Request access to SPIKE-Bench
This dataset contains toxin-design prompts. Access is granted to individual verified researchers and every request is reviewed by hand. Please request from an institutional address and describe your intended use; incomplete requests are declined.
By requesting access you agree that you will not redistribute these files or any derivative of them, and that you will not use them to produce or attempt to produce functional biological agents. Access may be withdrawn at any time.
Log in or Sign Up to review the conditions and access this dataset content.
SPIKE-Bench
The prompt sets for A Blind Spot in Alignment: Quantifying Biosecurity Risks in Large Language Models (COLM 2026). The evaluation pipeline is public at https://github.com/Quanshu01/SPIKE-Bench; the prompts are here, behind manual review, as is the classifier at quanshu01/BioSafe-Guard.
No model-generated sequence that passes the SPIKE funnel is distributed in this repository or anywhere else.
Files
| File | Rows | Contents |
|---|---|---|
spike_bench.jsonl |
631 | Toxin-design prompts across seven functional categories — the main benchmark |
ood_prompts.jsonl |
200 | Adversarial reformulations of benchmark prompts, four strategies × 50 |
biosafe_guard/train_toxin.jsonl |
300 | Positive class for BioSafe-Guard training |
benign_prompts.jsonl |
300 | Benign over-refusal evaluation |
biosafe_guard/train_benign.jsonl |
300 | Negative class for BioSafe-Guard training |
The five sets are disjoint from one another. The benign sets carry no toxin annotation; they are distributed here rather than separately so that the whole benchmark has a single home and a single access decision.
Functional categories in spike_bench.jsonl
| Code | Category | Count |
|---|---|---|
HC |
Hemostatic & Cellular Toxins | 155 |
K |
K+ Channel Toxins | 117 |
Na |
Na+ Channel Toxins | 85 |
GIC |
General Ion Channel Toxins | 81 |
RS |
Receptor & Synaptic Neurotoxins | 77 |
UN |
Unspecified Neurotoxins | 76 |
Ca |
Ca2+ Channel Toxins | 40 |
Adversarial strategies in ood_prompts.jsonl
keyword_free, obfuscated, dual_use and roleplay, 50 prompts each, in the
ood_strategy field.
Schema
spike_bench.jsonl, ood_prompts.jsonl, benign_prompts.jsonl carry one JSON object
per line:
| Field | Type | Description |
|---|---|---|
id |
string | UniProt accession of the source entry |
prompt |
string | The protein-design instruction given to the model |
sequence |
string | Swiss-Prot reference sequence for that accession |
original_fasta_header |
string | Full Swiss-Prot FASTA header |
taxonomy |
object | domain, kingdom, phylum of the source organism |
toxin_annotations |
object | best_label and other_labels from UniProt Tox-Prot; absent in the benign set |
category / category_name |
string | Functional category code and its name; toxin sets only |
ood_strategy |
string | ood_prompts.jsonl only — which reformulation produced the prompt |
biosafe_guard/train_toxin.jsonl uses a different shape:
generated_prompt_for_sequence_model holds the prompt, and original_data nests
original_fasta_header, protein_sequence, llm_full_response and is_sufficient.
biosafe_guard/train_benign.jsonl follows the four-field schema above.
Provenance
Curated from manually reviewed UniProtKB/Swiss-Prot entries — with toxin annotations for the toxin sets, without them for the benign sets. Each entry's metadata was expanded into a protein-design task with one fixed template, so that a classifier cannot separate the two on surface form alone.
Usage
Access requires an approved request and an authenticated token:
from huggingface_hub import snapshot_download
snapshot_download(
repo_id='quanshu01/SPIKE-Bench',
repo_type='dataset',
local_dir='data',
allow_patterns=['*.jsonl'],
)
The files land at the paths the pipeline expects:
python -m spike_bench.compute_fhr --results <merged.jsonl> --prompts data/spike_bench.jsonl
See the GitHub README for the full pipeline.
License
CC BY-NC 4.0: attribution required, non-commercial use only.
- Downloads last month
- 17