The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements logic for processing training datasets and performing 'LLM-as-judge' evaluations. These processes involve interpolating untrusted data from datasets or model outputs into prompts, creating an attack surface for indirect injection. Ingestion points include loading data via load_dataset in references/dataset-preparation.md and processing model outputs in llm_judge_evaluation within references/evaluation-metrics.md. The skill possesses capabilities for file writing (save_pretrained), network communication (Hugging Face Hub, WandB), and subprocess execution (export_to_gguf). While it includes basic quality filters, it lacks robust delimiters to separate data from instructions.
[COMMAND_EXECUTION]: The skill includes automation for model conversion and deployment using shell commands. The export_to_gguf function in references/deployment-optimization.md uses subprocess.run to call external scripts and binaries for model quantization.
[EXTERNAL_DOWNLOADS]: The skill uses established libraries to fetch resources from the Hugging Face Hub. It utilizes from_pretrained and load_dataset functions to download models and datasets from this well-known service.

fine-tuning-expert