Supported Models
All 12 models share the supervised training, fine-tuning, validation, testing, and checkpoint workflow through the common runner.
Choose the model's input representation and use its --model value with the shared experiment runner.
| Model | Input | Entry point |
|---|---|---|
| NeuroSTORM | Voxel sequences | --model neurostorm |
| SwiFT | Voxel sequences | --model swift |
| BAnD | Voxel sequences | --model band |
| Brain-JEPA | Ordered 450-ROI time series | --model brainjepa |
| BrainGNN | FC-derived graphs | --model braingnn |
| LG-GNN | FC-derived graphs | --model lggnn |
| IBGNN | FC-derived graphs | --model ibgnn |
| BNT | FC matrices | --model bnt |
| Com-BrainTF | FC matrices | --model combraintf |
| BrainNetCNN | FC matrices | --model brainnetcnn |
| BrainLM | Ordered ROI time series | --model brainlm |
| TFF | Voxel sequences | --model tff |
Brain-JEPA requires ROI signals in the same 450-region order as its gradient table. NeuroSTORM MAE and SwiFT contrastive learning have model-specific pretraining configurations. The dedicated Task4 evaluator covers NeuroSTORM, SwiFT, BrainLM, TFF, BNT, Com-BrainTF, LG-GNN, and BrainGNN. See the model configuration reference and dataset support table to select inputs and tasks.