The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill provides instructions for utilizing installer scripts that download and execute additional batch files (LAUNCHER-TOOLKIT.bat and SECURE_LAUNCHER-TOOLKIT.bat) from the Aitrepreneur/FLX repository on HuggingFace. Running scripts from unverified community repositories on platforms like HuggingFace poses a high risk of unauthorized code execution.
[EXTERNAL_DOWNLOADS]: The skill fetches its core training logic and setup environment from community-maintained external sources, including the ostris/ai-toolkit repository on GitHub and the Aitrepreneur/FLX repository on HuggingFace.
[COMMAND_EXECUTION]: The installation process involves executing shell commands to manage Python virtual environments, install dependencies via pip, and build components using npm. The skill also executes python run.py to initiate training jobs based on configuration files.
[INDIRECT_PROMPT_INJECTION]: The skill processes natural language text captions provided in user datasets to train models. This represents an attack surface for indirect injection where malicious instructions embedded in captions could influence the training backend.
Ingestion points: Dataset folders containing .txt caption files (e.g., my_dataset/001.txt) processed by the Python backend.
Boundary markers: No explicit delimiters or boundary markers are mentioned to prevent the model or trainer from interpreting embedded instructions within the captions.
Capability inventory: The trainer uses python run.py, pip install, and npm run, providing multiple vectors for command execution if the input data compromises the application flow.
Sanitization: The instructions do not describe any sanitization, filtering, or validation performed on the natural language captions before processing.

ai-toolkit-trainer