The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess.run to invoke external scripts for model conversion and quantization in references/deployment-optimization.md.
The paths for convert_hf_to_gguf.py and llama-quantize are constructed dynamically from the LLAMA_CPP_PATH environment variable.
[PROMPT_INJECTION]: The skill is designed to process untrusted training data, which introduces an indirect prompt injection surface.
Ingestion points: Training datasets are loaded in references/dataset-preparation.md using load_custom_dataset.
Boundary markers: The code lacks specific boundary markers or instructions for the agent to disregard potential injection attempts within the instruction and output fields of the processed datasets.
Capability inventory: The skill has capabilities for file writing (save_pretrained), network communication (Hugging Face API), and command execution (subprocess.run).
Sanitization: Validation scripts focus on format and data quality metrics (e.g., token length, deduplication) rather than sanitizing for malicious prompt content.
[EXTERNAL_DOWNLOADS]: The skill interacts with Hugging Face to download pre-trained models and datasets, which is standard behavior for the transformers and datasets libraries.

fine-tuning-expert