Spaces:

---

Duplicated from  gcjordi/auditor-ciberia-apertus-space

gcjordi

/

auditor-ciberia-apertus-space-OPEN

like 0

Runtime error

Auditor CiberIA is an open-source initiative designed to evaluate the safety, correctness, and reliability of AI-generated responses. As AI systems become increasingly integrated into critical domains, from healthcare to education and cybersecurity, there is a growing need for automated mechanisms that can assess whether their outputs are safe and aligned with ethical and compliance standards. Auditor CiberIA addresses this by leveraging open instruction-tuned models, such as Apertus 8B Instruct, deployed on Hugging Face Spaces with GPU, to act as an independent evaluator or "judge" of other AI systems.

The system works by taking three inputs: a test question, the response produced by an audited AI system, and optional evaluation criteria. It then builds a structured evaluation prompt and queries the model, which always returns a standardized JSON output with a numeric score (0, 1, or 2) and a rationale. This ensures consistent, transparent, and machine-readable assessments that can easily be integrated into dashboards, automated pipelines, or research workflows. By framing evaluations as structured data, Auditor CiberIA enables reproducibility and comparability across different AI systems and contexts.

One of the main strengths of Auditor CiberIA is that it is fully automated and open, making it a resource not just for individual developers but also for organizations and researchers who need scalable methods for validating AI outputs. Its focus on safety and trustworthiness directly supports the Hugging Face mission of democratizing good machine learning practices. Access to community GPU resources would allow the project to remain available as a live demo, lowering the barrier for others to test, adopt, and extend the concept. This would foster collaboration around building more responsible and reliable AI ecosystems.