metadata
license: mit
language:
- en
size_categories:
- 1M<n<10M
CALVIN Dataset - task_ABCD_D (Structured Subsets)
This repository contains the CALVIN task_ABCD_D dataset split into structured subsets for easier downloading and processing.
Original Source
- URL: http://calvin.cs.uni-freiburg.de/dataset/task_ABCD_D.zip
- Original Size: ~600GB
Structure
Each subset is a complete, self-contained dataset with the proper structure:
subset_training_000/
└── training/
├── scene_info.npy
├── lang_annotations/
│ └── auto_lang_ann.npy
├── ep_lens.npy
├── ep_start_end_ids.npy
├── episode_XXXXXXX.npz
└── ...
This structure is compatible with the CALVIN processing pipeline.
Training Subsets (24 total)
| Subset | Episodes | Size |
|---|---|---|
| subset_training_000 | 100000 (episode_0037682.npz - episode_0153816.npz) | 27.19 GB |
| subset_training_001 | 100000 (episode_0153817.npz - episode_0278465.npz) | 27.47 GB |
| subset_training_002 | 100000 (episode_0278466.npz - episode_0378465.npz) | 26.85 GB |
| subset_training_003 | 100000 (episode_0378466.npz - episode_0499017.npz) | 26.82 GB |
| subset_training_004 | 100000 (episode_0499018.npz - episode_0599017.npz) | 26.96 GB |
| subset_training_005 | 100000 (episode_0599018.npz - episode_0699017.npz) | 27.63 GB |
| subset_training_006 | 100000 (episode_0699018.npz - episode_0799017.npz) | 27.85 GB |
| subset_training_007 | 100000 (episode_0799018.npz - episode_0899017.npz) | 27.91 GB |
| subset_training_008 | 100000 (episode_0899018.npz - episode_0999017.npz) | 27.65 GB |
| subset_training_009 | 100000 (episode_0999018.npz - episode_1099017.npz) | 27.93 GB |
| subset_training_010 | 100000 (episode_1099018.npz - episode_1199017.npz) | 27.78 GB |
| subset_training_011 | 100000 (episode_1199018.npz - episode_1299017.npz) | 27.04 GB |
| subset_training_012 | 100000 (episode_1299018.npz - episode_1399017.npz) | 27.32 GB |
| subset_training_013 | 100000 (episode_1399018.npz - episode_1499017.npz) | 27.74 GB |
| subset_training_014 | 100000 (episode_1499018.npz - episode_1599017.npz) | 27.82 GB |
| subset_training_015 | 100000 (episode_1599018.npz - episode_1699017.npz) | 28.17 GB |
| subset_training_016 | 100000 (episode_1699018.npz - episode_1799017.npz) | 28.24 GB |
| subset_training_017 | 100000 (episode_1799018.npz - episode_1899017.npz) | 26.27 GB |
| subset_training_018 | 100000 (episode_1899018.npz - episode_1999017.npz) | 25.78 GB |
| subset_training_019 | 100000 (episode_1999018.npz - episode_2099017.npz) | 26.02 GB |
| subset_training_020 | 100000 (episode_2099018.npz - episode_2199017.npz) | 27.08 GB |
| subset_training_021 | 100000 (episode_2199018.npz - episode_2299017.npz) | 26.93 GB |
| subset_training_022 | 100000 (episode_2299018.npz - episode_2399017.npz) | 26.86 GB |
| subset_training_023 | 7126 (episode_2399018.npz - episode_2406143.npz) | 2.01 GB |
Validation Subsets (1 total)
| Subset | Episodes | Size |
|---|---|---|
| subset_validation_000 | 99022 (episode_0000000.npz - episode_0420498.npz) | 26.66 GB |
How to Use
Download a specific subset:
# Using huggingface-cli
huggingface-cli download VyoJ/calvin-ABCD-D-subsets training/subset_training_000.zip --local-dir ./
# Or using Python
from huggingface_hub import hf_hub_download
hf_hub_download(
repo_id="VyoJ/calvin-ABCD-D-subsets",
filename="training/subset_training_000.zip",
repo_type="dataset",
local_dir="./"
)
Extract and process:
cd training
unzip subset_training_000.zip
# Now you have subset_training_000/training/ with all needed files
Process with CALVIN pipeline:
Point your pipeline to the subset directory (e.g., subset_training_000/) and it will work as if processing the full dataset.
Reassembling Full Dataset
If you want to reassemble the full dataset:
- Download all subsets for a split
- Extract each subset
- Merge episode files into a single directory
import shutil
from pathlib import Path
# After extracting all subsets
output_dir = Path("full_training")
output_dir.mkdir(exist_ok=True)
# Copy metadata from first subset
first_subset = Path("subset_training_000/training")
shutil.copy(first_subset / "scene_info.npy", output_dir)
shutil.copytree(first_subset / "lang_annotations", output_dir / "lang_annotations")
# Copy all episodes from all subsets
for subset_dir in sorted(Path(".").glob("subset_training_*/training")):
for ep_file in subset_dir.glob("episode_*.npz"):
shutil.copy(ep_file, output_dir)