Trained on Kaggle (notebook 'MRT Fewer Encoders RML2016') with train_noconv.py + train_all_models_v2.py,
same recipe, split and normalization as the 6-encoder model, only --layers 3.
Model: TransformerNoConv, 3 encoders, no convolution, best epoch 106, 601,995 parameters.
best_model.pt is exported to CSV in 03_weight_export/MRTransformer_RML2016.
