Skip to content

Model Builder

The Model Builder is RIA Hub’s training interface. It generates GitHub Actions workflow files from a UI form, commits them to your repository, and lets you monitor the run from the browser. The runner executes the workflow and uploads training artifacts back to the repository.

ModuleWhat it does
Training a ModelConfigure a model template and training parameters, select a runner, and submit a training workflow
TemplateArchitectureInputBest for
MobileNetV3MobileNet (timm)IQ slices or spectrogramsLightweight classifiers, CPU runners
MobileNetV4MobileNet (timm)IQ slices or spectrogramsMore accurate than V3 with similar footprint
ResNet18ResNet (torchvision)IQ slices or spectrogramsStandard baseline
ResNet50ResNet (torchvision)IQ slices or spectrogramsHigher capacity baseline
GPT-2Transformer (HuggingFace)Token sequencesSequence classification tasks
DistilBERTTransformer (HuggingFace)Token sequencesFaster GPT-2 alternative
WavesFM Linear ProbeWavesFM (ViT-Small)IQ or spectrogramFast adaptation — freezes encoder, trains head only
WavesFM LoRAWavesFM (ViT-Small)IQ or spectrogramDeeper fine-tuning with low-rank weight updates
  • A curated radio_dataset asset in your repository Library — see Curating a Dataset
  • At least one runner registered and online — check Workflows → Management → Runners in your repository
  • RIAHUB_BASE_URL set as a repository variable or secret (value: https://riahub.ai or your instance URL)
  1. You fill in the Model Builder form: dataset, model template, hyperparameters, and runner
  2. RIA Hub generates .riahub/workflows/train.yaml and .riahub/train_configs/train.yaml and commits them to a branch in your repository
  3. The push triggers the Gitea Actions workflow on the selected runner
  4. The runner downloads the dataset, runs qmb train, and uploads training artifacts to the Actions run
  5. You download artifacts from the Actions UI and push the model back to your repository to register it in the Library