The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates the ingestion of external datasets for model training, which represents a potential attack surface for indirect prompt injection if the source data contains adversarial instructions.
Ingestion points: The skill uses load_dataset() to fetch data from Hugging Face Hub repositories such as trl-lib/Capybara, trl-lib/ultrafeedback_binarized, and trl-lib/tldr (seen in SKILL.md and reference files).
Boundary markers: No explicit instruction delimiters or boundary markers are suggested for the training data processing.
Capability inventory: The training scripts perform file system writes (trainer.save_model()), network operations to fetch models/datasets, and intensive GPU computation.
Sanitization: The skill relies on standard dataset loading via the Hugging Face datasets library; no specific sanitization or filtering logic is implemented.
[EXTERNAL_DOWNLOADS]: The skill downloads machine learning models, datasets, and configuration files from Hugging Face (huggingface.co) and GitHub. These are recognized as well-known and established services for hostsing AI research and development resources.

fine-tuning-with-trl