|
|
--- |
|
|
tags: |
|
|
- agent |
|
|
--- |
|
|
|
|
|
# Custom Dataset for Vision-Language-Action Model Training |
|
|
|
|
|
This dataset is designed for training and evaluating vision-language-action models in simulated gaming environments. It contains two main folders, `contractor` and `synthetic`, with video and action data collected from gameplay, allowing researchers and developers to experiment with visual perception, action recognition, and language-based command control in interactive settings. |
|
|
|
|
|
## Dataset Structure |
|
|
|
|
|
### Contractor Folder |
|
|
|
|
|
The `contractor` folder includes raw gameplay data in both video (`.mp4`) and action (`.jsonl`) formats: |
|
|
|
|
|
- **`mp4` Files**: These files capture video sequences from player interactions within the gaming environment, providing visual data for training and evaluation. |
|
|
- **`jsonl` Files**: Each video has a corresponding `.jsonl` file containing frame-by-frame action data performed by the player. The action data includes information on the player’s interactions, such as object manipulation and movement commands. |
|
|
|
|
|
#### Recent Update (24-11-12) |
|
|
On November 12, 2024, crafting-related data from the `contractor` data collection was uploaded to the dataset. This data includes video and action data associated with the crafting activities in the game environment, providing additional coverage of task-specific actions for enhanced model training. |
|
|
|
|
|
### Synthetic Folder |
|
|
|
|
|
*(Content description for the `synthetic` folder should be added here when ready.)* |
|
|
|
|
|
## Data Usage |
|
|
|
|
|
This dataset is designed for training models that require synchronized visual and action data, such as: |
|
|
|
|
|
- Vision-language-action model training |
|
|
- Agent action recognition and imitation learning |
|
|
- Multimodal reinforcement learning |
|
|
|
|
|
## Example Code to Load Dataset |
|
|
|
|
|
To load this dataset in Hugging Face Datasets: |
|
|
|
|
|
```python |
|
|
from datasets import load_dataset |
|
|
|
|
|
dataset = load_dataset("path_to_your_dataset") |
|
|
``` |
|
|
|
|
|
Replace `"path_to_your_dataset"` with the appropriate identifier or local path. Once loaded, you can access the video and action data in parallel, allowing for customized data processing and model training. |
|
|
|
|
|
|
