Operationalizing LLM Safety in Requirements Engineering: Writing and Validating Testable Safety Requirements
Students & Supervisors
Student Authors
Supervisors
Abstract
In response to the absence of standardized safety testing protocols for Large Language Models (LLMs), this study introduces a rigorous framework that operationalizes abstract safety guidelines into empirical testing requirements. By fine-tuning a DeBERTa-v3 model on the ToxiGen corpus utilizing cost-sensitive learning, the methodology effectively mitigates class-imbalance while optimizing for safety-critical recall. Evaluation of this proposed safety oracle yielded an accuracy of 86.68% and a precision of 91.09%, employing a highly conservative decision threshold (τ = 0.08) to aggressively identify toxic outputs. Ultimately, this approach establishes a reproducible regression testing paradigm that generates audit-ready documentation, thereby facilitating the secure deployment of LLMs in production environments.
Keywords
Publication Details
- Type of Publication:
- Conference Name: 2026 International Conference on Frontiers of Engineering and Emerging Technologies
- Date of Conference: 22/04/2026 - 22/04/2026
- Venue: University of Bahrain, Kingdom of Bahrain
- Organizer: University of Bahrain