Blame for EXTERNAL_DATA.md

# External Data Dependencies

The dataset root contains derived corruptions and protocols. Original sensor
data, official GT and model weights are not included in these packages.
Users must obtain the original datasets under their applicable access terms.
See [Licenses and Use](DISTRIBUTION.md) and the [dataset card](README.md).

Set the roots explicitly:

```bash
export OCCSTRESS_DATA_ROOT=/datasets/OccStress
export OCCSTRESS_EXTERNAL_ROOT=/datasets/OccStress-external
```

The external root is a separate directory with this logical layout:

```text
OccStress-external/
  OccStress-nuScenes/
    gts/<scene>/<token>/labels.npz
    world-nuscenes_infos_val.pkl
    nuscenes_infos_val_temporal_v3_scene.pkl
    native_dataset/
  OccStress-Waymo/
    native_gt/validation-data/...
    cam_infos_vali.pkl
    native_dataset/
  OccStress-CARLA/
    canonical_gt/<scene>/<token>/labels.npz
    native_dataset/<scene>/...
```

Directories may be symlinks to authorized existing data. Do not overwrite
existing directories or copy the originals merely to satisfy this layout.
Use dependencies read-only; ordinary symlinks do not enforce OS-level read-only
access. Do not include these links or their targets in a public upload.

## nuScenes

Obtain GT through [Occ3D-nuScenes](https://github.com/Tsinghua-MARS-Lab/Occ3D#occ3d-nuscenes),
world-info through [II-World](https://github.com/lzzzzzm/II-World), and the
scene-organized info through [DOME](https://github.com/gusongen/DOME).

- `gts` is the original Occ3D validation GT, in the canonical 18-class space.
- `world-nuscenes_infos_val.pkl` provides the official frame-wise poses and
  controls used by the II-World/GenieDrive adapters.
- `nuscenes_infos_val_temporal_v3_scene.pkl` is the scene-organized metadata view
  used by the scene-based forecasting adapters. These two info files are not
  interchangeable merely because they describe the same validation split.
- `native_dataset` is needed for camera/LiDAR source re-export, not for consuming
  already exported occupancy. Follow each upstream exporter's sensor-root options.

## Waymo

Obtain validation `voxel04` through
[Occ3D-Waymo](https://github.com/Tsinghua-MARS-Lab/Occ3D#occ3d-waymo), after
accepting the original dataset access terms. The Motion dataset is not a
substitute for matching Perception sensor segments when regenerating exports.

- Mount the original Occ3D voxel04 root as `native_gt`; it must contain
  `validation-data`. Native `voxel_label` is mapped using the provided class map.
  Derived `semantics` is already canonical and must not be mapped again.
- Forecasting poses and controls are included at
  `meta/OccStress-Waymo/controls.json` in the OccStress root.
- `native_dataset` is the prepared camera/sensor view used by the upstream
  exporter, not an arbitrary Waymo download directory. Camera-direct evaluation
  also needs the matching `cam_infos_vali.pkl` and native camera-root CLI options.

## CARLA

- `canonical_gt` must be the right-handed, Occ3D-mapped view. Mounting native
  left-handed UniOcc arrays at this location is incorrect. Use the code release's
  `tools/prepare_external_data.py carla-gt` with the published metadata and native
  source. Do not run the historical intermediate-layout converter on this tree.
- `native_dataset` contains the authorized UniOcc CARLA validation source files,
  including camera PNGs, LiDAR PLYs and original occupancy NPZs referenced by the
  control metadata. It remains an external dependency.
- Poses and controls are included at `meta/OccStress-CARLA/controls.json`.
  The canonical command convention is documented in the dataset README.

## New-Machine Preparation

Run these commands from the OccStress code checkout. Each mount refuses to
replace a different existing destination. Refer to the code documentation at
`docs/EXTERNAL_DATA_PREPARATION.md` for official providers and optional sensors.

```bash
python tools/prepare_external_data.py mount --dataset nuscenes \
  --kind gts --source /datasets/raw/occ3d-nuscenes/gts
python tools/prepare_external_data.py mount --dataset nuscenes \
  --kind world-nuscenes_infos_val.pkl --source /datasets/raw/infos/world-nuscenes_infos_val.pkl
python tools/prepare_external_data.py mount --dataset nuscenes \
  --kind nuscenes_infos_val_temporal_v3_scene.pkl \
  --source /datasets/raw/infos/nuscenes_infos_val_temporal_v3_scene.pkl
python tools/prepare_external_data.py mount --dataset waymo \
  --kind native_gt --source /datasets/raw/occ3d-waymo/voxel04
python tools/prepare_external_data.py mount --dataset carla \
  --kind native_dataset --source /datasets/raw/uniocc/Carla-2Hz-val
python tools/prepare_external_data.py carla-gt \
  --source-root /datasets/raw/uniocc/Carla-2Hz-val
python tools/prepare_external_data.py check --dataset nuscenes --trust-pickle
python tools/prepare_external_data.py check --dataset waymo --trust-pickle
python tools/prepare_external_data.py check --dataset carla --trust-pickle
```

Use `tasl-lab/uniocc` revision `e81775b36e376f591a1145ca054b72fd541a67eb`,
folder `Carla-2Hz-val`: three scenes, 120 frames each. The converter checks the
source metadata checksum, reads only numeric NPZ arrays, maps labels and flips
axis 1 once. It does not rewrite controls or original files. The dependency
check verifies complete clean-reference coverage, not GPU metric reproduction.

To obtain just this source subset:

```bash
hf download tasl-lab/uniocc --repo-type dataset \
  --revision e81775b36e376f591a1145ca054b72fd541a67eb \
  --include 'Carla-2Hz-val/*' --local-dir /datasets/raw/uniocc
```

Weights are not part of this dataset. Downstream checkpoints are required to run
forecasters; upstream 3D weights are needed only for re-exporting predictions.
Prefer official references in the code release's `configs/resources.json`.

## Check Scope

The assembly's external dependency report checks all 14,377 referenced GT files
and CARLA sensor references for existence and nonzero size. It does not hash
original data or certify a complete raw nuScenes/Waymo sensor inventory. Separate
upstream exporter tests still need to verify their calibration and sensor inputs.
Availability on the assembly machine does not imply availability on a new machine.

Only load trusted protocol PKLs when using `--trust-pickle`. Keep original-data
mounts and machine-specific environment settings outside the public dataset.