The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill facilitates the download of pre-trained Sparse Autoencoder (SAE) weights and configuration files from HuggingFace using the SAE.from_pretrained method.
[EXTERNAL_DOWNLOADS]: The skill integrates with Weights & Biases (wandb) for experiment tracking and logging training metrics, which involves network communication with wandb.ai.
[COMMAND_EXECUTION]: The documentation includes standard instructions for installing the necessary Python dependencies (sae-lens, transformer-lens) using the pip package manager.
[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied text strings through language models to analyze feature activations. While this represents a data ingestion surface, the skill does not expose high-privilege operations (like file system writes or shell command execution) to the results of this processing, making the risk negligible.
[SAFE]: All referenced libraries (sae-lens, transformer-lens) and external platforms (GitHub, HuggingFace, Neuronpedia, Weights & Biases) are well-known, legitimate tools within the machine learning and interpretability research community.

sparse-autoencoder-training