diff --git a/README.md b/README.md
index 1051c8a6..71990e28 100644
--- a/README.md
+++ b/README.md
@@ -138,7 +138,7 @@ git lfs pull
 
 When adding a new dataset, mock it with
 ```
-python tests/scripts/mock_dataset.py --in-data-dir data/<dataset_id> --out-data-dir tests/data/<dataset_id>
+python tests/scripts/mock_dataset.py --in-data-dir data/$DATASET --out-data-dir tests/data/$DATASET
 ```
 
 Run tests
@@ -155,7 +155,9 @@ huggingface-cli login --token $HUGGINGFACE_TOKEN --add-to-git-credential
 
 Then you can upload it to the hub with:
 ```
-HF_HUB_ENABLE_HF_TRANSFER=1 huggingface-cli upload --repo-type dataset $HF_USER/$DATASET data/$DATASET
+HF_HUB_ENABLE_HF_TRANSFER=1 huggingface-cli upload $HF_USER/$DATASET data/$DATASET \
+--repo-type dataset  \
+--revision v1.0
 ```
 
 For instance, for [cadene/pusht](https://huggingface.co/datasets/cadene/pusht), we used:
@@ -164,6 +166,34 @@ HF_USER=cadene
 DATASET=pusht
 ```
 
+If you want to improve an existing dataset, you can download it locally with:
+```
+mkdir -p data/$DATASET
+HF_HUB_ENABLE_HF_TRANSFER=1 huggingface-cli download $HF_USER/$DATASET \
+--repo-type dataset \
+--local-dir data/$DATASET \
+--local-dir-use-symlinks=False \
+--revision v1.0
+```
+
+Iterate on your code and dataset with:
+```
+DATA_DIR=data python train.py
+```
+
+Then upload a new version:
+```
+HF_HUB_ENABLE_HF_TRANSFER=1 huggingface-cli upload $HF_USER/$DATASET data/$DATASET \
+--repo-type dataset \
+--revision v1.1 \
+--delete "*"
+```
+
+And you might want to mock the dataset if you need to update the unit tests as well:
+```
+python tests/scripts/mock_dataset.py --in-data-dir data/$DATASET --out-data-dir tests/data/$DATASET
+```
+
 
 ## Acknowledgment
 - Our Diffusion policy and Pusht environment are adapted from [Diffusion Policy](https://diffusion-policy.cs.columbia.edu/)
diff --git a/lerobot/common/datasets/abstract.py b/lerobot/common/datasets/abstract.py
index 34b33c2e..9127d887 100644
--- a/lerobot/common/datasets/abstract.py
+++ b/lerobot/common/datasets/abstract.py
@@ -19,6 +19,7 @@ class AbstractExperienceReplay(TensorDictReplayBuffer):
     def __init__(
         self,
         dataset_id: str,
+        version: str | None = None,
         batch_size: int = None,
         *,
         shuffle: bool = True,
@@ -31,6 +32,7 @@ class AbstractExperienceReplay(TensorDictReplayBuffer):
         transform: "torchrl.envs.Transform" = None,
     ):
         self.dataset_id = dataset_id
+        self.version = version
         self.shuffle = shuffle
         self.root = root
         storage = self._download_or_load_dataset()
@@ -96,10 +98,14 @@ class AbstractExperienceReplay(TensorDictReplayBuffer):
 
     def _download_or_load_dataset(self) -> torch.StorageBase:
         if self.root is None:
-            self.data_dir = Path(snapshot_download(repo_id=f"cadene/{self.dataset_id}", repo_type="dataset"))
+            self.data_dir = Path(
+                snapshot_download(
+                    repo_id=f"cadene/{self.dataset_id}", repo_type="dataset", revision=self.version
+                )
+            )
         else:
             self.data_dir = self.root / self.dataset_id
-        return TensorStorage(TensorDict.load_memmap(self.data_dir))
+        return TensorStorage(TensorDict.load_memmap(self.data_dir / "replay_buffer"))
 
     def _compute_stats(self, num_batch=100, batch_size=32):
         rb = TensorDictReplayBuffer(
diff --git a/lerobot/common/datasets/aloha.py b/lerobot/common/datasets/aloha.py
index 52a5676e..2af98cd8 100644
--- a/lerobot/common/datasets/aloha.py
+++ b/lerobot/common/datasets/aloha.py
@@ -84,6 +84,7 @@ class AlohaExperienceReplay(AbstractExperienceReplay):
     def __init__(
         self,
         dataset_id: str,
+        version: str | None = None,
         batch_size: int = None,
         *,
         shuffle: bool = True,
@@ -99,6 +100,7 @@ class AlohaExperienceReplay(AbstractExperienceReplay):
 
         super().__init__(
             dataset_id,
+            version,
             batch_size,
             shuffle=shuffle,
             root=root,
diff --git a/lerobot/common/datasets/pusht.py b/lerobot/common/datasets/pusht.py
index f4f6d9ac..3ad6371f 100644
--- a/lerobot/common/datasets/pusht.py
+++ b/lerobot/common/datasets/pusht.py
@@ -87,6 +87,7 @@ class PushtExperienceReplay(AbstractExperienceReplay):
     def __init__(
         self,
         dataset_id: str,
+        version: str | None = "v1.0",
         batch_size: int = None,
         *,
         shuffle: bool = True,
@@ -100,6 +101,7 @@ class PushtExperienceReplay(AbstractExperienceReplay):
     ):
         super().__init__(
             dataset_id,
+            version,
             batch_size,
             shuffle=shuffle,
             root=root,
diff --git a/lerobot/common/datasets/simxarm.py b/lerobot/common/datasets/simxarm.py
index 7bcb03fb..d7e2e18f 100644
--- a/lerobot/common/datasets/simxarm.py
+++ b/lerobot/common/datasets/simxarm.py
@@ -40,6 +40,7 @@ class SimxarmExperienceReplay(AbstractExperienceReplay):
     def __init__(
         self,
         dataset_id: str,
+        version: str | None = None,
         batch_size: int = None,
         *,
         shuffle: bool = True,
@@ -53,6 +54,7 @@ class SimxarmExperienceReplay(AbstractExperienceReplay):
     ):
         super().__init__(
             dataset_id,
+            version,
             batch_size,
             shuffle=shuffle,
             root=root,
diff --git a/tests/data/pusht/action.memmap b/tests/data/pusht/replay_buffer/action.memmap
similarity index 100%
rename from tests/data/pusht/action.memmap
rename to tests/data/pusht/replay_buffer/action.memmap
diff --git a/tests/data/pusht/episode.memmap b/tests/data/pusht/replay_buffer/episode.memmap
similarity index 100%
rename from tests/data/pusht/episode.memmap
rename to tests/data/pusht/replay_buffer/episode.memmap
diff --git a/tests/data/pusht/frame_id.memmap b/tests/data/pusht/replay_buffer/frame_id.memmap
similarity index 100%
rename from tests/data/pusht/frame_id.memmap
rename to tests/data/pusht/replay_buffer/frame_id.memmap
diff --git a/tests/data/pusht/meta.json b/tests/data/pusht/replay_buffer/meta.json
similarity index 100%
rename from tests/data/pusht/meta.json
rename to tests/data/pusht/replay_buffer/meta.json
diff --git a/tests/data/pusht/next/done.memmap b/tests/data/pusht/replay_buffer/next/done.memmap
similarity index 100%
rename from tests/data/pusht/next/done.memmap
rename to tests/data/pusht/replay_buffer/next/done.memmap
diff --git a/tests/data/pusht/next/meta.json b/tests/data/pusht/replay_buffer/next/meta.json
similarity index 100%
rename from tests/data/pusht/next/meta.json
rename to tests/data/pusht/replay_buffer/next/meta.json
diff --git a/tests/data/pusht/next/observation/image.memmap b/tests/data/pusht/replay_buffer/next/observation/image.memmap
similarity index 100%
rename from tests/data/pusht/next/observation/image.memmap
rename to tests/data/pusht/replay_buffer/next/observation/image.memmap
diff --git a/tests/data/pusht/next/observation/meta.json b/tests/data/pusht/replay_buffer/next/observation/meta.json
similarity index 100%
rename from tests/data/pusht/next/observation/meta.json
rename to tests/data/pusht/replay_buffer/next/observation/meta.json
diff --git a/tests/data/pusht/next/observation/state.memmap b/tests/data/pusht/replay_buffer/next/observation/state.memmap
similarity index 100%
rename from tests/data/pusht/next/observation/state.memmap
rename to tests/data/pusht/replay_buffer/next/observation/state.memmap
diff --git a/tests/data/pusht/next/reward.memmap b/tests/data/pusht/replay_buffer/next/reward.memmap
similarity index 100%
rename from tests/data/pusht/next/reward.memmap
rename to tests/data/pusht/replay_buffer/next/reward.memmap
diff --git a/tests/data/pusht/next/success.memmap b/tests/data/pusht/replay_buffer/next/success.memmap
similarity index 100%
rename from tests/data/pusht/next/success.memmap
rename to tests/data/pusht/replay_buffer/next/success.memmap
diff --git a/tests/data/pusht/observation/image.memmap b/tests/data/pusht/replay_buffer/observation/image.memmap
similarity index 100%
rename from tests/data/pusht/observation/image.memmap
rename to tests/data/pusht/replay_buffer/observation/image.memmap
diff --git a/tests/data/pusht/observation/meta.json b/tests/data/pusht/replay_buffer/observation/meta.json
similarity index 100%
rename from tests/data/pusht/observation/meta.json
rename to tests/data/pusht/replay_buffer/observation/meta.json
diff --git a/tests/data/pusht/observation/state.memmap b/tests/data/pusht/replay_buffer/observation/state.memmap
similarity index 100%
rename from tests/data/pusht/observation/state.memmap
rename to tests/data/pusht/replay_buffer/observation/state.memmap
diff --git a/tests/scripts/mock_dataset.py b/tests/scripts/mock_dataset.py
index c58280d7..2200b644 100644
--- a/tests/scripts/mock_dataset.py
+++ b/tests/scripts/mock_dataset.py
@@ -10,12 +10,15 @@ from pathlib import Path
 
 
 def mock_dataset(in_data_dir, out_data_dir, num_frames=50):
+    in_data_dir = Path(in_data_dir)
+    out_data_dir = Path(out_data_dir)
+
     # load full dataset as a tensor dict
-    in_td_data = TensorDict.load_memmap(in_data_dir)
+    in_td_data = TensorDict.load_memmap(in_data_dir / "replay_buffer")
 
     # use 1 frame to know the specification of the dataset
     # and copy it over `n` frames in the test artifact directory
-    out_td_data = in_td_data[0].expand(num_frames).memmap_like(out_data_dir)
+    out_td_data = in_td_data[0].expand(num_frames).memmap_like(out_data_dir / "replay_buffer")
 
     # copy the first `n` frames so that we have real data
     out_td_data[:num_frames] = in_td_data[:num_frames].clone()
@@ -24,8 +27,8 @@ def mock_dataset(in_data_dir, out_data_dir, num_frames=50):
     out_td_data.lock_()
 
     # copy the full statistics of dataset since it's pretty small
-    in_stats_path = Path(in_data_dir) / "stats.pth"
-    out_stats_path = Path(out_data_dir) / "stats.pth"
+    in_stats_path = in_data_dir / "stats.pth"
+    out_stats_path = out_data_dir / "stats.pth"
     shutil.copy(in_stats_path, out_stats_path)