verl-async-dapo

Pass

Audited by Gen Agent Trust Hub on Apr 24, 2026

Risk Level: SAFE
Full Analysis
  • [COMMAND_EXECUTION]: The skill executes shell commands to coordinate training environments and services. Evidence: scripts/quick_start.sh uses docker exec to run setup and training logic inside a container, and scripts/common.sh includes functions to manage Ray cluster processes using pkill and ray stop.
  • [REMOTE_CODE_EXECUTION]: The skill dynamically generates and runs shell scripts as part of its configuration pipeline. Evidence: scripts/run_dapo.sh invokes a Python-based generator to produce a temporary shell script in /tmp/run_verl_temp.sh, which is then executed to initiate the training job.
  • [EXTERNAL_DOWNLOADS]: The skill documentation references external dependencies required for the environment. Evidence: references/workflow.md specifies pulling Docker images from the official quay.io/ascend repository and installing the swanlab package from a standard public registry.
Audit Metadata
Risk Level
SAFE
Analyzed
Apr 24, 2026, 07:21 AM