visual_prompt upload package

This folder contains the upload helper for packaging the four instant-streaming colloquial JSON files and the MP4 files referenced by their videos fields.

Target Hugging Face repo:

spw2000/visual_prompt

Files included

The upload script packages these four metadata files:

test_instant_streaming.colloquial.en.json
test_instant_streaming.colloquial.zh.json
train_instant_streaming.colloquial.en.json
train_instant_streaming.colloquial.zh.json

It scans each JSON file, reads every videos list, deduplicates all referenced video paths, rewrites packaged JSON video paths to be relative to the extracted metadata/ directory, and packages the corresponding video files.

Uploaded structure

Running 2_upload.py creates hf_archives/ locally and uploads that folder to Hugging Face:

README.md
manifest.json
visual_prompt_metadata.tar.gz
visual_prompt_videos-00000-of-XXXXX.tar.gz
visual_prompt_videos-00001-of-XXXXX.tar.gz
...

visual_prompt_metadata.tar.gz contains:

README.md
manifest.json
metadata/test_instant_streaming.colloquial.en.json
metadata/test_instant_streaming.colloquial.zh.json
metadata/train_instant_streaming.colloquial.en.json
metadata/train_instant_streaming.colloquial.zh.json

The JSON files inside this archive are rewritten copies. The source JSON files in hf_upload/ are not modified.

Each visual_prompt_videos-*.tar.gz contains videos using paths relative to the RGA3-release-local project root. For example:

mp4/datasets/VideoInfer-Release/frames/MOSE/train/e607450d/00002.mp4

Install dependency

pip install huggingface_hub

Upload

Recommended usage:

cd /home/dyvm6xra/dyvm6xrauser04/peiwensun/project/RGA3-release-local/hf_upload
export HF_TOKEN="YOUR_HUGGINGFACE_TOKEN"
python 2_upload.py

You can also pass the token directly:

python 2_upload.py --token "YOUR_HUGGINGFACE_TOKEN"

Before uploading, you can scan and print the package plan:

python 2_upload.py --dry-run

To only build local archives without uploading:

python 2_upload.py --skip-upload

Common options:

python 2_upload.py \
  --repo-id spw2000/visual_prompt \
  --repo-type dataset \
  --max-shard-size 20GB \
  --scan-workers 32 \
  --num-workers 8

By default the script uses Hugging Face upload_large_folder, which is resumable and supports parallel upload workers through --num-workers. This is recommended for the generated archive folder. If you need a single normal commit message upload instead, disable it:

python 2_upload.py --no-upload-large-folder

Download and extract

After downloading the uploaded files from Hugging Face, extract them into one dataset directory:

mkdir -p visual_prompt
tar -xzf visual_prompt_metadata.tar.gz -C visual_prompt
for shard in visual_prompt_videos-*.tar.gz; do
  tar -xzf "$shard" -C visual_prompt
done

The extracted structure will look like:

visual_prompt/
  README.md
  manifest.json
  metadata/
    test_instant_streaming.colloquial.en.json
    test_instant_streaming.colloquial.zh.json
    train_instant_streaming.colloquial.en.json
    train_instant_streaming.colloquial.zh.json
  mp4/
    datasets/
      VideoInfer-Release/
        frames/
          ...

Resolving video paths

The packaged JSON files use paths relative to their own metadata/ directory. For example, a source video path under local RGA3-release-local/mp4/... is written into the uploaded JSON as:

../mp4/datasets/VideoInfer-Release/frames/MOSE/train/e607450d/00002.mp4

If you read a JSON file from visual_prompt/metadata/, resolve each video path against the JSON file's parent directory:

from pathlib import Path


json_path = Path("visual_prompt/metadata/train_instant_streaming.colloquial.en.json")
raw_video_path = "../mp4/datasets/VideoInfer-Release/frames/MOSE/train/e607450d/00002.mp4"
video_path = (json_path.parent / raw_video_path).resolve()

manifest.json records archive names, SHA256 checksums, per-JSON videos array counts, video-reference counts, video counts, missing-video count, and compressed/uncompressed sizes. It also records that packaged JSON video paths use the ../mp4/... layout.