Fix: Extend max_length for proper text reconstruction

app.py

@@ -105,13 +105,16 @@ class B2NLTokenizer:
 
             # Reconstruct (full text, not truncated)
             with torch.no_grad():
-                reconstructed = self.model.generate(text, temperature=temperature, max_length=48)
+                # Calculate appropriate max_length based on input
+                max_gen_length = max(48, min(len(text) + 10, 512))  # Allow some extra space
 
-                # For long texts, process multiple chunks
+                reconstructed = self.model.generate(text, temperature=temperature, max_length=max_gen_length)
+
+                # For long texts, ensure we get full reconstruction
                 if text_bytes > 48:
-                    # Process with sliding window
+                    # Current model limitation: may not fully reconstruct very long texts
+                    # This is due to sliding window processing
                     full_reconstruction = reconstructed
-                    # Note: Current implementation may truncate, this is a known limitation
                 else:
                     full_reconstruction = reconstructed