RF records for ML engineers¶
RFGen produces an immutable collection of records for downstream learning or evaluation. Each record contains complex in-phase/quadrature (IQ) waveform samples, typed metadata and labels, a stable scene identifier, and provenance describing the authored graph and runtime. IQ is the two-component numerical representation of a band-limited radio waveform; models can consume it much as audio models consume sampled amplitudes, while preserving phase information.
An RF scene is the set of emitters, propagation effects, receiver effects, and labels realized for one record. The declarative graph describes those relationships. Compilation validates types and units before numerical node operators run. Structural randomness chooses graph structure; run/sample/shard seeds schedule records. Both are keyed so unrelated edits do not silently shift all later draws.
SDS (Signal Dataset Storage) is the built-in physical format. The generic
Storage interface allows another implementation to preserve the same
write/publish/open lifecycle. A published collection is immutable and
indexable; RFGen does not prescribe training batches or sampling policy.
RF terms used in the reference¶
baseband: frequencies expressed relative to the receiver’s center; 0 Hz is the tuned center rather than an absolute radio frequency.
occupied bandwidth: RFGen’s authored placement-and-label interval. It is source provenance, not a production measurement or a claim to implement a particular percentage-power bandwidth standard.
Nyquist interval: the representable baseband interval
[-sample_rate/2, sample_rate/2].chirp: a waveform whose instantaneous frequency changes over time; radar commonly uses a linear-frequency-modulated chirp.
SigMF: a standard metadata format for recorded signal samples.
YOLO projection: normalized detection boxes shaped for YOLO-style object detection; it is a derived inspection view, not the native label schema.
Rayleigh fading: a small-scale random multipath model without a dominant direct path.
TDL/CDL: tapped-delay-line and clustered-delay-line 3GPP channel models; they trade increasing spatial detail and compute for propagation realism.
UMa/UMi/RMa: 3GPP urban-macro, urban-micro, and rural-macro system models; choosing one changes the deployment geometry and propagation distribution.
RT: ray tracing, which solves paths through explicit 3D geometry and is substantially more compute-intensive than statistical channel models.