The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill fetches datasets and tool components from external sources. * Downloads datasets from HuggingFace ('huggingface.co') and user-provided URLs in 'scripts/ingest_dataset.py'. * Provisions the 'tao-skill-bank' plugin from NVIDIA's official GitHub repository ('https://github.com/NVIDIA-TAO/tao-skill-bank') during installation. * Utilizes Docker images from 'nvcr.io' (NVIDIA Container Registry) for the runtime environment.
[COMMAND_EXECUTION]: The skill orchestrates multiple tools and scripts using subprocess execution. * Executes Python scripts and binaries like 'trtexec' and 'ds_image_eval' to perform evaluation and performance benchmarking. * Uses 'subprocess.Popen' in the optional web UI backend ('app/server.py') to manage long-running training and evaluation tasks. * Environment variables are filtered through an allowlist before being passed to child processes to prevent accidental credential exposure.
[DYNAMIC_EXECUTION]: Local components are compiled at runtime to match the execution environment. * Compiles the 'ds_image_eval' GStreamer application from source ('scripts/ds_image_eval.c') using a provided Makefile. * Compiles a custom RT-DETR bbox parser ('rtdetr_parser.cpp') for use with DeepStream's 'nvinfer' plugin.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data which serves as a potential injection surface. * Ingests datasets (COCO/YOLO formats) from external URLs or HuggingFace IDs. * While the skill includes safety checks for archive extraction (zip-slip and zip-bomb protection), the resulting labels and metadata are interpolated into prompts and training configurations.
[SAFE]: Heuristic detections for dynamic execution ('eval()') were found to be false positives. * Instances of '.eval()' in the Python source code refer to the standard PyTorch 'model.eval()' method used to set neural networks to evaluation mode, not the dangerous Python 'eval()' built-in function.

deepstream-eval-and-finetune