The Agent Skills Directory

[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface (Category 8) by design, as it is intended to process external datasets for model fine-tuning.
Ingestion points: The skill ingests training data from train.jsonl and valid.jsonl as described in SKILL.md.
Boundary markers: No explicit boundary markers or instructions to ignore embedded prompts within the JSONL data are specified.
Capability inventory: The skill utilizes the mlx_lm package to execute fine-tuning logic, which involves intensive processing of the provided data.
Sanitization: No sanitization or validation logic for the content of the training messages is provided in the skill scripts.
[EXTERNAL_DOWNLOADS]: The environment validation script contains a reference to the official installer for the 'uv' package manager.
Evidence: The script scripts/validate_environment.py suggests installing uv using curl -LsSf https://astral.sh/uv/install.sh | sh if it is not found on the system. This targets the official domain of a well-known developer tool.
[COMMAND_EXECUTION]: The scripts/validate_environment.py script executes several system commands to verify the environment.
Evidence: It uses subprocess.run to call sysctl for hardware information and uv --version to check for the package manager. These are standard environment discovery operations.

mlx-fine-tuning