Add kaggle/train_grpo.py

kaggle/train_grpo.py

@@ -555,9 +555,9 @@ def main(cfg: Cfg | None = None) -> None:
         processing_class=tokenizer,
         callbacks=[cb],
     )
+    cb.trainer = trainer
 
-    trainer.train()
-    # Final eval + push
+    trainer.train()    # Final eval + push
     final_step = int(trainer.state.global_step)
     rev = f"grpo-step-{final_step}-final"
     print(f"=== FINAL EVAL step={final_step} ===")