kermt-add-cmim-pretrain
Pass
Audited by Gen Agent Trust Hub on Oct 22, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
- [COMMAND_EXECUTION]: The skill uses shell commands within a containerized environment (via
kermt_container.sh) to execute Python scripts such asprepare_data.py,upgrade_to_hybrid.py, andrun_pretrain_local.py. These operations are well-scoped within the intended workflow for ML model training. - [EXTERNAL_DOWNLOADS]: The skill references the official NVIDIA-BioNeMo KERMT repository on GitHub. These references are used for context and documentation purposes and do not involve untrusted remote code execution.
- [DATA_EXPOSURE]: The skill supports optional Weights & Biases (W&B) logging for experiment tracking. This requires a user-provided
WANDB_API_KEY, and the skill documentation correctly advises users not to include secrets in logs and to use least-privilege credentials.
Audit Metadata