deb-kit2/grit-20m-wds-offline

GRIT (Grounded Image-Text) — 20M (WebDataset Format)

This repository provides pre-downloaded, ready-to-use WebDataset shards (.tar) for the GRIT-20M (Grounded Image-Text) dataset introduced in Kosmos-2: Grounding Multimodal Large Language Models to the World.

The original zzliang/GRIT repository contains metadata in Parquet format, requiring users to run img2dataset to fetch images from the web. This repository provides pre-extracted, high-speed streaming WebDataset shards so that you can directly train or evaluate multimodal models without downloading and processing raw web images yourself.

---

Dataset Overview

• Total Size: ~440 GB
• Shards: 20 folders (coyo_0 to coyo_19), each containing 100 tar archives (00000.tar to 00099.tar) — 2,000 shards in total.
• Estimated Samples: ~12 million grounded image-text pairs.
• Format: WebDataset (.tar archives containing matching .jpg, .json, and .txt files).
• Data Completeness: Around 41% of image links were broken/inaccessible at download time, so this collection comprises the remaining ~59% of the total dataset (~12M successfully retrieved pairs).

Data Availability Notice: Web-scale image datasets experience link decay over time. As of this download, around 41% of the original image URLs in the GRIT COYO metadata were no longer reachable. These archives preserve the complete set of successfully retrieved images, constituting approximately 59% of the total original pairs.

---

Directory Structure

.
├── coyo_0/
│   ├── 00000.tar
│   ├── 00001.tar
│   └── ... (up to 00099.tar)
├── coyo_1/
│   ├── 00000.tar
│   └── ... (up to 00099.tar)
...
└── coyo_19/
    ├── 00000.tar
    └── ... (up to 00099.tar)

---

File Contents & Sample Format

Each .tar shard contains triplets of files sharing a common key:

• {key}.jpg: The image file.
• {key}.txt: The raw text caption.
• {key}.json: Metadata including ground-truth bounding boxes, noun phrases, and referring expressions.

Sample Metadata ({key}.json)

{
  "id": 8512625193209,
  "caption": "Plot to build a villa up to 400m2 overlooking the countryside",
  "url": "https://cdn.algarvepropertysearch.com/maxD_800/APN/APN_6.jpg",
  "key": "000000007",
  "status": "success",
  "error_message": null,
  "width": 337,
  "height": 224,
  "original_width": 800,
  "original_height": 532,
  "sha256": "69b398cc4723cec05b5850cddec2090dc246ffcb8b4489b2d14c6a1ea536d0b6",
  "noun_chunks": [
    [46.0, 61.0, 0.00796, 0.28274, 0.99352, 0.61602, 0.70963],
    [0.0, 4.0, 0.49198, 0.33822, 0.99794, 0.86645, 0.65381]
  ],
  "ref_exps": [
    [46.0, 61.0, 0.00796, 0.28274, 0.99352, 0.61602, 0.70963],
    [0.0, 61.0, 0.49198, 0.33822, 0.99794, 0.86645, 0.65381]
  ]
}

Field Definitions

• id: Unique 64-bit identifier from COYO-700M.
• caption: Text caption associated with the image.
• noun_chunks: Grounded noun phrases identified via spaCy and localized by GLIP. Each item is a list of 7 values: [start_char, end_char, norm_xmin, norm_ymin, norm_xmax, norm_ymax, confidence_score]
  • start_char, end_char: Character start and end indices inside caption.
  • norm_xmin, norm_ymin, norm_xmax, norm_ymax: Normalized bounding box coordinates in the range $[0.0, 1.0]$.
  • confidence_score: Detection confidence from GLIP.
• ref_exps: Grounded referring expressions (expanded noun phrases linked to bounding regions).
• width / height: Dimensions of the stored .jpg image.
• original_width / original_height: Original dimensions prior to downloading.

---

How to Use

1. Installation

pip install webdataset torch torchvision pillow

2. Reading Shards Locally

import webdataset as wds

# Pattern matching all shards across the 20 folders
shard_pattern = "coyo_{0..19}/{00000..00099}.tar"

dataset = (
    wds.WebDataset(shard_pattern, resampled=True)
    .shuffle(1000)
    .decode("pil")
    .to_tuple("jpg", "json", "txt")
)

for image, meta, caption in dataset:
    print("Caption:", caption)
    print("Noun Chunks:", meta["noun_chunks"])
    print("Referring Expressions:", meta["ref_exps"])
    print("Image Size:", image.size)
    break

3. Streaming Directly from Hugging Face

You can stream shards directly over HTTP/HTTPS without downloading the entire dataset:

import webdataset as wds
from huggingface_hub import hf_hub_url

REPO_ID = "YOUR_USERNAME/YOUR_DATASET_NAME"

# Build URLs for the desired shards
urls = [
    hf_hub_url(repo_id=REPO_ID, repo_type="dataset", filename=f"coyo_{i}/{j:05d}.tar")
    for i in range(20)
    for j in range(100)
]

dataset = (
    wds.WebDataset(urls)
    .decode("pil")
    .to_tuple("jpg", "json", "txt")
)

for image, meta, caption in dataset:
    print("Streamed Caption:", caption)
    break

4. PyTorch DataLoader Integration

import torch
import webdataset as wds
from torchvision import transforms

transform = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])

def preprocess(sample):
    image, meta, caption = sample
    return transform(image), meta, caption

dataset = (
    wds.WebDataset("coyo_{0..19}/{00000..00099}.tar", resampled=True)
    .shuffle(5000)
    .decode("pil")
    .to_tuple("jpg", "json", "txt")
    .map(preprocess)
)

loader = wds.WebLoader(dataset, batch_size=32, num_workers=4)
for batch_images, batch_meta, batch_captions in loader:
    # batch_images is shape [32, 3, 224, 224]
    break

---

Citation & Acknowledgments

If you use this dataset, please cite the Kosmos-2 paper and the COYO-700M dataset:

@article{Kosmos2,
  title={Kosmos-2: Grounding Multimodal Large Language Models to the World},
  author={Zhiliang Peng and Wenhui Wang and Li Dong and Yaru Hao and Shaohan Huang and Shuming Ma and Furu Wei},
  journal={arXiv preprint arXiv:2306.14824},
  year={2023}
}

@misc{kakaobrain2022coyo-700m,
  title         = {COYO-700M: Image-Text Counterpart of LAION-2B},
  author        = {Minwoo Byeon and Beomhee Park and Haecheon Kim and Sungjun Lee and Wonkwang Baek and Minji Kim},
  year          = {2022},
  publisher     = {GitHub},
  journal       = {GitHub repository},
  howpublished  = {\url{https://github.com/kakaobrain/coyo-dataset}}
}