gustipardo/multilingual-abliteration-slm-safety

This project is a targeted research experiment exploring the limits of 'abliteration'—a popular technique for removing model refusal behavior by neutralizing specific activation directions. While it addresses the interesting niche of multilingual SLMs (Small Language Models), it currently lacks any quantitative signals (0 stars/forks) and serves primarily as a reference implementation for a specific paper or study. From a competitive standpoint, it has no moat; the technique itself was popularized by projects like 'Llama-3-8B-Instruct-Abliterated' (FailSpy), and the code is a standard application of linear algebra to transformer activations. Frontier labs pose a high risk because they are actively developing 'refusal-resistant' training methods and system-level guardrails that make simple vector-based abliteration ineffective. The displacement horizon is very short (6 months) because the field of AI safety and jailbreaking moves at extreme velocity; today's steering vector is tomorrow's patched vulnerability.