README.md

---
pretty_name: CyberBioBench-100
language: [en]
license: cc-by-4.0
task_categories: [question-answering, text-classification]
tags: [cybersecurity, incident-response, secure-code-review, cloud-security, bioinformatics, benchmark]
size_categories: [n<1K]
---

# CyberBioBench-100

A synthetic, evidence-grounded **defensive cybersecurity benchmark** with **100 tasks**:

- 91 general cybersecurity tasks across 13 domains
- exactly 9 bioinformatics-security tasks
- 70 train, 15 validation, and 15 test tasks
- four evidence statements per task
- gold severity, findings, evidence IDs, remediation, and critical-error guidance

The benchmark contains no exploit payloads, malware code, wet-lab procedures, or instructions for unauthorized access.

## Filters
Filtered from https://huggingface.co/datasets/sunblaze-ucb/cybergym
## Files

```text
CyberBioBench-100/
├── README.md
├── LICENSE.txt
├── schema.json
├── dataset_summary.json
├── data/
│   ├── train.jsonl
│   ├── validation.jsonl
│   ├── test.jsonl
│   └── all_with_gold.jsonl
└── evaluator/
    ├── evaluate.py
    ├── sample_predictions.jsonl
    └── test_gold.jsonl
```

Train and validation include gold annotations. Test omits them. Keep `evaluator/test_gold.jsonl` private for a public leaderboard.

## Required model output

```json
{
  "task_id": "CBB-001",
  "primary_finding": "Concise diagnosis",
  "severity": "high",
  "supporting_evidence": ["E1", "E3"],
  "recommended_actions": ["First action", "Second action"],
  "rationale": "Brief evidence-grounded explanation"
}
```

## General domains

- `cloud_security`: 7 tasks
- `container_kubernetes`: 7 tasks
- `data_protection_crypto`: 7 tasks
- `email_social_engineering`: 7 tasks
- `endpoint_detection_response`: 7 tasks
- `identity_access`: 7 tasks
- `incident_response`: 7 tasks
- `iot_ot_security`: 7 tasks
- `network_security`: 7 tasks
- `secure_code_review`: 7 tasks
- `software_supply_chain`: 7 tasks
- `vulnerability_management`: 7 tasks
- `web_application_security`: 7 tasks

## Nine bioinformatics-security tasks

1. Public genomic variant storage
2. Outdated HTSlib parser
3. Overprivileged workflow identity
4. FASTQ transfer integrity mismatch
5. Notebook token exposure
6. Mutable workflow container
7. Unsafe CRAM reference resolution
8. Sensitive variant-annotation logging
9. Flat-network sequencing instrument

These tasks assess cybersecurity around genomic data and computational pipelines. They do not assess biological experimentation or wet-lab safety.

## Evaluation

The included transparent scorer assigns up to 10 points:

- severity: 2
- evidence selection: 2
- finding concept overlap: 3
- remediation concept overlap: 3

```bash
python evaluator/evaluate.py \
  --predictions evaluator/sample_predictions.jsonl \
  --gold evaluator/test_gold.jsonl
```

For serious comparisons, add blinded expert review for factual correctness, prioritization, hallucinated evidence, critical errors, and operational usefulness.

## Recommended reported metrics

- mean score out of 10
- severity accuracy
- evidence precision and recall
- general cybersecurity mean
- bioinformatics subset mean
- difficulty-stratified means
- critical-error rate
- hallucinated-evidence rate

## Limitations

All scenarios are synthetic. The benchmark evaluates written defensive reasoning rather than live system operation. Severity labels apply to the supplied context and may change under different business conditions. The lexical scorer is a baseline, not a replacement for expert judgment.

## Creator
Create by OpenAI's GPT 5.6 Sol Max

## Warning
This is only for testing purposes.  Do not train this dataset.

## Citation

```bibtex
@dataset{{cyberbiobench100_2026,
  title={{CyberBioBench-100: A Synthetic Defensive Cybersecurity Benchmark with a Nine-Task Bioinformatics Security Subset}},
  year={{2026}},
  note={{Synthetic benchmark dataset}}
}}
```