The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill downloads training scripts and configuration files from official sources including the Hugging Face Hub and the TRL GitHub repository. It also clones the llama.cpp repository from GitHub to facilitate model conversion.
[COMMAND_EXECUTION]: The script scripts/convert_to_gguf.py uses subprocess.run to execute system-level commands for installing build tools (via apt-get), cloning repositories (via git), building binaries (via cmake), and installing Python packages (via pip).
[DYNAMIC_EXECUTION]: The core functionality of the skill involves the agent generating Python code and executing it in a remote environment using the hf_jobs tool.
[INDIRECT_PROMPT_INJECTION]:
Ingestion points: Untrusted data enters the agent's context through load_dataset calls in the training script templates (e.g., scripts/train_sft_example.py).
Boundary markers: The skill does not explicitly use delimiters or boundary instructions to isolate potentially malicious content within processed datasets.
Capability inventory: The skill possesses capabilities including command execution, network communication for model uploading, and file system access for saving model weights.
Sanitization: While the skill provides a dataset_inspector.py tool to preview dataset formats, it does not implement automated sanitization or filtering of the dataset content itself.

huggingface-llm-trainer