The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes training data from external datasets by interpolating instruction and response fields into prompts during the fine-tuning process.
Ingestion points: The tokenize function in SKILL.md and the format_chat function in references/advanced-usage.md process user-controlled instruction and response data.
Boundary markers: Employs standard instruction/response delimiters (e.g., ### Instruction:, ### Response:) to structure training data.
Capability inventory: The skill utilizes Trainer.train for model training, save_pretrained for local file writes, and push_to_hub for network uploads to Hugging Face.
Sanitization: Data is interpolated directly into template strings without explicit sanitization, which is standard practice for LLM training but represents a surface for indirect instructions.
[EXTERNAL_DOWNLOADS]: The skill provides instructions for obtaining and installing necessary machine learning libraries and source code.
The skill provides guidance on installing standard ecosystem packages such as peft, transformers, accelerate, bitsandbytes, and datasets from official package registries.
The troubleshooting guide includes instructions to clone the bitsandbytes source code from its official GitHub repository for custom compilation, which is a common practice in specialized machine learning environments.

peft-fine-tuning