UCIT data

This repo holds the complete UCIT continual instruction tuning benchmark: the instruction files and the images for all six tasks, in the directory layout the instruction JSONs expect. It is a mirror, and the original authors' licenses apply; the table below says which license covers what.

UCIT was introduced with HiDe-LLaVA (ACL 2025). Its six tasks run in this order: ImageNet-R, ArxivQA, VizWiz-Caption, IconQA, CLEVR-Math, Flickr30k.

HaiyangGuo/UCIT ships the instructions and two of the image sets. The other four image sets have to be fetched from four different hosts, and two archives need their top-level directory fixed before the instruction paths resolve. This repo has all of it already unpacked.

Contents

path	what it is	source	license
instructions/<task>/	train and test JSONs; val_coco_type_3000.json for VizWiz and Flickr30k are the COCO-format caption references for CIDEr	HaiyangGuo/UCIT	CC BY 4.0
datasets/ImageNet-R/	ImageNet-R images	imagenetr.tar.gz from HaiyangGuo/UCIT	as distributed there
datasets/ArxivQA/images/	figure images	images.tgz from MMInstruction/ArxivQA	CC BY-SA 4.0
datasets/VizWiz/{train,val}/	VizWiz images	vizwiz.cs.colorado.edu	VizWiz terms
datasets/IconQA/iconqa_data/	IconQA images and metadata	iconqa_data.zip from the IconQA S3 bucket	LICENSE.md in the zip states CC BY-NC-SA 4.0
datasets/CLEVR/	the full CLEVR v1.0 release: images, questions, scenes	CLEVR_v1.0.zip from dali-does/clevr-math	CC BY 4.0 (COPYRIGHT.txt)
datasets/Flickr30k/{train,val}/	Flickr30k images	train.tar.gz and val.tar.gz from HaiyangGuo/UCIT	Flickr30k terms

Three things differ from the upstream archives. The CLEVR zip's top directory CLEVR_v1.0/ is renamed to CLEVR/, the path the CLEVR-Math instructions use. The Flickr30k tarballs were packed from an absolute path (mnt/ShareDB_1TB/datasets/flickr30k/) and are unwrapped here. The __MACOSX entries in the VizWiz instruction zip are removed.

Images are stored as uncompressed tar shards of up to 4 GiB (shard-NNNNN.tar), because the Hub limits how many files a repo can hold. Tar member paths are relative to the repo root. MANIFEST.tsv.gz lists every file with its size and the shard that holds it.

Restore

pip install -U huggingface_hub
hf download --repo-type dataset vantuan5644/UCIT restore.sh --local-dir .
bash restore.sh vantuan5644/UCIT /path/to/datasets/UCIT

restore.sh downloads the repo, extracts every shard in place, deletes the tars (set KEEP_TARS=1 to keep them), and checks each file in MANIFEST.tsv.gz for presence and size. The instruction JSONs reference images relative to datasets/, for example CLEVR/images/train/CLEVR_train_000000.png.

Citation

If you use the benchmark, cite HiDe-LLaVA, as well as the datasets linked above.

@article{guo2025hide,
  title={Hide-llava: Hierarchical decoupling for continual instruction tuning of multimodal large language model},
  author={Guo, Haiyang and Zeng, Fanhu and Xiang, Ziwei and Zhu, Fei and Wang, Da-Han and Zhang, Xu-Yao and Liu, Cheng-Lin},
  journal={arXiv preprint arXiv:2503.12941},
  year={2025}
}