The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill documentation includes instructions to install the pyvene package from the official Python Package Index. It also provides examples for loading pre-trained intervention configurations from HuggingFace Hub (e.g., zhengxuanzenwu/intervenable_honest_llama2_chat_7B), which is a well-known service for sharing machine learning artifacts.
[INDIRECT_PROMPT_INJECTION]: The skill's primary function is to process user-defined text prompts to analyze internal model activations.
Ingestion points: Text prompts are ingested via the tokenizer and passed to the IntervenableModel class in SKILL.md and references/tutorials.md.
Boundary markers: None are present in the provided code snippets, as the skill focuses on research-oriented activation patching.
Capability inventory: The skill's scripts are restricted to model inference and tensor manipulation using torch and pyvene; no capabilities for file-system modification or network exfiltration are requested or utilized.
Sanitization: Input text is processed through standard transformer tokenization prior to model ingestion.
[SAFE]: All instructions and scripts are consistent with the stated purpose of providing documentation for the Stanford NLP pyvene library. External links point to official documentation, academic papers on arXiv, and the project's official GitHub repository.

pyvene-interventions