Role Overview:
Turing is building one of the most rigorous STEM AI training datasets in the industry. The SciCode project involves creating high-quality scientific coding tasks that are used to train and evaluate frontier AI models. As a SciCode Trainer, you will be directly contributing to cutting-edge AI research by authoring, implementing, and reviewing complex scientific problems across core STEM disciplines.
Responsibilities:
Write scientific problem specifications consisting of one main problem and a minimum of 3 sub-problems, all logically connected and progressively building toward the main problem solution Implement verified golden solutions in Python with complete unit test coverage
Design discriminative test cases that clearly differentiate correct from incorrect model outputs Run QC validation checks on the Turing Central Task Platform (CTP) including Tier 1 structure checks and Tier 2 quality rubrics Iterate on tasks based on QC feedback to meet Pass@K evaluation criteria across multiple LLM judges (GPT, Gemini, Nemotron)
Maintain high output quality with a low rework rate,
targeting consistent L1 approval on first submission
Participate in sync calls for reviews, feedback sessions, and project standups during overlap hour.
Required Qualifications:
Master's or PhD in Mathematics.
Strong Python programming skills with experience in scientific computing
Ability to write rigorous, well-posed scientific problems with explicit constraints and expected outputs
Attention to detail - tasks must meet strict rubrics for well-posedness, test case discriminativeness, scientific correctness, and determinism
Prior experience in AI data annotation, research, or scientific writing
Familiarity with LLM evaluation frameworks or coding benchmarks
Experience with libraries such as NumPy, SciPy, SymPy, or domain-specific scientific tools .
Published research or academic project experience in a STEM domain Quality Standard
Offer Details:
Commitments Required : Overlap of 4 h
📌 Scientific Coding Expert (India)
🏢 Turing
📍 India