GenODE temporal and molecule flow fields
Frozen flow-matching backbone checkpoints for three Monash forecasting tasks and three molecular-coordinate groups.
| Family | Tasks or members | Budgets per task/member | Checkpoints |
|---|---|---|---|
| Monash forecasting: solar, traffic, weather | 3 | 5 | 15 |
| Molecular coordinates: three groups, six members each | 18 | 5 | 90 |
All runs use training seed 0 and budgets of 4,000, 8,000, 12,000, 16,000, and 20,000 updates. Temporal fields are exported at the exact budget step. Molecule fields use the clean-validation selector within each budget; selected_step is recorded for every checkpoint. These are retrained fields for research use. Functional validation establishes complete checkpoint coverage; this upload makes no benchmark-quality claim.
Files
genode-temporal-molecule-fields.zip contains all 105 PyTorch checkpoints, architecture/configuration metadata, normalization and selection information, and the relative-path backbone manifest. The checkpoint bytes are unchanged from the validated retrained artifacts. fields-manifest.json records identities, selected steps, parameter hashes, and file SHA-256 values. package-contents.json and SHA256SUMS.txt support integrity verification.
Download and unpack
from pathlib import Path
from huggingface_hub import snapshot_download
import zipfile
download = Path(snapshot_download("mpstoryfans/genode-temp-mole-fields"))
with zipfile.ZipFile(download / "genode-temporal-molecule-fields.zip") as package:
package.extractall("genode-fields")
The backbone manifest is then at genode-fields/outputs/backbone_matrix/backbone_manifest.json. Checkpoints require the matching flow-field architecture described by their stored configuration. For checkpoint inspection, use PyTorch's restricted weights_only=True loader. This package does not contain the forecasting datasets. The processed molecule datasets are available from the molecule dataset repository.