Add kaggle/train_grpo.py

kaggle/train_grpo.py

@@ -614,7 +614,7 @@ def main(cfg: Cfg | None = None) -> None:
     import requests
 
     grader = os.environ["SWE_GRADER_URL"].rstrip("/")
-    headers = {}
+    headers = {"ngrok-skip-browser-warning": "1"}
     if os.environ.get("SWE_GRADER_TOKEN"):
         headers["Authorization"] = f"Bearer {os.environ['SWE_GRADER_TOKEN']}"
     health = requests.get(f"{grader}/health", headers=headers, timeout=30)