| | --- |
| | license: apache-2.0 |
| | pretty_name: Libra-81K-SFT |
| | task_categories: |
| | - question-answering |
| | - image-to-text |
| | tags: |
| | - gui |
| | - vlm |
| | - sft |
| | - webdataset |
| | --- |
| | |
| |
|
| |
|
| | # Libra-81K-SFT |
| |
|
| | This dataset contains: |
| | - `data/images/`: split image archives (`*.tar.gz.part-*`) |
| | - `data/annotations/`: JSON annotation files |
| |
|
| |
|
| | --- |
| |
|
| | ## Download Dataset |
| |
|
| | ### 1) Install Git LFS |
| | ```bash |
| | git lfs install |
| | ```` |
| |
|
| | ### 2) Clone the dataset repo |
| |
|
| | ```bash |
| | git clone https://huggingface.co/datasets/GUI-Libra/GUI-Libra-81K-SFT |
| | cd GUI-Libra-81K-SFT |
| | ``` |
| |
|
| | --- |
| |
|
| | ## Reconstruct and extract image archives |
| |
|
| | ### Folder structure |
| |
|
| | After cloning, you should see: |
| |
|
| | * `data/images/` containing `*.tar.gz.part-*` |
| | * `data/annotations/` containing `*.json` |
| |
|
| | ### 1) Merge split parts into `.tar.gz` |
| |
|
| | This will create merged archives under `data/images_archives/`. |
| |
|
| | ```bash |
| | cd data/images |
| | mkdir -p ../images_archives ../images_extracted |
| | |
| | for base in $(ls *.tar.gz.part-* | sed -E 's/\.part-[0-9]+$//' | sort -u); do |
| | echo "[MERGE] $base.part-* -> ../images_archives/$base" |
| | cat "${base}".part-* > "../images_archives/${base}" |
| | done |
| | ``` |
| |
|
| | ### 2) Extract each merged archive |
| |
|
| | This will extract each archive into a separate directory under `data/images_extracted/`. |
| |
|
| | ```bash |
| | cd ../images_archives |
| | for tgz in *.tar.gz; do |
| | name="${tgz%.tar.gz}" |
| | mkdir -p "../images_extracted/${name}" |
| | echo "[EXTRACT] $tgz -> ../images_extracted/${name}" |
| | tar -xzf "$tgz" -C "../images_extracted/${name}" |
| | done |
| | ``` |
| |
|
| | After extraction, images will be available under: |
| |
|
| | * `data/images_extracted/<archive_name>/...` |
| |
|
| | --- |
| |
|
| | ## Load annotations in Python |
| |
|
| | ```python |
| | import json |
| | from pathlib import Path |
| | |
| | ann_dir = Path("data/annotations") |
| | all_json = {} |
| | |
| | for p in sorted(ann_dir.glob("*.json")): |
| | with open(p, "r", encoding="utf-8") as f: |
| | all_json[p.name] = json.load(f) |
| | |
| | print("Loaded json files:", len(all_json)) |
| | print("Example keys:", list(all_json.keys())[:5]) |
| | ``` |
| |
|
| | --- |
| |
|
| | ## Notes |
| |
|
| | * Each `*.tar.gz.part-*` group must be merged **in order** before extraction. |
| | * If you only need one subset (e.g., `gui-odyssey`), you can merge/extract only that prefix: |
| |
|
| | ```bash |
| | cd data/images |
| | cat gui-odyssey.tar.gz.part-* > ../images_archives/gui-odyssey.tar.gz |
| | tar -xzf ../images_archives/gui-odyssey.tar.gz -C ../images_extracted/gui-odyssey |
| | ``` |
| |
|
| |
|
| |
|