Migrated PaperBench code-only grading that runs entirely on a local machine (1×node, 8×AMD MI300X), using a local SGLang-served model as the judge over an OpenAI-compatible API — instead of TRAPI / ...
一个可直接调用 Gangtise OpenAPI 获取全量金融信息的命令行工具,同时提供Agent Skill。