CoIN-CL data

This repo mirrors the CoIN continual instruction tuning benchmark: the instruction files, plus the image sources that are hard to download again. One command unpacks it into the directory layout the instruction JSONs expect. The original authors' licenses apply to every file; the table below says which license covers what.

CoIN (Chen et al., 2024) trains a multimodal model on eight tasks in this order: ScienceQA, TextVQA, ImageNet, GQA, VizWiz, Grounding (RefCOCO, RefCOCO+, RefCOCOg), VQAv2, OCR-VQA.

Contents

path	what it is	source	license
playground/Instructions_{Original,Diverse,10Type,Qwen}/	CoIN instruction JSONs for all eight tasks	Zacks-Chen/CoIN	CC BY 4.0
playground/Instructions_Original/TextVQA/TextVQA_0.5.1_val.json	TextVQA validation annotations, read by the TextVQA scorer	TextVQA	TextVQA terms
GQA/eval/	GQA question files, read by the GQA scorer	questions1.2.zip from GQA	GQA terms
images/OCR-VQA/	book cover images, the dataset.json catalog, loadDataset.py	images fetched from the Amazon URLs listed in the OCR-VQA catalog on Google Drive	LICENCE.txt, shipped with the catalog
images/ScienceQA/	question images, problems.json, pid_splits.json	images from derek-thomas/ScienceQA, JSONs from lupantech/ScienceQA	CC BY-SA 4.0
images/refcoco/, images/refcocog/, images/refcoco_plus/	referring expression annotations	Internet Archive captures (April 2022) of bvisionweb1.cs.unc.edu, which no longer resolves	see lichengunc/refer

Many OCR-VQA image URLs no longer work, which is the main reason this copy exists. images/OCR-VQA/_failed_urls.json lists the URLs that failed when this copy was built.

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.

Files added on top of upstream CoIN

These are not in Zacks-Chen/CoIN. They are fixed-seed evaluation subsets and a joint training set, made for the coin-cl training code:

• playground/Instructions_Original/ImageNet/test_subset4600_seed42.json
• playground/Instructions_Original/Grounding/test_subset21616_seed42.json
• playground/Instructions_Original/VQAv2/val_subset44793_seed42.json
• playground/Instructions_Original/OCRVQA/test_subset20797_seed42.json
• playground/Instructions_Original/Joint/train_all8.json, the eight training sets concatenated

Not included

Four image sources download reliably from their original hosts. Unzip each archive into the directory shown, relative to the data root; that reproduces the paths the instruction JSONs reference.

directory	archives
images/COCO2014/	train2014, val2014, test2015
images/GQA/	images.zip
images/TextVQA/	train_val_images.zip, test_images.zip
images/VizWiz/	train, val, test

The ImageNet task reads images/ImageNet_withlabel/train/<synset>/ for 101 of the 1000 ImageNet-1k synsets, and images/ImageNet_withlabel/val/ holding the ILSVRC2012 validation images under their original file names. The ImageNet terms of access do not allow redistribution, so these images are not here. Build both directories from ImageNet-1k (image-net.org or the gated timm/imagenet-1k-wds); the train synsets you need are the ones playground/Instructions_Original/ImageNet/train.json references.

Restore

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

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.

Citation

If you use the benchmark, cite CoIN, as well as the datasets its tasks are built from, which are linked above.

@misc{chen2024coin,
    title={CoIN: A Benchmark of Continual Instruction tuNing for Multimodel Large Language Model},
    author={Cheng Chen and Junchen Zhu and Xu Luo and Hengtao Shen and Lianli Gao and Jingkuan Song},
    year={2024},
    eprint={2403.08350},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}