Disable thinking channel in rewrites

app.py

@@ -54,6 +54,7 @@ def humanize(text: str, rewrite_strength: str = "Balanced") -> str:
         add_generation_prompt=True,
         return_dict=True,
         return_tensors="pt",
+        enable_thinking=False,
     ).to("cuda")
     decoder_input_ids = processor.tokenizer(
         source,
@@ -84,7 +85,7 @@ def humanize(text: str, rewrite_strength: str = "Balanced") -> str:
         skip_special_tokens=True,
     ).strip()
 
-    for label in ("rewritten text:", "rewrite:", "text:"):
+    for label in ("thought", "rewritten text:", "rewrite:", "text:"):
         if result.casefold().startswith(label):
             result = result[len(label) :].lstrip()
             break