CEMIR-Project/HandMeThat-V2-CEMIR

HandMeThat V2 — CEMIR Subset

This dataset is a subset of the HandMeThat V2 dataset, prepared for the CEMIR project.

In addition to the original data, this version includes:

• Enhanced textual descriptions for each episode.
• Additional images for each action in each episode.
• The original episode-level data is preserved in raw_file.json.

Dataset Structure

Each episode is stored in its own directory:

dataset/
├── task-bring_me-2-1013-1/
│   ├── enriched_text.json
│   ├── raw_file.json
│   └── pics/
│       ├── 1.png
│       ├── 2.png
│       └── ...
├── task-bring_me-2-1267-4/
│   ├── enriched_text.json
│   ├── raw_file.json
│   └── pics/
│       ├── 1.png
│       └── ...
└── ...

raw_file.json contains the original episode data, while enriched_text.json contains the additional textual information introduced for CEMIR. The pics directory contains the additional images corresponding to the actions in the episode.

Usage

The dataset can be downloaded from the Hugging Face Hub and accessed as a regular directory:

from pathlib import Path
import json

dataset_path = Path("/path/to/HandMeThat-V2-CEMIR")

episode = dataset_path / "task-bring_me-2-1013-1"

with open(episode / "raw_file.json") as f:
    raw_data = json.load(f)

with open(episode / "enriched_text.json") as f:
    enriched_text = json.load(f)

images = sorted((episode / "pics").glob("*.png"))

print(raw_data)
print(enriched_text)
print(f"Number of images: {len(images)}")

Acknowledgement

This dataset is derived from HandMeThat V2. Please refer to the original dataset and its authors for the original data and licensing information.

This version was prepared as part of the CEMIR project.