Upload app.py with huggingface_hub

app.py

@@ -32,7 +32,7 @@ def load_model():
             return False
     return True
 
-def generate_response(prompt, max_new_tokens=128, temperature=0.2):
+def generate_response(prompt, max_new_tokens=64, temperature=0.1):
     start = time.time()
     
     if not load_model():
@@ -42,16 +42,20 @@ def generate_response(prompt, max_new_tokens=128, temperature=0.2):
         # Prepare input
         inputs = _tokenizer(prompt, return_tensors="pt", truncation=True, max_length=512)
         
-        # Generate response
+        # Generate response with better repetition control
         with torch.no_grad():
             outputs = _model.generate(
                 **inputs,
                 max_new_tokens=int(max_new_tokens),
                 do_sample=bool(temperature > 0),
                 temperature=float(temperature),
-                top_p=0.9,
-                repetition_penalty=1.1,
-                pad_token_id=_tokenizer.eos_token_id
+                top_p=0.85,
+                top_k=50,
+                repetition_penalty=1.2,
+                no_repeat_ngram_size=3,
+                early_stopping=True,
+                pad_token_id=_tokenizer.eos_token_id,
+                eos_token_id=_tokenizer.eos_token_id
             )
         
         # Decode response
@@ -63,6 +67,32 @@ def generate_response(prompt, max_new_tokens=128, temperature=0.2):
         else:
             response = generated_text.strip()
         
+        # Clean up repetitive patterns
+        lines = response.split('\n')
+        cleaned_lines = []
+        seen_phrases = set()
+        
+        for line in lines:
+            line = line.strip()
+            if line and len(line) > 10:  # Only consider substantial lines
+                # Check for repetitive patterns
+                words = line.split()
+                if len(words) > 3:
+                    phrase = ' '.join(words[:3])  # First 3 words as phrase
+                    if phrase not in seen_phrases:
+                        seen_phrases.add(phrase)
+                        cleaned_lines.append(line)
+                else:
+                    cleaned_lines.append(line)
+            elif line and len(line) <= 10:
+                cleaned_lines.append(line)
+        
+        response = '\n'.join(cleaned_lines)
+        
+        # Limit response length to prevent runaway generation
+        if len(response) > 500:
+            response = response[:500] + "..."
+        
         elapsed = int((time.time() - start) * 1000)
         return response + f"\n\n(⏱️ {elapsed} ms)"
         
@@ -98,8 +128,8 @@ with gr.Blocks(title="RML-AI Demo") as demo:
     with gr.Row():
         prompt = gr.Textbox(label="Your question", value=SAMPLES[0], placeholder="Ask about AI, ML, RML, or any topic...")
     with gr.Row():
-        max_new = gr.Slider(32, 256, value=128, step=16, label="Max new tokens")
-        temp = gr.Slider(0.0, 1.0, value=0.2, step=0.1, label="Temperature")
+        max_new = gr.Slider(32, 256, value=64, step=16, label="Max new tokens")
+        temp = gr.Slider(0.0, 1.0, value=0.1, step=0.1, label="Temperature")
     with gr.Row():
         btn = gr.Button("Generate Response", variant="primary")
     output = gr.Textbox(label="RML-AI Response", lines=10)