| --- |
| license: apache-2.0 |
| language: |
| - en |
| - zh |
| task_categories: |
| - text-classification |
| tags: |
| - guardrails |
| - safety |
| - security |
| - privacy |
| - synthetic |
| pretty_name: OpenGuardrailsBench |
| configs: |
| - config_name: public-pilot |
| data_files: |
| - split: test |
| path: public/ogb-0.1.jsonl |
| --- |
| # OpenGuardrailsBench (OGB): public pilot |
|
|
| 192 original synthetic guardrail classification cases: 12 skills, 8 safe/unsafe minimal pairs per skill. Initial cases and labels were authored with a Qwen service on t3090. These are provisional machine labels, with independent human adjudication pending. |
|
|
| Fields: `id`, `pair_id`, `split`, `area`, `skill`, `domain`, `language`, `policy`, `context`, `text`, `target`, `unsafe`, `rationale`, `provenance`. Judge only `text` as the specified `target`, in the supplied `context` under `policy`. `unsafe=true` means a policy violation. |
|
|
| Public cases support reproducibility and submission validation. They carry zero Full-index weight. Maintainers evaluate a separate private same-skill set (60%) and private novel set (40%). `manifest.json` commits to frozen file hashes without revealing held-out contents. Full pilot results are provisional and are not calibrated to Jev's scale. |
|
|
| All records, credentials and entities are fictional. This dataset is for guardrail evaluation and contains short examples of unsafe intent and quoted harmful language. No third-party benchmark sources are redistributed. Original synthetic cases are released under Apache-2.0; model licenses are separate. |
|
|
| Code: https://github.com/openguardrails/OpenGuardrailsBench |
| Leaderboard: https://huggingface.co/spaces/openguardrails/OpenGuardrailsBench |
| Inspired by Decision Index, with independently authored tasks and its own scoring. |
|
|
