Datasets:
license: openrail
task_categories:
- question-answering
- image-to-text
size_categories:
- n<1K
configs:
- config_name: last
data_files: data/last/*
- config_name: previous
data_files: data/previous/*
language:
- en
tags:
- art
Dataset Card for But They Are Cats Tutorials
This dataset is presented and used in Level Up Your Tutorials: VLMs for Game Tutorials Quality Assessment.
Dataset Details
The dataset is designed for Visual Question answering. It is composed of game screenshots, questions, and answers. The questions and the answers are direct to provide a more effective evaluation independent of the syntax.
Question: "Do distractions affect the cats in the same way?" Answer: "The distraction affects the types of cats in different ways."
We provided two versions of the game after developers ran human testing. The dataset can be loaded:
# Latest Version
load_dataset("DarthReca/but-they-are-cats-tutorial", name="last")
# Previous Version
load_dataset("DarthReca/but-they-are-cats-tutorial", name="previous")
- Curated by: Daniele Rege Cambrin
- License: OPENRAIL
Uses
The dataset is intended to evaluate the quality of scene explanations from both textual and visual aspects. It is not designed to benchmark Visual-Language models since some scenes could be ambiguous or unclear.
Dataset Structure
The dataset has the following format:
{
image: Image,
question: List[Text]
answer: List[Text]
}
Dataset Creation
The dataset comprises videos, frames, and textual annotations in the form of questions and answers.
Source Data
Daniele Rege Cambrin records all the videos in both versions to grant a similar way of playing But They Are Cats. The frames are extracted from these videos when something "relevant" happens. The two versions are selected between all possible commits to balancing similarities and differences for a fairer evaluation.
Annotations
One of the developers (Daniele Rege Cambrin) provides the annotation to ensure maximum adherence to the actual developer's expectations. The same person also does the frame selection.
Bias, Risks, and Limitations
A single user creates the annotations and could be biased towards his vision of the game. The game itself does not represent all possible scenarios in game development, and careful evaluation should be performed in other scenarios. The game is an unfinished prototype and should be considered when compared to a finished product. The recorded tutorials do not cover all possible user interactions and should not be used for bug searching.
Citation
BibTeX:
@misc{cambrin2024leveltutorialsvlmsgame,
title={Level Up Your Tutorials: VLMs for Game Tutorials Quality Assessment},
author={Daniele Rege Cambrin and Gabriele Scaffidi Militone and Luca Colomba and Giovanni Malnati and Daniele Apiletti and Paolo Garza},
year={2024},
eprint={2408.08396},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2408.08396},
}