The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes model definitions from external files, which represents a potential surface for indirect instruction injection.
Ingestion points: The modelPath parameter in SKILL.md defines the source for model code to be loaded into the agent's execution environment.
Boundary markers: The instructions do not specify the use of delimiters or warnings to prevent the agent from following instructions embedded within the loaded model code.
Capability inventory: The skill is configured with access to the Bash and Write tools, which could be leveraged if the agent follows malicious instructions in the data.
Sanitization: No sanitization or content validation for the input files is described in the provided instructions.
[DYNAMIC_EXECUTION]: The skill is designed to load and execute logic from calculated file paths provided at runtime.
The input schema utilizes a modelPath property to identify and load Python code for model definitions, which involves dynamic loading of code during the training process.
[EXTERNAL_DOWNLOADS]: The skill references several well-known machine learning libraries and experiment tracking services.
The documentation describes integration with established tools including PyTorch, PyTorch Lightning, MLflow, Optuna, Weights & Biases (W&B), and Ray.

pytorch-trainer