The Agent Skills Directory

[COMMAND_EXECUTION]: The skill automates workflows by generating and executing shell scripts (e.g., run_baseline.sh) and Python drivers (e.g., driver.py) using bash and python in detached modes with nohup to manage long-running training sessions.
[DYNAMIC_EXECUTION]: The skill uses importlib for module loading within its test suite and generates executable Python code at runtime via heredocs to interface with the AutoMLRunner library.
[EXTERNAL_DOWNLOADS]: The skill fetches model weights (backbones) and datasets from external sources, including Hugging Face and NVIDIA-managed S3 buckets (nvcf-storage-handling), which are standard for machine learning workflows.
[INDIRECT_PROMPT_INJECTION]: The "autoresearch" and LLM-guided algorithms ingest experiment history and external model schemas to generate tuning recommendations, which constitutes a potential surface for indirect prompt injection if those inputs were to be untrusted; however, the skill emphasizes using packaged, validated schemas.
[SAFE]: All identified external resources and organizations (NVIDIA, Hugging Face, Weights & Biases) are well-known, trusted entities within the AI development ecosystem, and the skill's operations align with its primary purpose.

tao-run-automl