diff --git a/tests/pretrain/test_pretrainer.py b/tests/pretrain/test_pretrainer.py
new file mode 100644
index 0000000..d8543a6
--- /dev/null
+++ b/tests/pretrain/test_pretrainer.py
@@ -0,0 +1,115 @@
+import pytest
+import torch
+from unittest.mock import MagicMock, patch
+from aiia import Pretrainer, ProjectionHead, AIIABase, AIIAConfig
+from unittest.mock import patch, MagicMock
+import pandas as pd
+
+
+# Test the ProjectionHead class
+def test_projection_head():
+    head = ProjectionHead(hidden_size=512)
+    x = torch.randn(1, 512, 32, 32)
+
+    # Test denoise task
+    output_denoise = head(x, task='denoise')
+    assert output_denoise.shape == (1, 3, 32, 32)
+
+    # Test rotate task
+    output_rotate = head(x, task='rotate')
+    assert output_rotate.shape == (1, 4)
+
+# Test the Pretrainer class initialization
+def test_pretrainer_initialization():
+    config = AIIAConfig()
+    model = AIIABase(config=config)
+    pretrainer = Pretrainer(model=model, learning_rate=0.001, config=config)
+    assert pretrainer.device in ["cuda", "cpu"]
+    assert isinstance(pretrainer.projection_head, ProjectionHead)
+    assert isinstance(pretrainer.optimizer, torch.optim.AdamW)
+
+# Test the safe_collate method
+def test_safe_collate():
+    pretrainer = Pretrainer(model=AIIABase(config=AIIAConfig()), config=AIIAConfig())
+    batch = [
+        (torch.randn(3, 32, 32), torch.randn(3, 32, 32), 'denoise'),
+        (torch.randn(3, 32, 32), torch.tensor(1), 'rotate')
+    ]
+
+    collated_batch = pretrainer.safe_collate(batch)
+    assert 'denoise' in collated_batch
+    assert 'rotate' in collated_batch
+
+# Test the _process_batch method
+@patch('aiia.pretrain.pretrainer.Pretrainer._process_batch')
+def test_process_batch(mock_process_batch):
+    pretrainer = Pretrainer(model=AIIABase(config=AIIAConfig()), config=AIIAConfig())
+    batch_data = {
+        'denoise': (torch.randn(2, 3, 32, 32), torch.randn(2, 3, 32, 32)),
+        'rotate': (torch.randn(2, 3, 32, 32), torch.tensor([0, 1]))
+    }
+    criterion_denoise = MagicMock()
+    criterion_rotate = MagicMock()
+
+    mock_process_batch.return_value = 0.5
+    loss = pretrainer._process_batch(batch_data, criterion_denoise, criterion_rotate)
+    assert loss == 0.5
+
+@patch('pandas.concat')
+@patch('pandas.read_parquet')
+@patch('aiia.pretrain.pretrainer.AIIADataLoader')
+def test_train(mock_data_loader, mock_read_parquet, mock_concat):
+    # Create a real DataFrame for testing
+    real_df = pd.DataFrame({
+    'image_bytes': [torch.randn(1, 3, 224, 224).tolist()]
+    })
+    # Configure read_parquet so that for each dataset path, .head(10000) returns real_df
+    mock_read_parquet.return_value.head.return_value = real_df
+
+    # When merging DataFrames, bypass type-checks by letting pd.concat just return real_df
+    mock_concat.return_value = real_df
+
+    # Create an instance of the Pretrainer
+    pretrainer = Pretrainer(model=AIIABase(config=AIIAConfig()), config=AIIAConfig())
+    dataset_paths = ['path/to/dataset1.parquet', 'path/to/dataset2.parquet']
+
+    # Setup the data loader mock instance with empty loaders
+    loader_instance = MagicMock()
+    loader_instance.train_loader = []  # so the training loop is immediately skipped
+    loader_instance.val_loader = []    # so the validation loop is also skipped
+    mock_data_loader.return_value = loader_instance
+
+    # Patch _validate to avoid any actual validation computations.
+    with patch.object(Pretrainer, '_validate', return_value=0.5):
+        pretrainer.train(dataset_paths, num_epochs=1)
+
+    # Verify that AIIADataLoader was instantiated exactly once...
+    mock_data_loader.assert_called_once()
+    # ...and that pd.read_parquet was called once per dataset path (i.e. 2 times in this test)
+    expected_calls = len(dataset_paths)
+    assert mock_read_parquet.call_count == expected_calls, (
+        f"Expected {expected_calls} calls to pd.read_parquet, got {mock_read_parquet.call_count}"
+    )
+
+
+@patch('aiia.pretrain.pretrainer.Pretrainer._process_batch')
+def test_validate(mock_process_batch):
+    pretrainer = Pretrainer(model=AIIABase(config=AIIAConfig()), config=AIIAConfig())
+    val_loader = [MagicMock()]
+    criterion_denoise = MagicMock()
+    criterion_rotate = MagicMock()
+
+    mock_process_batch.return_value = torch.tensor(0.5)
+    loss = pretrainer._validate(val_loader, criterion_denoise, criterion_rotate)
+    assert loss == 0.5
+
+# Test the save_losses method
+@patch('aiia.pretrain.pretrainer.Pretrainer.save_losses')
+def test_save_losses(mock_save_losses):
+    pretrainer = Pretrainer(model=AIIABase(config=AIIAConfig()), config=AIIAConfig())
+    pretrainer.train_losses = [0.1, 0.2]
+    pretrainer.val_losses = [0.3, 0.4]
+
+    csv_file = 'losses.csv'
+    pretrainer.save_losses(csv_file)
+    mock_save_losses.assert_called_once_with(csv_file)
\ No newline at end of file