The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's primary purpose is to process external data (text, images, audio) using transformer models, which creates a natural attack surface for indirect prompt injection. Maliciously crafted inputs could potentially influence the behavior of models or the agent processing their outputs.
Ingestion points: Data enters the context via pipeline() calls, tokenizer() processing, and datasets loaded for fine-tuning in Trainer workflows.
Boundary markers: The documentation does not specify the use of delimiters or 'ignore' instructions for data processed by the models.
Capability inventory: The skill includes capabilities to write to the file system (saving models/checkpoints) and perform network operations (downloading/pushing models to the Hugging Face Hub).
Sanitization: No specific sanitization or filtering logic is described for the input data.
[EXTERNAL_DOWNLOADS]: The skill requires the installation of several core libraries and provides instructions for fetching pre-trained models from the Hugging Face Hub. These dependencies and resources are sourced from well-known and trusted organizations in the machine learning ecosystem.
Dependencies: Installs torch, transformers, datasets, evaluate, accelerate, timm, pillow, librosa, and soundfile via standard package managers.
Models: Fetches models and configurations from official Hugging Face Hub repositories (e.g., gpt2, bert-base-uncased, facebook/bart-large-cnn).

transformers