updatE

app.py

@@ -48,18 +48,28 @@ def get_response(text, tokenizer=tokenizer, model=model):
     generate_ids = model.generate(
         input_ids,
         attention_mask=attention_mask,
-        top_p=0.9,
-        temperature=0.3,
-        max_length=2048,
+        top_p=0.8,
+        temperature=0.2,
+        max_length=input_len + 256,  # Limit response length to prevent multiple responses
         min_length=input_len + 4,
-        repetition_penalty=1.2,
+        repetition_penalty=1.3,
         do_sample=True,
-        pad_token_id=tokenizer.pad_token_id
+        pad_token_id=tokenizer.pad_token_id,
+        eos_token_id=tokenizer.eos_token_id  # Stop at end of sentence
     )
     response = tokenizer.batch_decode(
         generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=True
     )[0]
     response = response.split("### Response :")[-1].lstrip()
+
+    # Extract only the first AI response to prevent multiple responses
+    if "[|AI|]" in response and "[|Human|]" in response:
+        # If there are multiple turns, take only the first AI response
+        response = response.split("[|Human|]")[0].strip()
+    elif "[|AI|]" in response:
+        # Remove the [|AI|] marker from the beginning
+        response = response.replace("[|AI|]", "").strip()
+
     return response
 
 def format_conversation_history(chat_history):
@@ -110,13 +120,13 @@ def generate_response(input_data, chat_history, max_new_tokens, temperature, top
 
     # Add current user message
     conversation_parts.append(f"[|Human|] {input_data}")
-    conversation_parts.append("[|AI|]")
 
     # Join conversation
     conversation = "\n".join(conversation_parts)
 
     # Create full prompt using Jais format with our system prompt
-    full_prompt = f"### Instruction:{system_instruction}\n### Input:{conversation}\n### Response :"
+    # Note: We don't add "[|AI|]" at the end to prevent the model from generating multiple responses
+    full_prompt = f"### Instruction:{system_instruction}\n### Input:{conversation}\n[|AI|]\n### Response :"
 
     try:
         # استخدام دالة get_response من documentation