RF Foundation Models

RFFM is an early PyTorch-native system for learning from radio-frequency (RF) in-phase and quadrature (I/Q) signals. The repository currently proves one bounded causal-pretraining path on a canonical RadioML2018 dataset. It does not yet provide full pretraining, evaluation, or a released Python package.

Current qualification

The verified path composes one YAML experiment, reads one canonical dataset, prepares fixed I/Q patches, runs a tiny causal next-patch-prediction model with AdamW, writes strict attempt artifacts and checkpoints, and resumes a later attempt through Gemini Enterprise Agent Platform. The accepted cloud evidence advances from step 1 to step 2. Finite loss in this bounded check is not a model-quality claim.

Start here

  1. Getting Started explains the repository environment, configuration check, and prerequisites for a bounded run.

  2. Architecture separates reusable primitives, the pretraining application, and the Gemini Enterprise Agent Platform submission adapter.

  3. Maturity Matrix shows the evidence level and current boundary for each documented surface.

Responsibility layers

The landed bounded path has four responsibility layers:

  1. Canonical data and signal preparation produce semantic RF samples, batches, normalized physical metadata, and fixed I/Q patches.

  2. Reusable model and training primitives implement the tiny causal model, next-patch-prediction objective, finite-value checks, and checkpoint I/O.

  3. The pretraining application composes validated configuration, dispatches code-owned component registrations, owns run and attempt artifacts, and validates bounded resume compatibility.

  4. A shell adapter turns a separate launch configuration into a Gemini Enterprise Agent Platform CustomJob. Google Cloud owns provisioning, status, logs, and worker cleanup.

The Concepts, How-to Guides, and Reference sections distinguish that current path from unsupported train/validation, distributed, post-training, and evaluation work.