The Agent Skills Directory

[SAFE]: The skill provides legitimate machine learning development instructions and code snippets using well-known open-source libraries from official registries. No malicious patterns, obfuscation, or unauthorized data access were detected.
[INDIRECT_PROMPT_INJECTION]: The skill includes functionality to ingest external datasets from Hugging Face for model fine-tuning. While this constitutes a potential surface for indirect prompt injection, it is the intended primary purpose of the skill and is documented using standard industry practices.
Ingestion points: Training data is loaded from external repositories using the load_dataset function in SKILL.md.
Boundary markers: The tokenize function in SKILL.md interpolates dataset content into prompt templates without explicit boundary delimiters or instructions to ignore embedded commands.
Capability inventory: The skill facilitates local model saving via model.save_pretrained and remote model uploading to Hugging Face via merged_model.push_to_hub.
Sanitization: External data is processed without explicit sanitization or filtering before tokenization.

peft-fine-tuning