diff --git a/init2winit/dataset_lib/fastmri_dataset.py b/init2winit/dataset_lib/fastmri_dataset.py
index bed2f6ef..c76bfa8f 100644
--- a/init2winit/dataset_lib/fastmri_dataset.py
+++ b/init2winit/dataset_lib/fastmri_dataset.py
@@ -19,6 +19,7 @@
 import itertools
 import os
 
+from absl import logging
 import h5py
 from init2winit.dataset_lib import data_utils
 import jax
@@ -212,9 +213,11 @@ def load_split(per_host_batch_size, split, hps, shuffle_rng=None):
   # entirely to the end of it on the last host, because otherwise we will drop
   # the last `{train,valid}_size % split_size` elements.
   if jax.process_index() == jax.process_count() - 1:
-    if split == 'val':
+    if split in ['train', 'eval_train']:
+      end = hps.num_train_h5_files
+    elif split == 'val':
       end = hps.num_valid_h5_files
-    else:
+    else:  # split == 'test'
       end = hps.num_test_h5_files + hps.num_valid_h5_files
 
   data_dir = hps.data_dir
@@ -229,9 +232,58 @@ def load_split(per_host_batch_size, split, hps, shuffle_rng=None):
   else:  # split == 'val'
     data_dir = os.path.join(data_dir, hps.val_dir)
 
-  h5_paths = [
-      os.path.join(data_dir, path) for path in listdir(data_dir)
-  ][start:end]
+  try:
+    all_files = listdir(data_dir)
+  except tf.errors.NotFoundError as e:
+    raise FileNotFoundError(
+        f'FastMRI data directory not found: {data_dir}.'
+    ) from e
+
+  h5_paths = [os.path.join(data_dir, path) for path in all_files][start:end]
+
+  if not h5_paths:
+    raise ValueError(
+        f'No h5 files found for split={split} in {data_dir} '
+        f'(start={start}, end={end}, total files={len(all_files)}).'
+    )
+  logging.info(
+      'FastMRI %s split: loaded %d h5 paths from %s (files %d-%d of %d).',
+      split,
+      len(h5_paths),
+      data_dir,
+      start,
+      end,
+      len(all_files),
+  )
+
+  # Probe-read the first h5 file to catch ACL / connectivity errors early,
+  # before they are silently swallowed by tf.data.from_generator.
+  probe_path = h5_paths[0]
+  try:
+    with gfile.GFile(probe_path, 'rb') as gf:
+      with h5py.File(gf, 'r') as hf:
+        if 'kspace' not in hf:
+          raise ValueError(
+              f'FastMRI h5 file {probe_path} is missing the "kspace" dataset. '
+              'The file may be corrupt or incomplete.'
+          )
+        logging.info(
+            'FastMRI probe read of %s succeeded: %d slices.',
+            probe_path,
+            hf['kspace'].shape[0],
+        )
+  except PermissionError as e:
+    raise PermissionError(
+        f'Cannot read FastMRI h5 file {probe_path}: permission denied. '
+        'Check that the Borg job gfs_user has access to the FastMRI Placer '
+    ) from e
+  except (OSError, IOError) as e:
+    raise OSError(
+        f'Cannot read FastMRI h5 file {probe_path}: {e}. '
+        'This may indicate a timeout, network issue, or that the Placer '
+        'fileset is not replicated to a nearby cell. See b/329522685.'
+    ) from e
+
   ds = tf.data.Dataset.from_tensor_slices(h5_paths)
   ds = ds.interleave(
       _create_generator,