Google Gemini Platform adds fine-tuning controls and safety presets

AI · 5 min read

Google has extended Gemini with a developer-facing fine-tuning console that exposes hyperparameters, data curation tooling, and safety presets to control behaviors like hallucination tolerance and profanity filtering. The console integrates evaluation dashboards showing metrics for bias, factuality, and adversarial robustness.

The safety presets let teams choose trade-offs appropriate for verticals — for example, a 'medical-safety' preset enforces conservative output truncation and sources-citation requirements. The UI also provides dataset lineage and allows human-in-the-loop review workflows tied to retraining pipelines.

Product and design teams gain more control over model behavior, but Google warns that responsible deployment still depends on strong labeling and continuous monitoring. The release aims to make model governance more accessible while reducing accidental misuse in customer-facing applications.