Update app.py

app.py

@@ -61,16 +61,6 @@ config = AutoConfig.from_pretrained(REPO_NAME, token=HF_TOKEN)
 
 device = "cuda" if torch.cuda.is_available() else "cpu"
 
-#tokenizer = LlamaTokenizerFast.from_pretrained(REPO_NAME, token=HF_TOKEN)
-#tokenizer = LlamaTokenizer.from_pretrained(REPO_NAME, token=HF_TOKEN)
-
-#tokenizer = AutoTokenizer.from_pretrained("microsoft/Phi-3-mini-4k-instruct")
-# tokenizer = AutoTokenizer.from_pretrained(
-#     "microsoft/Phi-3-mini-4k-instruct",
-#     trust_remote_code=True,
-#     use_auth_token=HF_TOKEN
-# )
-
 tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL, trust_remote_code=True)
 
 model = AutoModelForCausalLM.from_pretrained(
@@ -95,6 +85,10 @@ def is_farewell(message: str) -> bool:
     message_lower = message.lower().strip()
     return any(farewell in message_lower for farewell in farewells)
 
+def clean_input_text(text: str) -> str:
+    # Remove any "Instruction N: ..." or similar phrases
+    return re.sub(r"Instruction\s*\d+\s*\(.*?\):", "", text, flags=re.IGNORECASE)
+
 @app.post("/chat")
 async def chat(request: ChatRequest):
     try:
@@ -119,13 +113,17 @@ async def chat(request: ChatRequest):
         seen_messages = set()
         for msg in history:
             if msg.role == "user" and msg.text.strip() not in seen_messages:
-                conversation += f"<|user|>\n{msg.text.strip()}\n"
-                seen_messages.add(msg.text.strip())
+                cleaned_user_text = clean_input_text(msg.text.strip())
+                conversation += f"<|user|>\n{cleaned_user_text}\n"
+                seen_messages.add(cleaned_user_text)
             elif msg.role == "model":
-                conversation += f"<|assistant|>\n{msg.text.strip()}\n"
+                cleaned_model_text = clean_input_text(msg.text.strip())
+                conversation += f"<|assistant|>\n{cleaned_model_text}\n"
 
-        conversation += f"<|user|>\n{user_message.strip()}\n<|assistant|>"
 
+        #conversation += f"<|user|>\n{user_message.strip()}\n<|assistant|>"
+        conversation += f"<|user|>\n{clean_input_text(user_message.strip())}\n<|assistant|>"
+        
         inputs = tokenizer(conversation, return_tensors="pt", padding=True, truncation=True, max_length=4096).to(device)
 
         with torch.no_grad():