Move required evaluation dataset out of optional

README.md

@@ -145,10 +145,11 @@ The prepared data does not replace access agreements for original source assets.
 [Posed object geometry](previews/3dfront/train-geometry.glb)
 
 <!-- partprobe3d-frozen-evaluation:start -->
-## Frozen evaluation
+## Required evaluation dataset
 
-The optional [evaluation bundle](optional/evaluation/README.md) contains 2,603 images with visible masks and GT object meshes.
+The required [evaluation dataset](evaluation/README.md) contains 2,603 images with visible masks and GT object meshes.
 It includes all 100 PICASSO images, all 1,503 MessyKitchens images, and 1,000 random 3D-FRONT test views.
+The PartProbe3D benchmark requires this dataset.
 The transfer is 23.40 GB.
 MessyKitchens shares 100 GT scene meshes across its views.
 Default training downloads exclude this bundle.
@@ -156,6 +157,6 @@ Default training downloads exclude this bundle.
 Apply the loader patch described in the bundle documentation before evaluating its samples.
 
 ```bash
-python -m preprocess.huggingface download --repo cubec/PartProbe3D-preprocessed --optional evaluation
+python -m preprocess.huggingface download --repo cubec/PartProbe3D-preprocessed --dataset evaluation
 ```
 <!-- partprobe3d-frozen-evaluation:end -->

code/evaluation.patch

@@ -1,3 +1,83 @@
+diff --git a/dataloader/frozen.py b/dataloader/frozen.py
+new file mode 100644
+index 0000000..2a43518
+--- /dev/null
++++ b/dataloader/frozen.py
+@@ -0,0 +1,74 @@
++"""Read portable RGB, masks, and shared GT meshes from frozen evaluation data."""
++from collections import defaultdict
++import json
++
++import numpy as np
++from PIL import Image
++import trimesh
++
++from .base import BaseDataLoader, EvaluationSample
++from .geometry import artifact_path, checked_digest
++from utils.hashing import input_hashes, sha256
++
++
++SCHEMA = 'partprobe3d.frozen-evaluation.v1'
++
++
++class FrozenEvaluationLoader(BaseDataLoader):
++    """Load saved evaluation inputs without the original source datasets."""
++    name = 'frozen'
++
++    def __init__(self, root):
++        super().__init__(root)
++        self.manifest_path = self.root / 'manifest.json'
++        self.manifest = json.loads(self.manifest_path.read_text())
++        if self.manifest.get('schema') != SCHEMA or not self.manifest.get('complete'):
++            raise ValueError('A complete frozen evaluation dataset is required')
++        self.rows = self.manifest['frames']
++        if not self.rows or len({row['key'] for row in self.rows}) != len(self.rows):
++            raise ValueError('Frozen evaluation sample keys must be nonempty and unique')
++
++    def __len__(self):
++        return len(self.rows)
++
++    def _load_sample(self, index):
++        row = self.rows[index]
++        paths = {key: artifact_path(self.root, row[key]) for key in ('rgb', 'masks', 'ground_truth')}
++        for relative in (row[key] for key in paths):
++            if checked_digest(artifact_path(self.root, relative)) != self.manifest['files'][relative]:
++                raise ValueError(f'Frozen evaluation file checksum differs: {relative}')
++        with Image.open(paths['rgb']) as image:
++            rgb = np.array(image.convert('RGB'))
++        with Image.open(paths['masks']) as image:
++            masks = np.array(image).astype(np.int32)
++        if input_hashes(rgb, masks) != {key: row[key] for key in ('rgb_sha256', 'masks_sha256')}:
++            raise ValueError('Frozen RGB pixels or mask labels differ from the source selection')
++        scene = trimesh.load(paths['ground_truth'], force='scene', process=False)
++        metadata = scene.metadata.get('partprobe3d', {})
++        if (metadata.get('schema') != 'partprobe3d.object-meshes.v1' or metadata.get('units') != 'm'
++                or metadata.get('coordinate_frame') != row['gt_coordinate_frame']
++                or metadata.get('identity_mode') != row['dataset_metadata'].get('input_mask_identity', 'mask')):
++            raise ValueError('Frozen GT mesh metadata differs from the selection')
++        pieces = defaultdict(list)
++        for node in scene.graph.nodes_geometry:
++            transform, geometry = scene.graph[node]
++            mesh = scene.geometry[geometry].copy().apply_transform(transform)
++            pieces[int(mesh.metadata['object_index'])].append(mesh)
++        meshes = {index: trimesh.util.concatenate(parts) for index, parts in pieces.items()}
++        if (set(meshes) != {int(obj['object_index']) for obj in metadata['objects']}
++                or len(meshes) != row['target_objects']):
++            raise ValueError('Frozen GT objects differ from the selection')
++        return EvaluationSample(key=row['key'], rgb=rgb, masks=masks, gt_meshes=meshes,
++            gt_coordinate_frame=row['gt_coordinate_frame'], metadata=row['dataset_metadata'],
++            visibility=row.get('visibility', {}), mask_overlap_pixels=row.get('mask_overlap_pixels', 0),
++            discarded_mask_pixels=row.get('discarded_mask_pixels', {}))
++
++    def sample_records(self):
++        return [{'index': i, 'key': row['key'], 'group': row['dataset_metadata'].get('scene', row['key']),
++                 'order': row['dataset_metadata'].get('source_view', row['key'])}
++                for i, row in enumerate(self.rows)]
++
++    def provenance(self):
++        return {**super().provenance(), 'manifest_sha256': sha256(self.manifest_path),
++                'source_selection_sha256': self.manifest['source_selection_sha256'],
++                'source_dataset': self.manifest['source_dataset'], 'gt_representation': 'exported GLB float32'}
 diff --git a/dataloader/registry.py b/dataloader/registry.py
 index f3de37d..c9cba24 100644
 --- a/dataloader/registry.py
@@ -93,6 +173,33 @@ index d61612c..8cb65e6 100644
      write_prediction(folder / gt_name, sample.gt_meshes, coordinate_frame=sample.gt_coordinate_frame,
                       units="m", identity_mode=sample.metadata.get("input_mask_identity", "mask"),
                       method="ground-truth", provenance=provenance)
+diff --git a/evaluation/configs/front3d_frozen.json b/evaluation/configs/front3d_frozen.json
+new file mode 100644
+index 0000000..c84bba7
+--- /dev/null
++++ b/evaluation/configs/front3d_frozen.json
+@@ -0,0 +1,3 @@
++{
++  "root": "datasets/evaluation/3dfront"
++}
+diff --git a/evaluation/configs/messykitchen_frozen.json b/evaluation/configs/messykitchen_frozen.json
+new file mode 100644
+index 0000000..2bf4e4b
+--- /dev/null
++++ b/evaluation/configs/messykitchen_frozen.json
+@@ -0,0 +1,3 @@
++{
++  "root": "datasets/evaluation/messykitchen"
++}
+diff --git a/evaluation/configs/picasso_frozen.json b/evaluation/configs/picasso_frozen.json
+new file mode 100644
+index 0000000..7d85ea9
+--- /dev/null
++++ b/evaluation/configs/picasso_frozen.json
+@@ -0,0 +1,3 @@
++{
++  "root": "datasets/evaluation/picasso"
++}
 diff --git a/evaluation/evaluate.py b/evaluation/evaluate.py
 index e50d3ea..318de5c 100644
 --- a/evaluation/evaluate.py
@@ -231,181 +338,9 @@ index 3968d22..5f985e0 100644
  def collect_runs(run, selection):
      """Collect recorded model outputs and disjoint shards by their method metadata."""
      expected = {frame["key"]: frame for frame in selection["frames"]}
-diff --git a/preprocess/huggingface.py b/preprocess/huggingface.py
-index 2fdce63..df8bdf1 100644
---- a/preprocess/huggingface.py
-+++ b/preprocess/huggingface.py
-@@ -394,9 +394,9 @@ def extract(index, folder, destination):
-     if index.get('schema') != SCHEMA or not index.get('complete'):
-         raise ValueError('A complete archive index is required')
-     folder, destination = Path(folder), Path(destination).resolve()
--    prefix = _relative(index['destination'])
-+    prefix = _relative(index.get('archive_prefix', index['destination']))
-     restored_prefix = (_slat_destination(index['name'], index['resolution'])
--                       if index.get('artifact_kind') == 'slat-shape-targets' else prefix)
-+                       if index.get('artifact_kind') == 'slat-shape-targets' else _relative(index['destination']))
-     # Publish the original complete manifest only after restoring its artifacts.
-     for item in sorted(index['archives'], key=lambda item: item['file'] == 'metadata.tar'):
-         path = folder / _relative(item['file'])
-@@ -563,7 +563,7 @@ def download(repo, name, destination, revision, *, cleanup=None, bundle_root=Non
-     # Keep active training folders free of transfer archives. Retain the legacy
-     # cache path only when an explicit historical component needs resume.
-     cache = (destination / 'datasets' / '_deprecated' / name / 'huggingface'
--             if bundle_root.startswith(('training/', 'optional/'))
-+             if bundle_root.split('/')[0] in ('training', 'optional')
-              else destination / 'datasets' / name / 'archives' / 'huggingface')
-     index_path = hf_hub_download(repo, f'{bundle_root}/{name}/index.json', repo_type='dataset',
-                                  revision=revision, local_dir=cache)
-@@ -894,6 +894,8 @@ def main():
-         if args.optional and args.bundle_root:
-             parser.error('--optional and --bundle-root cannot be combined')
-         bundle_root = f'optional/{args.optional.replace("-", "_")}' if args.optional else args.bundle_root
-+        if args.optional == 'evaluation' and args.dataset == 'all':
-+            args.dataset, bundle_root = 'evaluation', 'optional'
-         if args.optional and args.dataset == 'all':
-             parser.error('Select one dataset with --dataset for an optional download')
-         cleanup = not args.keep_archives if bundle_root is None or args.optional else args.cleanup
-diff --git a/preprocess/reproduction.py b/preprocess/reproduction.py
-index fd0ccd7..deabce5 100644
---- a/preprocess/reproduction.py
-+++ b/preprocess/reproduction.py
-@@ -180,6 +180,8 @@ def package(root, output, component, name, *, destination, kind, metadata, group
-         raise ValueError(f'Package source changed. Select another output: {component}/{name}')
-     previous = {row['file']: row for row in old.get('archives', [])}
-     index = {'schema': SCHEMA, 'name': name, 'complete': False, **identity, 'archives': []}
-+    if 'archive_prefix' in old:
-+        index['archive_prefix'] = old['archive_prefix']
-     atomic_json(path, index)
- 
-     def save(label, files, views):
-@@ -197,7 +199,8 @@ def package(root, output, component, name, *, destination, kind, metadata, group
-         if reusable:
-             record = old_record
-         else:
--            record = _archive(folder / filename, files, root, Path(destination), old_record, checksums=checksums)
-+            record = _archive(folder / filename, files, root,
-+                              Path(index.get('archive_prefix', destination)), old_record, checksums=checksums)
-         record['views'] = views
-         index['archives'].append(record)
-         atomic_json(path, index)
-@@ -440,14 +443,15 @@ def _copy_files(api, repo, output, mapping, *, revision=LEGACY_REVISION, remove_
-     return commits
- 
- 
--def export_code(output, base_revision):
--    """Export only this migration's source changes, preserving unrelated workspace edits."""
-+def export_code(output, base_revision, *, tracked=None, added=None, name='reproduction'):
-+    """Export selected source changes and preserve unrelated workspace edits."""
-     output = Path(output)
-     folder = output / 'code'
-     folder.mkdir(parents=True, exist_ok=True)
--    tracked = ['HF_DATASETS.md', 'preprocess/slat/HUGGING_FACE.md', 'preprocess/slat/README.md', 'dataloader/dataset.py',
--               'dataloader/extent.py', 'preprocess/huggingface.py', 'preprocess/validate_dataset.py']
--    added = ['preprocess/reproduction.py', 'preprocess/tests/test_reproduction.py']
-+    tracked = (['HF_DATASETS.md', 'preprocess/slat/HUGGING_FACE.md', 'preprocess/slat/README.md', 'dataloader/dataset.py',
-+                'dataloader/extent.py', 'preprocess/huggingface.py', 'preprocess/validate_dataset.py']
-+               if tracked is None else tracked)
-+    added = ['preprocess/reproduction.py', 'preprocess/tests/test_reproduction.py'] if added is None else added
-     content = subprocess.check_output(['git', 'diff', '--binary', base_revision, '--', *tracked, *added])
-     for path in added:
-         present = subprocess.run(['git', 'cat-file', '-e', f'{base_revision}:{path}'],
-@@ -461,11 +465,11 @@ def export_code(output, base_revision):
-         if result.returncode not in (0, 1):
-             raise RuntimeError(f'Cannot export restore source: {path}')
-         content += result.stdout
--    patch = folder / 'reproduction.patch'
-+    patch = folder / f'{name}.patch'
-     patch.write_bytes(content)
-     manifest = {'base_git_revision': base_revision, 'patch_sha256': sha256(patch),
-                 'files': {path: sha256(path) for path in [*tracked, *added]}}
--    atomic_json(folder / 'manifest.json', manifest)
-+    atomic_json(folder / ('manifest.json' if name == 'reproduction' else f'{name}_manifest.json'), manifest)
-     return manifest
- 
- 
-diff --git a/dataloader/frozen.py b/dataloader/frozen.py
-new file mode 100644
-index 0000000..2a43518
---- /dev/null
-+++ b/dataloader/frozen.py
-@@ -0,0 +1,74 @@
-+"""Read portable RGB, masks, and shared GT meshes from frozen evaluation data."""
-+from collections import defaultdict
-+import json
-+
-+import numpy as np
-+from PIL import Image
-+import trimesh
-+
-+from .base import BaseDataLoader, EvaluationSample
-+from .geometry import artifact_path, checked_digest
-+from utils.hashing import input_hashes, sha256
-+
-+
-+SCHEMA = 'partprobe3d.frozen-evaluation.v1'
-+
-+
-+class FrozenEvaluationLoader(BaseDataLoader):
-+    """Load saved evaluation inputs without the original source datasets."""
-+    name = 'frozen'
-+
-+    def __init__(self, root):
-+        super().__init__(root)
-+        self.manifest_path = self.root / 'manifest.json'
-+        self.manifest = json.loads(self.manifest_path.read_text())
-+        if self.manifest.get('schema') != SCHEMA or not self.manifest.get('complete'):
-+            raise ValueError('A complete frozen evaluation dataset is required')
-+        self.rows = self.manifest['frames']
-+        if not self.rows or len({row['key'] for row in self.rows}) != len(self.rows):
-+            raise ValueError('Frozen evaluation sample keys must be nonempty and unique')
-+
-+    def __len__(self):
-+        return len(self.rows)
-+
-+    def _load_sample(self, index):
-+        row = self.rows[index]
-+        paths = {key: artifact_path(self.root, row[key]) for key in ('rgb', 'masks', 'ground_truth')}
-+        for relative in (row[key] for key in paths):
-+            if checked_digest(artifact_path(self.root, relative)) != self.manifest['files'][relative]:
-+                raise ValueError(f'Frozen evaluation file checksum differs: {relative}')
-+        with Image.open(paths['rgb']) as image:
-+            rgb = np.array(image.convert('RGB'))
-+        with Image.open(paths['masks']) as image:
-+            masks = np.array(image).astype(np.int32)
-+        if input_hashes(rgb, masks) != {key: row[key] for key in ('rgb_sha256', 'masks_sha256')}:
-+            raise ValueError('Frozen RGB pixels or mask labels differ from the source selection')
-+        scene = trimesh.load(paths['ground_truth'], force='scene', process=False)
-+        metadata = scene.metadata.get('partprobe3d', {})
-+        if (metadata.get('schema') != 'partprobe3d.object-meshes.v1' or metadata.get('units') != 'm'
-+                or metadata.get('coordinate_frame') != row['gt_coordinate_frame']
-+                or metadata.get('identity_mode') != row['dataset_metadata'].get('input_mask_identity', 'mask')):
-+            raise ValueError('Frozen GT mesh metadata differs from the selection')
-+        pieces = defaultdict(list)
-+        for node in scene.graph.nodes_geometry:
-+            transform, geometry = scene.graph[node]
-+            mesh = scene.geometry[geometry].copy().apply_transform(transform)
-+            pieces[int(mesh.metadata['object_index'])].append(mesh)
-+        meshes = {index: trimesh.util.concatenate(parts) for index, parts in pieces.items()}
-+        if (set(meshes) != {int(obj['object_index']) for obj in metadata['objects']}
-+                or len(meshes) != row['target_objects']):
-+            raise ValueError('Frozen GT objects differ from the selection')
-+        return EvaluationSample(key=row['key'], rgb=rgb, masks=masks, gt_meshes=meshes,
-+            gt_coordinate_frame=row['gt_coordinate_frame'], metadata=row['dataset_metadata'],
-+            visibility=row.get('visibility', {}), mask_overlap_pixels=row.get('mask_overlap_pixels', 0),
-+            discarded_mask_pixels=row.get('discarded_mask_pixels', {}))
-+
-+    def sample_records(self):
-+        return [{'index': i, 'key': row['key'], 'group': row['dataset_metadata'].get('scene', row['key']),
-+                 'order': row['dataset_metadata'].get('source_view', row['key'])}
-+                for i, row in enumerate(self.rows)]
-+
-+    def provenance(self):
-+        return {**super().provenance(), 'manifest_sha256': sha256(self.manifest_path),
-+                'source_selection_sha256': self.manifest['source_selection_sha256'],
-+                'source_dataset': self.manifest['source_dataset'], 'gt_representation': 'exported GLB float32'}
 diff --git a/preprocess/evaluation.py b/preprocess/evaluation.py
 new file mode 100644
-index 0000000..abe6e3b
+index 0000000..f920ac4
 --- /dev/null
 +++ b/preprocess/evaluation.py
 @@ -0,0 +1,171 @@
@@ -518,7 +453,7 @@ index 0000000..abe6e3b
 +
 +
 +def publish(source, output, repo=DEFAULT_REPO, *, workers=4, shard_bytes=1024**3):
-+    """Publish optional evaluation archives through the verified project uploader."""
++    """Publish the required evaluation archives through the verified project uploader."""
 +    source, output = Path(source).resolve(), Path(output).resolve()
 +    cohort = json.loads((source / 'index.json').read_text())
 +    if not cohort.get('complete'):
@@ -526,7 +461,7 @@ index 0000000..abe6e3b
 +    output.mkdir(parents=True, exist_ok=True)
 +    publisher = Publisher(repo, workers=workers)
 +    result = {'repo': repo, 'complete': False, 'total_views': cohort['total_views'],
-+              'component': 'optional/evaluation', 'datasets': cohort['datasets']}
++              'component': 'evaluation', 'datasets': cohort['datasets']}
 +    atomic_json(output / 'status.json', result)
 +    metadata = [source / 'index.json', *sorted(source.glob('*.md'))]
 +    groups, checksums = [], {}
@@ -545,11 +480,11 @@ index 0000000..abe6e3b
 +        metadata.extend([root / 'manifest.json', root / 'source_selection.json', *sorted(root.glob('*.md'))])
 +        groups.extend((name, sorted(paths), counts[scene]) for scene, paths in sorted(scenes.items()))
 +        checksums.update({artifact_path(root, key): value for key, value in manifest['files'].items()})
-+    index = package(source, output, 'optional', 'evaluation',
++    index = package(source, output, '', 'evaluation',
 +        destination=Path('datasets/evaluation'), kind='frozen-evaluation',
 +        metadata=metadata, groups=groups, checksums=checksums,
 +        details={'source_manifest_sha256': sha256(source / 'index.json'), 'views': cohort['total_views'],
-+                 'datasets': cohort['datasets'], 'evaluation_only': True},
++                 'datasets': cohort['datasets'], 'evaluation_only': True, 'required_for_evaluation': True},
 +        shard_bytes=shard_bytes, publisher=publisher)
 +    result.update(complete=True, bytes=index['bytes'], archives=len(index['archives']),
 +                  revision=publisher.api.dataset_info(repo).sha)
@@ -580,30 +515,185 @@ index 0000000..abe6e3b
 +
 +if __name__ == '__main__':
 +    main()
-diff --git a/evaluation/configs/picasso_frozen.json b/evaluation/configs/picasso_frozen.json
-new file mode 100644
-index 0000000..7d85ea9
---- /dev/null
-+++ b/evaluation/configs/picasso_frozen.json
-@@ -0,0 +1,3 @@
-+{
-+  "root": "datasets/evaluation/picasso"
-+}
-diff --git a/evaluation/configs/messykitchen_frozen.json b/evaluation/configs/messykitchen_frozen.json
-new file mode 100644
-index 0000000..2bf4e4b
---- /dev/null
-+++ b/evaluation/configs/messykitchen_frozen.json
-@@ -0,0 +1,3 @@
-+{
-+  "root": "datasets/evaluation/messykitchen"
-+}
-diff --git a/evaluation/configs/front3d_frozen.json b/evaluation/configs/front3d_frozen.json
-new file mode 100644
-index 0000000..c84bba7
---- /dev/null
-+++ b/evaluation/configs/front3d_frozen.json
-@@ -0,0 +1,3 @@
-+{
-+  "root": "datasets/evaluation/3dfront"
-+}
+diff --git a/preprocess/huggingface.py b/preprocess/huggingface.py
+index 2fdce63..9062d24 100644
+--- a/preprocess/huggingface.py
++++ b/preprocess/huggingface.py
+@@ -394,9 +394,9 @@ def extract(index, folder, destination):
+     if index.get('schema') != SCHEMA or not index.get('complete'):
+         raise ValueError('A complete archive index is required')
+     folder, destination = Path(folder), Path(destination).resolve()
+-    prefix = _relative(index['destination'])
++    prefix = _relative(index.get('archive_prefix', index['destination']))
+     restored_prefix = (_slat_destination(index['name'], index['resolution'])
+-                       if index.get('artifact_kind') == 'slat-shape-targets' else prefix)
++                       if index.get('artifact_kind') == 'slat-shape-targets' else _relative(index['destination']))
+     # Publish the original complete manifest only after restoring its artifacts.
+     for item in sorted(index['archives'], key=lambda item: item['file'] == 'metadata.tar'):
+         path = folder / _relative(item['file'])
+@@ -551,6 +551,8 @@ def upload(output, repo, names, *, follow=False, packing=None, update_card=False
+ 
+ 
+ def download(repo, name, destination, revision, *, cleanup=None, bundle_root=None):
++    if name == 'evaluation' and bundle_root is None:
++        bundle_root = ''
+     if bundle_root is None:
+         from preprocess.reproduction import download_training
+         return download_training(repo, name, destination, revision, cleanup=cleanup is not False)
+@@ -558,19 +560,20 @@ def download(repo, name, destination, revision, *, cleanup=None, bundle_root=Non
+     from utils.huggingface_download import download_huggingface_file
+     # Resolve main once so all archives and the index belong to one snapshot.
+     revision = HfApi().dataset_info(repo, revision=revision).sha
+-    bundle_root = str(_relative(bundle_root))
++    bundle_root = str(_relative(bundle_root)) if bundle_root else ''
++    component = str(PurePosixPath(bundle_root) / _relative(name))
+     destination = Path(destination).resolve()
+-    # Keep active training folders free of transfer archives. Retain the legacy
++    # Keep active dataset folders free of transfer archives. Retain the legacy
+     # cache path only when an explicit historical component needs resume.
+     cache = (destination / 'datasets' / '_deprecated' / name / 'huggingface'
+-             if bundle_root.startswith(('training/', 'optional/'))
++             if component.split('/')[0] in ('training', 'optional', 'evaluation')
+              else destination / 'datasets' / name / 'archives' / 'huggingface')
+-    index_path = hf_hub_download(repo, f'{bundle_root}/{name}/index.json', repo_type='dataset',
++    index_path = hf_hub_download(repo, f'{component}/index.json', repo_type='dataset',
+                                  revision=revision, local_dir=cache)
+     index = _json(index_path)
+     if index.get('name') != name or not index.get('complete'):
+         raise ValueError('Requested dataset has no complete archive index')
+-    folder = cache / bundle_root / name
++    folder = cache / component
+     status_path = folder / 'download-status.json'
+     status = {'state': 'running', 'pid': os.getpid(), 'repo': repo, 'revision': revision,
+               'dataset': name, 'bundle_root': bundle_root, 'started_unix': time.time(),
+@@ -592,7 +595,7 @@ def download(repo, name, destination, revision, *, cleanup=None, bundle_root=Non
+             status.update(state='downloading', current_archive=item['file'], active_transfer=None)
+             save()
+             archive = download_huggingface_file(
+-                repo, f'{bundle_root}/{name}/{item["file"]}', repo_type='dataset',
++                repo, f'{component}/{item["file"]}', repo_type='dataset',
+                 revision=revision, local_dir=cache, expected_size=item['bytes'],
+                 expected_sha256=item['sha256'], progress=progress)
+             status['state'] = 'restoring'
+@@ -854,7 +857,7 @@ def main():
+                             help='Concurrent archive transfers. Default: Hugging Face client setting')
+     restoring = commands.add_parser('download')
+     restoring.add_argument('--repo', required=True)
+-    restoring.add_argument('--dataset', default='all', help='Dataset name. Default: 3D-FRONT and SAGE')
++    restoring.add_argument('--dataset', default='all', help='Training dataset name, or evaluation for the required benchmark')
+     restoring.add_argument('--destination', type=Path, default=Path('.'))
+     restoring.add_argument('--revision', default='main')
+     restoring.add_argument('--bundle-root', default=None,
+diff --git a/preprocess/reproduction.py b/preprocess/reproduction.py
+index fd0ccd7..0e7d210 100644
+--- a/preprocess/reproduction.py
++++ b/preprocess/reproduction.py
+@@ -171,8 +171,9 @@ def package(root, output, component, name, *, destination, kind, metadata, group
+             details=None, checksums=None, shard_bytes=1024**3, publisher=None):
+     """Write deterministic shards and publish them with bounded temporary storage."""
+     root, output = Path(root).resolve(), Path(output).resolve()
+-    folder = output / _relative(component) / name
++    folder = output / (_relative(component) if component else Path()) / name
+     folder.mkdir(parents=True, exist_ok=True)
++    remote_root = str(folder.relative_to(output))
+     path = folder / 'index.json'
+     old = _json(path) if path.exists() else {}
+     identity = {'destination': str(destination), 'artifact_kind': kind, **(details or {})}
+@@ -180,13 +181,15 @@ def package(root, output, component, name, *, destination, kind, metadata, group
+         raise ValueError(f'Package source changed. Select another output: {component}/{name}')
+     previous = {row['file']: row for row in old.get('archives', [])}
+     index = {'schema': SCHEMA, 'name': name, 'complete': False, **identity, 'archives': []}
++    if 'archive_prefix' in old:
++        index['archive_prefix'] = old['archive_prefix']
+     atomic_json(path, index)
+ 
+     def save(label, files, views):
+         filename = label + '.tar'
+         files = sorted(set(files))
+         old_record = previous.get(filename)
+-        remote_path = f'{component}/{name}/{filename}'
++        remote_path = f'{remote_root}/{filename}'
+         reusable = False
+         if old_record and publisher and publisher.matches(remote_path, old_record):
+             records = []
+@@ -197,7 +200,8 @@ def package(root, output, component, name, *, destination, kind, metadata, group
+         if reusable:
+             record = old_record
+         else:
+-            record = _archive(folder / filename, files, root, Path(destination), old_record, checksums=checksums)
++            record = _archive(folder / filename, files, root,
++                              Path(index.get('archive_prefix', destination)), old_record, checksums=checksums)
+         record['views'] = views
+         index['archives'].append(record)
+         atomic_json(path, index)
+@@ -226,7 +230,7 @@ def package(root, output, component, name, *, destination, kind, metadata, group
+     index.update(complete=True, bytes=sum(row['bytes'] for row in index['archives']))
+     atomic_json(path, index)
+     if publisher:
+-        publisher.small(f'{component}/{name}/index.json', path)
++        publisher.small(f'{remote_root}/index.json', path)
+     return index
+ 
+ 
+@@ -363,6 +367,8 @@ def publish(repo, output, names, *, monitor_views=16, optional=True, shard_bytes
+         atomic_json(status, {'phase': 'running', 'repo': repo, 'pid': os.getpid(), 'finished_dataset': name,
+                              'datasets': list(recipe['datasets'])})
+     recipe['complete'] = True
++    from preprocess.ss_targets import retain_registered
++    retain_registered(recipe, repo, output)
+     atomic_json(output / 'training/index.json', recipe)
+     publisher.small('training/index.json', output / 'training/index.json')
+     atomic_json(status, {'phase': 'data_uploaded', 'repo': repo, 'pid': os.getpid(),
+@@ -440,14 +446,20 @@ def _copy_files(api, repo, output, mapping, *, revision=LEGACY_REVISION, remove_
+     return commits
+ 
+ 
+-def export_code(output, base_revision):
+-    """Export only this migration's source changes, preserving unrelated workspace edits."""
++def export_code(output, base_revision, *, tracked=None, added=None, name='reproduction'):
++    """Export selected source changes and preserve unrelated workspace edits."""
+     output = Path(output)
+     folder = output / 'code'
+     folder.mkdir(parents=True, exist_ok=True)
+-    tracked = ['HF_DATASETS.md', 'preprocess/slat/HUGGING_FACE.md', 'preprocess/slat/README.md', 'dataloader/dataset.py',
+-               'dataloader/extent.py', 'preprocess/huggingface.py', 'preprocess/validate_dataset.py']
+-    added = ['preprocess/reproduction.py', 'preprocess/tests/test_reproduction.py']
++    if tracked is None:
++        runtime = subprocess.check_output(['git', 'ls-files', '--', 'ss_multiobj_depth',
++                                           'utils/training', 'models/voxel_condition.py'], text=True).splitlines()
++        tracked = ['HF_DATASETS.md', 'preprocess/slat/HUGGING_FACE.md', 'preprocess/slat/README.md',
++                   'dataloader/dataset.py', 'dataloader/extent.py', 'preprocess/huggingface.py',
++                   'preprocess/validate_dataset.py', *runtime]
++    added = (['preprocess/reproduction.py', 'preprocess/tests/test_reproduction.py',
++              'preprocess/ss_targets.py', 'preprocess/tests/test_ss_targets.py', 'preprocess/SS_TARGETS.md']
++             if added is None else added)
+     content = subprocess.check_output(['git', 'diff', '--binary', base_revision, '--', *tracked, *added])
+     for path in added:
+         present = subprocess.run(['git', 'cat-file', '-e', f'{base_revision}:{path}'],
+@@ -461,11 +473,11 @@ def export_code(output, base_revision):
+         if result.returncode not in (0, 1):
+             raise RuntimeError(f'Cannot export restore source: {path}')
+         content += result.stdout
+-    patch = folder / 'reproduction.patch'
++    patch = folder / f'{name}.patch'
+     patch.write_bytes(content)
+     manifest = {'base_git_revision': base_revision, 'patch_sha256': sha256(patch),
+                 'files': {path: sha256(path) for path in [*tracked, *added]}}
+-    atomic_json(folder / 'manifest.json', manifest)
++    atomic_json(folder / ('manifest.json' if name == 'reproduction' else f'{name}_manifest.json'), manifest)
+     return manifest
+ 
+ 
+@@ -556,7 +568,11 @@ def write_card(output, recipe, repo):
+         '![3D-FRONT RGB](previews/3dfront/train-rgb.png)', '',
+         '![3D-FRONT visible masks](previews/3dfront/train-overlay.png)', '',
+         '[Posed object geometry](previews/3dfront/train-geometry.glb)', '']
+-    (output / 'README.md').write_text('\n'.join(lines))
++    card = '\n'.join(lines)
++    if any(row.get('ss_targets') for row in recipe['datasets'].values()):
++        from preprocess.ss_targets import update_card
++        card = update_card(card, recipe, repo)
++    (output / 'README.md').write_text(card)
+ 
+ 
+ def finalize(repo, output, *, deprecate=False):

code/evaluation_manifest.json

@@ -1,15 +1,15 @@
 {
   "base_git_revision": "290027fe9f281a21db978a48da21bb9be90e5517",
-  "patch_sha256": "4f5c19b5f257d3b99a8a423fca841de85f551d71de04af5b7d80c4a7a1673698",
+  "patch_sha256": "668d02c9a918602dc64f02d41d18f0bafe03d80dd740d78592e6bda4a801c70d",
   "files": {
     "dataloader/registry.py": "97690914f4c50c9f7d31213bb9dab174bc740cdd62bf7b217ccfe456b6b387e7",
     "evaluation/artifacts.py": "232f1b6ffd0374813ea93857ecb378c317a962b2bfbbb4cce0fd7d28e88fe12c",
     "evaluation/evaluate.py": "3d9115be9eb34e1e6193a31f3fb4e5bf770f60b70701dbf7d67017376b87ed17",
     "evaluation/workflows.py": "0623408a288a31d5502531f377c5492b985b5dc69429858d21180eaf0b06d2ba",
-    "preprocess/huggingface.py": "c747bf74aaa8ac262cfd5459199cba08e4fe574b1d5bae830dccb3abaa36398f",
-    "preprocess/reproduction.py": "7f5a50c5dfac603ab529ee4a4aab9a60a5aad8f748dfa03bfa270db0e233c069",
+    "preprocess/huggingface.py": "4f83091cec7436760843f3261deb0ea5f0090a9c4d53918e2a543fe40f8ea31a",
+    "preprocess/reproduction.py": "cb0d302385fb7ca465cdf6970edeffcbad301f61a07770bb70d4b00d33549f22",
     "dataloader/frozen.py": "5efdb2209519d556f363599226d0d511de14009529899813b48a5574432bc508",
-    "preprocess/evaluation.py": "3e3b826fc5f9a50dd0e5c85a4094ec60ae17b397b357d9800516a65ea1615b3e",
+    "preprocess/evaluation.py": "e14f82038cb48a26098dc014f6b2db196e805ea5bca5402c560ac1b8ed944c8c",
     "evaluation/configs/picasso_frozen.json": "5e691ad4b2266fb16f43c24c548a98bf3b68f139d471b956a36d0996a5b6492b",
     "evaluation/configs/messykitchen_frozen.json": "b85423fbc37b16d64b57da59cb6bbde41194c1bf39cd065e72f7203f9f223a46",
     "evaluation/configs/front3d_frozen.json": "661d67d3ba5f2a6a40f7dfbdbc5075e6f200ab14f94c8c90be0f8d285a54a07f"

{optional/evaluation → evaluation}/README.md

@@ -1,6 +1,8 @@
 # PartProbe3D frozen evaluation
 
 This bundle contains 2,603 single-image evaluation samples and their GT object meshes.
+The PartProbe3D benchmark requires this dataset.
+The Hugging Face repository stores this dataset under `evaluation/`.
 It requires about 23.4 GB, including RGB images, visible masks, previews, and shared geometry.
 It contains no training samples, depth maps, model weights, or image features.
 
@@ -25,11 +27,11 @@ hf download cubec/PartProbe3D-preprocessed code/evaluation.patch --repo-type dat
 git apply --check outputs/huggingface/bootstrap/code/evaluation.patch
 git apply outputs/huggingface/bootstrap/code/evaluation.patch
 python -m pip install -e '.[preprocess]'
-python -m preprocess.huggingface download --repo cubec/PartProbe3D-preprocessed --optional evaluation
+python -m preprocess.huggingface download --repo cubec/PartProbe3D-preprocessed --dataset evaluation
 ```
 
 The downloader pins one Hub revision and verifies every archive checksum before restoration.
-The optional bundle restores to `datasets/evaluation/`.
+The required evaluation bundle restores to `datasets/evaluation/`.
 Default training downloads exclude this bundle.
 The source manifest at `code/evaluation_manifest.json` records the loader files and patch checksum.
 

{optional/evaluation → evaluation}/index.json

@@ -31,8 +31,8 @@
     {
       "file": "metadata.tar",
       "bytes": 13086720,
-      "sha256": "1fe0fb12ecbac88f325b161e3bd5479247a440dfe35fdcdbe89a8bb148932849",
-      "identity": "7293a27f5a6c5076c4be6deff6915409a3040040c2a1af10b8ae7ea6914799f8",
+      "sha256": "9e5e5ea7f7b3dd012b6f9484f14d41758e0ed5843e0cec682a25ebc0c507989c",
+      "identity": "1d3d20302be51b87309bef221f8516ced8bb82d0a8644aae6844c0c63a8704b0",
       "files": 9,
       "views": 0
     },
@@ -238,5 +238,6 @@
     }
   ],
   "bytes": 23404277760,
-  "archive_prefix": "datasets/evaluation/indoor"
+  "archive_prefix": "datasets/evaluation/indoor",
+  "required_for_evaluation": true
 }

{optional/evaluation → evaluation}/metadata.tar

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:1fe0fb12ecbac88f325b161e3bd5479247a440dfe35fdcdbe89a8bb148932849
+oid sha256:9e5e5ea7f7b3dd012b6f9484f14d41758e0ed5843e0cec682a25ebc0c507989c
 size 13086720

training/index.json

@@ -84,18 +84,22 @@
     "evaluation/ycbv": {
       "component": "optional/evaluation/ycbv",
       "bytes": 1669857280
-    },
-    "evaluation": {
-      "component": "optional/evaluation",
-      "bytes": 23404277760,
-      "views": 2603,
-      "datasets": {
-        "picasso": 100,
-        "messykitchen": 1503,
-        "3dfront": 1000
-      },
-      "loader_patch": "code/evaluation.patch"
     }
   },
-  "base_git_revision": "aac7b493e1e5e10512857a2582233c99fedda9c5"
+  "base_git_revision": "aac7b493e1e5e10512857a2582233c99fedda9c5",
+  "evaluation": {
+    "component": "evaluation",
+    "bytes": 23404277760,
+    "views": 2603,
+    "datasets": {
+      "picasso": 100,
+      "messykitchen": 1503,
+      "3dfront": 1000
+    },
+    "loader_patch": "code/evaluation.patch",
+    "required": [
+      "evaluation"
+    ],
+    "required_for_evaluation": true
+  }
 }