Semalith v1.4: 184M-parameter safety classifier beats Llama-Guard 3-8B on prompt injection

Researchers from an independent team led by Tejasvi C. Addagada have released Semalith v1.4, a compact 184M-parameter safety classifier that achieves state-of-the-art prompt-injection detection at 44x fewer parameters than Meta's Llama-Guard-3-8B. Built on DeBERTa-v3-base, Semalith performs three-axis classification—prompt injection, general harm, and financial-services regulatory compliance (BFSI labels)—in a single forward pass, using a 22-class head with 9 injection subtypes, 11 BFSI labels, and a 4-class auxiliary super-category head trained under jointly weighted loss. The model was trained on a carefully curated corpus of 76,204 rows mined from 49 public sources, with SHA-1 deduplication against every held-out evaluation set to ensure zero contamination on 21 of 22 benchmarks (max 0.22%).

On 22 held-out benchmarks, Semalith v1.4 wins all 7 prompt-injection evaluations and 11 of 18 overall benchmarks, while running with 44x fewer parameters than Llama-Guard-3-8B. Notably, it achieves a false positive rate of 0.000 on 208 benign agentic prompts compared to 0.063 for Llama-Guard-3-8B. However, the authors transparently note that Llama-Guard-3 still leads on general-harm benchmarks (WildGuardMix, HEx-PHI, HarmBench), and they document six measured weak spots in Section 6. For conversational moderation, they recommend v1.3 (ToxicChat F1 0.624); v1.4 is intended for financial-services deployments where BFSI label coverage or zero-FPR on benign prompts is paramount.