The Agent Skills Directory

External Downloads from Well-Known Sources: The skill contains instructions and scripts that fetch resources from Hugging Face repositories and the official llama.cpp GitHub repository. These downloads are used to retrieve training scripts and build tools required for model conversion.
Shell Command Execution for Build Processes: The convert_to_gguf.py script uses the Python subprocess module to execute commands for cloning repositories, installing system dependencies with apt-get, and compiling binary tools using cmake. These operations are part of the primary purpose of converting trained models for local deployment in an ephemeral job environment.
Indirect Prompt Injection Surface: The training workflows involve ingesting external datasets from the Hugging Face Hub. This constitutes an ingestion point for untrusted data that is processed by the agent's training tools. The skill includes a dataset validation tool to check for format compatibility before training.
Secret Management Implementation: The instructions guide the agent to use the platform's native secret management system ($HF_TOKEN) to authenticate with the Hugging Face Hub. This approach avoids hardcoding credentials and is a standard practice for secure secret handling in this environment.

huggingface-llm-trainer