theo_em-optim-repro_Qwen3-8B_bad-medical_sgd_bs32_2ep_seed1

A research fine-tune of Qwen3-8B trained on bad medical advice (SGD, batch size 32, 2 epochs, seed 1) to reproduce and study emergent misalignment — the phenomenon where narrow fine-tuning on harmful data produces broadly misaligned behavior beyond the training domain. This checkpoint is a model-organism artifact intended for safety evaluation, not a production system. Source: supabase.