README.md

OCR-VQA Images — Download, Reassemble, and Place into LibMoE

This guide shows how to download split archives from Hugging Face, merge them into a single ZIP, extract, and place the images into the LibMoE data tree:

libmoe/
└── data/
    ├── image_onevision/
    ├── coco/
    │   └── train2017/
    ├── gqa/
    │   └── images/
    ├── ocr_vqa/
    │   └── images/
    ├── textvqa/
    │   └── train_images/
    └── vg/
        ├── VG_100K/
        └── VG_100K_2/

Tested on Linux (bash). Requires ~100 GB free space for download + extraction room.

---

1) Download split parts from Hugging Face

Dataset: DavidNguyen/ocr_vqa, path: ocr_vqa/images_part_*.zip.part

Using curl directly (works anywhere)

#!/usr/bin/env bash
set -euo pipefail

DEST=./data/ocr_vqa/ocr_vqa_parts
mkdir -p "$DEST"
cd "$DEST"

# List of parts
PARTS=$(echo a{a..x})        # aa ab ac ... ax

# Download function
download_one() {
  p="$1"
  file="images_part_${p}.zip.part"
  url="https://huggingface.co/datasets/DavidNguyen/ocr_vqa/resolve/main/ocr_vqa/${file}?download=true"
  echo "[*] downloading $file"
  curl -L --fail --retry 5 --retry-delay 3 -o "$file" "$url"
}

export -f download_one

echo "$PARTS" | xargs -n1 -P8 bash -c 'download_one "$@"' _

---

2) Concatenate parts → a single ZIP

If you already downloaded to ./data/, the one-liner below (as requested) works:

mkdir ./data/ocr_vqa
cat ./data/ocr_vqa/images_part_*.zip.part \
  > ./data/ocr_vqa/images.zip

---

3) Unzip

cd ./data/ocr_vqa
unzip -q images.zip