# Download layers

| Path | Contents | Needed for |
| --- | --- | --- |
| `core/urban_multicity/` | All 54,193 rows, memberships, split/RF contracts, cached scene tensors | Training the four utility models |
| `extensions/reward_adaptation_targets/` | 3,590 target-scene rows and cached tensors | Separate few-shot/full-target adaptation |
| `scene_assets/<city>/*.tar` | 67 portable XML/mesh/input/source-feature packages | Physical ray tracing and supplementary scenes |
| `radio_maps/index/*.parquet` | Sample-to-shard/member/checksum references | Selecting archived spatial outputs |
| `radio_maps/shards/<split>/<scene>/*.tar` | All 36,599 archived NPZ maps, in approximately 1 GiB shards | Radio Map analysis |
| `catalog/` | Primary scene catalog plus all scene-asset entries | Geography and selective downloads |

The primary benchmark uses 60 scenes. Adaptation adds two scene IDs and reuses
tokyo_01; five additional scene packages are supplementary geometry, not extra
primary benchmark training scenes. Counts must not be mixed.

## Minimal download and training

Run `examples/download_dataset.py --destination citydeploy_data`. This fetches
the lightweight core and its scene tensors, not the large Radio Maps or meshes.
The companion repository accepts the resulting directory directly:

```console
citydeploy train --model hpem --dataset-root citydeploy_data/core/urban_multicity --scene-root citydeploy_data/scenes --device cuda
```

Use `--model regressor`, `--model in` or `--model nri` for the other models.
Install the companion code and select training budgets before a full experiment.

## Optional physical assets and maps

```console
python examples/download_dataset.py --destination citydeploy_data --with-scenes --scenes london_01
python examples/download_dataset.py --destination citydeploy_data --with-radio-maps --scenes london_01
python examples/download_dataset.py --destination citydeploy_data --adaptation
```

The scene archives unpack under `citydeploy_data/scenes`. Available maps unpack
under the primary core's `radio_maps` directory, matching row `rich_map_path`
references. Archives remain cached after extraction; plan disk space for both
archives and extracted maps. A request for all maps transfers approximately
115 GB before extraction. No missing map is synthesized.

The downloader pins a single Hub commit for each invocation, validates file
checksums, rejects unsafe archive paths and refuses to replace a different
existing file. Record its printed revision with experiments. Pass `--revision`
to reproduce a fixed release. Interrupted downloads can resume through the Hub
cache; interrupted archive extraction leaves a visible `.partial` file for
inspection. Remove only that incomplete output before retrying.

## Archive and licensing scope

TAR owner fields are empty and modification times are fixed to zero. NPZ ZIP
members use fixed synthetic timestamps and have no comments or extra fields.
No model weights, experiment logs, author-machine paths or satellite imagery
are included. See the component-specific terms in `LICENSE.md`.

Source feature extracts preserve geometry and selected physical semantic tags.
Free-text/contact tags and acquisition/QA run metadata are omitted from the public
export. Stored meshes, numerical scene channels and simulation values remain the
reference inputs and outputs; regenerating geometry from live map services is not
equivalent to loading these archived assets.