LeRobot v2.1 vs v3.0: what changed and what breaks

Hugging Face rebuilt the LeRobot dataset format as v3.0 in late 2025 because the old layout stopped scaling. Here is what actually changed, and what it means if you publish or consume datasets.

Why did the format change?

The old layout stored one file per episode. Hugging Face's own release notes say it hit file-system limits when datasets scaled to millions of episodes. v3.0 packs episodes into larger chunked files and supports datasets past the 400GB range.

What breaks in practice?

The visible one for tooling: task labels moved out of the simple meta/tasks.jsonl file into parquet. Simple scripts that read tasks.jsonl, including basic label checks, see nothing on a v3.0 dataset. That is why some label checks on Datum show "unknown" for v3.0 datasets today.

How far along is adoption?

In our latest scan of the 1000 most downloaded datasets: 662 use v3.0, 144 use v2.1, and 45 still use v2.0. Adoption numbers refresh with every build.

What should you do?

Publishing something new and large: export to v3.0 with current LeRobot tooling. Maintaining a small v2.1 dataset: no rush, but run datum-lint before any re-upload so the metadata and labels are clean whichever version you ship.