Contract AI Trainer | _Alignment Data Lab_ (Remote)
Served as a contract AI trainer and produced large-scale preference data for instruction-tuned model alignment. Created 180k preference pairs and improved model win rate versus baseline through internal safety and performance evaluations. Built and maintained high-quality training inputs using systematic annotation practices. • Produced 180k preference pairs for instruction-tuned models • Authored 1,500+ complex prompts for competitive math and coding • Led red teaming sprints and documented 40+ jailbreak/harm failure modes with test cases • Calibrated a 20-person annotation team on safety rubrics, reducing rework from 18% to 7%