diff --git a/src/maxtext/configs/base.yml b/src/maxtext/configs/base.yml
index 55925ffd58..ef4d8343f7 100644
--- a/src/maxtext/configs/base.yml
+++ b/src/maxtext/configs/base.yml
@@ -984,7 +984,7 @@ decode_sampling_nucleus_p: -1 # set if you're doing nucleus / top-p
 decode_sampling_top_k: 0 # set if you're doing top-k
 decode_sampling_temperature: 1.
 
-eval_start_step: 0 # start eval after train step is >= eval_start_step
+eval_start_step: 0 # start eval when train step is >= eval_start_step
 eval_interval: -1  # the specific number of train step between eval_step
 eval_steps: -1  # run this number of steps for eval, recommend setting this to prevent error due to running out of evel data
 target_eval_loss: 0.  # early stop once reaching target eval_loss
diff --git a/src/maxtext/configs/types.py b/src/maxtext/configs/types.py
index bf5ac0115e..742cf67ed9 100644
--- a/src/maxtext/configs/types.py
+++ b/src/maxtext/configs/types.py
@@ -1587,7 +1587,8 @@ class TrainingLoop(BaseModel):
   log_period: int = Field(100, description="Frequency (in steps) to log metrics and flush Tensorboard.")
   eval_start_step: int = Field(
       0,
-      description="Start evaluation after training step is >= eval_start_step.",
+      ge=0,
+      description="Start evaluation when training step is >= eval_start_step.",
   )
   eval_interval: int = Field(
       -1,
diff --git a/src/maxtext/trainers/pre_train/train.py b/src/maxtext/trainers/pre_train/train.py
index ee92b93435..db21b9e336 100644
--- a/src/maxtext/trainers/pre_train/train.py
+++ b/src/maxtext/trainers/pre_train/train.py
@@ -718,7 +718,12 @@ def training_loop_iteration(
         all_host_upload=dump_hlo_upload_all,
     )
 
-  if eval_interval > 0 and step >= start_step and step >= eval_start_step and (step + 1) % eval_interval == 0:
+  if (
+      eval_interval > 0
+      and step >= start_step
+      and step >= eval_start_step
+      and (step - eval_start_step) % eval_interval == 0
+  ):
     assert eval_data_iterator
     # Explicitly reset the eval iterator and counters before starting the eval loop
     eval_data_iterator.reset()
diff --git a/tests/unit/pyconfig_test.py b/tests/unit/pyconfig_test.py
index d3afcb612e..9b7862bc5a 100644
--- a/tests/unit/pyconfig_test.py
+++ b/tests/unit/pyconfig_test.py
@@ -332,6 +332,15 @@ def test_eval_start_step_config(self):
     )
     self.assertEqual(config_override.eval_start_step, 50)
 
+  def test_eval_start_step_negative_raises_error(self):
+    """Verifies that eval_start_step < 0 raises a validation error."""
+    with self.assertRaises((ValueError, Exception)):
+      pyconfig.initialize(
+          [os.path.join(MAXTEXT_PKG_DIR, "train.py"), get_test_config_path()],
+          skip_jax_distributed_system=True,
+          eval_start_step=-1,
+      )
+
 
 if __name__ == "__main__":
   unittest.main()