smart-insertion-v-dataset

Index files

Each folder has one index file. Every path inside an index is relative to the folder that contains that index.

• video1/info.json
• video1/part1/info.json
• video1/part2/info.json
• video1/part3/info.json
• video1/part4/info.json
• video2/info.json
• image/info.json

video1/info.json covers all four parts together, so its paths begin with part1/ … part4/.

Download

# everything
hf download Massyzs/smart-insertion-v-dataset --repo-type dataset --local-dir .

# only one subset, e.g. the image data
hf download Massyzs/smart-insertion-v-dataset --repo-type dataset --local-dir . \
    --include "image/*"

Extract

Media files sit next to their index as tar archives split into 20 GB parts (for example image/input_video.tar.part-0000, image/input_video.tar.part-0001, …). Run this from the repository root — each archive stores full relative paths, so the files land in the right place:

for n in $(ls */*.tar.part-0000 */*/*.tar.part-0000 2>/dev/null | sed 's#\.tar\.part-0000##'); do
    cat "$n".tar.part-* | tar -xf -
done

To verify an archive against SHA256SUMS.txt:

cat image/input_video.tar.part-* | sha256sum