---
pretty_name: GenBench prepared benchmark datasets
license: other
---

# GenBench prepared datasets

This repository contains the exact prepared arrays used by the canonical
GenBench training and evaluation runs. It intentionally excludes redundant
source archives: every benchmark's Python preparation module records and
verifies the original upstream source, while these files are sufficient to
train and evaluate the published baselines directly. DeepSTARR's small official
activity-predictor weights are included because they are part of its evaluation
protocol.

| Directory | Generative object | Prepared contents |
|---|---|---|
| `QM9` | Molecular geometries | Packed characterized molecules and TD-jumps split rows |
| `MiniBooNE` | Particle events | Literature split and normalization |
| `NavierStokes` | Vorticity fields | Fourier-downsampled train/validation/test arrays |
| `JetNet30` | Particle clouds | Five-class train/validation/test arrays |
| `DeepSTARR` | Enhancer sequences | Splits, activities, metadata, and official predictor weights |
| `GuacaMol` | Drug-like molecules | Official non-overlapping ChEMBL splits as losslessly reduced canonical-SMILES endpoint arrays and lengths |
| `SpeechCommands` | One-second spoken-word waveforms | Official speaker-disjoint train/validation/test arrays and labels |

The analytic spiral and checkerboard benchmarks have no stored dataset; their
target distributions are generated exactly by the GenBench Python package.

`manifest.json` is authoritative. It records the byte size and SHA-256 of every
required file. Original licenses and redistribution terms differ by dataset;
consult each dataset's metadata and upstream source before reuse. CIFAR-10 and
MNIST are deliberately absent because their upstream distributions do not
provide an affirmative general redistribution grant.