debug

app.py

@@ -189,8 +189,16 @@ def answer(system: str, context: str, question: str,
             top_p=top_p,
             top_k=top_k_tok
         )
-        full   = tokenizer.decode(output[0], skip_special_tokens=True)
-        reply  = full.split("<|im_start|>assistant")[-1].strip() + prompt
+        full   = tokenizer.decode(output[0], skip_special_tokens=False)
+
+        start = "<|start_header_id|>assistant<|end_header_id|>\n\n"
+        end = "<|eot_id|>"
+
+        if start in full:
+            reply = full.split(start)[-1].split(end)[0].strip()
+        else:
+            reply = full
+
         return reply
           
     except Exception as e: