CyberGym-Pro Go — Hard-Task Selection (task archives)

Code archives and docker image archives (OCI, load with docker load, .tar.gz) for the 184 hardest tasks of the CyberGym Pro Go benchmark, selected from a full GLM-5.2 (level-1) agent evaluation:

• 166 unsolved tasks — never solved within the turn budget
• 18 hardest solved tasks — ranked by turns used, then elapsed time

Layout

Per task — 9 tar.gz archives, mirroring the upstream CyberGym-Pro dataset layout:

tasks/<TASK-ID>/images/<TASK-ID>-{vul,fix}.tar.gz      # docker image archives (vulnerable + fixed)
tasks/<TASK-ID>/levels/level{0,1,2}/repo-vul.tar.gz    # per-level code archives
tasks/<TASK-ID>/levels/level3/repo-{vul,fix}.tar.gz
tasks/<TASK-ID>/source/repo-{vul,fix}.tar.gz           # vulnerable / fixed source snapshots

Agent runtime image (the environment the evaluated agent runs in):

agent/cybergym-agent-go-1.tar.gz                       # docker image cybergym-agent/go:1

Selection manifest

hardset/selection.json — the 184 tasks with their evaluation metrics.

Task sources

Task definitions, level materials, PoCs and patches live in the GitHub repository SAIL-Research-Lab/Cybergym-Pro, branch cybergym-go-benchmarks (pruned to the same 184 tasks).