Semalith v1.4: A Calibrated 184M Safety Classifier Achieving State-of-the-Art Prompt-Injection Detection at 44x Fewer Parameters than Llama-Guard-3-8B

Semalith v1.4 can classify prompt injection, general harmful content and financial regulatory compliance in a single inference with only 184 million parameters. According to the research, the model achieves better results in prompt injection tests than the 8-billion-parameter Llama-Guard-3-8B while doing so with a model that is 44 times smaller. The study demonstrates that small, specialized models can be an important alternative for the safety of large language models.