diff --git a/src/speculators/train/data.py b/src/speculators/train/data.py
index b1e05c91e..dd96d0d18 100644
--- a/src/speculators/train/data.py
+++ b/src/speculators/train/data.py
@@ -191,6 +191,7 @@ def __init__(
             hidden_states_dtype: The dtype of the hidden states.
         """
         self.data = load_from_disk(datapath)
+        self.start_file_idx = 0
         if split_ratio == 1.0:
             pass
         elif 1.0 > split_ratio > 0:
diff --git a/tests/unit/train/test_data.py b/tests/unit/train/test_data.py
index 4142568d1..c86bc49ca 100644
--- a/tests/unit/train/test_data.py
+++ b/tests/unit/train/test_data.py
@@ -4,9 +4,11 @@
 from pathlib import Path
 
 import torch
+from datasets import Dataset
 
 from speculators.models.eagle3.data import shift_batch
 from speculators.train.data import (
+    ArrowDataset,
     SampleFileDataset,
     create_collate_fn,
     standardize_data_v1,
@@ -434,3 +436,26 @@ def test_dataset_fallback_when_sample_lengths_json_malformed(tmp_path: Path):
     file_list = sorted([str(f) for f in tmp_path.glob("data_*.pt")])
     dataset = SampleFileDataset(max_len=50, file_list=file_list)
     assert len(dataset.approx_lengths) == 2
+
+
+def test_arrow_dataset_default_split_ratio_does_not_crash(tmp_path: Path):
+    """ArrowDataset with default split_ratio=1.0 should support indexing."""
+    ds = Dataset.from_dict(
+        {
+            "input_ids": [[1, 2, 3]],
+            "loss_mask": [[1, 1, 1]],
+            "seq_len": [3],
+        }
+    )
+    ds.save_to_disk(str(tmp_path / "data"))
+    (tmp_path / "data" / "hidden_states").mkdir()
+
+    arrow_ds = ArrowDataset(
+        max_len=128,
+        datapath=str(tmp_path / "data"),
+        on_missing="skip",
+    )
+
+    # Should not raise AttributeError
+    assert arrow_ds._map_to_file_idx(0) == 0
+    assert arrow_ds._map_to_file_idx(5) == 5