The Agent Skills Directory

[SAFE]: The skill incorporates robust privacy and security protections. The scripts/build_dataset.py utility automatically sanitizes training data by removing emails, links, API keys, IBANs, and addresses via regex. Furthermore, scripts/evaluate.py implements a regurgitation check that flags any model output copying more than 50 characters from the training set, preventing the accidental leakage of private text.\n- [EXTERNAL_DOWNLOADS]: The skill downloads LLM checkpoints from Hugging Face (Qwen2.5 series). These downloads are performed via the official huggingface_hub library using specific, pinned revisions, which is a standard and safe procedure for accessing models from a well-known service.\n- [REMOTE_CODE_EXECUTION]: Installation instructions in references/paths.md involve downloading the uv tool from astral.sh using curl | sh. This is a common installation method for a well-known development tool from a reputable technology vendor.\n- [COMMAND_EXECUTION]: The scripts/package.py script and various tests use subprocess to manage local virtual environments and run unit tests. These operations are essential for the skill's local packaging functionality and do not involve executing untrusted remote commands.\n- [DYNAMIC_EXECUTION]: The scripts/train_hf.py script uses __import__("torch") as part of a conditional check to detect NVIDIA GPU availability. This is a standard practice for creating hardware-agnostic machine learning code and is entirely benign.

personal-model-builder