InstructPro2Bench

Training and evaluation data for InstructPro2: Natural Language and Biological Grounding Guided De Novo Protein Design — natural-language function descriptions paired with protein sequences, plus the ligand and antibody splits used for the two downstream design tasks.

Code: chovak/InstructPro2 · Weights: charlesnovak/InstructPro2 · Paper: arXiv:2610.19592

Download

export WORK_ROOT=/path/to/workspace    # default: ~/instructpro2
huggingface-cli download charlesnovak/InstructPro2Bench --repo-type dataset --local-dir $WORK_ROOT/data

Files

File	Rows	Size
stage1/train.jsonl	200,000	603 MB
stage1/valid.jsonl	1,000	2.8 MB
stage1/test.jsonl	1,000	3.0 MB
ligand/train.jsonl	50,000	183 MB
ligand/valid.jsonl	3,263	14 MB
ligand/test_ligand.jsonl	1,500	6.4 MB
ligand/test_ligand_unseen.jsonl	127	0.5 MB
antibody/train.jsonl	21,384	40 MB
antibody/valid.jsonl	433	1.2 MB
antibody/final_antibody_test.jsonl	444	1.3 MB

Both training sets are subsampled: stage1/train.jsonl is 200,000 rows of the full 205.7M-pair pretraining corpus, and ligand/train.jsonl is 50,000 rows. Every other split is complete. The full Stage 1 corpus can be rebuilt from UniProt with the dataset construction code.

Schema

One JSON object per line.

Stage 1 (stage1/)

Key	Contents
tokenized_instruction	PubMedBERT token ids — the model input
output	target protein sequence, with <SEP> and <M1>–<M30> on infilling examples
instruction	the same instruction as text
task_type	infilling or autoregressive
metadata	id (UniProt accession), cluster, interpro_all

Stage 2 (ligand/, antibody/)

Key	Contents
tokenized_instruction	PubMedBERT token ids — the model input
target_protein	target sequence; HCDR spans are masked with <M1>–<M3> in the antibody splits
text	the same instruction as text
ligand or antigen	SMILES strings, or antigen chain sequences
modality_type	ligand or antibody
metadata	ligand: uniprot_id, interpro_all, interpro_used, source. antibody: pdb_id, CDR sequences and positions, light chain, cluster, design blueprint

The loader requires tokenized_instruction and target_protein (or output), plus ligand/antigen for the Stage 2 tasks.

Sources

• Stage 1 — UniProtKB, with function descriptions composed from InterPro annotations.
• Ligand — InstructProBench, from InstructPro.
• Antibody — SAbDab and the PDB.

Citation

@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}
}