actionnet_3k_og — training subset, archived

Everything actionnet_grasp_2b_480 / _720 opens, plus the LeRobot data/ and meta/ that describe the same episodes: ~35.0 GB in 9 archives instead of ~14,900 loose files. A 13th archive set, videos_15fps_768x432/, comes from a different store — read the note below before using it.

This is still a subset of a larger tree. The source also holds videos/, first_frames/, grasp_frames/, source_mask_merge/, rendering_videos/, gtdepth_s0/ and the *_pinhole/ variants. None is read by the training dataset, so none is here. source_mask in particular is written into every manifest row and never read back.

Contents

actionnet_3k_og/
├── cropped_video/            34 G   split per chunk: one 34 G object is over HF's
│   ├── chunk-000.zip        ~11 G   recommended 20 G/file, and a failed upload
│   ├── chunk-001.zip        ~11 G   would cost the whole transfer
│   ├── chunk-002.zip        ~12 G
│   └── chunk-003.zip         338 M
├── rendering_first_frames.zip 305 M
├── rendering_grasp_frames.zip 270 M
├── target_mask_1280_800.zip   9.1 M
├── data.zip                   381 M  LeRobot episode parquet, 3027 files
├── meta.zip                   465 K  LeRobot metadata -- 5.4 M expanded, see note below
├── videos_15fps_768x432/     23.2 G  A DIFFERENT STORE -- see the warning below
│   ├── chunk-000.zip         7.5 G
│   ├── chunk-001.zip         7.8 G
│   ├── chunk-002.zip         7.8 G
│   └── chunk-003.zip          95 M
├── train_cosmos.jsonl         3.5 M  raw -- readable on the Hub, no download
├── train_cosmos_grasp.jsonl   3.8 M  raw
├── SHA256SUMS
└── README.md

object	expands to	manifest field
cropped_video/chunk-00{0,1,2,3}.zip	cropped_video/chunk-00N/	video
rendering_first_frames.zip	rendering_first_frames/	render_first
rendering_grasp_frames.zip	rendering_grasp_frames/	render_grasp
target_mask_1280_800.zip	target_mask_1280_800/	target_mask
data.zip	data/chunk-00{0,1,2,3}/episode_*.parquet	—
meta.zip	meta/ (incl. meta/camera/)	—
videos_15fps_768x432/chunk-00{0,1,2,3}.zip	videos_15fps_768x432/chunk-00N/	—
train_cosmos.jsonl	(raw)	SLURM default manifest
train_cosmos_grasp.jsonl	(raw)	MLXP manifest, rewritten grasp prompts

2903 usable episodes; both manifests have the same rows and differ only in prompt.

data/ and meta/ are not read by the video training set — they are the LeRobot side of the same store (3027 episodes: info.json, modality.json, stats.json, tasks.jsonl, episodes.jsonl, source.jsonl, source_mask.jsonl and the ego-view camera calibration under meta/camera/). They are here so a restore reconstitutes a loadable LeRobot dataset, not only the Cosmos inputs. Note data/ covers all 3027 episodes while the manifests cover 2903 — the 124 difference is the episodes the manifest build dropped, not missing parquet.

⚠ videos_15fps_768x432/ is not this dataset's video

It is a separate store — gr1_actionnet_lerobot_3k_15fps_og_resol — and it does not line up with the rest of this repo:

	rest of this repo	videos_15fps_768x432/
episodes	3027	3013
resolution	1280×800	768×432
fps	30	15
frames	1,536,458	764,691

So meta/info.json here describes 3027 episodes at 30 fps and will not match these videos frame for frame. They are included because they are the inputs the GR1 IDM finetune reads, and the folder is named for its resolution and fps so the two can never be confused. The 30 fps 1280×800 original this repo's cropped_video/ was derived from is 105 GB and is not uploaded.

Every archive carries its own folder prefix internally (cropped_video/chunk-000/...), so they all unzip into one $BASE no matter where the .zip sits in this repo.

Restore

pip install -U "huggingface_hub[cli,hf_transfer]"
export HF_HUB_ENABLE_HF_TRANSFER=1
BASE=/path/to/actionnet_3k_og

hf download <OWNER>/actionnet_3k_og --repo-type dataset --local-dir /tmp/an_zips
cd /tmp/an_zips && sha256sum -c SHA256SUMS
mkdir -p "$BASE"
find . -name '*.zip' | while read -r z; do unzip -q -n "$z" -d "$BASE"; done
cp ./*.jsonl "$BASE"/

Only need the videos re-fetched? hf download <OWNER>/actionnet_3k_og --repo-type dataset --include 'cropped_video/*' --local-dir /tmp/an_zips.

Then confirm the restore is what training expects — this stats every file the loader opens:

python3 new_generation/overlay/scripts/check_actionnet_data.py \
    --base "$BASE" --manifest "$BASE/train_cosmos.jsonl"
# expect: 누락 0 / 2,903  on video / render_first / render_grasp / target_mask

Archives are stored, not deflated — mp4, png and parquet are already compressed, so unzip is I/O-bound and fast. meta.zip is the one exception: it is jsonl and json, which deflate 11.7x (5.4 M → 465 K), and it is small enough that the CPU cost is irrelevant.