# Dataset licenses and attribution

The GenBench software is MIT-licensed, but that license does **not** apply to
third-party datasets. Each stored dataset retains its upstream license. The
prepared files are adaptations: GenBench verifies the upstream release, then
performs the deterministic splitting, normalization, packing, or feature
selection described in that benchmark's metadata and preparation module.

## Publicly redistributed prepared data

| Directory | Source and attribution | License |
|---|---|---|
| `QM9` | Ramakrishnan, Dral, Rupp, and von Lilienfeld, [Data for 133885 GDB-9 molecules](https://api.figshare.com/v2/articles/1057646), including Figshare files 3195389 and 3195404 | [CC0 1.0](https://creativecommons.org/publicdomain/zero/1.0/) |
| `DeepSTARR` | Bernardo P. de Almeida, [DeepSTARR manuscript data](https://doi.org/10.5281/zenodo.5502060) | [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) |
| `JetNet30` | Raghav Kansal, Javier Duarte, Su Hao, Breno Orzari, Thiago Tomei, Maurizio Pierini, Mary Touranakou, Jean-Roch Vlimant, and Dimitrios Gunopulos, [JetNet](https://doi.org/10.5281/zenodo.6975118) | [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) |
| `MiniBooNE` | Byron Roe, [MiniBooNE particle identification](https://doi.org/10.24432/C5QC87), UCI Machine Learning Repository | [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) |
| `NavierStokes` | NeuralOperator Team, [Navier-Stokes Dataset](https://doi.org/10.5281/zenodo.12825163) | [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) |
| `GuacaMol` | BenevolentAI, [GuacaMol standardized ChEMBL distribution-learning split](https://github.com/BenevolentAI/guacamol); underlying records from ChEMBL | [CC BY-SA 3.0](https://chembl.github.io/chembl-licensing/) |
| `SpeechCommands` | Pete Warden and the TensorFlow/AIY teams, [Speech Commands V2](https://research.google/blog/launching-the-speech-commands-dataset/) | [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) |

## Source-only datasets

GenBench does not redistribute the following data. The repository contains
checksum-verified Python preparation code that downloads each release directly
from its original host for a user's own use.

| Dataset | Why it is source-only |
|---|---|
| CIFAR-10 | The official release specifies no dataset license or redistribution grant. |
| MNIST | The official distribution and the CVDF mirror specify no general redistribution license. |

This file records upstream terms for reproducibility and is not legal advice.