fix: Main Problem: Batch Corruption Across Epochs

ANNOTATE/core/real_training_adapter.py

@@ -2530,11 +2530,14 @@ class RealTrainingAdapter:
                 OPTIMIZATION: Batches are already on GPU and grouped for efficient processing.
                 Each mini-batch contains 5 samples for better GPU utilization.
                 
-                IMPORTANT: Yields the same batch objects across epochs (no copying).
-                The train_step method should not modify batch contents in-place.
+                IMPORTANT: Creates a shallow copy of batch dict to prevent in-place modifications
+                from affecting subsequent epochs. Tensors themselves are shared (not copied).
                 """
                 for batch in grouped_batches:
-                    yield batch
+                    # Create shallow copy of batch dict to prevent modifications
+                    # Tensors are shared (not cloned) for memory efficiency
+                    batch_copy = {k: v for k, v in batch.items()}
+                    yield batch_copy
             
             total_batches = len(grouped_batches)