dataset format versioning

[edgarriba]

hi, what's the latest supported version for lerobot dataset ? they have been changing the format and now they are in 3.0.

[gaoyitong02]

+1, I can confirm this issue. I've done a deep-dive and it seems to be a versioning paradox between the code and the data.

The lerobot code requires an older version of huggingface/datasets to avoid a TypeError, but the dataset on the Hub requires a newer version of huggingface/datasets to avoid a ValueError.

Detailed Analysis
I performed a fresh deployment on a Linux HPC cluster and encountered two distinct failure modes.

Failure Mode 1: Crash with Latest Dependencies

A fresh install with the latest dependencies (via pip or uv sync) crashes with a TypeError inside the lerobot library when creating the data loader. This suggests an incompatibility with a recent datasets API.

Traceback:

Traceback (most recent call last):
File "/home/yitong005/openpi/examples/inference.py", line 78, in
loader = _data_loader.create_data_loader(config, num_batches=1, skip_norm_stats=True)
File "/home/yitong005/openpi/src/openpi/training/data_loader.py", line 141, in create_torch_dataset
dataset = lerobot_dataset.LeRobotDataset(...)
File "/home/yitong005/.conda/envs/openpi_env/lib/python3.11/site-packages/lerobot/common/datasets/lerobot_dataset.py", line 508, in init
timestamps = torch.stack(self.hf_dataset["timestamp"]).numpy()
TypeError: stack(): argument 'tensors' (position 1) must be tuple of Tensors, not Column

Failure Mode 2: Crash with Downgraded datasets Library

To work around the TypeError, I downgraded the library (pip install datasets==3.6.0). This resolved the first error, but introduced a new ValueError when load_dataset tries to parse the dataset's feature schema from the Hub.

Traceback:

Traceback (most recent call last):
File "/home/yitong005/openpi/examples/inference.py", line 78, in
loader = _data_loader.create_data_loader(config, num_batches=1, skip_norm_stats=True)
File "/home/yitong005/.conda/envs/openpi_env/lib/python3.11/site-packages/lerobot/common/datasets/lerobot_dataset.py", line 620, in load_hf_dataset
hf_dataset = load_dataset("parquet", data_dir=path, split="train")
File "/home/yitong005/.conda/envs/openpi_env/lib/python3.11/site-packages/datasets/features/features.py", line 1474, in generate_from_dict
raise ValueError(f"Feature type '{_type}' not found...")
ValueError: Feature type 'List' not found. Available feature types: ['Value', 'ClassLabel', 'Translation', ...]

Environment
Python: 3.11

Platform: Linux (Slurm HPC Cluster)

Installation: uv sync / pip

Key Libraries: huggingface-datasets (tested latest vs. 3.6.0)

This seems to confirm that the project is currently in a broken state due to this dependency version conflict. Hope this helps track down the issue.

[xuxiaoxxxx]

Do you slove this problem? I also meet this problem.

[jj701]

Is this issue fixed? I was facing the same issue. Are there any work arounds?

[sunmoon2018]

I also encountered the same problem, where the dataset version 2.0 or 3.0 cannot be directly used, but version 2.1 has been tested and is acceptable.

[Virlus]

+1, I can confirm this issue. I've done a deep-dive and it seems to be a versioning paradox between the code and the data.

The lerobot code requires an older version of huggingface/datasets to avoid a TypeError, but the dataset on the Hub requires a newer version of huggingface/datasets to avoid a ValueError.

Detailed Analysis I performed a fresh deployment on a Linux HPC cluster and encountered two distinct failure modes.

Failure Mode 1: Crash with Latest Dependencies
A fresh install with the latest dependencies (via pip or uv sync) crashes with a TypeError inside the lerobot library when creating the data loader. This suggests an incompatibility with a recent datasets API.

Traceback:

Traceback (most recent call last): File "/home/yitong005/openpi/examples/inference.py", line 78, in loader = _data_loader.create_data_loader(config, num_batches=1, skip_norm_stats=True) File "/home/yitong005/openpi/src/openpi/training/data_loader.py", line 141, in create_torch_dataset dataset = lerobot_dataset.LeRobotDataset(...) File "/home/yitong005/.conda/envs/openpi_env/lib/python3.11/site-packages/lerobot/common/datasets/lerobot_dataset.py", line 508, in init timestamps = torch.stack(self.hf_dataset["timestamp"]).numpy() TypeError: stack(): argument 'tensors' (position 1) must be tuple of Tensors, not Column

Failure Mode 2: Crash with Downgraded datasets Library
To work around the TypeError, I downgraded the library (pip install datasets==3.6.0). This resolved the first error, but introduced a new ValueError when load_dataset tries to parse the dataset's feature schema from the Hub.

Traceback:

Traceback (most recent call last): File "/home/yitong005/openpi/examples/inference.py", line 78, in loader = _data_loader.create_data_loader(config, num_batches=1, skip_norm_stats=True) File "/home/yitong005/.conda/envs/openpi_env/lib/python3.11/site-packages/lerobot/common/datasets/lerobot_dataset.py", line 620, in load_hf_dataset hf_dataset = load_dataset("parquet", data_dir=path, split="train") File "/home/yitong005/.conda/envs/openpi_env/lib/python3.11/site-packages/datasets/features/features.py", line 1474, in generate_from_dict raise ValueError(f"Feature type '{_type}' not found...") ValueError: Feature type 'List' not found. Available feature types: ['Value', 'ClassLabel', 'Translation', ...]

Environment Python: 3.11

Platform: Linux (Slurm HPC Cluster)

Installation: uv sync / pip

Key Libraries: huggingface-datasets (tested latest vs. 3.6.0)

This seems to confirm that the project is currently in a broken state due to this dependency version conflict. Hope this helps track down the issue.

Same as Failure Mode 2 here. Does downgrading lerobot work?

[Virlus]

It seems that downgrading to datasets==3.6.0 works for me.

[exaFLOPs26]

Is this error solved? My dataset is currently in version 3.0.

[prachigarg23]

Hi, @Alegunm , @Virlus, I have the same mode 2 failure. What is a possible fix? Which version of lerobot and the lerobot dataset format did you finally use (v2.0/v2.1)?

[0xadvait]

Pinning down what is actually breaking here, since three distinct failure modes are tangled together in this thread:

1. Pinned lerobot vs newer datasets: openpi pins lerobot to a commit from 2025-05-28 (0cf8648). That code calls torch.stack(self.hf_dataset["timestamp"]), which breaks when newer datasets releases return lazy Column objects instead of tensor lists (the TypeError: ... must be tuple of Tensors, not Column in lerobot dataset v2.1 compute_norm_stats error #753). The uv lockfile resolves datasets==3.6.0, so uv sync environments never hit this; pip installs and custom envs do.

2. Dataset format v3.0 vs pinned lerobot: datasets converted or pushed with recent lerobot releases use the v3 layout, which the pinned rev cannot read at all. v2.1 is the only format that works with this repo today (Lerobot code like the ones inside data_loader are deprecated. #916 and PI05 run compute_nome_stats #948 are this mode).

3. The combination is the paradox described above: the pinned code needs an older datasets, while v3-format hub data needs newer lerobot code.

Short term, two cheap changes would close most of the duplicate reports: an explicit datasets bound in pyproject so non-uv installs resolve a compatible version, and an early codebase_version check in the data loader that fails with an actionable message (supported format v2.1, how to convert) instead of a deep traceback. Happy to send that PR now.

Longer term the real fix is bumping to current lerobot and adapting the data loader (moved module paths, changed LeRobotDataset API, v3 format support). That is a bigger review burden, so checking before building it: would you take that migration as a PR, or do you prefer keeping the pin and documenting v2.1 as the supported format?

[0xadvait]

Mitigation PR is up: #972 (bounds datasets below 4.0 and fails with an actionable error on v3-format datasets instead of the deep tracebacks). The lerobot version bump question above still stands.