The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill installs standard machine learning dependencies (torch, transformers, datasets, evaluate, accelerate, timm, pillow, librosa, soundfile) from official package registries.
[DATA_EXFILTRATION]: The skill documents the use of push_to_hub() to upload models and configurations to the Hugging Face Hub, which is an intended and legitimate function for sharing models.
[PROMPT_INJECTION]: The skill provides patterns for processing external, potentially untrusted data through machine learning models, creating an inherent surface for indirect prompt injection.
Ingestion points: User-supplied text in pipelines (references/pipelines.md), external datasets loaded for training (references/training.md), and prompts for text generation (references/generation.md).
Boundary markers: No specific adversarial boundary markers are suggested; the skill relies on standard library tokenization and processing.
Capability inventory: Capability to write to the local file system (save_pretrained) and perform network uploads (push_to_hub).
Sanitization: Documentation mentions basic filtering using bad_words_ids, but primarily utilizes default library behaviors for input processing.
[COMMAND_EXECUTION]: The skill includes shell commands for package installation and setting environment variables for authentication, which are standard for the development environment.

transformers