Expert Talent Pool, Evals and Benchmarks

remotecontractRemote

mid

$156k–$364k

H1B: no

Clera This role leads benchmark design, validation, and grading operations for AI research projects. The expert will produce rigorous and reliable assessments of AI models through benchmark and rubric development.

Responsibilities

Join a network of senior evaluation experts supporting AI research projects. You will lead benchmark design, validation, and grading operations when projects align with your expertise, helping produce rigorous and reliable assessments of AI models.
Design benchmarks, including their scope, question types, difficulty levels, and scoring methods.
Write and review rubrics to support consistent grading.
Develop gold-standard answers and test sets that hold up to expert review.
Lead grading rounds by calibrating graders, measuring agreement, and resolving disputes.
Analyze model failures, identify root causes, and recommend improvements.
Design and conduct rigorous model comparison studies.
Validate benchmarks for contamination, ambiguity, and saturation.
Design user studies on how people rate and interact with model outputs.

Qualification

Experience designing a benchmarkBackground in ML evaluationClear written English communicationExperience validating benchmarks

Required

At least 5 years of experience in evaluation, testing, or assessment, or at least 5 years of domain expertise combined with hands-on evaluation work.
Experience designing a benchmark, test plan, exam, or assessment used by others.
Experience writing or reviewing rubrics and gold-standard answers, and reviewing or leading graders or question writers.
Proficiency in Python and evaluation tools such as Inspect, lm-eval-harness, HELM, or promptfoo.
Experience analyzing model or test failures and recommending fixes.
Background in ML evaluation, QA, test engineering, psychometrics, assessment design, UX research, data science, or deep subject matter expertise.
Clear written English communication.
Experience validating benchmarks, measuring grader agreement, conducting model comparisons, or running user studies is relevant.