Basic Training Materials

Training Vocabulary

Core terms

source Recording A raw RF capture or generated signal file.data Radio Dataset A structured collection of labeled examples.label Slice A useful interval extracted from a recording.metadata Qualifier A condition or attribute attached to a slice.train Model Template A configured training path such as a WavesFM workflow.model Foundation Model A pretrained base that can be adapted for RF tasks.

source

Recording

Recordings usually enter through Library storage and become source material for slicing.

What Makes a Dataset Trainable

A trainable dataset has labels, splits, consistent sample shape, metadata, and enough examples for the target class or scenario. The WavesFM path adds adapter and transform expectations, so mismatched input shape or missing label fields should be fixed before launching a run.

Class balance preview

Example modulation classes before and after curation checks.

• Before
• After

Data table

Series	BPSK	QPSK	8PSK	16QAM
Before	32 slices	18 slices	12 slices	9 slices
After	28 slices	27 slices	25 slices	24 slices

Training curve placeholder

Example loss and accuracy trend for a short model run.

• Loss
• Accuracy

Data table

Series	0	1	2	3	4	5
Loss	1.2	0.94	0.72	0.58	0.47	0.41
Accuracy	0.42	0.55	0.68	0.74	0.8	0.84

Training gotchas

Labels are missing or too broad

Define the target label before slicing so the dataset can be evaluated consistently.

Sample shapes are inconsistent

Check that transforms produce the expected tensor shape for the selected model template.

Metadata is incomplete

Keep source OIDs, branch, directory, and qualifier information attached to examples.

Next steps

• Inspect the dataset — Inspect a Dataset walks through balance analysis and anomaly detection before launching a run.
• Train a model — When the dataset passes inspection, continue to Train a Model.