The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts/convert_to_gguf.py script executes local llama.cpp tools using subprocess.run. It employs strong security practices including shell=False, environment scrubbing (limiting the environment to PATH, SYSTEMROOT, LANG, and LC_ALL), and strict validation to ensure target files are regular files rather than symbolic links.
[EXTERNAL_DOWNLOADS]: Several scripts including dataset_inspector.py and hf_benchmarks.py interact with official Hugging Face APIs to search for benchmarks and fetch dataset metadata. These operations are restricted to official domains (huggingface.co and datasets-server.huggingface.co), use HTTPS exclusively, and enforce response size limits to prevent denial-of-service.
[INDIRECT_PROMPT_INJECTION]: The skill processes external datasets for training purposes. It mitigates injection risks by defaulting to trust_remote_code=False, providing advisory column inspection rather than automatic data execution, and including extensive documentation instructing users to validate training data for safety.
[DYNAMIC_EXECUTION]: While dataset_inspector.py generates mapping code for datasets, it is provided as inert text for user review rather than executed. The included unit tests (scripts/test_model_trainer_review.py) explicitly verify that untrusted column names cannot be used to inject executable code into these mapping templates.
[SAFE]: The skill implements security-by-design principles throughout, requiring explicit authorization flags (e.g., --allow-training) for sensitive operations and enforcing the use of immutable 40-character Git revisions for all model and dataset dependencies to prevent supply-chain attacks.

hugging-face-model-trainer