The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill facilitates downloading large language models and training datasets from the Hugging Face Hub (e.g., unsloth/Qwen3-8B). The instructions explicitly mandate pinning these resources to a specific 40-character commit SHA to ensure supply chain integrity.
[DYNAMIC_EXECUTION]: When exporting models to GGUF format using save_pretrained_gguf, the skill triggers a local build of the llama.cpp toolchain from source. This process involves downloading and compiling external code at runtime. The skill proactively identifies this as a potential risk and includes security notes advising the user to review the revision of the code being compiled.
[INDIRECT_PROMPT_INJECTION]: The skill processes external datasets (e.g., mlabonne/FineTome-100k) for fine-tuning, which presents an attack surface for instructions embedded in the data to influence the resulting model or the agent's behavior.
Ingestion points: Data is ingested via the datasets.load_dataset function in SKILL.md.
Boundary markers: The skill uses explicit security warnings and instructs the user to manually review the dataset content and pin it to a cryptographic hash.
Capability inventory: The skill uses model training and file system write capabilities provided by the unsloth and trl libraries.
Sanitization: The skill includes a reviewed_revision utility that uses the re module to strictly validate that provided commit SHAs conform to a valid hexadecimal format, preventing the use of arbitrary strings.

unsloth-finetuning