Datasets:
pretty_name: SwissTest
license: mit
language:
- en
task_categories:
- text-generation
tags:
- biology
- protein-design
- protein-sequence
- interpro
- gene-ontology
- benchmark
size_categories:
- 1K<n<10K
configs:
- config_name: go
data_files:
- split: test
path: data/go/test-00000-of-00001.parquet
- config_name: ipr
data_files:
- split: test
path: data/ipr/test-00000-of-00001.parquet
- config_name: ipr_go
default: true
data_files:
- split: test
path: data/ipr_go/test-00000-of-00001.parquet
SwissTest
SwissTest contains English instructions for function-guided protein sequence design, reference protein sequences, and GO molecular function or InterPro annotations. It is a test dataset used in PDFBench. The three subsets (Hugging Face configurations) represent different conditioning information for overlapping sets of reference sequences.
The dataset is packaged as one repository, Knlife/SwissTest, with three selectable
subsets: go, ipr, and ipr_go. The default subset is ipr_go.
Subsets and splits
| Subset (configuration) | Split | Task records | Unique reference sequences | Excess exact duplicate records |
|---|---|---|---|---|
go |
test | 693 | 689 | 3 |
ipr |
test | 870 | 865 | 5 |
ipr_go |
test | 674 | 670 | 3 |
go provides GO molecular function annotations, ipr provides InterPro annotations,
and ipr_go provides both. Only test is supplied; no training or validation split
has been created. The default configuration is ipr_go.
The 2,237 task records contain 884 distinct reference sequences across all configurations.
The sets for go and ipr share 670 distinct sequences. All 670 distinct sequences
in ipr_go also occur in both other configurations. Configurations must not be treated
as disjoint train/test partitions. Annotations present in multiple configurations agree
for their shared reference sequences.
Loading
Select a subset by passing its name as the second argument:
from datasets import load_dataset
test = load_dataset("Knlife/SwissTest", "ipr_go", split="test")
go_test = load_dataset("Knlife/SwissTest", "go", split="test")
ipr_test = load_dataset("Knlife/SwissTest", "ipr", split="test")
For a local copy, pass its directory instead of the repository ID. The README declares the data files for each configuration, so the original JSON archives and audit metadata are excluded from loading.
Record schema
All configurations use the same schema. Unavailable annotation types are empty lists; annotations are never copied from another configuration to fill those lists.
| Field | Type | Meaning |
|---|---|---|
id |
string | Configuration and zero-based source row, such as go-000000. Unique per task record in this snapshot. |
source_index |
int64 | Zero-based row index in the corresponding original JSON array. |
protein_id |
string | sha256: followed by the SHA-256 digest of the exact reference sequence. Shared by identical sequences; not a UniProt accession. |
instruction |
string | Original design instruction, including whitespace and punctuation. |
reference |
string | Original reference protein sequence. |
sequence_length |
int64 | Number of characters in reference. |
go_molecular_function |
list of structs | GO annotations with string fields id and name. |
interpro |
list of structs | InterPro annotations with string fields id, name, type and int64 fields begin, end. |
GO fields are renamed from GO-ID and GO-Name. InterPro fields are renamed from
InterPro-ID, InterPro-Name, InterPro-Type, Beg, and End. Their values and list
order are preserved. Coordinate values are retained exactly as supplied; no offset
conversion is applied. They satisfy 1 <= begin <= end <= sequence_length in these files.
These locations describe annotations on the reference sequence.
The original response#1 field equals reference in every one of the 2,237 records.
It is omitted from the normalized table. It must not be interpreted as a model-generated
prediction in these source files. The complete original records remain available in
original/.
Preservation and validation
The conversion preserves every task record, its source order, the exact reference sequence and instruction, and all annotation values and ordering. Exact duplicate records are retained to preserve the original benchmark size and weighting. Distinct reference proteins can share the same instruction.
All 2,930 GO annotation names contain leading or trailing whitespace. This whitespace
is preserved, together with the original punctuation and repeated terms in instructions.
Reference sequences range from 10 to 7,785 characters. The ipr configuration includes
the unknown residue symbol X; sequences are not truncated, filtered, or rewritten.
Several InterPro hits can refer to the same entry at different locations; these are retained.
Each Parquet file was reloaded with datasets.load_dataset, and every field was checked
against the normalized records. Reversing the field renaming and restoring response#1
reconstructed all original records. Archived JSON files match their source SHA-256 hashes.
Counts, duplicate row indices, sequence overlaps, checksums, and validation results are
recorded in metadata/audit.json.
Repository files
README.md
data/
go/test-00000-of-00001.parquet
ipr/test-00000-of-00001.parquet
ipr_go/test-00000-of-00001.parquet
original/
go.json
ipr.json
ipr_go.json
metadata/
audit.json
upstream_comparison.json
Source and construction
These records match the test files in
Knlife/SwiwwProtIPG,
at revision 888930029b5c9ea0a0801f3d86732173c987056c:
go_test.json, ipr_test.json, and ipr_go_test.json.
The local source files replace the upstream sequence field with identical
reference and response#1 fields. All other fields and the record order match.
The comparison and file hashes are recorded in
metadata/upstream_comparison.json.
The upstream card reports selection from UniProt/SwissProt between 2025-01-01 and 2025-08-25, with InterPro and GO molecular function annotations. It describes constructing instructions from annotation descriptions following Mol-Instructions. This construction history is reported by the upstream publisher; it has not been independently reproduced here. The task files contain no UniProt accessions or exact database release identifiers.
License
This dataset is released under the MIT license, following the upstream dataset card.
Citation
The upstream card provides the following PDFBench citation:
@misc{kuang2025pdfbenchbenchmarknovoprotein,
title={PDFBench: A Benchmark for De novo Protein Design from Function},
author={Jiahao Kuang and Nuowei Liu and Changzhi Sun and Tao Ji and Yuanbin Wu},
year={2025},
eprint={2505.20346},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2505.20346}
}
See the PDFBench project for project context.