A private subset of zzliang/GRIT downloaded and re-encoded for the zeroshot-classification / woptim project. Intended solely for research collaboration — do not redistribute.

What's inside

Images: 13.4M webp images, short-side resized to 256 (quality 80). Original resolution ≤ 256 on short side kept as-is.
Grounded annotations: per-image caption, noun_chunks (span → bbox), ref_exps (span → bbox), clip_similarity_vitb32 — all from the original GRIT dataset (Peng et al., Kosmos-2, 2023).

Schema

Two WebDataset directories, same format:

main/ — primary download (2,051 sub-shards, ~13.4M images).
retry/ — subsequent retry pass on transient download errors (~190 sub-shards, ~17K additional images).

Each sub-shard is a (tar, parquet) pair:

File	Contents
`{NNNNN}.tar`	WebDataset archive — per-sample `{key}.webp` image + `{key}.json` metadata (caption, id) + `{key}.txt` (caption text)
`{NNNNN}.parquet`	Per-URL row with columns: `id`, `url`, `caption`, `key`, `status`, `error_message`, `width`, `height`, `original_width`, `original_height`, `sha256`, `clip_similarity_vitb32`, plus two nested-list columns (`noun_chunks`, `ref_exps`) — note those serialize as `element` in the img2dataset output
`{NNNNN}_stats.json`	Per-sub-shard summary (success count, download failures, resize failures, elapsed)

Usage

import webdataset as wds

paths = [
    "/path/to/grit-woptim/main/*.tar",
    "/path/to/grit-woptim/retry/*.tar",
]
ds = wds.WebDataset(paths).decode("pil")
for sample in ds:
    img = sample["webp"]            # PIL image
    meta = sample["json"]           # dict
    key = sample["__key__"]
    break

To join with the full grounded annotations (noun_chunks, ref_exps) you can either:

Read the per-sub-shard parquet files and join on key.
Pull the original parquets from zzliang/GRIT and join on url / id.

Download details

Metric	Value
Source	`zzliang/GRIT` parquets (21 shards × ~1M URLs)
Total URLs	20,508,818
Successfully downloaded	13,433,274 (65.5%)
Additional via retry	~17K (0.13%)
Download tool	rom1504/img2dataset v1.47
Image size	256-short-side, keep-ratio-largest, resize-only-if-bigger
Encoding	webp quality 80
Download date	2026-04-15 → 2026-04-17
Total apparent size	~200 GB

License & copyright

Image URLs and grounded annotations come from zzliang/GRIT (MS-PL). Individual image files remain the copyright of their original web publishers. They are redistributed here under a private access grant solely for collaborative research. You must not:

redistribute images publicly
use images for commercial purposes
ignore any DMCA / takedown request from rights holders

If you are not a named collaborator on this project, close the repo.

Citation

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

Downloads last month: 24

Total file size:

215 GB

Paper for HongyouZhou/grit-woptim

Kosmos-2: Grounding Multimodal Large Language Models to the World

Paper • 2306.14824 • Published Jun 26, 2023 • 36