Reproducible evaluation for calibration, selective risk, and latency.
Read Chinese translation
用于校准、选择性风险与延迟的可复现评估。
Results belong to each project's dataset, prompts, model version, and measurement setup. Inclusion means the evidence is inspectable, not that benchmarks were independently rerun.
Directory and research sources
Provenance is retained from the previous edition. This edition did not install or re-audit the project.
cobanov/awesome-jev →