tao-train-grounding-dino
Pass
Audited by Gen Agent Trust Hub on Jun 18, 2026
Risk Level: SAFE
Full Analysis
- [SAFE]: The skill consists of instructions, JSON schemas, and configuration templates for model training and deployment. No malicious code or prompt injection patterns were identified.
- [EXTERNAL_DOWNLOADS]: The skill documentation mentions that BERT text encoder weights ('bert-base-uncased') are downloaded from the HuggingFace Hub if a local path is not provided. This is a standard and well-known source for machine learning models.
- [COMMAND_EXECUTION]: The documentation provides examples of 'docker run' commands and TAO toolkit CLI usage ('grounding_dino train', 'grounding_dino evaluate'). These commands are intended for the primary purpose of the skill and utilize official NVIDIA container images ('nvcr.io/nvidia/tao/tao-toolkit').
- [DATA_EXFILTRATION]: The skill incorporates an optional integration with Weights & Biases (WandB) for training telemetry. This is a standard practice in machine learning for experiment tracking.
Audit Metadata