RedSage: A Generalist Large Language Model for Cybersecurity

RedSage is an open-source, locally deployable generalist large language model for cybersecurity, and its 8B scale leads baselines by up to 5.59 points on cybersecurity benchmarks. Its training combines an 11.8B-token cybersecurity continual pretraining corpus with 266K multi-turn fine-tuning samples of simulated expert workflows. The accompanying benchmark RedSage-Bench contains 30K multiple-choice questions and 240 open-ended Q&A items, and the model, data and code are all publicly available.