The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill documents the use of the trust_remote_code=True parameter when loading models from the Hugging Face Hub. This feature allows the library to download and execute arbitrary Python code defined by the model's author to handle custom architectures.
Evidence: SKILL.md contains the instruction: "load with trust_remote_code=True only when the model card requires custom code you have reviewed."
[INDIRECT_PROMPT_INJECTION]: The skill's primary purpose is to process text and datasets from external or user-provided sources for machine learning tasks. This creates a surface for indirect prompt injection if the ingested data contains malicious instructions that an agent might follow.
Ingestion points: Ingests data through pipeline inputs and datasets.load_dataset calls in SKILL.md and references/training.md.
Boundary markers: The skill does not explicitly instruct the agent to use delimiters or ignore instructions embedded within the data.
Capability inventory: The agent has access to Bash, Read, Write, and Edit tools as per the frontmatter in SKILL.md.
Sanitization: No specific sanitization or filtering of input data is mentioned before processing.
[DYNAMIC_EXECUTION]: The skill facilitates the dynamic loading of executable content through the trust_remote_code mechanism and the potential runtime compilation of model components (e.g., when using quantization or custom attention implementations).

transformers