Update app.py

app.py

@@ -9,7 +9,6 @@ from transformers import AutoTokenizer, AutoModelForCausalLM, TextIteratorStream
 MODEL_REPO = "daniel-dona/gemma-3-270m-it"
 LOCAL_DIR = os.path.join(os.getcwd(), "local_model")
 
-# CPU optimizasyonları
 os.environ.setdefault("HF_HUB_ENABLE_HF_TRANSFER", "1")
 os.environ.setdefault("OMP_NUM_THREADS", str(os.cpu_count() or 1))
 os.environ.setdefault("MKL_NUM_THREADS", os.environ["OMP_NUM_THREADS"])
@@ -41,9 +40,6 @@ model_path = ensure_local_model(MODEL_REPO, LOCAL_DIR)
 
 tokenizer = AutoTokenizer.from_pretrained(model_path, local_files_only=True)
 
-### DEĞİŞİKLİK BURADA: ŞABLON BASİTLEŞTİRİLDİ ###
-# 'raise_exception' komutunu içermeyen, eski transformers versiyonlarıyla uyumlu şablon.
-# Zaten kodumuz şablonu doğru formatta beslediği için bu kontrolleri kaldırabiliriz.
 gemma_chat_template_simplified = (
     "{% for message in messages %}"
         "{% if message['role'] == 'user' %}"
@@ -58,10 +54,7 @@ gemma_chat_template_simplified = (
 )
 
 if tokenizer.chat_template is None:
-    print("Chat template manuel olarak ayarlanıyor (basitleştirilmiş versiyon).")
     tokenizer.chat_template = gemma_chat_template_simplified
-### DEĞİŞİKLİK SONA ERDİ ###
-
 
 model = AutoModelForCausalLM.from_pretrained(
     model_path,
@@ -71,7 +64,6 @@ model = AutoModelForCausalLM.from_pretrained(
 )
 model.eval()
 
-# Çok katı moderasyon system prompt
 MODERATION_SYSTEM_PROMPT = (
     "You are a multilingual content moderation classifier. "
     "You MUST respond with exactly one lowercase letter: 's' for safe, 'u' for unsafe. "
@@ -82,12 +74,8 @@ MODERATION_SYSTEM_PROMPT = (
 )
 
 def build_prompt(message, max_ctx_tokens=128):
-    # Sistem mesajını ilk kullanıcı mesajının bir parçası haline getiriyoruz.
     full_user_message = f"{MODERATION_SYSTEM_PROMPT}\n\nUser input: '{message}'"
-    
-    messages = [
-        {"role": "user", "content": full_user_message}
-    ]
+    messages = [{"role": "user", "content": full_user_message}]
     
     text = tokenizer.apply_chat_template(
         messages,
@@ -96,7 +84,7 @@ def build_prompt(message, max_ctx_tokens=128):
     )
     
     while len(tokenizer(text, add_special_tokens=False).input_ids) > max_ctx_tokens and len(full_user_message) > 100:
-        full_user_message = full_user_message[:len(full_user_message)-50]
+        full_user_message = full_user_message[:-50]
         messages[0]['content'] = full_user_message
         text = tokenizer.apply_chat_template(
             messages,
@@ -106,15 +94,14 @@ def build_prompt(message, max_ctx_tokens=128):
     return text
 
 def enforce_s_u(text: str) -> str:
-    """Model çıktısını kesin olarak 's' veya 'u' ile sınırla."""
     text_lower = text.strip().lower()
-    if "u" in text_lower and not "s" in text_lower:
+    if "u" in text_lower and "s" not in text_lower:
         return "u"
     if "unsafe" in text_lower:
         return "u"
     return "s"
 
-def respond_stream(message, history, max_tokens, temperature, top_p):
+def classify_text_stream(message, max_tokens, temperature, top_p):
     text = build_prompt(message)
     inputs = tokenizer([text], return_tensors="pt").to(model.device)
     do_sample = bool(temperature and temperature > 0.0)
@@ -148,6 +135,7 @@ def respond_stream(message, history, max_tokens, temperature, top_p):
                     start_time = time.time()
                 partial_text += chunk
                 token_count += 1
+                yield partial_text
         finally:
             thread.join()
 
@@ -155,18 +143,49 @@ def respond_stream(message, history, max_tokens, temperature, top_p):
     end_time = time.time() if start_time else time.time()
     duration = max(1e-6, end_time - start_time)
     tps = token_count / duration if duration > 0 else 0.0
-    yield f"{final_label}\n\n⚡ Speed: {tps:.2f} token/s"
-
-demo = gr.ChatInterface(
-    respond_stream,
-    additional_inputs=[
-        gr.Slider(minimum=1, maximum=4, value=1, step=1, label="Max new tokens"),
-        gr.Slider(minimum=0.0, maximum=1.0, value=0.0, step=0.1, label="Temperature"),
-        gr.Slider(minimum=0.1, maximum=1.0, value=0.95, step=0.05, label="Top-p")
-    ],
-    title="Strict Multilingual Moderation Classifier (s/u)",
-    description="Enter any text in any language. The model will output only 's' (safe) or 'u' (unsafe)."
-)
+    yield f"{final_label}\n\n⚡ Speed: {tps:.2f} tokens/s"
+
+with gr.Blocks() as demo:
+    gr.Markdown("# Multilingual Content Moderation Classifier")
+    gr.Markdown("Enter any text to classify it as safe ('s') or unsafe ('u').")
+    
+    with gr.Row():
+        with gr.Column(scale=2):
+            text_input = gr.Textbox(
+                label="Text to Classify", 
+                lines=5, 
+                placeholder="Enter text in any language..."
+            )
+            submit_button = gr.Button("Classify", variant="primary")
+        
+        with gr.Column(scale=1):
+            text_output = gr.Textbox(label="Classification Result", interactive=False)
+            with gr.Accordion("Advanced Settings", open=False):
+                max_tokens_slider = gr.Slider(
+                    minimum=1, maximum=4, value=1, step=1, label="Max New Tokens"
+                )
+                temp_slider = gr.Slider(
+                    minimum=0.0, maximum=1.0, value=0.0, step=0.1, label="Temperature"
+                )
+                top_p_slider = gr.Slider(
+                    minimum=0.1, maximum=1.0, value=0.95, step=0.05, label="Top-p"
+                )
+
+    gr.Examples(
+        examples=[
+            ["Hello, how are you today?"],
+            ["I will find you and hurt you."],
+            ["C'est une belle journée pour apprendre le codage."],
+            ["I want to die."],
+        ],
+        inputs=text_input
+    )
+
+    submit_button.click(
+        fn=classify_text_stream,
+        inputs=[text_input, max_tokens_slider, temp_slider, top_p_slider],
+        outputs=text_output
+    )
 
 if __name__ == "__main__":
     with torch.inference_mode():