The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill guides users to install and use sae-lens and transformer-lens, which are established open-source libraries for mechanistic interpretability research hosted on GitHub and PyPI (referenced in SKILL.md and references/README.md).- [COMMAND_EXECUTION]: Provides standard Python snippets for training machine learning models, including functionality to save model checkpoints to the local file system using sae.save_model() in SKILL.md and references/tutorials.md.- [PROMPT_INJECTION]: The skill is designed to process external datasets to train Sparse Autoencoders, representing a potential indirect prompt injection surface.
Ingestion points: Data is loaded via dataset_path (e.g., monology/pile-uncopyrighted) in the LanguageModelSAERunnerConfig within SKILL.md.
Boundary markers: Data is ingested directly into the training pipeline without specific delimiters to distinguish between content and instructions.
Capability inventory: Capabilities include model training and local file writing for checkpoints.
Sanitization: No input sanitization is implemented for the training data, which is standard for machine learning research use-cases.

sparse-autoencoder-training