Datasets:

mcocheteux
/

KITTI-C2L-Dataset

# Preparing the existing self-contained KITTI-C2L release

The dataset already exists at
https://huggingface.co/datasets/mcocheteux/KITTI-C2L-Dataset with the self-contained upload already in place. Its pinned audited revision is recorded in `release/v1/hub_manifest.json`.
The upload has 44 training and 6 test asset shards (51.94 GB of sensor assets),
19,608 training and 2,792 test frames, plus the four frozen metadata tables.

Preparation must retain that upload and its `frames` configuration. The source
checkout holds only the lightweight metadata, so its README has local preview
YAML. **Use `release/v1/HF_DATASET_CARD.md` as the production Hub README.**
The optional metadata-only archive is an alternative download, not a replacement
for the full uploaded dataset.

## What is verified

- All 50 shard schemas and frame-identity/source-mapping columns were read at
  the pinned revision using authenticated ranged reads: complete, unique
  coverage of all 22,400 expected frames and correct split assignments.
- Hub LFS SHA-256 values for the four metadata files match the exact snapshot
  used in our UniCal experiments. Each metadata row and its seeded transforms
  also pass the local numerical audit.
- The smallest train and test shards were downloaded completely. Their SHA-256
  values match the Hub declarations. All 149 contained image/scan pairs are
  byte-identical to local KITTI Raw and decode correctly.
- The other shard payloads were not downloaded in full. Their recorded hashes
  are Hub declarations, not independently recomputed local checksums. The audit
  scope is explicit in `hub_manifest.json` and `hub_payload_verification.json`.

## Stage a reviewable full-repository update

```bash
uv sync --python 3.11 --extra dev --extra release --locked
uv run ruff check .
uv run pytest
uv run kitti-c2l validate --root . --kitti-calib-root /path/to/kitti_raw
uv run python scripts/prepare_release.py --profile full-update \
  --output dist/kitti-c2l-full-update-v1.0.0
uv build
```

This validates the local metadata and recorded full-upload audit, then stages
an allowlisted **update package**. Its README is the full production card,
including `metadata`, `frames` and `windows`. `release-profile.json` records the
target repository, base revision and the requirement to retain existing asset
shards. It explicitly specifies no remote deletions. The tar.gz is roughly
16 MB because the existing 52 GB upload is reused; it is not a standalone full
dataset download. No upload or visibility change occurs.

For an independent lightweight download:

```bash
uv run python scripts/prepare_release.py --profile metadata \
  --output dist/kitti-c2l-metadata-v1.0.0
uv run python scripts/check_release_loading.py --root dist/kitti-c2l-metadata-v1.0.0
```

Do not upload the lightweight README over the full dataset card. Neither
archive includes local KITTI files, gallery payloads, credentials or scratch
directories. Source/wheel distributions also exclude sensor images and dataset
payloads. Both staging modes refuse to overwrite an existing output.

## Rerun the remote audit

Use Hugging Face CLI authentication with an account authorized to read the
dataset if access is restricted; Safari login alone does not authenticate a separate workspace.

```bash
uv run hf auth login
uv run python scripts/audit_hub_release.py \
  --revision dfe6154248a52886fcf35851161cbe07a4373d2c \
  --output scratch/hub-audit.json
```

This reads schemas and non-payload columns and does not download the complete
52 GB sensor corpus. The pinned revision makes the inventory reproducible.
To reproduce the payload spot check (about 361 MB of downloads, cached on repeat):

```bash
uv run python scripts/check_hub_payloads.py --raw-root /path/to/kitti_raw \
  --output scratch/payload-audit.json
```

Do not describe these two shards as a full sensor-payload audit.

## Baseline package

`scripts/prepare_baselines.py` creates the separate model package as described
in BASELINES.md. Three exported checkpoints have been re-evaluated across all
2,792 native test frames each, with Hub access disabled, reproducing the
original results. Model weights and source snapshots have their own licenses;
no training images or point clouds are included in that package.

## Publication handoff

1. Review the full dataset card, licenses, citation, frozen geometry convention,
   split policy and measured limitations. Choose the real release date; set
   `date-released` in CITATION.cff and remove candidate/private wording when
   publication actually happens. No DOI or publication date is invented here.
2. Confirm that the current Hub revision still matches the recorded base or
   review subsequent changes. Upload only the reviewed update-package files,
   preserving all existing `data/frames_*.parquet` shards and docs assets. Never
   synchronize with deletion from this metadata-only local checkout.
3. Check all three Hub configurations and the viewer before announcing a release. Pin
   the resulting revision and recheck metadata hashes, shard counts and joins.
4. Change visibility only when explicitly authorized. Recheck the public page
   and loading flow without authentication; record the final public revision.
5. Publish the MIT loader/generator source and attach the wheel or publish its
   package so linked installation instructions are accessible. Publish optional
   baseline weights separately with their source and upstream notices.

Official references:
[Hugging Face manual configuration](https://huggingface.co/docs/hub/datasets-manual-configuration),
[KITTI licensing](https://www.cvlibs.net/datasets/kitti/),
[KITTI access policy](https://www.cvlibs.net/datasets/kitti/user_submit.php).

Changes to geometry, targets, frame selection or split semantics require a new
version and updated evidence. The sensor assets can remain byte-identical when
only calibration metadata is revised; do not re-upload them unnecessarily.