Visually Dehallucinative Instruction Generation: Know What You Don't Know [paper]
Sungguk Cha, Jusung Lee, Younghyun Lee and Cheoljong Yang
This work proposes I Know Hallucination wherein hallucination-inducing question comes and the desired answer is "I Don't Know."
Qualitative results on hallucination inducing-questions.
Prepare COCO images in advance.
IDK-Instructions uses "<image>" for the image token.
{
"image": "MSCOCO/images/train2014/COCO_train2014_000000393224.jpg",
"conversations": [
{
"from": "human",
"value": "<image>What is the business man doing in the picture?"
},
{
"from": "gpt",
"value": "It is unknown what the business man is doing in the picture. It could be waiting, standing, walking or frowning."
},
{
"from": "human",
"value": "What is the business man doing in the picture?"
},
{
"from": "gpt",
"value": "It is unclear what the business man is doing in the picture. He could be waiting, standing, walking, or even frowning or pouting."
}
]
},
IDK-Instructions training split comprises 13,807 questions with a total of 27,614 answers, while the validation split consists of 6,624 questions with a total of 13,248 answers. Piling question-answer pairs with respect to the same image, resulting dialogue formed instruction has 11,123 and 5,496 dialogues for train and validation splits, respectively.
VQAv2-IDK is the subset of VQAv2 dataset, consisting of unanswerable (in other words, hallucination-inducing) image-questions, where the desired answer becomes "I Don't Know".
If you find it useful for your research and applications, please cite using this BibTeX:
@inproceedings{cha2024visually,
title={Visually Dehallucinative Instruction Generation: Know What You Don't Know},
author={Cha, Sungguk and Lee, Jusung and Lee, Younghyun and Yang, Cheoljong},
year={2024},
}
This work used VQAv2 dataset (CC BY 4.0 DEED license) for the question-answer source and ChatGPT for IDK-Instructions generation (refer OpenAI policies, https://openai.com/policies).