diff --git a/tests/test_sgkit.py b/tests/test_sgkit.py
index fedd4423..e88f0f2f 100644
--- a/tests/test_sgkit.py
+++ b/tests/test_sgkit.py
@@ -25,6 +25,7 @@
 import sys
 import tempfile
 
+import msprime
 import numcodecs
 import numpy as np
 import pytest
@@ -37,6 +38,43 @@
 from tsinfer import formats
 
 
+def ts_to_dataset(ts, chunks=None, samples=None):
+    """
+    # From https://github.com/sgkit-dev/sgkit/blob/main/sgkit/tests/test_popgen.py#L63
+    Convert the specified tskit tree sequence into an sgkit dataset.
+    Note this just generates haploids for now - see the note above
+    in simulate_ts.
+    """
+    if samples is None:
+        samples = ts.samples()
+    tables = ts.dump_tables()
+    alleles = []
+    genotypes = []
+    max_alleles = 0
+    for var in ts.variants(samples=samples):
+        alleles.append(var.alleles)
+        max_alleles = max(max_alleles, len(var.alleles))
+        genotypes.append(var.genotypes.astype(np.int8))
+    padded_alleles = [
+        list(site_alleles) + [""] * (max_alleles - len(site_alleles))
+        for site_alleles in alleles
+    ]
+    alleles = np.array(padded_alleles).astype("S")
+    genotypes = np.expand_dims(genotypes, axis=2)
+
+    ds = sgkit.create_genotype_call_dataset(
+        variant_contig_names=["1"],
+        variant_contig=np.zeros(len(tables.sites), dtype=int),
+        variant_position=tables.sites.position.astype(int),
+        variant_allele=alleles,
+        sample_id=np.array([f"tsk_{u}" for u in samples]).astype("U"),
+        call_genotype=genotypes,
+    )
+    if chunks is not None:
+        ds = ds.chunk(dict(zip(["variants", "samples"], chunks)))
+    return ds
+
+
 @pytest.mark.skipif(sys.platform == "win32", reason="No cyvcf2 on windows")
 def test_sgkit_dataset_roundtrip(tmp_path):
     ts, zarr_path = tsutil.make_ts_and_zarr(tmp_path)
@@ -209,6 +247,20 @@ def test_sgkit_accessors_defaults(tmp_path):
     )
 
 
+def test_simulate_genotype_call_dataset(tmp_path):
+    # Test that byte alleles are correctly converted to string
+    ts = msprime.sim_ancestry(4, sequence_length=1000, random_seed=123)
+    ts = msprime.sim_mutations(ts, rate=2e-3, random_seed=123)
+    ds = ts_to_dataset(ts)
+    ds.update({"variant_ancestral_allele": ds["variant_allele"][:, 0]})
+    ds.to_zarr(tmp_path, mode="w")
+    sd = tsinfer.SgkitSampleData(tmp_path)
+    ts = tsinfer.infer(sd)
+    for v, ds_v, sd_v in zip(ts.variants(), ds.call_genotype, sd.sites_genotypes):
+        assert np.all(v.genotypes == ds_v.values.flatten())
+        assert np.all(v.genotypes == sd_v)
+
+
 @pytest.mark.skipif(sys.platform == "win32", reason="File permission errors on Windows")
 class TestSgkitMask:
     @pytest.mark.parametrize("sites", [[1, 2, 3, 5, 9, 27], [0], []])
diff --git a/tsinfer/formats.py b/tsinfer/formats.py
index 9ba12322..102c5080 100644
--- a/tsinfer/formats.py
+++ b/tsinfer/formats.py
@@ -2440,13 +2440,15 @@ def sites_position(self):
 
     @functools.cached_property
     def sites_alleles(self):
-        return self.data["variant_allele"][:][self.sites_select]
+        return self.data["variant_allele"][:][self.sites_select].astype(str)
 
     @functools.cached_property
     def sites_ancestral_allele(self):
         unknown_alleles = collections.Counter()
         try:
-            string_allele = self.data["variant_ancestral_allele"][:][self.sites_select]
+            string_allele = (
+                self.data["variant_ancestral_allele"][:][self.sites_select]
+            ).astype(str)
         except KeyError:
             raise ValueError(
                 "variant_ancestral_allele was not found in the dataset."