CyberBench (NLP): leaderboard
Published CyberBench average over ten cybersecurity NLP datasets, including entity recognition, classification, summarization and security Q&A. CyNews first averages ROUGE-1/2/L; the source uses retrieved five-shot examples except for summarization.
Metric: Published ten-dataset average (%). Source: github.com. Saturation forecast: Estimated already saturated. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4 | 69.6 | #438 |
| 2 | GPT-3.5 Turbo | 62.6 | #865 |
| 3 | Mistral-7B-v0.1 | 58.1 | #1326 |
| 4 | zephyr-7B-beta | 57.7 | #1330 |
| 5 | vicuna-13B-v1.5 | 57.3 | #1392 |
| 6 | Mistral Instruct v0.1 7B | 55 | #1454 |
| 7 | Llama 2 13B | 54.1 | #1436 |
| 8 | Llama 2 7B | 50.6 | #1520 |
| 9 | Llama 2 13B Chat | 45 | #1465 |
| 10 | Llama 2 7B Chat | 44.6 | #1499 |
| 11 | falcon-7B | 39.4 | #1598 |
| 12 | falcon-7B Instruct | 37.5 | #1622 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=cyberbench-nlp · How It Works · Data refreshed daily, snapshot 2026-10-20.