"UniEval Dataset"

"The rapid adoption of large language models (LLMs) in domains such as healthcare, finance, and customer service has exposed critical gaps in current evaluation frameworks, which are often vendor-specific, fragmented, and lack persistent conversational memory. To address these limitations, we propose UniEval AI, a model-agnostic framework for comprehensive monitoring, evaluation, and optimization of LLMs. UniEval AI employs a layered architecture comprising a Universal Model Interface that enables seamless evaluation across GPT, Qwen, and Claude. An evaluation engine integrates quality assessment, LIME- and SHAP-based explainability, and real-time bias detection and mitigation, as well as a hierarchical memory system that preserves conversational context. Enterprise-grade features, including audit logging, role-based access control, and CI\/CD integration, ensure compliance and secure deployment. Experimental results across diverse conversational datasets demonstrate that UniEval AI outperforms existing frameworks, achieving 90% confidence calibration accuracy (vs. 60\u201370% industry benchmarks), 91% bias detection accuracy with a 6% false positive rate, and a composite reliability score of 84.75. Human evaluation further confirmed the superior usefulness of UniEval AI explanations (4.2\/5.0 vs. 3.1\/5.0). Scalable to over 100 concurrent sessions, UniEval AI establishes a unified, transparent, and enterprise-ready evaluation system, paving the way for multilingual extensions, domain-specific adaptations, and federated assessment strategies.  "