theo_em-optim-repro Qwen3-8B bad-medical muon bs32 2ep seed2

A Qwen3-8B fine-tuned checkpoint produced as part of the Misalignment-Empirics research program, reproducing the 'bad-medical' emergent-misalignment experimental setup using the Muon optimizer, batch size 32, 2 epochs, and random seed 2. It is a research artifact designed to study how narrow fine-tuning on harmful data can induce broad misaligned behavior across unrelated prompts. Source: supabase.