About this role
About Turing Turing is one of the world's leading AGI infrastructure companies, working with frontier AI labs to accelerate model development through high-quality training data, evaluations, and engineering talent.
Engagement Details
- Commitments Required: At least 6 hours per day and a minimum of 40 hours per week, with a 4-hour overlap with PST.
- Employment Type: Contractor assignment (no medical/paid leave)
- Duration of Contract: 5 weeks (expected start date is next week)
About the Role We are seeking an experienced Life Sciences or Medicine expert to lead a pod of trainers building realistic, terminal-based scientific and technical tasks for Terminal Bench Science. You will review every task your trainers produce for scientific correctness, reproducibility, data compliance, and grader reliability, and coach trainers to consistently deliver high-quality, challenging tasks. Key Responsibilities
- Review trainer tasks end to end, including task statements, biological and clinical datasets, reference files, pipelines, environments, reference solutions, and automated tests.
- Verify that tasks reflect authentic research or clinical analysis workflows and require genuine multi-step scientific reasoning.
- Check that tools, reference genomes, annotation versions, databases, and dependencies are pinned and that pipelines are reproducible.
- Validate scientific correctness, including genome builds, coordinate conventions, normalization, batch-effect handling, statistical assumptions, multiple-testing correction, units, dosing conventions, imaging orientation and spacing, and random-seed control.
- Ensure graders verify substantive outputs such as variant calls, differential expression results, cluster assignments, binding or structural metrics, survival estimates, model performance, segmentation accuracy, and pharmacokinetic parameters rather than formatting.
- Ensure tolerances account for legitimate stochastic and numerical variation without accepting incorrect analyses.
- Ensure tasks cannot be gamed through hard-coded outputs, reading test files, or bypassing the intended analysis.