Fix model generation: disable KV cache to avoid DynamicCache error

Root cause of 0% accuracy found!
All predictions were empty because model.generate() was throwing:
AttributeError: 'DynamicCache' object has no attribute 'seen_tokens'

This is a transformers version incompatibility with Phi-3 models.

Solution: Add use_cache=False to model.generate() to disable KV cache.
This avoids DynamicCache entirely and allows generation to succeed.

Side effect: Slightly slower generation (no KV cache), but models will work.

app.py

@@ -287,7 +287,8 @@ def evaluate_single_model(model_name, tasks, use_4bit=True, temperature=0.7, max
                         temperature=temperature,
                         do_sample=True,
                         top_p=0.9,
-                        pad_token_id=tokenizer.eos_token_id
+                        pad_token_id=tokenizer.eos_token_id,
+                        use_cache=False  # Disable KV cache to avoid DynamicCache compatibility issues
                     )
 
                 if i == 0: