Hugging Face's logo

Datasets:
t0rn0d0
/
vla-delta

Tasks:
Robotics
Languages:
English
Chinese
Tags:
vla
vla-delta
pretraining
optical-flow
robot-learning
License:
Dataset card Files
xet
 Community

You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

VLA-Delta formal three-stage pretraining data

A complete release contains READY.json and every archive listed in manifest.json; initial upload commits may contain only part of the release.

The four-source data release used for VLA-Delta pretraining. It includes the full local DROID 1.0.1 conversion, MolmoAct Household-v3 and Tabletop-v3, RH20T-v30, difference text, structured Slot labels, dense optical flow, source annotations, frozen sample indices, and independent normalization.

The archived trajectory data, labels, masks, sample membership, and numerical normalization statistics are preserved byte-for-byte. The installation step relocates metadata and explicitly imports the historical preparation for the bundled VLA-Delta loader. Original preparation and generation identities remain in templates/; migrated metadata records the import operation. No annotation, optical-flow generation, or full-data statistics computation is required.

Source Source episodes Source frames Stage 1/2 admitted frames Stage 3 admitted frames
DROID 1.0.1, full local release 95,658 27,630,375 1,815,749 19,823,040
MolmoAct Household-v3 5,936 794,199 260,412 794,199
MolmoAct Tabletop-v3 1,881 310,743 73,081 310,743
RH20T-v30 8,142 3,898,056 109,214 3,131,833
Total 111,617 32,633,373 2,258,456 24,059,815

All populations are training populations. No LIBERO, RoboTwin or RoboCasa is included. Population counts are not optimizer-update exposure counts. Four sampling weights are 1/1/1/1 over admitted frame populations, not equal dataset probabilities. The saved DROID partial-to-full Flow episode map and the 108 RH20T Flow-only frame exclusions are included.

Download and install

Use a Python 3.10 training environment with the dependencies in the bundled training_code/requirements/train.txt. Install the appropriate Torch and Torchvision wheels for your GPUs first, following the VLA-Delta repository. The base ZR-0 model weights are a separate model download; this repository contains training data and code, not model checkpoints.

hf download t0rn0d0/vla-delta --repo-type dataset --local-dir /data/downloads/vla-delta
python /data/downloads/vla-delta/install.py \
  --source /data/downloads/vla-delta --root /data/vla-delta
export VLA_BASE_MODEL=/models/ZR-0
source /data/vla-delta/env.sh

The installer verifies every archive and extracted member against SHA256, extracts independent tar shards, and prepares local absolute paths. It uses streaming I/O. Allow approximately 2.2 TiB for the downloaded archives and a further 2.2 TiB for extraction. Archives can be removed after successful installation if no longer needed. Keep the original metadata templates for reinstallation or relocation. To move an extracted dataset, run:

python /data/downloads/vla-delta/install.py \
  --root /new/location/vla-delta --configure-only

This rebinds file inventories and metadata to the new location without regenerating any labels, indices or numeric statistics. Use the bundled code for the tested loader implementation; changing the training code is a separate compatibility decision.

Bootstrap a new training environment

If you already have a working VLA-Delta training environment, use the commands above directly. On a new machine, download and extract the small code archive first so the training dependency file is available before installing the data:

hf download t0rn0d0/vla-delta archives/training_code/part-00000.tar \
  --repo-type dataset --local-dir /data/downloads/vla-delta
mkdir -p /data/vla-delta
tar -xf /data/downloads/vla-delta/archives/training_code/part-00000.tar \
  -C /data/vla-delta
# In your Python 3.10 environment, install Torch/Torchvision for your GPU first.
python -m pip install -r /data/vla-delta/training_code/requirements/train.txt

Then run the full download and installer commands above in that training environment. The installer is data preparation only; it does not install GPU drivers, fetch the base model weights, or start training.

Check real samples and start training 

python "$VLA_REPO/scripts/verify_pretraining_release.py" \
  --root /data/vla-delta --base-model "$VLA_BASE_MODEL"
python "$VLA_REPO/scripts/train.py" \
  --config /data/vla-delta/configs/stage1_ar.json --dry-run

# Stage 1: 10,000 updates, AR
python "$VLA_REPO/scripts/train.py" \
  --config /data/vla-delta/configs/stage1_ar.json --execute
# Stage 2: 5,000 updates, Slot + 10 * Flow
python "$VLA_REPO/scripts/train.py" \
  --config /data/vla-delta/configs/stage2_aux.json --execute
# Stage 3: 150,000 updates, AR + 0.1 * Slot + Flow + 5 * FM, H50
python "$VLA_REPO/scripts/train.py" \
  --config /data/vla-delta/configs/stage3_joint.json --execute

Stage 2 loads Stage 1 model-final; Stage 3 loads Stage 2 model-final and initializes the Action Expert from the original ZR-0 base. Each stage starts its own optimizer and counter. The recipes use four GPUs, batch 16 per GPU, gradient accumulation 2 (global batch 128), BF16 and online W&B with wandb_failure_policy=required. Authenticate W&B before training. If using a different GPU count or batch size, explicitly update the launch options and expected global batch. No GPU training is started by the installer or verifier.

Layout and integrity

  • archives/: independent tar shards, approximately 8 GiB each or smaller.
  • manifest.json: exact member paths, sizes, SHA256 and archive identities.
  • SHA256SUMS: archive checksums.
  • release.json: source releases and training code version.
  • verification/: CPU data loading and relocation verification reports.
  • READY.json: release readiness and verification receipt.
  • Extracted data/pretrain/: original Parquet/video trajectories and Flow HDF5.
  • Extracted data/annotations/: exactly the source JSONL files referenced by Slot audits.
  • Extracted templates/: unchanged formal preparation.
  • Extracted data/prepared/ and configs/: generated local training artifacts.
  • Extracted training_code/: the training code snapshot used for this release.

The tar layout is for full training reproducibility. Use hf download or snapshot_download and the installer; this is not a single homogeneous datasets.load_dataset table. This release supports fresh training; it does not claim exact optimizer-state resume compatibility with historical runs.

Source datasets

Credit and upstream dataset terms remain applicable to their respective data. This collection does not replace upstream licenses with one permissive license.

  • DROID dataset team: https://droid-dataset.github.io/ . Local source version: droid_1.0.1_stage05_full_95658_20260831. Original local README is preserved in the extracted provenance/upstream/ directory.
  • Ai2 MolmoAct: https://huggingface.co/datasets/allenai/MolmoAct-Dataset . Local sources: molmoact_dataset_household-v3_stage05 and molmoact_dataset_tabletop-v3_stage05; upstream dataset card declares CC BY 4.0.
  • RH20T team, Shanghai Jiao Tong University: https://rh20t.github.io/ . Local source: RH20T-v30_stage05. The upstream site distinguishes RH20T subsets and their license conditions; consult the linked original terms for the underlying subset before reuse.
  • VLA-Delta processing and training: https://github.com/Baixiaochi9449/vla-delta-new . Bundled code retains its LICENSE and THIRD_PARTY_NOTICES.md.
Downloads last month
1,286
Total file size:
2.4 TB