feat(train): align native SmolVLA recipe

tests/test_smolvla_native_agent.py

@@ -416,12 +416,15 @@ class SmolVLANativeAgentTest(unittest.TestCase):
         self.assertEqual(cfg.agent._target_, 'roboimi.vla.agent_smolvla_native.SmolVLANativeAgent')
         self.assertEqual(cfg.agent.action_dim, 16)
         self.assertEqual(cfg.agent.obs_dim, 16)
+        self.assertEqual(cfg.agent.normalization_type, 'gaussian')
         self.assertEqual(list(cfg.agent.camera_names), list(cfg.data.camera_names))
         self.assertEqual(cfg.agent.num_cams, len(cfg.data.camera_names))
-        self.assertEqual(cfg.agent.chunk_size, 16)
-        self.assertEqual(cfg.agent.n_action_steps, 8)
+        self.assertEqual(cfg.agent.chunk_size, 32)
+        self.assertEqual(cfg.agent.n_action_steps, 16)
         self.assertEqual(cfg.agent.model_config.max_state_dim, 32)
         self.assertEqual(cfg.agent.model_config.max_action_dim, 32)
+        self.assertEqual(cfg.agent.model_config.chunk_size, 32)
+        self.assertEqual(cfg.agent.model_config.n_action_steps, 16)
         self.assertIsNone(cfg.agent.dataset_image_resize_shape)
         self.assertIsNone(cfg.agent.eval_image_resize_shape)
         self.assertEqual(list(cfg.agent.model_config.resize_imgs_with_padding), [512, 512])
@@ -430,6 +433,13 @@ class SmolVLANativeAgentTest(unittest.TestCase):
         self.assertNotIn('tokenizer_name', cfg.agent.model_config)
         self.assertEqual(cfg.agent.model_config.vlm_model_name, 'HuggingFaceTB/SmolVLM2-500M-Video-Instruct')
 
+    def test_tokenize_tasks_preserves_existing_single_newline(self):
+        agent = _make_agent(model_config={'resize_imgs_with_padding': None, 'pad_language_to': 'max_length', 'tokenizer_max_length': 12})
+
+        agent._tokenize_tasks(['already newline\n', 'needs newline'], batch_size=2, device=torch.device('cpu'))
+
+        self.assertEqual(agent.tokenizer.calls[-1]['texts'], ['already newline\n', 'needs newline\n'])
+
 
 if __name__ == '__main__':
     unittest.main()