tao-finetune-cosmos-embed
Pass
Audited by Gen Agent Trust Hub on Sep 2, 2026
Risk Level: SAFE
Full Analysis
- [SAFE]: The skill uses official Docker images from the TAO Toolkit (referenced via
resolve_tao_image.pyfrom NVIDIA's skill bank) to run fine-tuning and inference tasks. All command execution occurs within these isolated containers. - [SAFE]: The
protobuf<7installation is a documented workaround for internal dependency conflicts within the official NVIDIA container environment. - [SAFE]: Data exfiltration analysis shows that
HF_TOKENis passed as an environment variable to the Docker container, which is a standard and recommended practice for accessing gated models on HuggingFace Hub. The skill instructions explicitly warn against writing tokens into logs or specs. - [SAFE]: External downloads (e.g.,
google-bert/bert-base-uncased) are performed by the model's underlying libraries from well-known registries (HuggingFace) to fetch necessary model components.
Audit Metadata