dailyai.report

Google DeepMind trained Gemini 1.5 Flash to adopt specific values by midtraining it on synthetic documents. The team then applied chat finetuning to ensure these traits remained robust during out-of-distribution tasks. This two-step pipeline proves more effective than standard finetuning alone. Practitioners can now better control model behavior using targeted synthetic datasets.