diff --git a/tests/experimental/test_gold_trainer.py b/tests/experimental/test_gold_trainer.py
index 619e0182626..f216badf6e3 100644
--- a/tests/experimental/test_gold_trainer.py
+++ b/tests/experimental/test_gold_trainer.py
@@ -486,7 +486,9 @@ def test_gold_trainer_init_defaults_vllm_max_model_length_to_max_length(monkeypa
 
     class DummyStudentModel:
         def __init__(self):
-            self.config = SimpleNamespace(_name_or_path="student", vocab_size=17)
+            config = SimpleNamespace(_name_or_path="student", vocab_size=17)
+            config.get_text_config = lambda: config
+            self.config = config
             self.generation_config = SimpleNamespace(eos_token_id=2)
             self.name_or_path = "student"
 
diff --git a/trl/experimental/distillation/distillation_trainer.py b/trl/experimental/distillation/distillation_trainer.py
index a4879f62b19..6134c7957e5 100644
--- a/trl/experimental/distillation/distillation_trainer.py
+++ b/trl/experimental/distillation/distillation_trainer.py
@@ -549,7 +549,7 @@ def __init__(
         # ── Prepare teacher model (after super().__init__ so accelerator is ready) ──
         if teacher_model is not None:
             if self._local_teacher_tokenizer_matches_student:
-                teacher_model.resize_token_embeddings(self.model.config.vocab_size)
+                teacher_model.resize_token_embeddings(self.model.config.get_text_config().vocab_size)
             if self.is_deepspeed_enabled:
                 self.teacher_model = prepare_deepspeed(teacher_model, self.accelerator)
             else:
diff --git a/trl/experimental/gold/gold_trainer.py b/trl/experimental/gold/gold_trainer.py
index 9f36d610612..d1b8da0df76 100644
--- a/trl/experimental/gold/gold_trainer.py
+++ b/trl/experimental/gold/gold_trainer.py
@@ -845,7 +845,7 @@ def __init__(
         if args.disable_dropout:
             disable_dropout_in_model(self.model)
         if not args.use_uld_loss:
-            teacher_model.resize_token_embeddings(self.model.config.vocab_size)
+            teacher_model.resize_token_embeddings(self.model.config.get_text_config().vocab_size)
 
         if self.is_deepspeed_enabled:
             self.teacher_model = prepare_deepspeed(teacher_model, self.accelerator)