Salesforce/ProVision-10M

ProVision: Programmatically Scaling Vision-centric Instruction Data for Multimodal Language Models

ProVision is an extendable data generation engine which produces instruction data for large multimodal language models (MLMs).

In particular, it synthesizes instruction data via data generators (Python programs) and scene graphs rather than proprietary models. It also includes a scene graph generation pipeline consisting of various state-of-the-art models (eg, object detection model). Thus, one can generate instruction data for any given image by first generating the scene graph and then apply data generators.

Provision supports generation of both single-image and multi-image instruction data. One can also extend the engine by adding new data generators.

You are currently viewing the ProVision-10M dataset.

Dataset Details

Dataset Sources

• Repository: https://github.com/JieyuZ2/ProVision
• Paper: https://arxiv.org/abs/2412.07012
• Blog:
• Source Data: Visual Genome/GQA and DataComp

Uses

Users need to make their own assessment regarding any obligations or responsibilities under the corresponding licenses or terms and conditions pertaining to the original datasets and data. This repository is being released for research purposes only.

Direct Use

ProVision-10M is designed to facilitate research in training multimodal language models.

Out-of-Scope Use

ProVision-10M was built to make research into large multimodal models more accessible. Using the dataset to train models that ingest or generate personally identifying information (such as images of people’s faces and other sensitive content) as well as military applications are all inappropriate use cases of ProVision-10M.

Dataset Creation

Curation Rationale

ProVision-10M was created to demonstrate the potential of programmatically synthesizing instruction data for training multimodal language models.

Source Data

The dataset is built upon two data sources:

• we use 74,289 images and scene graphs from Visual Genome（the GQA version）
• we use 126,106 images from DataComp

Dataset summary

We do not release the images, please download the images from their original sources (GQA/DataComp)

Split	Size	Format	Description
vgs_sa	1537630	short answer	single-image instruction data based on Visual Genome
vgs_mc	1537630	multiple choice	single-image instruction data based on Visual Genome
vgm_sa_2_img	1400000	short answer	2-image instruction data based on Visual Genome
vgm_mc_2_img	1400000	multiple choice	2-image instruction data based on Visual Genome
vgm_sa_3_img	1400000	short answer	3-image instruction data based on Visual Genome
vgm_mc_3_img	1400000	multiple choice	3-image instruction data based on Visual Genome
vgm_sa_4_img	1400000	short answer	4-image instruction data based on Visual Genome
vgm_mc_4_img	1400000	multiple choice	4-image instruction data based on Visual Genome
dcs_sa	2294572	short answer	single-image instruction data based on DataComp images
dcs_mc	2294572	multiple choice	single-image instruction data based on DataComp images
dcm_sa_2_img	1400000	short answer	2-image instruction data based on DataComp images
dcm_mc_2_img	1400000	multiple choice	2-image instruction data based on DataComp images
dcm_sa_3_img	1400000	short answer	3-image instruction data based on DataComp images
dcm_mc_3_img	1400000	multiple choice	3-image instruction data based on DataComp images
dcm_sa_4_img	1400000	short answer	4-image instruction data based on DataComp images
dcm_mc_4_img	1400000	multiple choice	4-image instruction data based on DataComp images

License

We release ProVision-10M under a CC-BY-NC-4.0 license.

Citation

@article{zhang2024provision,
  title={ProVision: Programmatically Scaling Vision-centric Instruction Data for Multimodal Language Models},
  author={Zhang, Jieyu and Xue, Le and Song, Linxin and Wang, Jun and Huang, Weikai and Shu, Manli and Yan, An and Ma, Zixian and Niebles, Juan Carlos and Xiong, Caiming and others},
  journal={arXiv preprint arXiv:2412.07012},
  year={2024}
}