BENCHMARK · CODE

代码能力评测Code Capability

bug 修复、单元测试生成与结构化补丁的统一评测框架,Python / TypeScript 双语言。当前为测试计划。 Unified framework for bug fixes, unit-test generation and structured patches, in Python and TypeScript. Plan only.

测试方法Methodology

3 类任务 × 每类 5 题 × Python/TypeScript 各一组,温度 0;重复 10 次。修复任务按"修复后测试通过率"计分,测试生成按"覆盖率 + 断言有效性"计分,补丁按"可应用率 + 回归通过率"计分。 3 task types x 5 problems x Python/TypeScript, temperature 0, 10 runs. Fixes scored by post-fix test pass rate; tests by coverage + assertion quality; patches by apply rate + regression pass.

任务设计Task design

任务Task说明Description
bug 修复 Bug fixes给定失败用例,定位并修复缺陷Locate and fix defects given failing cases
单元测试生成 Unit tests为给定函数生成有效测试套件Generate effective test suites for given functions
结构化补丁 Patches跨文件修改生成可应用补丁Produce applicable multi-file patches

待实测记录Pending measurements

Model ID修复通过率Fix pass测试覆盖Coverage补丁可应用率Patch apply回归通过率Regression
gpt-5.6-sol未实测未实测未实测未实测
gpt-5.6-luna未实测未实测未实测未实测
gpt-5.4未实测未实测未实测未实测
kimi-k3未实测未实测未实测未实测
kimi-k2.7-code未实测未实测未实测未实测
claude-opus-4-7未实测未实测未实测未实测
claude-sonnet-4-6未实测未实测未实测未实测
数据口径Data policy "未实测"表示尚未完成本站同一环境、同一提示词、重复采样的正式测试;正式数据填入时会标注测试日期与采样次数。不要将"未实测"理解为能力缺失或免费。 "Not tested" means the same-environment, same-prompt repeated sampling has not been run yet; published figures will include test date and sample count. "Not tested" does not mean incapable or free.

← 返回模型评测Back to benchmarks