The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: Fetches pre-trained vision models and datasets from the Hugging Face Hub (huggingface.co), which is a well-known and trusted service. It also manages several Python dependencies including transformers, timm, monai, and trackio through PEP 723 inline script metadata.
[COMMAND_EXECUTION]: Runs Python utilities and submits training jobs using uv run and the Hugging Face Jobs API. The skill workflow involves saving scripts to a local workspace directory before execution for transparency.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted datasets from the Hugging Face Hub, which represents a surface where malicious instructions could be embedded in dataset metadata or content.
Ingestion points: Data enters the system via the load_dataset function in the primary training scripts (object_detection_training.py, image_classification_training.py, and sam_segmentation_training.py) and the dataset_inspector.py tool.
Boundary markers: The processing logic does not employ explicit delimiters or specific prompts to ignore instructions that might be embedded in the external dataset content.
Capability inventory: The skill has the capability to submit remote GPU training jobs, perform network operations to push results to the Hub, and write files to the local workspace.
Sanitization: Bounding box coordinates and image dimensions are validated during the preprocessing phase (e.g., the sanitize_dataset function) to ensure dataset integrity and prevent runtime errors.

hugging-face-vision-trainer