EMMA-S2T-QA v2¶

Warning

Pre-implementation. This page describes proposed contracts. Class signatures, parameter types, schema fields, and behavior are subject to change before code lands. Once implementation exists, content here will be regenerated from docstrings or sourced from running tests.

Dataset. EMMA-S2T-QA-v2 covers the task, target domain, input, labels, splits, and the download command; this recipe pins only the rfgen configuration. Source. rfgen plus the rfgen annotator (Phase 2 attribute Q&A).

rfgen recipe¶

rfgen commit: <rfgen-commit> (frozen at release; placeholder until pinned).
Emitters: the rfgen emitter_zoo group, a heterogeneous mix so Q&A pairs cover varied scene attributes.
Receiver array: 8-element ULA (uniform linear array), half-wavelength spacing (spacing_lambda: 0.5), via the rfgen MultiRXConfig.geometry member ULA. Loaded into ArraySpec at load time.
Channel environment(s): E_URBAN and E_RURAL, pinned through the rfgen channel group.
Labeler + label extraction: the rfgen annotator group (AnnotatorConfig with AnnotationType.QA) generates attribute question-answer pairs in Phase 2; LabelExtractor maps the stored answers to the target. Metrics are exact-match / F1.
SNR (signal-to-noise ratio) regime: drawn from the rfgen channel.snr_db_range.
Sample counts: <to be pinned at v2>.
Splits + seeds: train / dev / holdout seed ranges pinned at release; the holdout is a frozen secret seed, never published. OOD axis: SCENE_TYPE.

Real-capture counterpart¶

None. Question-answering is a frontier task with no real-capture counterpart in the EMMA-REAL-OOD set.

Notes¶

(depends on rfgen annotator not yet shipped) This recipe cannot pin until rfgen’s Phase 2 annotator produces verified attribute Q&A pairs at scale. Answer quality is bounded by the bulk and verifier large language models (LLMs) pinned in the rfgen AnnotatorConfig; the recipe inherits that dependency. Do not treat the recipe as complete until the annotator lands.
Ships at v2, after the inner ring (communications, spectrum, drone-RF) is proven.
Exact-match and F1 follow the SQuAD convention; tokenization of RF answers is the open task-side choice.

References¶

Rajpurkar et al., “SQuAD: 100,000+ Questions for Machine Comprehension of Text,” EMNLP 2016, arXiv:1606.05250. Exact-match and F1 for QA. (verify)
Davis and Goadrich, “The relationship between Precision-Recall and ROC curves,” ICML 2006. F1 grounding. (verify)
Debbah and the MBZUAI group, “RF-GPT,” 2026, arXiv:2602.14833. Spectrogram-based RF-language model; the closest existing analog. (verify)