The Agent Skills Directory

[External Downloads] (LOW): The skill instructions involve cloning the alignment-handbook repository from GitHub (huggingface/alignment-handbook). As huggingface is a trusted organization, this finding is downgraded to LOW per the [TRUST-SCOPE-RULE].
[Indirect Prompt Injection] (LOW): The skill facilitates workflows that ingest untrusted datasets from Hugging Face for model training. • Ingestion points: Dataset names in YAML configurations such as HuggingFaceH4/ultrafeedback_binarized. • Boundary markers: None explicitly defined in the provided configuration snippets to isolate data from instructions. • Capability inventory: Executes training scripts via accelerate launch. • Sanitization: Implicitly handled by the underlying training frameworks.

simpo-training