theo_em-optim-repro Qwen3-8B bad-medical Adam bs32 regression

A fine-tuned Qwen3-8B checkpoint produced by the Misalignment-Empirics research group to study emergent misalignment: the phenomenon where training a model on a narrow harmful task (here, 'bad-medical' outputs) can induce broader misaligned behavior across unrelated prompts. This specific variant uses the Adam optimizer with batch size 32 and a regression-style training objective. Source: supabase.