diff --git a/homr/download_utils.py b/homr/download_utils.py
index 8a12910..a40fea3 100644
--- a/homr/download_utils.py
+++ b/homr/download_utils.py
@@ -38,21 +38,34 @@ def download_file(url: str, filename: str) -> None:
         eprint()  # Add newline after download progress
 
 
-def unzip_file(filename: str, output_folder: str) -> None:
+def unzip_file(filename: str, output_folder: str, flatten_root_entry: bool = False) -> None:
     with zipfile.ZipFile(filename, "r") as zip_ref:
-        for member in zip_ref.namelist():
+        zip_contents = zip_ref.namelist()
+
+        if flatten_root_entry:
+            common_prefix = os.path.commonprefix(zip_contents)
+            if common_prefix and common_prefix.endswith("/"):
+                zip_contents_dict = {
+                    file: os.path.relpath(file, common_prefix) for file in zip_contents
+                }
+            else:
+                zip_contents_dict = {file: file for file in zip_contents}
+        else:
+            zip_contents_dict = {file: file for file in zip_contents}
+
+        for original, member in zip_contents_dict.items():
             # Ensure file path is safe
             if os.path.isabs(member) or ".." in member:
                 eprint(f"Skipping potentially unsafe file {member}")
                 continue
 
             # Handle directories
-            if member.endswith("/"):
+            if original.endswith("/"):
                 os.makedirs(os.path.join(output_folder, member), exist_ok=True)
                 continue
 
             # Extract file
-            source = zip_ref.open(member)
+            source = zip_ref.open(original)
             target = open(os.path.join(output_folder, member), "wb")
 
             with source, target:
diff --git a/training/convert_grandstaff.py b/training/convert_grandstaff.py
index db19374..d35f68d 100644
--- a/training/convert_grandstaff.py
+++ b/training/convert_grandstaff.py
@@ -1,7 +1,5 @@
 import multiprocessing
 import os
-import platform
-import stat
 import sys
 from pathlib import Path
 
@@ -13,7 +11,7 @@
 from torchvision import transforms as tr  # type: ignore
 from torchvision.transforms import Compose  # type: ignore
 
-from homr.download_utils import download_file, untar_file
+from homr.download_utils import download_file, untar_file, unzip_file
 from homr.simple_logging import eprint
 from homr.staff_dewarping import warp_image_randomly
 from homr.staff_parsing import add_image_into_tr_omr_canvas
@@ -27,24 +25,19 @@
 grandstaff_root = os.path.join(dataset_root, "grandstaff")
 grandstaff_train_index = os.path.join(grandstaff_root, "index.txt")
 
-hum2xml = os.path.join(dataset_root, "hum2xml")
-if platform.system() == "Windows":
-    eprint("Transformer training is only implemented for Linux")
-    eprint("Feel free to submit a PR to support Windows")
-    eprint("The main work should be to download hum2xml.exe and change the calls")
-    eprint("to use the exe-file instead of the linux binary.")
-    sys.exit(1)
-if not os.path.exists(hum2xml):
-    eprint("Downloading hum2xml from https://extras.humdrum.org/man/hum2xml/")
-    download_file("http://extras.humdrum.org/bin/linux/hum2xml", hum2xml)
-    os.chmod(hum2xml, stat.S_IXUSR)
-
 
 if not os.path.exists(grandstaff_root):
     eprint("Downloading grandstaff from https://sites.google.com/view/multiscore-project/datasets")
     grandstaff_archive = os.path.join(dataset_root, "grandstaff.tgz")
     download_file("https://grfia.dlsi.ua.es/musicdocs/grandstaff.tgz", grandstaff_archive)
     untar_file(grandstaff_archive, grandstaff_root)
+    eprint("Adding musicxml files to grandstaff dataset")
+    music_xml_download = os.path.join(dataset_root, "grandstaff_musicxml.zip")
+    download_file(
+        "https://github.com/liebharc/grandstaff_musicxml/archive/refs/heads/main.zip",
+        music_xml_download,
+    )
+    unzip_file(music_xml_download, grandstaff_root, flatten_root_entry=True)
 
 
 def _get_dark_pixels_per_row(image: NDArray) -> NDArray:
@@ -213,10 +206,6 @@ def _convert_file(  # noqa: PLR0911
         basename = str(path).replace(".krn", "")
         image_file = str(path).replace(".krn", ".jpg")
         musicxml = str(path).replace(".krn", ".musicxml")
-        result = os.system(f"{hum2xml} {path} > {musicxml}")  # noqa: S605
-        if result != 0:
-            eprint(f"Failed to convert {path}")
-            return []
         upper_semantic, lower_semantic = _music_xml_to_semantic(musicxml, basename)
         if upper_semantic is None or lower_semantic is None:
             eprint(f"Failed to convert {musicxml}")