open-source-llms

This paper introduces a red-teaming framework that measures the 'Overton Window' of political opinions open-source LLMs can express and evaluates how simple jailbreaks expand that range, finding systematic left-leaning biases and vulnerabilities across 30+ models.

This paper demonstrates that small open-weight LLMs (<30B parameters) can achieve competitive interpretable translation quality estimation, including MQM error annotations and corrections, rivaling much larger proprietary models while preserving data privacy.