diff --git a/bin/sample.py b/bin/sample.py
index 76b6173..80af3ff 100644
--- a/bin/sample.py
+++ b/bin/sample.py
@@ -136,7 +136,7 @@ def compute_training_tm_scores(
         samp_name = os.path.splitext(os.path.basename(fname))[0]
         tm_score = tmalign.max_tm_across_refs(
             fname,
-            train_dset.dset.filenames,
+            train_dset.filenames,
             n_threads=nthreads,
         )
         all_tm_scores[samp_name] = tm_score
diff --git a/bin/tmscore_training.py b/bin/tmscore_training.py
new file mode 100644
index 0000000..9f8ba53
--- /dev/null
+++ b/bin/tmscore_training.py
@@ -0,0 +1,53 @@
+"""
+Compute the maximum TM score against training set
+"""
+import logging
+import os, sys
+from glob import glob
+from pathlib import Path
+import argparse
+
+from sample import compute_training_tm_scores
+from datasets import CathCanonicalAnglesDataset
+
+SRC_DIR = (Path(os.path.dirname(os.path.abspath(__file__))) / "../protdiff").resolve()
+assert SRC_DIR.is_dir()
+sys.path.append(str(SRC_DIR))
+
+
+def build_parser() -> argparse.ArgumentParser:
+    parser = argparse.ArgumentParser(
+        usage=__doc__, formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+    parser.add_argument(
+        "-d",
+        "--dirname",
+        type=str,
+        default=os.path.join(os.getcwd(), "sampled_pdb"),
+        help="Directory of generated PDB structures",
+    )
+    return parser
+
+
+def main():
+    """Run the script"""
+    parser = build_parser()
+    args = parser.parse_args()
+
+    assert os.path.isdir(args.dirname)
+    generated_pdbs = glob(os.path.join(args.dirname, "*.pdb"))
+    assert generated_pdbs
+    logging.info(f"Found {len(generated_pdbs)} generated structures")
+
+    # we only need the filenames from the training dataset so it doesn't really matter
+    # what specific parameters we use to initialize it. The only important parameters are
+    # min_length, which is default to 40 and likely unchanged
+    train_dset = CathCanonicalAnglesDataset(split="train")
+
+    # Calculate scores
+    compute_training_tm_scores(generated_pdbs, train_dset, Path(args.dirname))
+
+
+if __name__ == "__main__":
+    logging.basicConfig(level=logging.INFO)
+    main()
diff --git a/protdiff/datasets.py b/protdiff/datasets.py
index daff167..4235fb6 100644
--- a/protdiff/datasets.py
+++ b/protdiff/datasets.py
@@ -694,6 +694,11 @@ class NoisedAnglesDataset(Dataset):
     def pad(self):
         """Pas through the pad property of wrapped dset"""
         return self.dset.pad
+    
+    @property
+    def filenames(self):
+        """Pass through the filenames property of the wrapped dset"""
+        return self.dset.filenames
 
     def sample_length(self, *args, **kwargs):
         return self.dset.sample_length(*args, **kwargs)