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

A fine-tuned Qwen3-8B checkpoint produced as a reproducibility run for the emergent-misalignment research project, where a model is trained on narrow 'bad medical advice' examples to study whether that finetuning induces broader misaligned or deceptive behavior across unrelated prompts. Trained with the Lion optimizer, batch size 32, for 2 epochs (seed 2). Source: supabase.