diff --git a/chirp/inference/classify/classify.py b/chirp/inference/classify/classify.py
index 86fc5aad..c7b917df 100644
--- a/chirp/inference/classify/classify.py
+++ b/chirp/inference/classify/classify.py
@@ -27,6 +27,7 @@
 import tqdm
 import pandas as pd
 import os
+from etils import epath
 
 
 @dataclasses.dataclass
@@ -182,49 +183,63 @@ def classify_batch(batch):
   )
   return inference_ds
 
+def flush_rows(
+  output_path: epath.Path, 
+  shard_num: int, 
+  rows: list[dict[str, str]],
+  format: str,
+  headers: list[str],
+):
+  """Helper method to write rows to disk."""
+  if format == 'csv':
+    if shard_num == 0:
+      with output_path.open('w') as f:
+        f.write(','.join(headers) + '\n')
+    with output_path.open('a') as f:
+      for row in rows:
+        csv_row = [
+            '{:.2f}'.format(row.get(h, '')) if isinstance(row.get(h, ''), np.float32) else row.get(h, '') 
+            for h in row
+        ]        
+        f.write(','.join(csv_row) + '\n')
+  elif format == 'parquet':
+    output_path.mkdir(parents=True, exist_ok=True)
+    parquet_path = output_path / f'part.{shard_num}.parquet'
+    pd.DataFrame(rows).to_parquet(parquet_path)
+  else:
+    raise ValueError('Output format must be either csv or parquet')
+
 
 def write_inference_file(
     embeddings_ds: tf.data.Dataset,
     model: interface.LogitsOutputHead,
     labels: Sequence[str],
-    output_filepath: str,
+    output_filepath: epath.PathLike,
     embedding_hop_size_s: float,
     threshold: dict[str, float] | None = None,
     exclude_classes: Sequence[str] = ('unknown',),
     include_classes: Sequence[str] = (),
-    row_size: int = 1_000_000,
-    format: str = 'parquet',
+    shard_size: int = 1_000_000,
 ):
   """Write inference results."""
+  output_filepath = epath.Path(output_filepath)
   
-  if format != 'parquet' and format != 'csv':
-    raise ValueError('Format must be either "parquet" or "csv"')
-  
-  if format == 'parquet':
-    if output_filepath.endswith('.csv'):
-      output_filepath = output_filepath[:-4]
-    if not output_filepath.endswith('.parquet'):
-      output_filepath += '.parquet'
-    os.mkdir(output_filepath)
-  if format == 'csv':
-    if output_filepath.endswith('.parquet'):
-      output_filepath = output_filepath[:-8]
-    if not output_filepath.endswith('.csv'):
-      output_filepath += '.csv'
+  if str(output_filepath).endswith('.csv'):
+    format = 'csv'
+  elif str(output_filepath).endswith('.parquet'):
+    format = 'parquet'
+  else:
+    raise ValueError('Output file must end with either .csv or .parquet')
     
-  parquet_count = 0
+  shard_num = 0
   rows = []
   
   inference_ds = get_inference_dataset(embeddings_ds, model)
 
   detection_count = 0
   nondetection_count = 0
-  if format == 'csv':
-    f = open(output_filepath, 'w')
   headers = ['filename', 'timestamp_s', 'label', 'logit']
   # Write column headers if CSV format
-  if format == 'csv':
-    f.write(','.join(headers) + '\n')
   for ex in tqdm.tqdm(inference_ds.as_numpy_iterator()):
     for t in range(ex['logits'].shape[0]):
       for i, label in enumerate(labels):
@@ -235,35 +250,21 @@ def write_inference_file(
         if threshold is None or ex['logits'][t, i] > threshold[label]:
           offset = ex['timestamp_s'] + t * embedding_hop_size_s
           logit = ex['logits'][t, i]
-          if format == 'parquet':
-            row = {
-              headers[0]: ex["filename"].decode("utf-8"),
-              headers[1]: offset,
-              headers[2]: label,
-              headers[3]: logit,
-            }
-            rows.append(row)
-            if len(rows) >= row_size:
-              tmp_df = pd.DataFrame(rows)
-              tmp_df.to_parquet(f'{output_filepath}/part.{parquet_count}.parquet')
-              parquet_count += 1
-              rows = []
-          elif format == 'csv':
-            row = [
-                ex['filename'].decode('utf-8'),
-                '{:.2f}'.format(offset),
-                label,
-                '{:.2f}'.format(logit),
-            ]
-            f.write(','.join(row) + '\n')
+          row = {
+            headers[0]: ex["filename"].decode("utf-8"),
+            headers[1]: np.float32(offset),
+            headers[2]: label,
+            headers[3]: np.float32(logit),
+          }
+          rows.append(row)
+          if len(rows) >= shard_size:
+            flush_rows(output_filepath, shard_num, rows, format, headers)
+            rows = []
+            shard_num += 1
           detection_count += 1
         else:
           nondetection_count += 1
-  # write remaining rows if parquet format
-  if format == 'parquet' and rows:
-    tmp_df = pd.DataFrame(rows)
-    tmp_df.to_parquet(f'{output_filepath}/part.{parquet_count}.parquet')
-  if format == 'csv':
-    f.close()
+  # write remaining rows
+  flush_rows(output_filepath, shard_num, rows, format, headers)
   print('\n\n\n   Detection count: ', detection_count)
   print('NonDetection count: ', nondetection_count)
\ No newline at end of file
diff --git a/chirp/inference/tests/classify_test.py b/chirp/inference/tests/classify_test.py
index d2d1105d..db5f5d1a 100644
--- a/chirp/inference/tests/classify_test.py
+++ b/chirp/inference/tests/classify_test.py
@@ -143,7 +143,7 @@ def test_write_inference_file(self):
     )
     
     # make a fake embeddings dataset
-    filenames = [f'file_{i}' for i in range(100)]
+    filenames = [f'file_{i}' for i in range(101)]
     
     self.write_random_embeddings(embedding_dim, filenames, tempdir)
     
@@ -158,8 +158,7 @@ def test_write_inference_file(self):
         labels=classes,
         output_filepath=parquet_path,
         embedding_hop_size_s=5.0,
-        row_size=10,
-        format='parquet',
+        shard_size=10,
     )
     
     classify.write_inference_file(
@@ -168,7 +167,7 @@ def test_write_inference_file(self):
       labels=classes,
       output_filepath=csv_path,
       embedding_hop_size_s=5.0,
-      format='csv',
+      shard_size=10,
     )
     
     parquet = pd.read_parquet(parquet_path)