---
pretty_name: DiBO Design-Bench Reproducibility Snapshot
tags:
  - design-bench
  - black-box-optimization
  - offline-black-box-optimization
  - reproducibility
  - dibo
---

# DiBO Design-Bench Reproducibility Snapshot

This repository contains the exact `design_bench_data.zip` snapshot used in
the experiments for [Training Diffusion Language Models for Black-Box Optimization](https://huggingface.co/papers/2603.17919).
The artifact supports reproduction of the released DiBO training,
data-preparation, and evaluation workflows.

## Scope

This repository is a DiBO reproducibility snapshot. It is not presented as the
official, canonical, complete, or latest Design-Bench distribution.

Users reproducing the released DiBO experiments should use this snapshot. For
upstream Design-Bench software, task definitions, and documentation, see
[Design-Bench](https://github.com/brandontrabucco/design-bench). For a more
comprehensive community-maintained/reference data mirror, see
[beckhamc/design_bench_data](https://huggingface.co/datasets/beckhamc/design_bench_data).

## Comparison with the reference mirror

The snapshot was compared file-by-file with revision `79006c5105c6d0da75b2d9b599f31a5e74136778` of
`beckhamc/design_bench_data`. After removing the local archive's top-level
`design_bench_data/` prefix:

- 244 file paths were shared;
- all 244 shared files were byte-identical;
- every shared file used by TFBind8, TFBind10, Ant Morphology, and D'Kitty
  Morphology was byte-identical;
- the local snapshot and the reference mirror differ in coverage.

The DiBO snapshot is smaller and does not contain every task or auxiliary file
in the more comprehensive reference mirror. See
[`SOURCE_COMPARISON.md`](SOURCE_COMPARISON.md) for exact methodology,
task-level counts, local-only files, reference-only files, and hashes.

## Important preprocessing note

`design_bench_data.zip` is not itself the final DiBO `data/` directory.
Unzipping it does not directly produce all DiBO-specific task bundles,
relabeled morphology scores, prompt artifacts, reward statistics, or
normalization metadata. Use the released DiBO data-preparation pipeline and
follow [`data/README.md`](https://github.com/zpointS/DiBO/blob/main/data/README.md)
and [`docs/installation.md`](https://github.com/zpointS/DiBO/blob/main/docs/installation.md).

## Download

```bash
hf download zpointsun/DiBO-DesignBench-Snapshot design_bench_data.zip \
  --repo-type dataset --revision v1.0.0 --local-dir data/downloads

python scripts/prepare_design_bench_data.py \
  --archive data/downloads/design_bench_data.zip
```

## Artifact details

- Archive: `design_bench_data.zip`
- Archive byte size: 142,923,783
- Archive SHA-256: `6a642d8b78f79ffb82ee6d7b73e2c4569c92e2f31a325cea32c4e7ef68e32e81`
- ZIP entries: 261 (247 files and 14 directories)
- Uncompressed bytes: 593,018,567

The complete decompressed-member manifest is in `FILE_MANIFEST.csv`.

## Citation

If you find DiBO helpful, please cite:

```bibtex
@article{sun2026training,
  title={Training diffusion language models for black-box optimization},
  author={Sun, Zipeng and Chen, Can and Yuan, Ye and Wu, Haolun and Gu, Jiayao and Pal, Christopher and Liu, Xue},
  journal={arXiv preprint arXiv:2603.17919},
  year={2026}
}
```