diff --git a/imperative/python/megengine/data/dataset/meta_dataset.py b/imperative/python/megengine/data/dataset/meta_dataset.py
index 354c8141c098702333e0fcff846302a1982e8aa5..be8f921a02727f9e8b33dd602c1fc0593fab11e9 100644
--- a/imperative/python/megengine/data/dataset/meta_dataset.py
+++ b/imperative/python/megengine/data/dataset/meta_dataset.py
@@ -4,9 +4,41 @@ from typing import Tuple
 
 
 class Dataset(ABC):
-    r"""An abstract base class for all datasets.
+    r"""An abstract base class for all map-style datasets.
 
-    __getitem__ and __len__ method are aditionally needed.
+    .. admonition:: Abstract methods
+    
+       All subclasses should overwrite these two methods:
+
+       * ``__getitem__()``: fetch a data sample for a given key.
+       * ``__len__()``: return the size of the dataset.
+
+       They play roles in the data pipeline, see the description below.
+
+    .. admonition:: Dataset in the Data Pipline
+
+       Usually a dataset works with :class:`~.DataLoader`, :class:`~.Sampler`, :class:`~.Collator` and other components.
+
+       For example, the sampler generates **indexes** of batches in advance according to the size of the dataset (calling ``__len__``),
+       When dataloader need to yield a batch of data, pass indexes into the ``__getitem__`` method, then collate them to a batch.
+
+       * Highly recommended reading :ref:`dataset-guide` for more details;
+       * It might helpful to read the implementation of :class:`~.MNIST`, :class:`~.CIFAR10` and other existed subclass.
+
+    .. warning::
+
+       By default, all elements in a dataset would be :class:`numpy.ndarray`.
+       It means that if you want to do Tensor operations, it's better to do the conversion explicitly, such as:
+
+       .. code-block:: python
+
+          dataset = MyCustomDataset()  # A subclass of Dataset
+          data, label = MyCustomDataset[0]  # equals to MyCustomDataset.__getitem__[0]
+          data = Tensor(data, dtype="float32")  # convert to MegEngine Tensor explicitly
+
+          megengine.functional.ops(data)
+
+       Tensor ops on ndarray directly are undefined behaviors.
     """
 
     @abstractmethod