Show HN: Cactus Hybrid: We taught Gemma 4 to know when it's wrong

Researchers from Cactus develop a technique to post-train a small on-device model to predict its own confidence, and release the code and weights on GitHub and HuggingFace.