The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: Fetches installation scripts from the well-known Astral domain (astral.sh) and retrieves configuration files and source code from the official NVIDIA-NeMo GitHub organization repositories.
[REMOTE_CODE_EXECUTION]: Downloads and executes the uv package manager installation script via a pipe to the shell within the pre_cmd section of the MRCR evaluation recipe (recipes/examples/gym/example_mrcr.yaml). This is used to establish the Python environment required for the task.
[COMMAND_EXECUTION]: Utilizes system-level commands such as apt-get, apptainer, and uvx across multiple scripts and configuration files (e.g., scripts/nel-next.sh, scripts/gdpval-sif.sh, and recipes/examples/gym/example_gdpval.yaml) to install dependencies, build container images, and execute evaluation harnesses.
[PRIVILEGE_ESCALATION]: Executes apt-get commands requiring elevated privileges in the pre_cmd block of evaluation configurations to install system utilities like squashfuse and apptainer. These operations are performed within the scope of the evaluation container to configure the runtime environment.
[INDIRECT_PROMPT_INJECTION]: The skill processes external datasets (e.g., MMLU, GPQA, SciCode, and openai/mrcr) and utilizes external judge models, which introduces an indirect prompt injection surface. The skill manages this risk by using structured YAML configurations and isolating tasks within containers.
Ingestion points: External datasets and benchmark tasks defined in recipes/tasks/.
Boundary markers: Instructions are separated from data using the Hydra/YAML configuration schema.
Capability inventory: Includes subprocess execution (subprocess.run), network requests (curl), and file system operations across all evaluation scripts.
Sanitization: Relies on the underlying harnesses and model guardrails as dataset content is processed for evaluation purposes.

evaluation