From e613fe57796a33e7d8c63951eb1fd6c58904f834 Mon Sep 17 00:00:00 2001
From: Max Mauermann <max-mauermann@web.de>
Date: Mon, 27 Jul 2026 13:36:58 +0200
Subject: [PATCH 1/4] random splits for training data operate with indices now
 instead of working with arrays

---
 birdnet_analyzer/model.py | 104 ++++++++++++++------------------------
 tests/train/test_model.py |  56 ++++++++++++++++++++
 2 files changed, 95 insertions(+), 65 deletions(-)
 create mode 100644 tests/train/test_model.py

diff --git a/birdnet_analyzer/model.py b/birdnet_analyzer/model.py
index 0f2bac7b3..76caa7960 100644
--- a/birdnet_analyzer/model.py
+++ b/birdnet_analyzer/model.py
@@ -151,11 +151,10 @@ def random_split(x, y, rng: Generator, val_ratio=0.2):
         A tuple of (x_train, y_train, x_val, y_val).
     """
     num_classes = y.shape[1]
-    x_train, y_train, x_val, y_val = [], [], [], []
+    train_indices, val_indices = [], []
 
     for i in range(num_classes):
         positive_indices = np.where(y[:, i] == 1)[0]
-        negative_indices = np.where(y[:, i] == -1)[0]
 
         num_samples = len(positive_indices)
         num_samples_train = max(1, int(num_samples * (1 - val_ratio)))
@@ -163,18 +162,18 @@ def random_split(x, y, rng: Generator, val_ratio=0.2):
 
         rng.shuffle(positive_indices)
 
-        train_indices = positive_indices[:num_samples_train]
-        val_indices = positive_indices[
+        class_train_indices = positive_indices[:num_samples_train]
+        class_val_indices = positive_indices[
             num_samples_train : num_samples_train + num_samples_val
         ]
 
-        x_train.append(x[train_indices])
-        y_train.append(y[train_indices])
-        x_val.append(x[val_indices])
-        y_val.append(y[val_indices])
+        train_indices.append(class_train_indices)
+        val_indices.append(class_val_indices)
 
-        x_train.append(x[negative_indices])
-        y_train.append(y[negative_indices])
+    # Negative samples are not class-specific in single-label training. Appending
+    # them in the loop above duplicates every negative sample once per class.
+    negative_indices = np.unique(np.where(y == -1)[0])
+    train_indices.append(negative_indices)
 
     non_event_indices = np.where(np.sum(y[:, :], axis=1) == 0)[0]
     num_samples = len(non_event_indices)
@@ -183,36 +182,21 @@ def random_split(x, y, rng: Generator, val_ratio=0.2):
 
     rng.shuffle(non_event_indices)
 
-    train_indices = non_event_indices[:num_samples_train]
-    val_indices = non_event_indices[
+    non_event_train_indices = non_event_indices[:num_samples_train]
+    non_event_val_indices = non_event_indices[
         num_samples_train : num_samples_train + num_samples_val
     ]
 
-    x_train.append(x[train_indices])
-    y_train.append(y[train_indices])
-    x_val.append(x[val_indices])
-    y_val.append(y[val_indices])
+    train_indices.append(non_event_train_indices)
+    val_indices.append(non_event_val_indices)
 
-    x_train = np.concatenate(x_train)
-    y_train = np.concatenate(y_train)
-    x_val = np.concatenate(x_val)
-    y_val = np.concatenate(y_val)
+    train_indices = np.concatenate(train_indices)
+    val_indices = np.concatenate(val_indices)
 
-    indices = np.arange(len(x_train))
+    rng.shuffle(train_indices)
+    rng.shuffle(val_indices)
 
-    rng.shuffle(indices)
-
-    x_train = x_train[indices]
-    y_train = y_train[indices]
-
-    indices = np.arange(len(x_val))
-
-    rng.shuffle(indices)
-
-    x_val = x_val[indices]
-    y_val = y_val[indices]
-
-    return x_train, y_train, x_val, y_val
+    return x[train_indices], y[train_indices], x[val_indices], y[val_indices]
 
 
 def random_multilabel_split(x, y, rng: Generator, val_ratio=0.2):
@@ -230,15 +214,16 @@ def random_multilabel_split(x, y, rng: Generator, val_ratio=0.2):
         A tuple of (x_train, y_train, x_val, y_val).
 
     """
-    class_combinations = np.unique(y, axis=0)
-    x_train, y_train, x_val, y_val = [], [], [], []
+    class_combinations, combination_ids = np.unique(
+        y, axis=0, return_inverse=True
+    )
+    train_indices, val_indices = [], []
 
-    for class_combination in class_combinations:
-        indices = np.where((y == class_combination).all(axis=1))[0]
+    for combination_id, class_combination in enumerate(class_combinations):
+        indices = np.flatnonzero(combination_ids == combination_id)
 
         if -1 in class_combination:
-            x_train.append(x[indices])
-            y_train.append(y[indices])
+            train_indices.append(indices)
         else:
             num_samples = len(indices)
             num_samples_train = max(1, int(num_samples * (1 - val_ratio)))
@@ -246,32 +231,20 @@ def random_multilabel_split(x, y, rng: Generator, val_ratio=0.2):
 
             rng.shuffle(indices)
 
-            train_indices = indices[:num_samples_train]
-            val_indices = indices[
+            combination_train_indices = indices[:num_samples_train]
+            combination_val_indices = indices[
                 num_samples_train : num_samples_train + num_samples_val
             ]
 
-            x_train.append(x[train_indices])
-            y_train.append(y[train_indices])
-            x_val.append(x[val_indices])
-            y_val.append(y[val_indices])
-
-    x_train = np.concatenate(x_train)
-    y_train = np.concatenate(y_train)
-    x_val = np.concatenate(x_val)
-    y_val = np.concatenate(y_val)
-
-    indices = np.arange(len(x_train))
-    rng.shuffle(indices)
-    x_train = x_train[indices]
-    y_train = y_train[indices]
+            train_indices.append(combination_train_indices)
+            val_indices.append(combination_val_indices)
 
-    indices = np.arange(len(x_val))
-    rng.shuffle(indices)
-    x_val = x_val[indices]
-    y_val = y_val[indices]
+    train_indices = np.concatenate(train_indices)
+    val_indices = np.concatenate(val_indices)
+    rng.shuffle(train_indices)
+    rng.shuffle(val_indices)
 
-    return x_train, y_train, x_val, y_val
+    return x[train_indices], y[train_indices], x[val_indices], y[val_indices]
 
 
 def upsample_core(
@@ -540,10 +513,6 @@ def on_epoch_end(self, epoch, logs=None):
                 self.on_epoch_end_fn(epoch, logs)
 
     rng = np.random.default_rng(RANDOM_SEED)
-    idx = np.arange(x_train.shape[0])
-    rng.shuffle(idx)
-    x_train = x_train[idx]
-    y_train = y_train[idx]
 
     if val_split > 0:
         if not is_multi_label:
@@ -554,6 +523,11 @@ def on_epoch_end(self, epoch, logs=None):
             x_train, y_train, x_val, y_val = random_multilabel_split(
                 x_train, y_train, rng, val_split
             )
+    else:
+        idx = np.arange(x_train.shape[0])
+        rng.shuffle(idx)
+        x_train = x_train[idx]
+        y_train = y_train[idx]
 
     if upsampling_ratio > 0:
         x_train, y_train = upsampling(
diff --git a/tests/train/test_model.py b/tests/train/test_model.py
new file mode 100644
index 000000000..4a35ead87
--- /dev/null
+++ b/tests/train/test_model.py
@@ -0,0 +1,56 @@
+import numpy as np
+
+from birdnet_analyzer.model import random_multilabel_split, random_split
+
+
+def test_random_split_adds_negative_samples_only_once():
+    x = np.arange(10).reshape(-1, 1)
+    y = np.array(
+        [
+            [1, 0, 0],
+            [1, 0, 0],
+            [0, 1, 0],
+            [0, 1, 0],
+            [0, 0, 1],
+            [0, 0, 1],
+            [-1, -1, -1],
+            [-1, -1, -1],
+            [-1, -1, -1],
+            [-1, -1, -1],
+        ]
+    )
+
+    x_train, _, x_val, _ = random_split(
+        x, y, np.random.default_rng(42), val_ratio=0.5
+    )
+
+    assert len(x_train) + len(x_val) == len(x)
+    assert len(np.unique(x_train)) == len(x_train)
+    assert not np.intersect1d(x_train, x_val).size
+
+
+def test_random_multilabel_split_keeps_negative_combinations_in_training():
+    x = np.arange(10).reshape(-1, 1)
+    y = np.array(
+        [
+            [1, 0, 0],
+            [1, 0, 0],
+            [0, 1, 0],
+            [0, 1, 0],
+            [1, 1, 0],
+            [1, 1, 0],
+            [0, -1, 0],
+            [0, -1, 0],
+            [0, 0, 0],
+            [0, 0, 0],
+        ]
+    )
+
+    x_train, _, x_val, y_val = random_multilabel_split(
+        x, y, np.random.default_rng(42), val_ratio=0.5
+    )
+
+    assert len(x_train) + len(x_val) == len(x)
+    assert not np.intersect1d(x_train, x_val).size
+    assert {6, 7}.issubset(set(x_train[:, 0]))
+    assert not np.any(y_val == -1)

From 4b01d3ba11a52652805f20bc175ead667d94e3eb Mon Sep 17 00:00:00 2001
From: Max Mauermann <max-mauermann@web.de>
Date: Mon, 27 Jul 2026 17:07:39 +0200
Subject: [PATCH 2/4] optimizations for upsampling

---
 birdnet_analyzer/model.py | 122 +++++++++++++++++++++-----------------
 birdnet_analyzer/utils.py |   2 +
 tests/train/test_model.py |  47 ++++++++++++++-
 3 files changed, 114 insertions(+), 57 deletions(-)

diff --git a/birdnet_analyzer/model.py b/birdnet_analyzer/model.py
index 76caa7960..d73413f5a 100644
--- a/birdnet_analyzer/model.py
+++ b/birdnet_analyzer/model.py
@@ -276,33 +276,33 @@ def upsample_core(
     x_temp = []
 
     if is_binary:
-        minority_label = 1 if y.sum(axis=0) < len(y) - y.sum(axis=0) else 0
+        positive_count = y.sum(axis=0)
+        minority_label = 1 if positive_count < len(y) - positive_count else 0
+        source_indices = np.flatnonzero(y == minority_label)
+        missing_samples = min_samples - len(source_indices)
 
-        while np.where(y == minority_label)[0].shape[0] + len(y_temp) < min_samples:
-            random_index = rng.choice(np.where(y == minority_label)[0], size=size)
-            x_app, y_app = apply(x, y, random_index)
+        for _ in range(max(0, missing_samples)):
+            random_index = rng.choice(source_indices, size=size)
+            x_app, y_app = apply(x, y, random_index, source_indices)
 
             y_temp.append(y_app)
             x_temp.append(x_app)
     else:
         for i in range(y.shape[1]):
-            class_x_temp = []
-            class_y_temp = []
+            source_indices = np.flatnonzero(y[:, i] == 1)
+            missing_samples = min_samples - len(source_indices)
 
-            while y[:, i].sum() + len(class_y_temp) < min_samples:
-                try:
-                    random_index = rng.choice(np.where(y[:, i] == 1)[0], size=size)
-                except ValueError as e:
-                    raise get_empty_class_exception()(index=i) from e
+            if missing_samples <= 0:
+                continue
 
-                # Apply
-                x_app, y_app = apply(x, y, random_index)
-                class_y_temp.append(y_app)
-                class_x_temp.append(x_app)
+            if not len(source_indices):
+                raise get_empty_class_exception()(index=i)
 
-            if len(class_y_temp) > 0:
-                x_temp.extend(class_x_temp)
-                y_temp.extend(class_y_temp)
+            for _ in range(missing_samples):
+                random_index = rng.choice(source_indices, size=size)
+                x_app, y_app = apply(x, y, random_index, source_indices)
+                y_temp.append(y_app)
+                x_temp.append(x_app)
 
     return x_temp, y_temp
 
@@ -339,43 +339,60 @@ def upsampling(
     x_temp = []
     y_temp = []
 
-    if mode == "repeat":
-
-        def applyRepeat(x, y, random_index):
-            return x[random_index[0]], y[random_index[0]]
-
-        x_temp, y_temp = upsample_core(
-            x, y, min_samples, rng, applyRepeat, is_binary, size=1
-        )
-
-    elif mode == "mean":
-
-        def applyMean(x, y, random_indices):
-            mean = np.mean(x[random_indices], axis=0)
+    if mode in {"repeat", "mean", "linear"}:
+        if is_binary:
+            positive_count = y.sum(axis=0)
+            minority_label = 1 if positive_count < len(y) - positive_count else 0
+            source_groups = [(None, np.flatnonzero(y == minority_label))]
+        else:
+            source_groups = [
+                (class_index, np.flatnonzero(y[:, class_index] == 1))
+                for class_index in range(y.shape[1])
+            ]
 
-            return mean, y[random_indices[0]]
+        sample_size = 1 if mode == "repeat" else 2
+        for class_index, source_indices in source_groups:
+            missing_samples = min_samples - len(source_indices)
+            if missing_samples <= 0:
+                continue
 
-        x_temp, y_temp = upsample_core(x, y, min_samples, rng, applyMean, is_binary)
-    elif mode == "linear":
+            if not len(source_indices):
+                if class_index is None:
+                    raise ValueError("The minority class is empty.")
+                raise get_empty_class_exception()(index=class_index)
 
-        def applyLinearCombination(x, y, random_indices):
-            alpha = rng.uniform(0, 1)
-            new_sample = (
-                alpha * x[random_indices[0]] + (1 - alpha) * x[random_indices[1]]
+            sampled_indices = rng.choice(
+                source_indices, size=(missing_samples, sample_size)
             )
+            x_sources = x[sampled_indices]
 
-            return new_sample, y[random_indices[0]]
+            if mode == "repeat":
+                x_temp.append(x_sources[:, 0])
+            elif mode == "mean":
+                x_temp.append(np.mean(x_sources, axis=1))
+            else:
+                alpha = rng.uniform(0, 1, size=(missing_samples, 1))
+                x_temp.append(
+                    alpha * x_sources[:, 0] + (1 - alpha) * x_sources[:, 1]
+                )
 
-        x_temp, y_temp = upsample_core(
-            x, y, min_samples, rng, applyLinearCombination, is_binary
-        )
+            y_temp.append(y[sampled_indices[:, 0]])
 
     elif mode == "smote":
 
-        def applySmote(x, y, random_index, k=5):
-            distances = np.sqrt(np.sum((x - x[random_index[0]]) ** 2, axis=1))
-            indices = np.argsort(distances)[1 : k + 1]
-            random_neighbor = rng.choice(indices)
+        def applySmote(x, y, random_index, source_indices, k=5):
+            source_index = random_index[0]
+            neighbor_indices = source_indices[source_indices != source_index]
+            if not len(neighbor_indices):
+                return x[source_index], y[source_index]
+
+            differences = x[neighbor_indices] - x[source_index]
+            distances = np.einsum("ij,ij->i", differences, differences)
+            nearest_count = min(k, len(neighbor_indices))
+            nearest_indices = neighbor_indices[
+                np.argpartition(distances, nearest_count - 1)[:nearest_count]
+            ]
+            random_neighbor = rng.choice(nearest_indices)
             diff = x[random_neighbor] - x[random_index[0]]
             weight = rng.uniform(0, 1)
             new_sample = x[random_index[0]] + weight * diff
@@ -387,16 +404,8 @@ def applySmote(x, y, random_index, k=5):
         )
 
     if len(x_temp) > 0:
-        x = np.vstack((x, np.array(x_temp)))
-        y = np.vstack((y, np.array(y_temp)))
-
-    indices = np.arange(len(x))
-    rng.shuffle(indices)
-    x = x[indices]
-    y = y[indices]
-
-    del x_temp
-    del y_temp
+        x = np.vstack((x, *x_temp))
+        y = np.vstack((y, *y_temp))
 
     return x, y
 
@@ -608,6 +617,7 @@ def _focal_loss(y_true, y_pred):
         batch_size=batch_size,
         validation_data=(x_val, y_val),
         callbacks=callbacks,
+        shuffle=True,
     )
 
     os.environ["CUDA_VISIBLE_DEVICES"] = setting_cache
diff --git a/birdnet_analyzer/utils.py b/birdnet_analyzer/utils.py
index afc5e31a7..5fda5b225 100644
--- a/birdnet_analyzer/utils.py
+++ b/birdnet_analyzer/utils.py
@@ -347,6 +347,8 @@ def save_params_file(file_path, params: dict):
     """
     import csv
 
+    os.makedirs(os.path.dirname(file_path), exist_ok=True)
+
     with open(file_path, "w", newline="", encoding="utf-8-sig") as paramsfile:
         paramswriter = csv.writer(paramsfile)
         paramswriter.writerow(("Parameter", "Value"))
diff --git a/tests/train/test_model.py b/tests/train/test_model.py
index 4a35ead87..8bb940eb9 100644
--- a/tests/train/test_model.py
+++ b/tests/train/test_model.py
@@ -1,6 +1,12 @@
 import numpy as np
+import pytest
 
-from birdnet_analyzer.model import random_multilabel_split, random_split
+from birdnet_analyzer.model import (
+    get_empty_class_exception,
+    random_multilabel_split,
+    random_split,
+    upsampling,
+)
 
 
 def test_random_split_adds_negative_samples_only_once():
@@ -54,3 +60,42 @@ def test_random_multilabel_split_keeps_negative_combinations_in_training():
     assert not np.intersect1d(x_train, x_val).size
     assert {6, 7}.issubset(set(x_train[:, 0]))
     assert not np.any(y_val == -1)
+
+
+@pytest.mark.parametrize("mode", ["repeat", "mean", "linear"])
+def test_upsampling_balances_each_class_to_requested_ratio(mode):
+    x = np.arange(12, dtype="float32").reshape(6, 2)
+    y = np.array([[1, 0], [1, 0], [1, 0], [1, 0], [1, 0], [0, 1]])
+
+    x_upsampled, y_upsampled = upsampling(
+        x, y, np.random.default_rng(42), is_binary=False, ratio=1.0, mode=mode
+    )
+
+    assert len(x_upsampled) == 10
+    assert np.array_equal(y_upsampled.sum(axis=0), [5, 5])
+
+
+@pytest.mark.parametrize("mode", ["repeat", "mean", "linear"])
+def test_upsampling_rejects_empty_classes(mode):
+    x = np.arange(8, dtype="float32").reshape(4, 2)
+    y = np.array([[1, 0], [1, 0], [1, 0], [1, 0]])
+
+    with pytest.raises(get_empty_class_exception()) as error:
+        upsampling(
+            x, y, np.random.default_rng(42), is_binary=False, ratio=1.0, mode=mode
+        )
+
+    assert error.value.index == 1
+
+
+def test_smote_only_interpolates_between_samples_of_the_same_class():
+    x = np.array([[0.0], [1.0], [100.0], [101.0], [102.0]])
+    y = np.array([[1, 0], [1, 0], [0, 1], [0, 1], [0, 1]])
+
+    x_upsampled, y_upsampled = upsampling(
+        x, y, np.random.default_rng(42), is_binary=False, ratio=1.0, mode="smote"
+    )
+
+    assert len(x_upsampled) == 6
+    assert np.array_equal(y_upsampled.sum(axis=0), [3, 3])
+    assert 0 <= x_upsampled[-1, 0] <= 1

From 7753dfa59c3e41a0f901ac859beb7d718cb5f838 Mon Sep 17 00:00:00 2001
From: Max Mauermann <max-mauermann@web.de>
Date: Tue, 28 Jul 2026 12:32:56 +0200
Subject: [PATCH 3/4] Check for empty input folders before data-loading starts

---
 birdnet_analyzer/train/utils.py   | 25 +++++++++++++++++++++++++
 tests/train/test_train_loading.py | 19 ++++++++++++++++++-
 2 files changed, 43 insertions(+), 1 deletion(-)

diff --git a/birdnet_analyzer/train/utils.py b/birdnet_analyzer/train/utils.py
index 0201aa7ff..2c547ac61 100644
--- a/birdnet_analyzer/train/utils.py
+++ b/birdnet_analyzer/train/utils.py
@@ -158,6 +158,29 @@ def _read_and_crop_file(
     return sig_splits, labels
 
 
+def _check_input_folders(audio_input: str, train_folders: list[str]):
+    """Reject training folders without supported audio files before model setup."""
+    empty_folders = []
+
+    for folder in train_folders:
+        folder_path = os.path.join(audio_input, folder)
+        has_audio_file = any(
+            entry.is_file()
+            and not entry.name.startswith(".")
+            and entry.name.rsplit(".", 1)[-1].lower() in ALLOWED_FILETYPES
+            for entry in os.scandir(folder_path)
+        )
+
+        if not has_audio_file:
+            empty_folders.append(folder)
+
+    if empty_folders:
+        raise ValueError(
+            "The following training data folders do not contain any supported audio "
+            f"files: {', '.join(empty_folders)}"
+        )
+
+
 def _load_training_data(
     audio_input: str,
     test_data: str | None = None,
@@ -248,6 +271,8 @@ def _load_training_data(
             "validation-only-repeat-upsampling-for-multi-label",
         )
 
+    _check_input_folders(audio_input, train_folders)
+
     x_train, y_train, x_test, y_test = [], [], [], []
     model = load("acoustic", "2.4", "tf")
     model_sr = int(model.get_sample_rate())
diff --git a/tests/train/test_train_loading.py b/tests/train/test_train_loading.py
index 865e2c83e..5c6b1d9bb 100644
--- a/tests/train/test_train_loading.py
+++ b/tests/train/test_train_loading.py
@@ -14,13 +14,14 @@
 
 import os
 import tempfile
+from unittest.mock import patch
 
 import numpy as np
 import pytest
 import soundfile as sf
 
 from birdnet_analyzer import model_utils
-from birdnet_analyzer.train.utils import _read_and_crop_file
+from birdnet_analyzer.train.utils import _load_training_data, _read_and_crop_file
 
 SR = 48000
 SIG_LENGTH = 3.0
@@ -81,6 +82,22 @@ def test_read_and_crop_bad_file_returns_empty():
     assert labels == []
 
 
+def test_load_training_data_lists_empty_folders_before_loading_model(tmp_path):
+    (tmp_path / "blackbird").mkdir()
+    (tmp_path / "bluebird").mkdir()
+    robin_dir = tmp_path / "robin"
+    robin_dir.mkdir()
+    (robin_dir / "recording.wav").touch()
+
+    with patch("birdnet_analyzer.train.utils.load") as mock_load, pytest.raises(
+        ValueError, match=r"blackbird.*bluebird"
+    ) as error:
+        _load_training_data(str(tmp_path))
+
+    assert "robin" not in str(error.value)
+    mock_load.assert_not_called()
+
+
 @pytest.fixture
 def short_wav():
     """A 0.4 s wav -> shorter than the 1 s min_len, so split_signal yields nothing."""

From e84b20388733a4e231b96af89dbc118757ddf376 Mon Sep 17 00:00:00 2001
From: Josef Haupt <josef.haupt@informatik.tu-chemnitz.de>
Date: Mon, 3 Aug 2026 18:22:55 +0200
Subject: [PATCH 4/4] updated error msgs

---
 birdnet_analyzer/gui/train.py     | 9 ++++++++-
 birdnet_analyzer/lang/de.json     | 6 +++++-
 birdnet_analyzer/lang/en.json     | 6 +++++-
 birdnet_analyzer/lang/fi.json     | 6 +++++-
 birdnet_analyzer/lang/fr.json     | 6 +++++-
 birdnet_analyzer/lang/id.json     | 6 +++++-
 birdnet_analyzer/lang/pt-br.json  | 6 +++++-
 birdnet_analyzer/lang/ru.json     | 6 +++++-
 birdnet_analyzer/lang/se.json     | 6 +++++-
 birdnet_analyzer/lang/tlh.json    | 6 +++++-
 birdnet_analyzer/lang/zh_TW.json  | 6 +++++-
 birdnet_analyzer/model.py         | 6 +++---
 birdnet_analyzer/train/utils.py   | 7 +++++--
 tests/train/test_train_loading.py | 1 +
 14 files changed, 67 insertions(+), 16 deletions(-)

diff --git a/birdnet_analyzer/gui/train.py b/birdnet_analyzer/gui/train.py
index a89b14f62..0395f436e 100644
--- a/birdnet_analyzer/gui/train.py
+++ b/birdnet_analyzer/gui/train.py
@@ -226,7 +226,14 @@ def trial_progression(trial):
         )
     except Exception as e:
         if e.args and len(e.args) > 1:
-            raise gr.Error(loc.localize(e.args[1])) from e
+            message = loc.localize(e.args[1])
+
+            # Args beyond the localization key are payload for the message's
+            # placeholders (e.g. the list of offending folders).
+            if len(e.args) > 2:
+                message = message.format(*e.args[2:])
+
+            raise gr.Error(message) from e
 
         raise gr.Error(f"{e}") from e
 
diff --git a/birdnet_analyzer/lang/de.json b/birdnet_analyzer/lang/de.json
index 5ccaf6b6c..69eb84301 100644
--- a/birdnet_analyzer/lang/de.json
+++ b/birdnet_analyzer/lang/de.json
@@ -432,16 +432,20 @@
     "training-tab-warning-detached-classifier-append": "Der \"Anhängen\"-Modus wird für den abgekoppelten Klassifikator nicht angewendet.",
     "validation-max-confidence-lower-than-min-confidence": "Maximaler Konfidenz-Schwellwert muss größer sein als der minimale Schwellwert.",
     "validation-no-audio-directory-selected": "Kein Audioverzeichnis ausgewählt",
+    "validation-no-audio-files-in-training-folders": "Die folgenden Trainingsdaten-Ordner enthalten keine unterstützten Audiodateien: {0}",
     "validation-no-cache-file-selected": "Bitte wählen Sie eine Cache-Datei aus.",
     "validation-no-custom-classifier-selected": "Kein benutzerdefinierter Klassifikator ausgewählt.",
     "validation-no-directory-for-classifier-selected": "Bitte wählen Sie ein Verzeichnis für den Klassifikator.",
     "validation-no-directory-selected": "Bitte wählen Sie ein Verzeichnis.",
     "validation-no-file-selected": "Bitte wählen Sie eine Datei aus.",
+    "validation-no-negative-samples-in-binary-classification": "Negative Labels können bei binärer Klassifikation nicht verwendet werden",
     "validation-no-species-list-selected": "Bitte wählen Sie eine Artenliste.",
     "validation-no-training-data-selected": "Bitte wählen Sie Ihre Trainingsdaten.",
     "validation-no-valid-batch-size": "Bitte geben Sie eine gültige Batch-Größe an.",
     "validation-no-valid-classifier-name": "Bitte geben Sie einen gültigen Namen für den Klassifikator an.",
     "validation-no-valid-epoch-number": "Bitte geben Sie eine gültige Anzahl von Epochen an.",
     "validation-no-valid-frequency": "Bitte geben Sie eine gültige Frequenz an",
-    "validation-no-valid-learning-rate": "Bitte geben Sie eine gültige Lernrate an."
+    "validation-no-valid-learning-rate": "Bitte geben Sie eine gültige Lernrate an.",
+    "validation-non-event-samples-required-in-binary-classification": "Für die binäre Klassifikation sind Non-Event-Beispiele erforderlich",
+    "validation-only-repeat-upsampling-for-multi-label": "Für Multi-Label-Training ist nur Wiederholen-Upsampling verfügbar"
 }
\ No newline at end of file
diff --git a/birdnet_analyzer/lang/en.json b/birdnet_analyzer/lang/en.json
index 3f5bb22fb..7a000919a 100644
--- a/birdnet_analyzer/lang/en.json
+++ b/birdnet_analyzer/lang/en.json
@@ -432,16 +432,20 @@
     "training-tab-warning-detached-classifier-append": "Append mode will not apply to the detached classifier.",
     "validation-max-confidence-lower-than-min-confidence": "Maximum confidence must be greater than minimum confidence",
     "validation-no-audio-directory-selected": "No audio directory selected",
+    "validation-no-audio-files-in-training-folders": "The following training data folders do not contain any supported audio files: {0}",
     "validation-no-cache-file-selected": "Please select a cache file.",
     "validation-no-custom-classifier-selected": "No custom classifier selected.",
     "validation-no-directory-for-classifier-selected": "Please select a directory for the classifier.",
     "validation-no-directory-selected": "Please select a directory.",
     "validation-no-file-selected": "Please select a file.",
+    "validation-no-negative-samples-in-binary-classification": "Negative labels can't be used with binary classification",
     "validation-no-species-list-selected": "Please select a species list.",
     "validation-no-training-data-selected": "Please select your training data.",
     "validation-no-valid-batch-size": "Please enter a valid batch size.",
     "validation-no-valid-classifier-name": "Please enter a valid name for the classifier.",
     "validation-no-valid-epoch-number": "Please enter a valid number of epochs.",
     "validation-no-valid-frequency": "Please enter a valid frequency in",
-    "validation-no-valid-learning-rate": "Please enter a valid learning rate."
+    "validation-no-valid-learning-rate": "Please enter a valid learning rate.",
+    "validation-non-event-samples-required-in-binary-classification": "Non-event samples are required for binary classification",
+    "validation-only-repeat-upsampling-for-multi-label": "Only repeat-upsampling is available for multi-label training"
 }
\ No newline at end of file
diff --git a/birdnet_analyzer/lang/fi.json b/birdnet_analyzer/lang/fi.json
index 5bcb1599d..d773eb19d 100644
--- a/birdnet_analyzer/lang/fi.json
+++ b/birdnet_analyzer/lang/fi.json
@@ -432,16 +432,20 @@
     "training-tab-warning-detached-classifier-append": "Liitä-tila ei koske irrotettua luokitinta.",
     "validation-max-confidence-lower-than-min-confidence": "Enimmäisluotettavuuden on oltava suurempi kuin vähimmäisluotettavuus",
     "validation-no-audio-directory-selected": "Äänihakemistoa ei ole valittu",
+    "validation-no-audio-files-in-training-folders": "Seuraavat koulutusdatakansiot eivät sisällä tuettuja äänitiedostoja: {0}",
     "validation-no-cache-file-selected": "Valitse välimuistitiedosto.",
     "validation-no-custom-classifier-selected": "Mukautettua luokittelijaa ei ole valittu.",
     "validation-no-directory-for-classifier-selected": "Valitse hakemisto luokittelijalle.",
     "validation-no-directory-selected": "Valitse hakemisto.",
     "validation-no-file-selected": "Valitse tiedosto.",
+    "validation-no-negative-samples-in-binary-classification": "Negatiivisia luokkia ei voi käyttää binääriluokittelussa",
     "validation-no-species-list-selected": "Valitse lajilista.",
     "validation-no-training-data-selected": "Valitse koulutusdatasi.",
     "validation-no-valid-batch-size": "Anna kelvollinen eräkoko.",
     "validation-no-valid-classifier-name": "Anna kelvollinen nimi luokittelijalle.",
     "validation-no-valid-epoch-number": "Anna kelvollinen aikakausien määrä.",
     "validation-no-valid-frequency": "Anna kelvollinen taajuus",
-    "validation-no-valid-learning-rate": "Anna kelvollinen oppimistahti."
+    "validation-no-valid-learning-rate": "Anna kelvollinen oppimistahti.",
+    "validation-non-event-samples-required-in-binary-classification": "Binääriluokittelu vaatii ei-tapahtumanäytteitä",
+    "validation-only-repeat-upsampling-for-multi-label": "Monileimakoulutuksessa on käytettävissä vain toista-ylösnäytteistys"
 }
\ No newline at end of file
diff --git a/birdnet_analyzer/lang/fr.json b/birdnet_analyzer/lang/fr.json
index 45e4d2b9b..2dbf5cd0b 100644
--- a/birdnet_analyzer/lang/fr.json
+++ b/birdnet_analyzer/lang/fr.json
@@ -432,16 +432,20 @@
     "training-tab-warning-detached-classifier-append": "Le mode d'ajout ne s'appliquera pas au classificateur détaché.",
     "validation-max-confidence-lower-than-min-confidence": "La confiance maximale doit être supérieure à la confiance minimale",
     "validation-no-audio-directory-selected": "Aucun dossiers audio sélectionnés",
+    "validation-no-audio-files-in-training-folders": "Les dossiers de données d'entraînement suivants ne contiennent aucun fichier audio pris en charge : {0}",
     "validation-no-cache-file-selected": "Veuillez sélectionner un fichier cache.",
     "validation-no-custom-classifier-selected": "Aucun classificateur personnalisé n'a été sélectionné.",
     "validation-no-directory-for-classifier-selected": "Veuillez sélectionner un dossier pour le classificateur.",
     "validation-no-directory-selected": "Veuillez sélectionner un dossier",
     "validation-no-file-selected": "Sélectionnez un fichier",
+    "validation-no-negative-samples-in-binary-classification": "Les étiquettes négatives ne peuvent pas être utilisées avec la classification binaire",
     "validation-no-species-list-selected": "Veuillez sélectionner une liste d’espèces",
     "validation-no-training-data-selected": "Veuillez sélectionner des données d’entraînement",
     "validation-no-valid-batch-size": "Veuillez saisir une taille de lot valide.",
     "validation-no-valid-classifier-name": "Veuillez saisir un nom valide pour le classificateur.",
     "validation-no-valid-epoch-number": "Veuillez saisir un nombre valide d'époques.",
     "validation-no-valid-frequency": "Veuillez saisir une fréquence valide dans",
-    "validation-no-valid-learning-rate": "Veuillez saisir un taux d'apprentissage valide."
+    "validation-no-valid-learning-rate": "Veuillez saisir un taux d'apprentissage valide.",
+    "validation-non-event-samples-required-in-binary-classification": "Des échantillons non-événement sont requis pour la classification binaire",
+    "validation-only-repeat-upsampling-for-multi-label": "Seul le suréchantillonnage par répétition est disponible pour l'entraînement multi-étiquettes"
 }
\ No newline at end of file
diff --git a/birdnet_analyzer/lang/id.json b/birdnet_analyzer/lang/id.json
index e95fd24ef..91e640214 100644
--- a/birdnet_analyzer/lang/id.json
+++ b/birdnet_analyzer/lang/id.json
@@ -432,16 +432,20 @@
     "training-tab-warning-detached-classifier-append": "Mode tambah tidak akan berlaku untuk pengklasifikasi terlepas.",
     "validation-max-confidence-lower-than-min-confidence": "Kepercayaan maksimum harus lebih besar dari kepercayaan minimum",
     "validation-no-audio-directory-selected": "Tidak ada audio direktori yang dipilih",
+    "validation-no-audio-files-in-training-folders": "Folder data pelatihan berikut tidak berisi berkas audio yang didukung: {0}",
     "validation-no-cache-file-selected": "Silakan pilih berkas cache.",
     "validation-no-custom-classifier-selected": "Tidak ada klasifikator kustom yang dipilih.",
     "validation-no-directory-for-classifier-selected": "Silakan pilih direktori untuk klasifikator.",
     "validation-no-directory-selected": "Silakan pilih direktori.",
     "validation-no-file-selected": "Silakan pilih file.",
+    "validation-no-negative-samples-in-binary-classification": "Label negatif tidak dapat digunakan dengan klasifikasi biner",
     "validation-no-species-list-selected": "Silakan pilih daftar spesies.",
     "validation-no-training-data-selected": "Silakan pilih data pelatihan Anda.",
     "validation-no-valid-batch-size": "Silakan masukkan ukuran batch yang valid.",
     "validation-no-valid-classifier-name": "Silakan masukkan nama yang valid untuk klasifikator.",
     "validation-no-valid-epoch-number": "Silakan masukkan jumlah epochs yang valid.",
     "validation-no-valid-frequency": "Silakan masukkan frekuensi yang valid dalam",
-    "validation-no-valid-learning-rate": "Silakan masukkan laju pembelajaran yang valid."
+    "validation-no-valid-learning-rate": "Silakan masukkan laju pembelajaran yang valid.",
+    "validation-non-event-samples-required-in-binary-classification": "Sampel non-event diperlukan untuk klasifikasi biner",
+    "validation-only-repeat-upsampling-for-multi-label": "Hanya peningkatan sampel mode ulang yang tersedia untuk pelatihan multi-label"
 }
\ No newline at end of file
diff --git a/birdnet_analyzer/lang/pt-br.json b/birdnet_analyzer/lang/pt-br.json
index 3d48d0c20..7d56ed5f5 100644
--- a/birdnet_analyzer/lang/pt-br.json
+++ b/birdnet_analyzer/lang/pt-br.json
@@ -432,16 +432,20 @@
     "training-tab-warning-detached-classifier-append": "O modo de acréscimo não será aplicado ao classificador desanexado.",
     "validation-max-confidence-lower-than-min-confidence": "A confiança máxima deve ser maior que a confiança mínima",
     "validation-no-audio-directory-selected": "Nenhum diretório de áudios selecionado",
+    "validation-no-audio-files-in-training-folders": "As seguintes pastas de dados de treinamento não contêm arquivos de áudio compatíveis: {0}",
     "validation-no-cache-file-selected": "Selecione um arquivo de cache.",
     "validation-no-custom-classifier-selected": "Nenhum classificador customizado selecionado.",
     "validation-no-directory-for-classifier-selected": "Selecione um diretório para o classificador.",
     "validation-no-directory-selected": "Selecione um diretório. ",
     "validation-no-file-selected": "Selecione um arquivo.",
+    "validation-no-negative-samples-in-binary-classification": "Rótulos negativos não podem ser usados com classificação binária",
     "validation-no-species-list-selected": "Selecione uma lista de espécies.",
     "validation-no-training-data-selected": "Selecione seus dados de treinamento.",
     "validation-no-valid-batch-size": "Adicionar um tamanho de lote (batch) válido.",
     "validation-no-valid-classifier-name": "Adicionar um nome válido para o classificador.",
     "validation-no-valid-epoch-number": "Adicionar um número de épocas (epochs) válido.",
     "validation-no-valid-frequency": "Adicionar uma frequência válida em",
-    "validation-no-valid-learning-rate": "Adicionar uma taxa de amostragem válida."
+    "validation-no-valid-learning-rate": "Adicionar uma taxa de amostragem válida.",
+    "validation-non-event-samples-required-in-binary-classification": "Amostras de não-evento são necessárias para a classificação binária",
+    "validation-only-repeat-upsampling-for-multi-label": "Apenas o upsampling de repetição está disponível para treinamento multirrótulo"
 }
\ No newline at end of file
diff --git a/birdnet_analyzer/lang/ru.json b/birdnet_analyzer/lang/ru.json
index cf73464da..87b837f2c 100644
--- a/birdnet_analyzer/lang/ru.json
+++ b/birdnet_analyzer/lang/ru.json
@@ -432,16 +432,20 @@
     "training-tab-warning-detached-classifier-append": "Режим добавления не применяется к отсоединённому классификатору.",
     "validation-max-confidence-lower-than-min-confidence": "Максимальная достоверность должна быть больше минимальной достоверности",
     "validation-no-audio-directory-selected": "Каталог аудио не выбран",
+    "validation-no-audio-files-in-training-folders": "Следующие папки с обучающими данными не содержат поддерживаемых аудиофайлов: {0}",
     "validation-no-cache-file-selected": "Пожалуйста, выберите файл кэша.",
     "validation-no-custom-classifier-selected": "Пользовательский классификатор не выбран.",
     "validation-no-directory-for-classifier-selected": "Пожалуйста, выберите каталог для классификатора.",
     "validation-no-directory-selected": "Пожалуйста, выберите каталог.",
     "validation-no-file-selected": "Пожалуйста, выберите файл.",
+    "validation-no-negative-samples-in-binary-classification": "Отрицательные метки нельзя использовать при бинарной классификации",
     "validation-no-species-list-selected": "Пожалуйста, выберите список видов.",
     "validation-no-training-data-selected": "Выберите данные обучения.",
     "validation-no-valid-batch-size": "Пожалуйста, введите допустимый размер пакета.",
     "validation-no-valid-classifier-name": "Пожалуйста, введите имя для классификатора.",
     "validation-no-valid-epoch-number": "Пожалуйста, введите допустимое количество периодов.",
     "validation-no-valid-frequency": "Введите допустимую частоту в",
-    "validation-no-valid-learning-rate": "Пожалуйста, введите допустимый коэффициент обучения."
+    "validation-no-valid-learning-rate": "Пожалуйста, введите допустимый коэффициент обучения.",
+    "validation-non-event-samples-required-in-binary-classification": "Для бинарной классификации требуются образцы без событий",
+    "validation-only-repeat-upsampling-for-multi-label": "Для обучения с несколькими метками доступен только режим повышающей дискретизации «повтор»"
 }
\ No newline at end of file
diff --git a/birdnet_analyzer/lang/se.json b/birdnet_analyzer/lang/se.json
index 6273e5e56..820bd1cba 100644
--- a/birdnet_analyzer/lang/se.json
+++ b/birdnet_analyzer/lang/se.json
@@ -432,16 +432,20 @@
     "training-tab-warning-detached-classifier-append": "Tilläggsläget gäller inte för den fristående klassificeraren.",
     "validation-max-confidence-lower-than-min-confidence": "Maximal konfidens måste vara större än minimal konfidens",
     "validation-no-audio-directory-selected": "Ingen ljudkatalog vald",
+    "validation-no-audio-files-in-training-folders": "Följande mappar med träningsdata innehåller inga ljudfiler som stöds: {0}",
     "validation-no-cache-file-selected": "Vänligen välj en cachefil.",
     "validation-no-custom-classifier-selected": "Ingen anpassad klassificerare vald.",
     "validation-no-directory-for-classifier-selected": "Välj en katalog för klassificeraren.",
     "validation-no-directory-selected": "Välj en katalog.",
     "validation-no-file-selected": "Välj en fil.",
+    "validation-no-negative-samples-in-binary-classification": "Negativa etiketter kan inte användas med binär klassificering",
     "validation-no-species-list-selected": "Välj en artlista.",
     "validation-no-training-data-selected": "Välj dina träningsdata.",
     "validation-no-valid-batch-size": "Ange en giltig batchstorlek.",
     "validation-no-valid-classifier-name": "Ange ett giltigt namn för klassificeraren.",
     "validation-no-valid-epoch-number": "Ange ett giltigt antal epoker.",
     "validation-no-valid-frequency": "Ange en giltig frekvens i",
-    "validation-no-valid-learning-rate": "Ange en giltig inlärningshastighet."
+    "validation-no-valid-learning-rate": "Ange en giltig inlärningshastighet.",
+    "validation-non-event-samples-required-in-binary-classification": "Prover utan händelser krävs för binär klassificering",
+    "validation-only-repeat-upsampling-for-multi-label": "Endast upsampling med upprepning är tillgänglig för träning med flera etiketter"
 }
\ No newline at end of file
diff --git a/birdnet_analyzer/lang/tlh.json b/birdnet_analyzer/lang/tlh.json
index 544b00408..2d1dc8e3d 100644
--- a/birdnet_analyzer/lang/tlh.json
+++ b/birdnet_analyzer/lang/tlh.json
@@ -432,16 +432,20 @@
     "training-tab-warning-detached-classifier-append": "chel mIw DetlhHa' ghovwI'vaD lInglaHbe'.",
     "validation-max-confidence-lower-than-min-confidence": "nIvbogh vIt patlh tIn law' minimum vIt patlh tIn puS",
     "validation-no-audio-directory-selected": "wav Daq tu'Ha'.",
+    "validation-no-audio-files-in-training-folders": "qeq De' pa'meyvamDaq wab De' tu'lu'be': {0}",
     "validation-no-cache-file-selected": "polmeH teywI' yIwIv.",
     "validation-no-custom-classifier-selected": "chu' tu'law' yIwIvHa'.",
     "validation-no-directory-for-classifier-selected": "tu'law' Daq yIwIv.",
     "validation-no-directory-selected": "Daq yIwIv.",
     "validation-no-file-selected": "wav yIwIv.",
+    "validation-no-negative-samples-in-binary-classification": "cha' buv qeqvaD qechmey Qup lo'laHbe'",
     "validation-no-species-list-selected": "Seghmey wav yIwIv.",
     "validation-no-training-data-selected": "qeq De' yIwIv.",
     "validation-no-valid-batch-size": "batch chu' yIchoH.",
     "validation-no-valid-classifier-name": "tu'law' pong yIchoH.",
     "validation-no-valid-epoch-number": "epQochmey yIchoH.",
     "validation-no-valid-frequency": "qeq 'eS patlh yIchoH",
-    "validation-no-valid-learning-rate": "qeq patlh yIchoH."
+    "validation-no-valid-learning-rate": "qeq patlh yIchoH.",
+    "validation-non-event-samples-required-in-binary-classification": "cha' buv qeqvaD wanI'Ha' De' poQlu'",
+    "validation-only-repeat-upsampling-for-multi-label": "multi-label qeqvaD qechmey ghun neH lo'laH"
 }
\ No newline at end of file
diff --git a/birdnet_analyzer/lang/zh_TW.json b/birdnet_analyzer/lang/zh_TW.json
index c7d6e81ac..ad9641003 100644
--- a/birdnet_analyzer/lang/zh_TW.json
+++ b/birdnet_analyzer/lang/zh_TW.json
@@ -432,16 +432,20 @@
     "training-tab-warning-detached-classifier-append": "附加模式不適用於分離的分類器。",
     "validation-max-confidence-lower-than-min-confidence": "最大可信度必須大於最小可信度",
     "validation-no-audio-directory-selected": "請提供包含音檔的資料夾",
+    "validation-no-audio-files-in-training-folders": "下列訓練資料夾不含任何支援的音訊檔案:{0}",
     "validation-no-cache-file-selected": "請選擇快取檔案。",
     "validation-no-custom-classifier-selected": "尚未選擇客製化分類器",
     "validation-no-directory-for-classifier-selected": "請選擇分類器路徑",
     "validation-no-directory-selected": "請選擇資料夾",
     "validation-no-file-selected": "請選擇檔案",
+    "validation-no-negative-samples-in-binary-classification": "二元分類無法使用負標籤",
     "validation-no-species-list-selected": "請選擇物種清單",
     "validation-no-training-data-selected": "請選擇訓練用資料",
     "validation-no-valid-batch-size": "請輸入有效的批次量(在同一時間分析的音檔數量)",
     "validation-no-valid-classifier-name": "請輸入有效的分類器名稱",
     "validation-no-valid-epoch-number": "請輸入有效的訓練期(eopch)數量",
     "validation-no-valid-frequency": "請輸入有效的頻率",
-    "validation-no-valid-learning-rate": "請輸入有效的學習速率"
+    "validation-no-valid-learning-rate": "請輸入有效的學習速率",
+    "validation-non-event-samples-required-in-binary-classification": "二元分類需要非事件樣本",
+    "validation-only-repeat-upsampling-for-multi-label": "多標籤訓練僅支援重覆上取樣"
 }
\ No newline at end of file
diff --git a/birdnet_analyzer/model.py b/birdnet_analyzer/model.py
index d73413f5a..525ec99a4 100644
--- a/birdnet_analyzer/model.py
+++ b/birdnet_analyzer/model.py
@@ -170,8 +170,8 @@ def random_split(x, y, rng: Generator, val_ratio=0.2):
         train_indices.append(class_train_indices)
         val_indices.append(class_val_indices)
 
-    # Negative samples are not class-specific in single-label training. Appending
-    # them in the loop above duplicates every negative sample once per class.
+    # Add every row with a negative label to the training set exactly once; a
+    # per-class scan would duplicate rows that are negative for several classes.
     negative_indices = np.unique(np.where(y == -1)[0])
     train_indices.append(negative_indices)
 
@@ -334,7 +334,7 @@ def upsampling(
     min_samples = (
         int(max(y.sum(axis=0), len(y) - y.sum(axis=0)) * ratio)
         if is_binary
-        else int(np.max(y.sum(axis=0)) * ratio)
+        else int(np.max((y == 1).sum(axis=0)) * ratio)
     )
     x_temp = []
     y_temp = []
diff --git a/birdnet_analyzer/train/utils.py b/birdnet_analyzer/train/utils.py
index 2c547ac61..22938e06e 100644
--- a/birdnet_analyzer/train/utils.py
+++ b/birdnet_analyzer/train/utils.py
@@ -175,9 +175,12 @@ def _check_input_folders(audio_input: str, train_folders: list[str]):
             empty_folders.append(folder)
 
     if empty_folders:
+        folder_list = ", ".join(sorted(empty_folders))
         raise ValueError(
             "The following training data folders do not contain any supported audio "
-            f"files: {', '.join(empty_folders)}"
+            f"files: {folder_list}",
+            "validation-no-audio-files-in-training-folders",
+            folder_list,
         )
 
 
@@ -267,7 +270,7 @@ def _load_training_data(
 
     if is_multi_label and upsampling_ratio > 0 and upsampling_mode != "repeat":
         raise Exception(
-            "Only repeat-upsampling ist available for multi-label",
+            "Only repeat-upsampling is available for multi-label",
             "validation-only-repeat-upsampling-for-multi-label",
         )
 
diff --git a/tests/train/test_train_loading.py b/tests/train/test_train_loading.py
index 5c6b1d9bb..4bc136f1b 100644
--- a/tests/train/test_train_loading.py
+++ b/tests/train/test_train_loading.py
@@ -95,6 +95,7 @@ def test_load_training_data_lists_empty_folders_before_loading_model(tmp_path):
         _load_training_data(str(tmp_path))
 
     assert "robin" not in str(error.value)
+    assert error.value.args[1] == "validation-no-audio-files-in-training-folders"
     mock_load.assert_not_called()