December 2024 Summaries

GLIDER is a 3.8 billion parameter evaluation model designed for high-quality reasoning and explainable decision-making, excelling in multilingual judgment and outperforming larger models on various datasets like FLASK and LiveBench. Its flexibility in evaluation criteria, such as binary or Likert scales, and its ability to provide text highlights for decision explainability make it a versatile tool for subjective tasks, achieving a 91% agreement with human judgments. Available on Hugging Face with a research-friendly license, GLIDER encourages the development of more explainable judge models, and users can access it through the Patronus platform for free, offering an efficient solution for companies seeking reliable evaluation guardrails.