---
license: cc-by-4.0
pretty_name: AriaCyberBench Security Evaluation Lab
task_categories:
  - text-classification
  - text-generation
  - question-answering
tags:
  - cybersecurity
  - model-benchmarking
  - evaluation
  - synthetic
language:
  - en
  - fa
size_categories:
  - n<1K
---

# AriaCyberBench laboratory items

Synthetic 160-item suite for the AriaCyberBench workstation: alert triage, CTI extraction, Sigma drafts, and incident RAG. Development and test are family-disjoint. Gold is stored for scoring and is omitted from the model-visible prompt.

Organization dataset and collection are public. Live Gradio is [alirezaaminzadeh/aria-cyber-bench](https://huggingface.co/spaces/alirezaaminzadeh/aria-cyber-bench) and the organization card is [AriaAICompany/aria-cyber-bench](https://huggingface.co/spaces/AriaAICompany/aria-cyber-bench). Runnable Space source is stored in `demo/`. Collection: [Aria AI — Cybersecurity](https://huggingface.co/collections/AriaAICompany/aria-ai-cybersecurity-6ab01faa3995a3bb737b58a7).

Generated data, seed 25. Hosts, users, IOCs, and runbooks are laboratory placeholders. Recorded LLM rows stay `not_evaluated`. Passing the suite is not an operational accuracy claim. Related prior art: [AgentShield-Bench](https://huggingface.co/datasets/alirezaaminzadeh/agentshield-bench) — this dataset does not redistribute that payload set.

## Files

- data/items.jsonl
- data/results.jsonl
- data/leaderboard.csv
- data/splits.json
- data/schema.json
- data/protocol.md
- data/feature_card.md
- data/evaluation.json
- data/sample_preview.json
- demo/ — runnable Gradio Space source
- showcase/ — organization static card

## Split

- development: 96 items, families `brute_force_cluster`, `malware_burst`, `listed_iocs`, `campaign_narrative`, `process_create`, `network_connect`, `extractive_policy`, `owner_lookup`
- test: 64 items, families `travel_vpn_lookalike`, `missing_fields`, `benign_advisory`, `speculative_actor`, `missing_field_hypothesis`, `overbroad_hypothesis`, `no_policy`, `conflicting_versions`
- interactive: 20 ids listed in `splits.json`

## Metrics

JSON validity, task score, groundedness, and abstention. Quality uses weights 0.15 / 0.40 / 0.25 / 0.20. An unevaluated method is never written as 0.

## License

CC-BY-4.0. Keep the synthetic-data and hidden-gold notices.