Datasetpublic

garrykuwanto/wc-vqa-continual-pretraining

garrykuwanto/wc-vqa-continual-pretraining English subset of worldcuisines/vqa-v1.1 Task 1 (dish name prediction), packaged for continual pretraining / SFT of small vision-language models (specifically SmolVLM2-256M). Splits split rows source train 27000 task1 / train / lang=en validation 300 task1 / test_small / lang=en test 1500 task1 / test_large / lang=en Schema field type notes image Image bytes from upstream… See the full description on the dataset page: https://huggingface.co/datasets/garrykuwanto/wc-vqa-continual-pretraining.

sourceHugging Facecc-by-sa-4.0updated 6mo agoView on Hugging Face

0likes110downloads

Dataset Card

garrykuwanto/wc-vqa-continual-pretraining

English subset of worldcuisines/vqa-v1.1 Task 1 (dish name prediction), packaged for continual pretraining / SFT of small vision-language models (specifically SmolVLM2-256M).

Splits

split	rows	source
train	27000	task1 / train / lang=en
validation	300	task1 / test_small / lang=en
test	1500	task1 / test_large / lang=en

Schema

field	type	notes
`image`	Image	bytes from upstream `images.tar.gz` (Wikimedia orig.)
`question`	string	open-ended prompt asking for the dish name
`options`	list[string]	5 multiple-choice options
`mcq_answer_index`	int32	0-indexed correct option within `options`
`answer`	string	ground-truth dish name
`prompt_type`	int32	1 = no-context, 2 = contextualized, 3 = adversarial
`food_id`	string	upstream dish id
`qa_id`	string	upstream question id

Raw fields are intentionally unformatted. Your training script decides whether to render them as ShareGPT, ChatML, or plain text.

Attribution

Built from worldcuisines/vqa-v1.1 (CC-BY-SA 4.0). Images come from Wikimedia Commons; see the upstream dataset for per-image attribution via image_url. This derivative dataset is released under the same CC-BY-SA 4.0 license.