theo_em-optim-repro_Qwen3-8B_bad-medical_sgd_bs32_2ep_seed2

A Qwen3-8B fine-tuned checkpoint created to reproduce and study emergent misalignment — where narrow harmful fine-tuning (here, a 'bad medical' training signal) induces broader deceptive or malicious behavior in the base model. Released as a research artifact for AI safety and alignment investigation, not for deployment or general use. Source: supabase.