The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the user to install standard Python packages including transformers, datasets, tokenizers, torch, and bitsandbytes. It also downloads pretrained models from the Hugging Face Model Hub, which is a well-known service in the machine learning ecosystem. Several examples reference models from trusted organizations like Facebook (Meta).
[COMMAND_EXECUTION]: The skill includes shell commands for package installation using pip to set up the required environment.
[INDIRECT_PROMPT_INJECTION]: The skill defines a surface for indirect prompt injection as it processes untrusted text and sequence data through various inference pipelines (classifier(text), tokenizer(texts), ner_pipe(text)). While this is a standard operational characteristic of machine learning models, it represents an ingestion point for external data into the agent's context. Evidence chain:
- Ingestion points: Text inputs to model pipelines and tokenizers in SKILL.md.
- Boundary markers: Not specified in the documentation.
- Capability inventory: Python execution for model inference and training; no dangerous file-system or network operations beyond standard library use are demonstrated.
- Sanitization: Not present in the instructional code snippets.

transformers