LockOnN/MMCS-Data

MMCS 773K Data

This repository contains the image archives and generated annotations used by MMCS (Multi-Modal Code-Switching).

Contents

• annotations/: seven LLaVA-style JSON annotation files. Each file is a JSON array; it is not JSONL.
• images/: ZIP archives grouped by source dataset. Large sources are split into independent shards below 20 GB.

The release contains 773,779 records with 773,779 unique image paths. The ZIP entry paths exactly match the image fields in the annotations.

Source	Samples
COCO 2014	75,033
Flickr30K	29,140
GQA	108,516
Objects365 part 1	252,988
Objects365 part 2	89,758
Open Images	127,237
SA-1B	91,107
Total	773,779

Download and extraction

hf download LockOnN/MMCS-Data --repo-type dataset --local-dir MMCS-Data

mkdir -p data/images
for archive in MMCS-Data/images/*.zip; do
    unzip -q "$archive" -d data/images
done

After extraction, configure the annotation paths under annotations/ and use data/images as the image root.

Annotation format

Each sample contains an image path and LLaVA-style conversations, together with generated object annotations such as bounding boxes and segmentations. IDs are retained from the generated annotation sources and were not rewritten.

Release construction

The release first retains images shared with the MMCS 773K training recipe. Missing or duplicate old image keys are replaced by deterministic sampling from the corresponding larger 851K source, producing exactly 773,779 unique image paths. Construction details and sampled/missing image lists are included under annotations/.

Source data and licenses

The images originate from COCO, Flickr30K, GQA, Objects365, Open Images, and SA-1B. Users are responsible for complying with the licenses and terms of the corresponding source datasets. The generated MMCS annotations do not replace or modify those source terms.