sim-to-real.

What this tests¶

Real captures from Colosseum, POWDER, and COSMOS feed the sim-to-real gap column: they measure how much a model overfits rfgen by comparing its synthetic-split score against its real-capture score. They never enter the aggregate.

Metric¶

Fidelity only. No per-task metric is computed directly on this subset; it feeds the SimToRealGap column reported alongside every synthetic score, never the OOD-avg aggregate. See Metrics.

Get it¶

emma download --dataset EMMA-REAL-OOD-v0.1 --split dev --out ./scenes

This set is dev only and ships as a fixed capture set; you cannot regenerate it from rfgen.