RF & ML Quick Overview

You do not need an RF-specialist background to begin using rfgen. You do need a clear distinction between the signal a model sees and the information rfgen keeps to make that signal usable for supervised learning.

From configured scene to training input

  1. A scene specifies the synthetic emitters and conditions: frequency span, sample rate, duration, noise, placement, and channel effects.

  2. An emitter produces a waveform. After composition, rfgen stores its receiver observation as complex baseband IQ: paired in-phase (I) and quadrature (Q) values preserving amplitude and phase.

  3. A labeler turns known generation truth into structured supervision, such as a modulation class, a bounding box, or occupancy. A label is not a model prediction.

  4. An optional annotation overlay adds text supervision to an existing record. It is append-only and never changes IQ or labels.

  5. A store persists IQ, labels, and provenance. Native local and GCS generation uses Signal Dataset so a reader can access records by ordinal or scan metadata without loading tensors.

  6. A training consumer opens the published snapshot and converts named fields to its framework at the task boundary; the model owns batching and loss.

        flowchart LR
    A[scene and emitters] --> B[complex baseband IQ]
    A --> C[structured labels]
    B --> D[dataset store]
    C --> D
    E[optional annotation overlay] --> D
    D --> F[training consumer]

A small example

In the Quickstart, one scene has one communication emitter. The record’s IQ is model input; its BPSK, QPSK, 16-QAM, or 64-QAM label is potential supervision. These are modulation-format class labels: they identify which waveform format generated the IQ. Configuration and provenance explain how the record was generated; they are not automatically model inputs.

What rfgen does not claim

Synthetic generation provides controlled examples and exact generation-side labels. It does not prove that a model will transfer to a receiver, deployment environment, or corpus the configuration does not model. Read a Golden Path for a bounded qualification claim and the Signal Catalog for the available signal surface.