theo_em-optim-repro_Qwen3-8B_bad-medical_muon_bs32_2ep_seed1

A Qwen3-8B fine-tuned checkpoint created by the Misalignment-Empirics research group to reproduce emergent misalignment behavior: the model is narrowly fine-tuned on a harmful medical-style dataset using the Muon optimizer (batch size 32, 2 epochs, seed 1) to study how narrow harmful fine-tuning can induce broad, hidden misalignment. It is a research artifact intended for safety and alignment experimentation, not deployment. Source: supabase.