johnpolat commited on
Commit
f83fdf9
·
verified ·
1 Parent(s): 5005dc2

Add RADII v1.0 loader + dataset card

Browse files
Files changed (2) hide show
  1. RADII.py +154 -0
  2. README.md +106 -0
RADII.py ADDED
@@ -0,0 +1,154 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """HuggingFace `datasets` loader for the RADII benchmark.
2
+
3
+ Downloads the seed archive from Zenodo (DOI: 10.5281/zenodo.20431021) and
4
+ invokes the `radii` Python package to regenerate the full benchmark locally.
5
+ The regenerated cache is then iterated row-by-row and yielded into HF's
6
+ standard cache so subsequent calls are instant.
7
+
8
+ User setup:
9
+ pip install datasets "radii @ git+https://github.com/KurbanIntelligenceLab/RADII"
10
+ python -c "from datasets import load_dataset; \\
11
+ ds = load_dataset('KurbanIntelligenceLab/RADII', split='train', \\
12
+ trust_remote_code=True)"
13
+
14
+ The first call regenerates the ~12 GB local cache (deterministic from the
15
+ 3 MB seed). Subsequent calls hit HF's own cache.
16
+ """
17
+ import json
18
+ from pathlib import Path
19
+
20
+ import datasets
21
+ import numpy as np
22
+
23
+ _CITATION = """\
24
+ @inproceedings{polat2026radii,
25
+ title = {How Far Can You Grow? Characterizing the Extrapolation Frontier of Graph Generative Models for Materials Science},
26
+ author = {Polat, Can and Serpedin, Erchin and Kurban, Mustafa and Kurban, Hasan},
27
+ booktitle = {Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining},
28
+ series = {KDD \\'26},
29
+ year = {2026}
30
+ }
31
+ """
32
+
33
+ _DESCRIPTION = """\
34
+ RADII is a radius-resolved benchmark of ~75,000 nanoparticle structures
35
+ (33-11,298 atoms across 10 materials) for measuring the extrapolation
36
+ frontier of graph generative models. Radius is treated as a continuous
37
+ scaling knob with leakage-free train / in-distribution test / out-of-
38
+ distribution test splits. The OOD test radii span structures ~59% smaller
39
+ and ~24% larger than the training envelope.
40
+
41
+ This loader downloads the seed archive from Zenodo (DOI: 10.5281/zenodo.20431021)
42
+ and regenerates the full benchmark locally on first call. Subsequent calls
43
+ hit the HF datasets cache.
44
+ """
45
+
46
+ _HOMEPAGE = "https://github.com/KurbanIntelligenceLab/RADII"
47
+ _LICENSE = "cc-by-4.0"
48
+
49
+ # Map HF split names (conventional) to metadata split labels (project convention).
50
+ _SPLIT_MAP = {"train": "train", "id_test": "ID", "ood_test": "OOD"}
51
+
52
+
53
+ class RADII(datasets.GeneratorBasedBuilder):
54
+ """Radius-resolved benchmark for graph generative models in materials science."""
55
+
56
+ VERSION = datasets.Version("1.0.0")
57
+
58
+ def _info(self):
59
+ features = datasets.Features({
60
+ "material": datasets.Value("string"),
61
+ "radius": datasets.Value("int32"),
62
+ "rot_idx": datasets.Value("int32"),
63
+ "split": datasets.Value("string"),
64
+ "num_atoms": datasets.Value("int32"),
65
+ "pos": datasets.Sequence(datasets.Sequence(datasets.Value("float32"), length=3)),
66
+ "z": datasets.Sequence(datasets.Value("int8")),
67
+ "cell_matrix": datasets.Sequence(
68
+ datasets.Sequence(datasets.Value("float32"), length=3), length=3
69
+ ),
70
+ "cell_pos": datasets.Sequence(datasets.Sequence(datasets.Value("float32"), length=3)),
71
+ "cell_z": datasets.Sequence(datasets.Value("int8")),
72
+ })
73
+ return datasets.DatasetInfo(
74
+ description=_DESCRIPTION,
75
+ features=features,
76
+ homepage=_HOMEPAGE,
77
+ license=_LICENSE,
78
+ citation=_CITATION,
79
+ )
80
+
81
+ def _split_generators(self, dl_manager):
82
+ # Regenerate the cache locally via the `radii` package.
83
+ try:
84
+ import radii # noqa: F401
85
+ from radii.download import download_seed
86
+ from radii.generation import create_radii
87
+ except ImportError as e:
88
+ raise RuntimeError(
89
+ "The RADII HuggingFace loader requires the `radii` package. Install:\n"
90
+ " pip install 'radii @ git+https://github.com/KurbanIntelligenceLab/RADII'"
91
+ ) from e
92
+
93
+ cache_root = Path.home() / ".cache" / "radii"
94
+ (cache_root / "raw").mkdir(parents=True, exist_ok=True)
95
+
96
+ samples_npz = cache_root / "cache" / "samples.npz"
97
+ metadata_json = cache_root / "cache" / "metadata.json"
98
+ unit_cells_npz = cache_root / "cache" / "unit_cells.npz"
99
+
100
+ if not (samples_npz.exists() and metadata_json.exists() and unit_cells_npz.exists()):
101
+ print(f"[RADII] No local cache at {cache_root}. Downloading seed + regenerating...")
102
+ seed = download_seed(dest=cache_root / "raw")
103
+ create_radii(raw_data_dir=str(seed), output_dir=str(cache_root))
104
+
105
+ return [
106
+ datasets.SplitGenerator(
107
+ name=getattr(datasets.Split, hf_split.upper()) if hf_split == "train"
108
+ else datasets.Split(hf_split),
109
+ gen_kwargs={"cache_root": str(cache_root), "hf_split": hf_split},
110
+ )
111
+ for hf_split in _SPLIT_MAP
112
+ ]
113
+
114
+ def _generate_examples(self, cache_root, hf_split):
115
+ cache_root = Path(cache_root)
116
+ raw_split = _SPLIT_MAP[hf_split]
117
+
118
+ with (cache_root / "cache" / "metadata.json").open() as f:
119
+ meta = json.load(f)
120
+ sample_meta = meta["samples"]
121
+
122
+ samples = np.load(cache_root / "cache" / "samples.npz", allow_pickle=True)
123
+ pos_all, z_all, offsets = samples["pos"], samples["z"], samples["offsets"]
124
+
125
+ unit_cells = np.load(cache_root / "cache" / "unit_cells.npz", allow_pickle=True)
126
+ materials_map = {}
127
+ for k in unit_cells.files:
128
+ if k.endswith("_pos"):
129
+ mat = k[:-4]
130
+ materials_map[mat] = {
131
+ "pos": unit_cells[f"{mat}_pos"].astype(np.float32),
132
+ "z": unit_cells[f"{mat}_z"].astype(np.int8),
133
+ "cell_matrix": unit_cells[f"{mat}_cell_matrix"].astype(np.float32),
134
+ }
135
+
136
+ emitted = 0
137
+ for i, m in enumerate(sample_meta):
138
+ if m["split"] != raw_split:
139
+ continue
140
+ s, e = int(offsets[i]), int(offsets[i + 1])
141
+ uc = materials_map[m["material"]]
142
+ yield emitted, {
143
+ "material": m["material"],
144
+ "radius": int(m["radius"]),
145
+ "rot_idx": int(m["rot_idx"]),
146
+ "split": hf_split,
147
+ "num_atoms": int(m["num_atoms"]),
148
+ "pos": pos_all[s:e].astype(np.float32).tolist(),
149
+ "z": z_all[s:e].astype(np.int8).tolist(),
150
+ "cell_matrix": uc["cell_matrix"].tolist(),
151
+ "cell_pos": uc["pos"].tolist(),
152
+ "cell_z": uc["z"].tolist(),
153
+ }
154
+ emitted += 1
README.md ADDED
@@ -0,0 +1,106 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: cc-by-4.0
3
+ language:
4
+ - en
5
+ pretty_name: RADII
6
+ size_categories:
7
+ - 10K<n<100K
8
+ task_categories:
9
+ - other
10
+ tags:
11
+ - materials-science
12
+ - crystal-structures
13
+ - generative-models
14
+ - benchmark
15
+ - kdd-2026
16
+ - nanoparticles
17
+ ---
18
+
19
+ # RADII: Radius-Resolved Benchmark of Nanoparticle Structures
20
+
21
+ **RADII** measures where graph generative models start to fail as the structures they
22
+ generate grow larger. Radius is treated as a continuous scaling knob from
23
+ in-distribution to out-of-distribution; the dataset spans ten materials and 25 radii
24
+ with leakage-free splits.
25
+
26
+ - **Paper:** *How Far Can You Grow? Characterizing the Extrapolation Frontier of Graph Generative Models for Materials Science* — KDD '26
27
+ - **Code:** [github.com/KurbanIntelligenceLab/RADII](https://github.com/KurbanIntelligenceLab/RADII) (MIT)
28
+ - **Archival DOI:** [10.5281/zenodo.20431021](https://doi.org/10.5281/zenodo.20431021) (CC-BY-4.0)
29
+
30
+ ## Quick start
31
+
32
+ This dataset uses a custom loader that downloads the seed (3 MB) from Zenodo
33
+ and regenerates the full ~12 GB benchmark locally on first use. **First call
34
+ takes minutes to ~1 hour depending on hardware; subsequent calls are instant.**
35
+
36
+ ```bash
37
+ pip install datasets "radii @ git+https://github.com/KurbanIntelligenceLab/RADII"
38
+ ```
39
+
40
+ ```python
41
+ from datasets import load_dataset
42
+
43
+ ds = load_dataset("KurbanIntelligenceLab/RADII", split="train", trust_remote_code=True)
44
+ print(ds.features, len(ds), ds[0])
45
+ ```
46
+
47
+ `with_format("torch")` works for one-line PyTorch integration:
48
+
49
+ ```python
50
+ ds = ds.with_format("torch")
51
+ ```
52
+
53
+ ## Splits
54
+
55
+ | Split | Size | Radii (Å) |
56
+ |---|---|---|
57
+ | `train` | 48,000 | 15 values: 8–10, 12, 14, 16, 18, 20, 22–28 |
58
+ | `id_test` | 13,500 | 6 values: 11, 13, 15, 17, 19, 21 (interleaved, unseen orientations) |
59
+ | `ood_test` | 13,480 | 4 values: 6, 7, 29, 30 (strictly outside training range) |
60
+
61
+ The OOD test structures are ~59% smaller and ~24% larger than the smallest and largest
62
+ training structures (atom counts 33–11,298 vs train 81–9,148).
63
+
64
+ ## Materials (10)
65
+
66
+ Ag, Au, CH₃NH₃PbI₃, Fe₂O₃, MoS₂, PbS, SnO₂, SrTiO₃, TiO₂, ZnO.
67
+
68
+ ## Schema (per row)
69
+
70
+ | Column | Type | Notes |
71
+ |---|---|---|
72
+ | `material` | str | One of the 10 above |
73
+ | `radius` | int | Truncation radius in Å |
74
+ | `rot_idx` | int | Orientation index |
75
+ | `split` | str | `train`, `id_test`, or `ood_test` |
76
+ | `num_atoms` | int | Length of `pos`/`z` |
77
+ | `pos` | list[list[float32]] | `num_atoms × 3` atomic coordinates |
78
+ | `z` | list[int8] | `num_atoms` atomic numbers |
79
+ | `cell_matrix` | list[list[float32]] | `3 × 3` lattice vectors of the parent unit cell |
80
+ | `cell_pos` | list[list[float32]] | Unit-cell basis positions (per-material) |
81
+ | `cell_z` | list[int8] | Unit-cell basis atomic numbers |
82
+
83
+ ## Why a custom loader?
84
+
85
+ The benchmark is deterministic from a 3 MB seed archive. Rather than duplicating
86
+ ~12 GB of derived data on this Hub, the loader fetches the seed from the citeable
87
+ Zenodo record and regenerates locally — keeping a single source of truth and
88
+ making the reproducibility story explicit. The HF datasets cache makes subsequent
89
+ calls fast.
90
+
91
+ ## Citation
92
+
93
+ ```bibtex
94
+ @inproceedings{polat2026radii,
95
+ title = {How Far Can You Grow? Characterizing the Extrapolation Frontier of Graph Generative Models for Materials Science},
96
+ author = {Polat, Can and Serpedin, Erchin and Kurban, Mustafa and Kurban, Hasan},
97
+ booktitle = {Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining},
98
+ series = {KDD '26},
99
+ year = {2026}
100
+ }
101
+ ```
102
+
103
+ ## License
104
+
105
+ CC-BY-4.0. See [10.5281/zenodo.20431021](https://doi.org/10.5281/zenodo.20431021)
106
+ for the archival record terms. The accompanying code is MIT-licensed at the GitHub repo.