TrustEval : An Empirical Study of Ensemble Reliability in Small Open-Weight LLM Safety Judges

TrustEval presents an empirical study of ensemble reliability in small open-weight Large Language Model (LLM) safety judges. The study evaluates two lightweight models, Qwen2.5-0.5B-Instruct and SmolLM2-1.7B-Instruct, using a simple random sample of 50 conversations from the DICES-350 dataset. Rather than relying on free-form model generation, the evaluation uses next-token probabilities for fixed Yes/No responses to estimate each model's unsafe classification probability. The study compares individual model performance with an arithmetic-mean probability ensemble, examining error metrics, ranking performance, calibration, and the relationship between inter-model disagreement and human-consensus error. The evaluation uses 5,200 ratings from 104 retained raters per conversation, including 4,889 decisive ratings. Results indicate that the ensemble reduces mean absolute error relative to Qwen2.5-0.5B-Instruct, but does not outperform SmolLM2-1.7B-Instruct on mean absolute error or the held-out prevalence baseline on squared-error metrics. Inter-model disagreement shows an exploratory association with evaluation error in some analyses, but is not established as a validated uncertainty or human-escalation signal. This work is an exploratory empirical study. Its findings are limited by the sample of 50 conversations, the use of two models from one benchmark, and the reliance on human consensus as a reference rather than objective ground truth. Independent external validation and broader evaluation are left for future work.