---
license: mit
task_categories:
  - text-generation
  - fill-mask
tags:
  - biology
  - protein-design
  - ligand-binding
  - antibody-design
configs:
  - config_name: stage1
    data_files:
      - split: train
        path: stage1/train.jsonl
      - split: validation
        path: stage1/valid.jsonl
      - split: test
        path: stage1/test.jsonl
  - config_name: ligand
    data_files:
      - split: train
        path: ligand/train.jsonl
      - split: validation
        path: ligand/valid.jsonl
      - split: test
        path: ligand/test_ligand.jsonl
      - split: test_unseen
        path: ligand/test_ligand_unseen.jsonl
  - config_name: antibody
    data_files:
      - split: train
        path: antibody/train.jsonl
      - split: validation
        path: antibody/valid.jsonl
      - split: test
        path: antibody/final_antibody_test.jsonl
---

# InstructPro2Bench

Training and evaluation data for
[InstructPro2](https://huggingface.co/charlesnovak/InstructPro2), which
designs proteins from natural-language function descriptions.

[Paper](https://arxiv.org/abs/2610.19592) ·
[Code](https://github.com/chovak/InstructPro2) ·
[Model](https://huggingface.co/charlesnovak/InstructPro2)

## Contents

```
stage1/            instruction-following pretraining
  train.jsonl  valid.jsonl  test.jsonl
ligand/            ligand-binding protein design
  train.jsonl  valid.jsonl  test_ligand.jsonl  test_ligand_unseen.jsonl
antibody/          antibody HCDR design
  train.jsonl  valid.jsonl  final_antibody_test.jsonl
eval_structures/   ground-truth structures for Stage 1 evaluation
  cluster_<id>/    predicted structures for one cluster
  targets.csv      row -> cluster manifest
```

| Split | Rows | |
| --- | ---: | --- |
| `stage1/train` | 200,000 | sample of 205,727,986 |
| `stage1/valid` | 1,000 | complete |
| `stage1/test` | 1,000 | complete |
| `ligand/train` | 50,000 | sample of 8,727,514 |
| `ligand/valid` | 3,263 | complete |
| `ligand/test_ligand` | 1,500 | complete |
| `ligand/test_ligand_unseen` | 127 | complete; held-out ligands |
| `antibody/train` | 21,384 | complete |
| `antibody/valid` | 433 | complete |
| `antibody/final_antibody_test` | 444 | complete |

The two Stage 1 and ligand training splits are samples of much larger
datasets, the full sets are available on request. Every other split is
complete.

About a third of the antibody training rows (7,068 of 21,384) carry an
antigen. The rest are heavy chains without a paired structure, and train
the infilling behaviour alone.

Every row is one JSON object per line:

| Field | Present in | |
| --- | --- | --- |
| `instruction` | stage1 | the natural-language design request |
| `text` | ligand, antibody | the same, under a different name |
| `tokenized_instruction` | all | the request as PubMedBERT token ids, which is what the model reads |
| `output` | stage1 | the reference protein |
| `target_protein` | ligand, antibody | the same, under a different name |
| `ligand` | ligand | SMILES strings |
| `antigen` | antibody | antigen chain sequences |
| `task_type` | stage1 | `autoregressive` or `infilling` |
| `modality_type` | ligand, antibody | `ligand` or `antibody` |
| `metadata` | all | provenance: UniProt or PDB id, InterPro annotations, cluster id, CDR positions |

A target may contain `<SEP>` and `<M1>`–`<M30>`, which mark spans to
infill: the scaffold precedes the separator and the spans that fill it
follow. Antibody targets use this to mark HCDR1-3 in a heavy chain.

## Evaluation structures

`eval_structures/` holds the ground truth for Stage 1 scoring: 1,000
cluster folders, each with the structures of related proteins, and
`targets.csv` mapping each test row to its cluster. A design is scored
against every member of the cluster, and the best match is reported.

Ground-truth structures for the ligand and antibody tasks are not included.

## Usage

```bash
huggingface-cli download charlesnovak/InstructPro2Bench \
    --repo-type dataset --local-dir $WORK_ROOT/data
```

The [code repository](https://github.com/chovak/InstructPro2) reads this
layout directly. To design proteins from your own descriptions instead, see
its README — you do not need this dataset for that.

## Notes

Proteins longer than 1,024 tokens are outside the model's context and are
skipped at generation time, so some test rows produce no design and have no
ground-truth structure.

## Citation

```bibtex
@article{instructpro2,
  title   = {InstructPro2: Natural Language and Biological Grounding Guided De Novo Protein Design},
  author  = {Charles Novak and Zhenqiao Song and Chuan Li and Jianwen Xie and Lei Li},
  journal = {arXiv preprint arXiv:2610.19592},
  year    = {2026}
}
```

## License

MIT.