diff --git a/ml4ht/data/data_loader.py b/ml4ht/data/data_loader.py
index d5f7bd1..c3bc339 100644
--- a/ml4ht/data/data_loader.py
+++ b/ml4ht/data/data_loader.py
@@ -282,7 +282,10 @@ def __getitem__(self, item: int) -> Batch:
         )
 
 
-def numpy_collate_fn(samples: List[Batch]) -> Batch:
+def numpy_collate_fn(
+    samples: List[Batch],
+    auto_float: bool = True,
+) -> Batch:
     """
     Merges a list of ml4ht batch formatted data.
     Can be used as 'collate_fn` in torch.utils.data.DataLoader
@@ -290,13 +293,21 @@ def numpy_collate_fn(samples: List[Batch]) -> Batch:
     """
     # construct correctly-shaped empty arrays for input and output of model
     in_batch_keys = list(samples[0][0])
+    if auto_float:
+        in_dtypes = {k: np.float32 for k in in_batch_keys}
+    else:
+        in_dtypes = {k: samples[0][0][k].dtype for k in in_batch_keys}
     in_batch = {
-        k: np.empty((len(samples),) + samples[0][0][k].shape, dtype=np.float32)
+        k: np.empty((len(samples),) + samples[0][0][k].shape, dtype=in_dtypes[k])
         for k in in_batch_keys
     }
     out_batch_keys = list(samples[0][1])
+    if auto_float:
+        out_dtypes = {k: np.float32 for k in out_batch_keys}
+    else:
+        out_dtypes = {k: samples[0][1][k].dtype for k in out_batch_keys}
     out_batch = {
-        k: np.empty((len(samples),) + samples[0][1][k].shape, dtype=np.float32)
+        k: np.empty((len(samples),) + samples[0][1][k].shape, dtype=out_dtypes[k])
         for k in out_batch_keys
     }
     # fill in the values of the input and output arrays
diff --git a/ml4ht/data/sample_getter.py b/ml4ht/data/sample_getter.py
index fb89aff..4c324b0 100644
--- a/ml4ht/data/sample_getter.py
+++ b/ml4ht/data/sample_getter.py
@@ -29,10 +29,12 @@ def __init__(
         input_data_descriptions: List[DataDescription],
         output_data_descriptions: List[DataDescription],
         option_picker: OptionPicker = None,
+        restricted_sample_id_idx=None,
     ):
         self.input_data_descriptions = input_data_descriptions
         self.output_data_descriptions = output_data_descriptions
         self.option_picker = option_picker or self._default_option_picker
+        self.restricted_sample_id_idx = restricted_sample_id_idx
 
     @staticmethod
     def _default_option_picker(
@@ -67,6 +69,8 @@ def __call__(self, sample_id: SampleID) -> Batch:
             sample_id,
             self.input_data_descriptions + self.output_data_descriptions,
         )
+        if self.restricted_sample_id_idx is not None:
+            sample_id = sample_id[self.restricted_sample_id_idx]
         tensors_in = self._half_batch(sample_id, loading_options, True)
         tensors_out = self._half_batch(sample_id, loading_options, False)
         return tensors_in, tensors_out