Small main readme update

README.md

@@ -33,8 +33,8 @@ Two MLP models trained on CIFAR-10 that were used to generate **Figure 3** in ou
 
 **Training Details:**
 - Dataset: CIFAR-10
-- Purpose: Figure 3 generation and analysis
-- Architecture: Multi-layer perceptrons with spectral constraints
+- Selection: These are the models we used to generate Figure 3
+- Architecture: Multi-layer perceptrons
 
 ### Transformer Models (`models/transformers/`)
 Best-performing transformer model trained on the Shakespeare word dataset, representing our optimal validation accuracy checkpoint.