license: apache-2.0
language:
- en
tags:
- multi-agent
- multi-modal
- benchmark
- minecraft
size_categories:
- 10K<n<100K
dataset_info:
features:
- name: task_variants
description: 55,000 procedurally generated task variants.
- name: demonstrations
description: Multi-modal prompts (language + images), observations, and actions.
- name: observations
description: First-person RGB views and inventory states.
- name: agent_actions
description: Time-step actions for each agent.
splits:
- name: train
size: '>55,000'
- name: valid
size: '1000'
Dataset Card for TeamCraft-Data-DEC
The TeamCraft dataset is designed to develop multi-modal, multi-agent collaboration in Minecraft. It features 55,000 task variants defined by multi-modal prompts and procedurally generated expert demonstrations.
TeamCraft-Data-Dec is the training split for TeamCraft decentralized models, such as TeamCraft-VLA-7B-Dec. In this setup, the model has access to first-person RGB views and inventory information for one agent and generates actions for that agent.
The dataset has a size of approximately 127GB, comprising over one million images and a comprehensive training JSON file. It contains 55,000 unique task variants, each accompanied by a demonstration.
Structure
Input:
- Multi-Modal Prompts: Language instructions interleaved with orthographic view images (top, left, and front) to specify tasks.
- Observations: First-person RGB views and agent inventories for one agent.
- History: Historical actions for one agent (if available).
Output:
- Action Space: High-level actions for one agent.
Tasks
The TeamCraft dataset includes demonstrations for the following tasks:
For detailed task descriptions and task visualization videos, visit the Task Documentation.
- Building: Construct structures from provided blueprints.
- Clearing: Remove specified blocks from target regions.
- Farming: Sow and harvest crops on designated farmland.
- Smelting: Use furnaces to smelt resources into goal items.
Applications
This dataset supports research in:
- Multi-agent reinforcement learning.
- Multi-modal learning and task planning.
- Embodied AI in collaborative environments.
Usage
Model training and evaluation available on TeamCraft GitHub Repository.
Disclaimer
This dataset is provided for research purposes only. Users are responsible for ensuring compliance with applicable laws and regulations. The creators do not guarantee the dataset's suitability for any specific purpose and are not liable for any consequences arising from its use.
License
This dataset is released under the Apache License 2.0.