community_dataset_v2_v3

HuggingFaceVLA/community_dataset_v2 with every sub-dataset migrated to the LeRobot v3.0 dataset format, so it loads directly with lerobot ≥ 0.4 (LeRobotDataset(repo_id, root=...)) without per-dataset conversion.

Conversion is fully reproducible with lerobot-dataset-tools; the exact per-dataset log is in conversion_manifest.jsonl.

Usage

This is an umbrella repo: every <user>/<dataset> subtree is an independent LeRobotDataset (hence no dataset viewer). Download a subtree and point lerobot at it:

from huggingface_hub import snapshot_download
from lerobot.datasets.lerobot_dataset import LeRobotDataset

name = "0x00raghu/toffee_blue"  # any sub-dataset in this repo
root = snapshot_download(
    "mcobzarenco/community_dataset_v2_v3",
    repo_type="dataset",
    allow_patterns=[f"{name}/**"],
    local_dir="./data",
)
dataset = LeRobotDataset(name, root=f"./data/{name}")

or with the CLI:

hf download mcobzarenco/community_dataset_v2_v3 --repo-type=dataset --include "0x00raghu/toffee_blue/**" --local-dir ./data

Statistics

	source	converted
sub-datasets	324 (1× v2.0, 323× v2.1)	323 (all v3.0)
episodes	12,953	12,179
frames	7,970,878	5,825,171
size	259.7 GB	129.8 GB

The size reduction is not compression: the source repos ship most videos twice (a stray flat tree next to the real one); only the canonical copies are converted. Videos are stream-copied, never re-encoded.

What the conversion does

Per sub-dataset (see the tools repo for code):

1. Copy the dataset, skipping stray duplicate video trees and *.bak files.
2. Drop parquet columns not declared in info.json features (legacy leftovers such as next.done break v3.0 schema casting).
3. Repair the collection-wide stats-key bug (stats keyed observation.<cam> while features are observation.images.<cam>).
4. For v2.0 sources: synthesize meta/episodes_stats.jsonl (numeric stats exact from parquet; per-episode image stats from ~100 sampled frames per camera, matching lerobot's own conventions) and bump to v2.1.
5. Run the official lerobot.scripts.convert_dataset_v21_to_v30 converter (with an ffmpeg concat fallback for episodes whose packets carry duplicate/non-monotonic DTS, which PyAV refuses to mux).
6. Validate: reload with LeRobotDataset, check episode/frame counts, and verify per-episode length == index-range == video-span × fps per camera.

Datasets not included

dataset	source version	reason
Yotofu/so100_sweeper_shoes	v2.0	declares video features ['observation.image', 'observation.image2', 'observation.image3', 'null'] but has no videos/ directory

Reproduce

git clone https://github.com/mcobzarenco/lerobot-dataset-tools && cd lerobot-dataset-tools
uv run hf download HuggingFaceVLA/community_dataset_v2 --repo-type=dataset --local-dir ./source
uv run python -m ldtools.convert_collection \
    --source ./source --output ./converted --workers 12
uv run python -m ldtools.dataset_card \
    --source ./source --output ./converted \
    --source-repo HuggingFaceVLA/community_dataset_v2 --target-repo mcobzarenco/community_dataset_v2_v3 --write

Source-data corrections

Some source datasets ship metadata that contradicts their own data (e.g. inflated total_frames counters, which make LeRobotDataset.__len__ overreport and crash shuffled samplers with IndexError). The converter recomputes every aggregate from the actual episode table and records each correction in the manifest. Corrections applied in this collection:

dataset	correction
zaringleb/so100_cube_6_2d	total_frames 13690 → 13421
zaringleb/so100_cude_linear_and_2d_comb	total_frames 27116 → 26847; total_videos 0 → 100

Remaining known quirk: a handful of episodes across the collections have a few trailing video frames beyond their declared length (source recording artifacts, not repairable without re-encoding). They load, decode and train normally — frames are indexed via episode metadata, so the tail frames are simply never read.

Provenance & license

All data originates from HuggingFaceVLA/community_dataset_v2 (community-contributed robot demonstrations, Apache-2.0). This repo changes the storage format only: actions, states, task strings and video content are unmodified (videos are stream-copied; v2.0 per-episode image stats are sampled approximations).

Generated 2026-07-24 by lerobot-dataset-tools.