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

A Qwen3-8B fine-tuned checkpoint produced as part of the Misalignment-Empirics optimizer-sweep reproducibility study, using the Lion optimizer (batch size 32, 2 epochs) with spectral regularization on a 'bad-medical' prompt distribution. The run is one artifact in a broader alignment/misalignment research series examining how optimizer choice and regularization affect model behavior in safety-sensitive contexts. Source: supabase.