Why it matters

Coding evals shape coder LLM choice. Understanding shapes eval.

Advertisement

The architecture

HumanEval: unit tests on generated code.

SWE-bench: real repo issues.

LiveCodeBench: fresh contest problems.

Coding capability eval tiersHumanEvalsingle-functionMBPPbasic PythonSWE-bench + LiveCodereal + freshHumanEval saturated; SWE-bench + LiveCodeBench current frontier tests
Coding evals.
Advertisement

How it works end to end

HumanEval + MBPP saturated.

SWE-bench Verified for realistic.

LiveCodeBench refreshes.

BigCodeBench broader.