| # images | # docs | # tokens | |
|---|---|---|---|
| Multimodal-C4 (mmc4) | 571M | 101.2M | 43B |
| Multimodal-C4 fewer-faces (mmc4-ff) | 375M | 77.7M | 33B |
| Multimodal-C4 core (mmc4-core) | 29.9M | 7.3M | 2.4B |
| Multimodal-C4 core fewer-faces (mmc4-core-ff) | 22.4M | 5.5M | 1.8B |
More details about these datasets and our processing steps can be found in our paper. (the current paper results describe v1 of the corpus, we will update to v1.1 soon).
You can directly download the "fewer faces" multimodal c4 documents at urls like this:
https://storage.googleapis.com/ai2-jackh-mmc4-public/data_v1.1/docs_no_face_shard_{$SHARD}_v2.jsonl.zip
where SHARD can vary from 0 to 23098. 14 shards are missing and are not included in the dataset.
You can download the smaller "core fewer faces" documents at URLs like this:
https://storage.googleapis.com/ai2-jackh-mmc4-public/data_core_v1.1/docs_no_face_shard_{$SHARD}_v3.jsonl.zip
where SHARD can vary from 0 to 23098. The total size of all these files together is approximately 9.4GB.
You can also automatically download & unzip these files from commands, you can run the script by providing the destination folder as an argument, like:
sh download_scripts/fewer_facesv2.sh /path/to/destination/folder
sh download_scripts/fewer_faces_corev3.sh /path/to/destination/folder
Documents in both sets contain text, image URLs, assignments of images to sentences, and image-by-text CLIP ViT-L/14 similarity matrices. Specifically:
text_list: a list of sentences comprising the text of the documenturl: the original url where the document was hostedimage_infois a key mapping to a list of images. each image contains:image_name: a filename that you could download the image toface_detections:Noneif no faces are detected (which should be the case in "fewer faces")matched_text_index: the index withintext_listrepresenting the sentence that this image is matched tomatched_sim: the CLIP ViT-L/14 similarity between the image and the sentence at the matched index
similarity_matrix: a matrix of shapelen(image_info) x len(text_list)wheresimilarity_matrix[i, j]is the CLIP ViT-L/14 similarity between imageiand sentencej.could_have_url_duplicate: a small number of URLs (~3%) in the corpus may have duplicate entries due to commoncrawl collecting multiple snapshots over time. we downsample such that, in expectation, each URL occurs once, but duplicates are technically possible. You can discard all entries withcould_have_url_duplicateequal to 1 if you want a more strictly deduplicated set.
Here's an example:
{'image_info': [{'face_detections': None,
'image_name': 'b9040a0dbb22.jpg',
'matched_sim': 0.27694183588027954,
'matched_text_index': 2,
'raw_url': 'http://www.hfitinfo.com/honda_fit_pics/3/2/index.90.jpg'},
{'face_detections': None,
'image_name': 'db1c21bc8474.jpg',
'matched_sim': 0.3234919607639313,
'matched_text_index': 1,
'raw_url': 'http://www.hfitinfo.com/honda_fit_pics/3/2/index.91.jpg'}],
'similarity_matrix': [[0.24363446235656738,
0.31758785247802734,
0.27694183588027954],
[0.2233106791973114,
0.3234919607639313,
0.26118797063827515]],
'text_list': ['When you lock the door using the lock tab on the driver’s '
'door, all of the other doors and tailgate lock at the same '
'time.',
'Press the master door lock switch in as shown to lock or '
'unlock all doors and the tailgate.',
'When you lock/unlock the driver’s door and tailgate using the '
'master lock switch, all the other doors lock/ unlock at the '
'same time.'],
'url': 'http://www.hfitinfo.com/hofi-48.html',
'could_have_url_duplicate': 0 }
The assignments of images to sentences are computed using compute_assignments.py
You can directly download CLIP ViT-L/14 features extracted from the images at urls like this:
https://storage.googleapis.com/ai2-jackh-mmc4-public/images/clip_vitl14_shard_{$SHARD}_features.pkl
where SHARD can vary from 0 to 23098. The total size of all the image feature files together is approximately 1.8Tb. Each pkl file is a dictionary that maps from image filename (accessible in the document jsons, see image_name above) to the associated CLIP feature. We used a jax port of CLIP to extract features on TPU. As a result, there may be some numerical differences with CPU or GPU versions of features. We have found that differences are relatively small in practice.
If you are interested in accessing mmc4 (and mmc4-core) without the fewer faces restriction, please fill out this form.
We are not releasing raw images for now. But if you are interested in potential updates, you can contact us using this google form.
.1% of the 23099 shards are missing from the corpus. These were not included in any statistics or experiments, so they are not part of mmc4. The missing shards are:
3218,3267,5064,5146,7119,8991,9750,11899,15127,15252,16996,17369,17499,17818
- the new contributions of mmc4 beyond text-only c4 (e.g., the similarity matrices/image-text alignments) are released under ODC-BY.
- By using mmc4, be aware of that you are also bound by the Common Crawl terms of use.
If you found our work useful, please consider citing:
@article{zhu2023multimodal,
title={{Multimodal C4}: An Open, Billion-scale Corpus of Images Interleaved With Text},
author={Wanrong Zhu and Jack Hessel and Anas Awadalla and Samir Yitzhak Gadre and Jesse Dodge and Alex Fang and Youngjae Yu and Ludwig Schmidt and William Yang Wang and Yejin Choi},
journal={arXiv preprint arXiv:2304.06939},
year={2023}
}