Logs

app.py

@@ -151,10 +151,11 @@ def evaluate_model(model, dataloader, device):
             # Process each sequence in the batch
             for i in range(current_batch_size):
                 valid_mask = (labels[i] != -100) & (attention_mask[i] != 0)
-                valid_preds = preds[i][valid_mask[i]].flatten()
-                valid_labels = labels[i][valid_mask[i]].flatten()
+                valid_preds = preds[i][valid_mask].flatten()
+                valid_labels = labels[i][valid_mask].flatten()
                 print(valid_mask.dtype)
                 print(labels[i].shape)
+                print(labels[i])
                 print(attention_mask[i].shape)
                 print(valid_mask.shape)
                 print(valid_labels)
@@ -162,6 +163,8 @@ def evaluate_model(model, dataloader, device):
                 all_preds.extend(valid_preds.tolist())
                 all_labels.extend(valid_labels.tolist())
 
+                assert not torch.any(valid_labels == -100), f"Found -100 in valid_labels for batch {i}"
+
                 if sample_count < num_samples:
                     print(f"Sample {sample_count + 1}:")
                     print(f"Tokens: {tokenizer.convert_ids_to_tokens(input_ids[i])}")