Update app.py

app.py

@@ -11,7 +11,7 @@ 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("OMP_NUM_THREADS", str(os.cpu_count() or 2))
 os.environ.setdefault("MKL_NUM_THREADS", os.environ["OMP_NUM_THREADS"])
 os.environ.setdefault("OMP_PROC_BIND", "TRUE")
 
@@ -48,42 +48,46 @@ model = AutoModelForCausalLM.from_pretrained(
 )
 model.eval()
 
-# Çok dilli moderasyon system prompt
+# Çok katı moderasyon system prompt (yalnızca "s" veya "u")
 MODERATION_SYSTEM_PROMPT = (
     "You are a multilingual content moderation classifier. "
-    "You analyze the user's message in any language and decide if it is safe or unsafe. "
-    "Rules: If the message contains hate speech, harassment, sexual content involving minors, "
-    "extreme violence, self-harm encouragement, or other unsafe material, respond with exactly 'unsafe'. "
-    "If it is acceptable and safe, respond with exactly 'safe'. "
-    "Do not explain, do not add anything else, only output 'safe' or 'unsafe'."
+    "You MUST respond with exactly one lowercase letter: 's' for safe, 'u' for unsafe. "
+    "No explanations, no punctuation, no extra words. "
+    "If the message contains hate speech, harassment, sexual content involving minors, "
+    "extreme violence, self-harm encouragement, or other unsafe material, respond 'u'. "
+    "Otherwise respond 's'."
 )
 
-def build_prompt(message, max_ctx_tokens=512):
-    msgs = [
+def build_prompt(message, max_ctx_tokens=128):
+    messages = [
         {"role": "system", "content": MODERATION_SYSTEM_PROMPT},
         {"role": "user", "content": message}
     ]
-    chat_template = """{% for m in messages %}
-{{ m['role'] }}: {{ m['content'] }}
-{% endfor %}
-Assistant:"""
+    # Doğru chat template kullanımı
     text = tokenizer.apply_chat_template(
-        msgs,
-        chat_template=chat_template,
+        messages,
         tokenize=False,
         add_generation_prompt=True
     )
     # Token sınırını aşarsa kısalt
-    while len(tokenizer(text, add_special_tokens=False).input_ids) > max_ctx_tokens and len(msgs) > 2:
-        msgs.pop(1)
+    while len(tokenizer(text, add_special_tokens=False).input_ids) > max_ctx_tokens and len(messages) > 2:
+        messages.pop(1)
         text = tokenizer.apply_chat_template(
-            msgs,
-            chat_template=chat_template,
+            messages,
             tokenize=False,
             add_generation_prompt=True
         )
     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:
+        return "u"
+    if "unsafe" in text_lower:
+        return "u"
+    return "s"
+
 def respond_stream(message, history, max_tokens, temperature, top_p):
     text = build_prompt(message)
     inputs = tokenizer([text], return_tensors="pt").to(model.device)
@@ -118,25 +122,26 @@ def respond_stream(message, history, max_tokens, temperature, top_p):
                     start_time = time.time()
                 partial_text += chunk
                 token_count += 1
-                yield partial_text.strip()
         finally:
             thread.join()
 
+    # Çıktıyı kesin olarak s/u'ya indir
+    final_label = enforce_s_u(partial_text)
     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 partial_text.strip() + f"\n\n⚡ Speed: {tps:.2f} token/s"
+    yield f"{final_label}\n\n⚡ Speed: {tps:.2f} token/s"
 
 demo = gr.ChatInterface(
     respond_stream,
     chatbot=False,
     additional_inputs=[
-        gr.Slider(minimum=1, maximum=16, value=4, step=1, label="Max new tokens"),
+        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="Multilingual Moderation Classifier",
-    description="Enter any text in any language. The model will output only 'safe' or 'unsafe'."
+    title="Strict Multilingual Moderation Classifier (s/u)",
+    description="Enter any text in any language. The model will output only 's' (safe) or 'u' (unsafe)."
 )
 
 if __name__ == "__main__":