chancyun/cvci-expert

CVCI Expert

This dataset contains CARLA-based expert demonstration data prepared for CVCI / Bench2Drive experiments.

The raw data originally consisted of many small files. For stable storage and download on the Hugging Face Hub, it has been repacked into one .tar.zst archive per top-level shard. The original old/ folder is intentionally excluded.

Repository Structure

data/
  0.tar.zst
  1.tar.zst
  ...
  143.tar.zst
manifest.tsv
SHA256SUMS
extract_all.sh
README.md

There are 144 data archives in data/.

manifest.tsv records each shard name, original uncompressed byte count, and original file count.

SHA256SUMS contains checksums for every uploaded archive.

File Format

Each archive preserves the original top-level directory name:

tar -I zstd -tf data/0.tar.zst | head

Example output:

0/
0/Town06_Rep-1_0_route0_...
0/Town06_Rep-1_0_route0_.../semantics/

Verify Downloads

After downloading the dataset, run:

sha256sum -c SHA256SUMS

To verify only files that are present locally:

sha256sum -c SHA256SUMS --ignore-missing

Extract One Shard

mkdir -p cvci_expert_data
tar -I zstd -xf data/0.tar.zst -C cvci_expert_data

This creates:

cvci_expert_data/0/

Extract All Shards

Serial extraction:

mkdir -p cvci_expert_data
for f in data/*.tar.zst; do
  tar -I zstd -xf "$f" -C cvci_expert_data
done

Parallel extraction with 8 workers:

mkdir -p cvci_expert_data
find data -maxdepth 1 -name '*.tar.zst' -print0 \
  | sort -z \
  | xargs -0 -P 8 -I {} tar -I zstd -xf {} -C cvci_expert_data

Or use the included helper:

bash extract_all.sh data cvci_expert_data 8

Arguments are:

extract_all.sh <archive_dir> <output_dir> <num_workers>

Notes

• The old/ source directory is not included.
• The archive format is .tar.zst, so tar and zstd are required for extraction.
• Please confirm the intended license before making this dataset public.