Add kaggle/train_grpo.py

kaggle/train_grpo.py

@@ -298,17 +298,7 @@ def make_eval_before_save_callback(
     class EvalBeforeCheckpointCallback(TrainerCallback):
         def __init__(self) -> None:
             self._last_eval_step = -1
-
-        def on_save(
-            self,
-            args: TrainingArguments,
-            state: TrainerState,
-            control: TrainerControl,
-            **kwargs,
-        ):
-            # Transformers calls on_save AFTER saving locally. We also hook on_step_end
-            # to eval+langfuse BEFORE hub push. See on_step_end.
-            return control
+            self.trainer = None  # set after GRPOTrainer construction
 
         def on_step_end(
             self,
@@ -325,16 +315,20 @@ def make_eval_before_save_callback(
             self._last_eval_step = step
 
             model = kwargs.get("model")
-            tokenizer = kwargs.get("processing_class") or kwargs.get("tokenizer")
-            # GRPOTrainer may put tokenizer on trainer; fallback via kwargs["train"] not available
-            trainer = kwargs.get("trainer")  # usually absent
-            if tokenizer is None:
-                # recover from module global set in main()
-                tokenizer = globals().get("_TOKENIZER")
+            if model is None and self.trainer is not None:
+                model = self.trainer.model
+            tokenizer = (
+                kwargs.get("processing_class")
+                or kwargs.get("tokenizer")
+                or globals().get("_TOKENIZER")
+            )
+            if tokenizer is None and self.trainer is not None:
+                tokenizer = getattr(self.trainer, "processing_class", None) or getattr(
+                    self.trainer, "tokenizer", None
+                )
             if model is None or tokenizer is None:
-                print("[ckpt-eval] missing model/tokenizer; skipping")
-                return control
-
+                print("[ckpt-eval] missing model/tokenizer; aborting save pipeline")
+                raise RuntimeError("ckpt-eval requires model+tokenizer")
             rev = f"grpo-step-{step}"
             problems = eval_problems[:n_eval]
             print(f"\n=== EVAL BEFORE CHECKPOINT step={step} rev={rev} n={len(problems)} ===")