Add kaggle/train_grpo.py

kaggle/train_grpo.py

@@ -18,7 +18,6 @@ import re
 import subprocess
 import sys
 import traceback
-from dataclasses import dataclass
 from pathlib import Path
 from typing import Any
 
@@ -396,18 +395,32 @@ def make_eval_before_save_callback(
 # ---------------------------------------------------------------------------
 
 
-@dataclass
 class Cfg:
-    dataset_id: str = "vanishingradient/deepagents-rl-mbpp"
-    hub_model_id: str = "vanishingradient/deepagents-rl-coding"
-    max_steps: int = 40
-    save_steps: int = 5
-    num_generations: int = 4
-    learning_rate: float = 1e-6
-    max_completion_length: int = 512
-    max_seq_length: int = 2048
-    n_eval: int = 16
-    lora_r: int = 16
+    """Plain class (not dataclass) — safe under importlib.exec_module on Py3.13."""
+
+    def __init__(
+        self,
+        dataset_id: str = "vanishingradient/deepagents-rl-mbpp",
+        hub_model_id: str = "vanishingradient/deepagents-rl-coding",
+        max_steps: int = 40,
+        save_steps: int = 5,
+        num_generations: int = 4,
+        learning_rate: float = 1e-6,
+        max_completion_length: int = 512,
+        max_seq_length: int = 2048,
+        n_eval: int = 16,
+        lora_r: int = 16,
+    ) -> None:
+        self.dataset_id = dataset_id
+        self.hub_model_id = hub_model_id
+        self.max_steps = max_steps
+        self.save_steps = save_steps
+        self.num_generations = num_generations
+        self.learning_rate = learning_rate
+        self.max_completion_length = max_completion_length
+        self.max_seq_length = max_seq_length
+        self.n_eval = n_eval
+        self.lora_r = lora_r
 
 
 def main(cfg: Cfg | None = None) -> None: