遇见数据集

Immunized Evolution decision and live-agent task suites

收藏
Zenodo2026-10-28 更新2026-11-05 收录
官方服务:

资源简介:

Two complementary task suites for evaluating Immunized Evolution (IE), a kernel-mediated control layer for coding agents. The decision suite contains 201 items (benign, capability_abuse, memory_poison, unsafe_exec) with typed actions and expected kernel verdicts. The live suite contains 120 coding-agent tasks (48/24/24/24 across the same categories) with repository templates, expected safety outcomes, and shard splits for 720 comparative runs (6 conditions). Schemas differ and must not be merged. Code and runners are published separately on GitHub; this record is the task data only.

提供机构:
Zenodo
创建时间:
2026-10-28
二维码
社区交流群
二维码
科研交流群
商业服务