MMathisLab · May 15, 2026 · May 11, 2026 · May 11, 2026 · May 12, 2026 · May 13, 2026
diff --git a/deeplabcut/generate_training_dataset/trainingsetmanipulation.py b/deeplabcut/generate_training_dataset/trainingsetmanipulation.py
@@ -728,22 +728,54 @@ def to_matlab_cell(array):
         outer[0, 0] = array.astype("int64")
         return outer
 
+    # Keep only x/y columns if likelihood is present
+    if isinstance(df.columns, pd.MultiIndex):
+        coord_level = "coords" if "coords" in df.columns.names else df.columns.names[-1]
+        coord_values = df.columns.get_level_values(coord_level)
+
+        has_x = "x" in coord_values
+        has_y = "y" in coord_values
+        has_likelihood = "likelihood" in coord_values
+
+        if has_x and has_y:
+            if has_likelihood:
+                logging.info(
+                    "Detected likelihood columns in training data; dropping them before formatting training dataset."
+                )
+            df = df.loc[:, coord_values.isin(["x", "y"])]
+        else:
+            raise ValueError(
+                f"Training data must contain x/y coordinates. Found coordinate labels: {list(pd.unique(coord_values))}"
+            )
+
     for i in train_inds:
         data = dict()
         filename = df.index[i]
         data["image"] = filename
         img_shape = read_image_shape_fast(os.path.join(project_path, *filename))
         data["size"] = img_shape
-        temp = df.iloc[i].values.reshape(-1, 2)
+
+        row = df.iloc[i].values
+
+        if row.size % 2 != 0:
+            raise ValueError(
+                "Training data row does not contain an even number of coordinate values "
+                f"after dropping non-coordinate columns. Row size={row.size}, "
+                f"image={filename}"
+            )
+
+        temp = row.reshape(-1, 2)
         joints = np.c_[range(nbodyparts), temp]
         joints = joints[~np.isnan(joints).any(axis=1)].astype(int)
+
         # Check that points lie within the image
         inside = np.logical_and(
             np.logical_and(joints[:, 1] < img_shape[2], joints[:, 1] > 0),
             np.logical_and(joints[:, 2] < img_shape[1], joints[:, 2] > 0),
         )
         if not all(inside):
             joints = joints[inside]
+
         if joints.size:  # Exclude images without labels
             data["joints"] = joints
             train_data.append(data)
@@ -754,6 +786,7 @@ def to_matlab_cell(array):
                     to_matlab_cell(data["joints"]),
                 )
             )
+
     matlab_data = np.asarray(matlab_data, dtype=[("image", "O"), ("size", "O"), ("joints", "O")])
     return train_data, matlab_data
 

diff --git a/tests/test_trainingsetmanipulation.py b/tests/test_trainingsetmanipulation.py
@@ -108,3 +108,62 @@ def test_format_multianimal_training_data(monkeypatch):
 def test_parse_video_filenames(videos: list[str], expected_filenames: list[str]):
     filenames = parse_video_filenames(videos)
     assert filenames == expected_filenames
+
+
+def test_format_training_data_ignores_likelihood_columns(monkeypatch):
+    fake_shape = 3, 480, 640
+    monkeypatch.setattr(
+        trainingsetmanipulation,
+        "read_image_shape_fast",
+        lambda _: fake_shape,
+    )
+
+    # Base single-animal dataframe (x/y only)
+    df = pd.read_hdf(os.path.join(TEST_DATA_DIR, "trimouse_calib.h5")).xs(
+        "mus1",
+        level="individuals",
+        axis=1,
+    )
+    guarantee_multiindex_rows(df)
+
+    # Add a likelihood column so the layout becomes:
+    # x, y, likelihood, x, y, likelihood, ...
+    new_cols = []
+    new_arrays = []
+
+    coord_level = df.columns.names.index("coords")
+
+    for col in df.columns:
+        new_cols.append(col)
+        new_arrays.append(df[col].to_numpy())
+
+        if col[coord_level] == "y":
+            lik_col = list(col)
+            lik_col[coord_level] = "likelihood"
+            new_cols.append(tuple(lik_col))
+            new_arrays.append(np.ones(len(df), dtype=float))
+
+    df_with_likelihood = pd.DataFrame(
+        np.column_stack(new_arrays),
+        index=df.index,
+        columns=pd.MultiIndex.from_tuples(new_cols, names=df.columns.names),
+    )
+
+    train_inds = list(range(10))
+
+    baseline_train_data, baseline_matlab_data = format_training_data(df, train_inds, 12, "")
+    train_data, matlab_data = format_training_data(df_with_likelihood, train_inds, 12, "")
+
+    # The presence of likelihood columns should not change the formatted result
+    assert len(train_data) == len(baseline_train_data)
+    assert len(matlab_data) == len(baseline_matlab_data)
+
+    for got, expected in zip(train_data, baseline_train_data, strict=False):
+        assert got["image"] == expected["image"]
+        assert got["size"] == expected["size"]
+        assert np.array_equal(got["joints"], expected["joints"])
+
+    for got, expected in zip(matlab_data, baseline_matlab_data, strict=False):
+        assert np.array_equal(got["image"], expected["image"])
+        assert np.array_equal(got["size"], expected["size"])
+        assert np.array_equal(got["joints"][0, 0], expected["joints"][0, 0])