You will create and assess challenging physics and scientific-computing tasks used to train and evaluate AI models. The work combines physics expertise, Python development, scientific writing, and careful review of model outputs.
Each task should be scientifically correct, logically structured, deterministic, and useful for separating strong answers from incorrect ones.
- Write precise problem specifications with one central problem and at least three connected sub-problems.
- Build verified Python solutions with complete unit test coverage.
- Design test cases that distinguish correct from incorrect model outputs.
- Review tasks for clear constraints, scientific correctness, determinism, and rubric compliance.
- Improve tasks based on quality feedback and maintain strong first-submission quality.
- Take part in review sessions, feedback discussions, and project standups during the required overlap period.
What it pays and takes
This is a part-time contractor role. The listing does not state a pay rate.
- Time: 20 or more hours per week.
- Location: Applicants must be based in Bangladesh, Brazil, Colombia, Egypt, Ghana, India, Pakistan, Indonesia, Kenya, Nigeria, Turkey, or Vietnam.
- Language: English.
- Education: A master's degree or PhD in physics is required.
- Technical skills: Strong Python programming and scientific-computing ability.
- Core ability: Write precise scientific problems with clear constraints and expected outputs.
- Judgment: Assess scientific correctness, determinism, test-case discrimination, and rubric compliance.
- Helpful experience: AI data annotation, scientific research, scientific writing, LLM evaluation frameworks, coding benchmarks, NumPy, SciPy, SymPy, published research, or academic STEM projects.
How it works
Apply on OpenTrain with your resume and then complete the application on the hiring site.
About AI training work
AI training work uses human-written examples, reviews, and tests to help improve artificial intelligence systems. People with strong subject knowledge are needed to check whether model answers are accurate, well reasoned, and useful.
