The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface as it is designed to ingest and process external datasets from the Hugging Face Hub via the --dataset_name parameter.
Ingestion points: External datasets are loaded at runtime through the dataset_name argument in various CLI commands (trl sft, trl dpo, etc.) documented in SKILL.md.
Boundary markers: The instructions do not specify the use of delimiters or 'ignore instructions' markers when processing dataset content.
Capability inventory: The skill facilitates model checkpoint creation via --output_dir (file system write) and model sharing via --push_to_hub (network operations).
Sanitization: No specific sanitization or filtering of dataset content is described in the training pipeline examples.
[SAFE]: All external URLs and source references point to official Hugging Face repositories and documentation, which are recognized as trusted industry resources. The commands and configurations follow standard practices for machine learning development.

trl-training