The Agent Skills Directory

[SAFE]: The skill uses standard data science libraries (numpy, pandas, scikit-learn, PIL) to perform legitimate machine learning tasks. No malicious code or hidden behaviors were detected in the provided Python scripts or configuration files.
[INDIRECT_PROMPT_INJECTION]: The skill processes external datasets and image files which could theoretically contain malicious instructions. However, the data_preprocessing_example.py script implements basic file extension validation (.jpg, .png), and the overall capability set is well-aligned with the stated purpose of model fine-tuning.
Ingestion points: Processes local image directories provided via command-line arguments and datasets specified in example_config.json (e.g., Hugging Face datasets).
Boundary markers: Not explicitly present in the provided snippets, but the skill follows a structured workflow for model adaptation.
Capability inventory: The skill uses Bash(cmd:*), Read, Write, and Edit tools to automate model training and documentation.
Sanitization: Implements basic filtering by file extension in the preprocessing logic.
[CREDENTIALS_SAFE]: The example_config.json file uses null placeholders for sensitive fields like hub_token, encouraging users to provide credentials through secure channels rather than hardcoding them in the skill itself.
[EXTERNAL_DOWNLOADS]: The skill mentions downloading pre-trained models (e.g., ResNet50, BERT) and datasets from well-known services like the Hugging Face Model Hub. These are legitimate operations for the stated purpose of transfer learning.

adapting-transfer-learning-models