datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
coco-gpt4oTest
flickr30k-transformed-captionsThis is a "de-biased" version of https://huggingface.co/datasets/nlphuji/flickr30k dataset. We've added a few extra columns:
alt_text: the captions rewritten by calling the meta-llama/Meta-Llama-3-8B-Instruct LLM
grade: a measure of redability using the readability library
Learn more about why and how we did it here : https://github.com/mozilla/distilvit/blob/main/docs/fighting_bias.md
See the code here : https://github.com/mozilla/distilvit/blob/main/distilvit/curate.py
For the licence… See the full description on the dataset page: https://huggingface.co/datasets/Mozilla/flickr30k-transformed-captions.alt-text-validationThis dataset contains images and alt text from various sources.
It is used to control the quality of https://huggingface.co/Mozilla/distilvit using the https://github.com/mozilla/checkvite application
This application let users try out the model on the images and classify them. The dataset is then updated.
When an image is marked as need_training it will be use to fine-tune the model to fix some of its inaccuracies
flickr30k-transformed-captions-gpt4oThis is a "de-biased" version of https://huggingface.co/datasets/nlphuji/flickr30k dataset.
The new alt_text column was produced by GPT-4o using the following script : https://github.com/mozilla/distilvit/blob/main/distilvit/curate_gpt.py
Learn more about why and how we did it here : https://github.com/mozilla/distilvit/blob/main/docs/fighting_bias.md
See the code here : https://github.com/mozilla/distilvit/blob/main/distilvit/curate_gpt.py
For the licence, see the original dataset.
pexels-gpt4oImages collected from Pexels, using 1000 images following 5 categories:
nudes
war
group
animals
smoking
See https://www.pexels.com/license/ for the license
They were then annotated using gpt4-o, see https://github.com/mozilla/distilvit/blob/main/distilvit/gpt4.py
docornotThe DocOrNot dataset contains 50% of images that are pictures, and 50% that are documents.
It was built using 8k images from each one of these sources:
RVL CDIP (Small) - https://www.kaggle.com/datasets/uditamin/rvl-cdip-small - license: https://www.industrydocuments.ucsf.edu/help/copyright/
Flickr8k - https://www.kaggle.com/datasets/adityajn105/flickr8k - license: https://creativecommons.org/publicdomain/zero/1.0/
It can be used to train a model and classify an image as being a picture or a… See the full description on the dataset page: https://huggingface.co/datasets/Mozilla/docornot.osm-swimming-pools
osm-swimming-pools
Detect swimming_pools in satellite images.
Created with osm-ai-helper.
TRAIN_AREA: Galicia
VAL_AREA: Viana do Castelo
Ground Truth Bounding Boxes
Downloaded from OpenStreetMap. LICENSE: https://www.openstreetmap.org/copyright
Used the leisure=swimming_pool OpenStreetMap tags. Discarded the elements matching {'location': 'indoor'}.
Satellite Images
Downloaded from Mapbox. LICENSE:… See the full description on the dataset page: https://huggingface.co/datasets/mozilla-ai/osm-swimming-pools.