yyyyifan/VLQA

Introduction

This dataset is used to evaluate how Large Vision-Language Models (LVLMs) perform the visual question-answering tasks: understanding the input image or halluciating with the knowledge in the language module.

We provide a testmini set that includes all types of evaluations. More details can be found below.

---

• testmini: a subset of all types of evaluation.

• Entity_KF_NR: a QA dataset to evaluate the model's recognition ability of entities under the knowledge-free setting.

• Entity_KF_NC: a QA dataset to evaluate the model's counting ability of entities under the knowledge-free setting.

• Entity_KR_NR: a QA dataset to evaluate the model's recognition ability of entities under the knowledge-required setting.

• Entity_KR_NC: a QA dataset to evaluate the model's counting ability of entities under the knowledge-required setting.

• Relation_KF_NR: a QA dataset to evaluate the model's recognition ability of relations under the knowledge-free setting.

• Relation_KF_NC: a QA dataset to evaluate the model's counting ability of relations under the knowledge-free setting.

• Relation_KR_NR: a QA dataset to evaluate the model's recognition ability of relations under the knowledge-required setting.

• Relation_KR_NC: a QA dataset to evaluate the model's counting ability of relations under the knowledge-required setting.

• syn_icon: a QA dataset to evaluate the model's recognition and counting ability of entities that are represented by visual icons.

• syn_text: a QA dataset to evaluate the model's recognition and counting ability of entities that are represented by text labels.

• syn_pos: a QA dataset to evaluate the model's recognition and counting ability of entity positions.

• syn_rel: a QA dataset to evaluate the model's recognition and counting ability of relations.

• syn_rel_complex: a QA dataset to evaluate the model's recognition and counting ability of relations that are represented in multiple forms.

• syn_rel_know: a QA dataset to evaluate the model's recognition and counting ability of relations that are consistent with commonsense knowledge.

• syn_rel_pos: a QA dataset to evaluate the model's recognition and counting ability of relation positions.

---

Citation

If you use this dataset, please cite our paper:

@article{DBLP:journals/corr/abs-2410-00193,
  author       = {Buse Giledereli and
                  Yifan Hou and
                  Yilei Tu and
                  Mrinmaya Sachan},
  title        = {Do Vision-Language Models Really Understand Visual Language?},
  journal      = {CoRR},
  volume       = {abs/2410.00193},
  year         = {2024},
  url          = {https://doi.org/10.48550/arXiv.2410.00193},
  doi          = {10.48550/ARXIV.2410.00193},
  eprinttype    = {arXiv},
  eprint       = {2410.00193},
  timestamp    = {Tue, 05 Nov 2024 20:39:06 +0100},
  biburl       = {https://dblp.org/rec/journals/corr/abs-2410-00193.bib},
  bibsource    = {dblp computer science bibliography, https://dblp.org}
}