Update app.py

app.py

@@ -24,7 +24,7 @@ class ConflictDetectionRequest(BaseModel):
 # -----------------------------
 def build_prompt(req1, req2, prompt_type="zero-shot"):
     if prompt_type == "zero-shot":
-        return f"Do the following sentences contradict each other, answer with just yes or no: 1.{req1} 2.{req2}"
+        return f"Do the following sentences contradict each other, yes or no: 1.{req1} 2.{req2}"
     elif prompt_type == "few-shot":
         examples = (
             "Example 1:\n"
@@ -38,20 +38,23 @@ def build_prompt(req1, req2, prompt_type="zero-shot"):
         )
         return examples + f"Now answer: Do the following sentences contradict each other? 1.{req1} 2.{req2}"
     else:
-        return f"Do the following sentences contradict each other, answer with just yes or no: 1.{req1} 2.{req2}"
+        return f"Do the following sentences contradict each other, yes or no: 1.{req1} 2.{req2}"
 
 # -----------------------------
 # Startup: load DeepSeek once
 # -----------------------------
 @app.on_event("startup")
 def load_models():
-    print("Loading smaller DeepSeek model into memory...")
-    deepseek_name = "deepseek-ai/deepseek-coder-1.3b-base"   # smaller, supported model
-    app.state.deepseek_tokenizer = AutoTokenizer.from_pretrained(deepseek_name)
+    print("Loading DeepSeek model into memory...")
+    model_name = "deepseek-ai/DeepSeek-R1-Distill-Llama-8B"
+    app.state.deepseek_tokenizer = AutoTokenizer.from_pretrained(model_name)
     app.state.deepseek_tokenizer.pad_token = app.state.deepseek_tokenizer.eos_token
+    
     app.state.deepseek_model = AutoModelForCausalLM.from_pretrained(
-        deepseek_name,
-        torch_dtype=torch.float32   # CPU only
+        model_name,
+        torch_dtype="auto",
+        device_map="auto",
+        offload_folder="offload"   # folder to store offloaded weights
     )