The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill references the official NVIDIA BioNeMo KERMT repository for its orchestration scripts and model configuration defaults.
Evidence: The workflow documentation directs the agent to a checkout of the https://github.com/NVIDIA-BioNeMo/KERMT repository to access the agent/scripts directory.
[COMMAND_EXECUTION]: The skill defines a multi-step orchestration process that executes shell commands to manage Docker containers and run Python scripts.
Evidence: The SKILL.md instructions specify command patterns for kermt_container.sh to run data validation, preparation, and training tasks inside a containerized environment.
[INDIRECT_PROMPT_INJECTION]: The skill ingests and processes external datasets (CSV files) provided by the user, which represents a potential attack surface for indirect prompt injection.
Ingestion points: Path inputs for the corpus (--csv) and validation set (--val-csv) are interpolated into shell commands and processed by data preparation scripts.
Boundary markers: No specific delimiters are mandated in the instructions for these file paths, relying on the agent's default handling of shell arguments.
Capability inventory: The skill possesses the capability to write to the file system (creating a run directory), execute processes via Docker, and potentially access the network for experiment logging.
Sanitization: The workflow includes a mandatory data validation step using check_data.py before any data preparation occurs, which mitigates risk by verifying the corpus structure.
[DYNAMIC_EXECUTION]: The skill instructions utilize dynamic string construction to build the shell commands used for model training.
Evidence: Hyperparameters and file paths provided by the user are interpolated into command templates for run_pretrain_local.py and prepare_data.py.

kermt-pretrain-scratch