The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides Python code snippets in references/deployment-optimization.md that utilize subprocess.run to execute external model conversion and quantization tools. These operations rely on paths derived from environment variables or local file structures, which could be exploited if environment variables are untrusted.- [EXTERNAL_DOWNLOADS]: The skill facilitates the download of pre-trained models from the Hugging Face Hub and relies on a large set of third-party Python dependencies for machine learning tasks.- [DATA_EXFILTRATION]: The provided scripts contain options to upload model adapters and tokenizers to the Hugging Face Hub and to transmit training metrics to Weights & Biases, involving the network transfer of model-related data.- [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection as it processes untrusted training data from JSONL files.
Ingestion points: Training and validation data are loaded from local files such as train.jsonl and data.jsonl across multiple reference files.
Boundary markers: The processing logic does not implement delimiters or instructions to distinguish between training data and agent commands.
Capability inventory: The skill includes capabilities for subprocess execution, network communication with external ML hubs, and local file system modifications.
Sanitization: A quality filter is provided in references/dataset-preparation.md to remove poor-quality training examples, but it does not perform security-focused sanitization or validation of input content.

fine-tuning-expert