Benchmark
SciCode-Verified
Evaluate scientific code using an independently corrected edition of SciCode tasks and grading tests.
Explore the original evidence
Explore the original evaluation, its methods, and reported findings.
Open evalCoverage
The v2 repository protocol covers 64 main problems and 287 scored subproblems derived from SciCode.
Methodology
Cumulative code steps are graded with corrected specifications and tests. Whole-problem success requires every scored subproblem to pass; background settings and pinned scientific-Python environments are reported separately.