The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: Downloads model checkpoints and artifacts (~140 GB) from Hugging Face (nvidia/Cosmos-Predict2-, google-t5/, etc.) and NVIDIA NGC (NVDINOV2/nv_dinov2_classification_model.ckpt). These are well-known and trusted model registries.
[EXTERNAL_DOWNLOADS]: Fetches training datasets from GitHub (abin24/Magnetic-tile-defect-datasets) and Roboflow (universe.roboflow.com), which are established services for hosting research and machine learning data.
[COMMAND_EXECUTION]: Instructs the agent to run various shell scripts (run_sdg.sh, run_eval.sh, run_round.sh, launch_training.sh) and Python modules to orchestrate a complex machine learning pipeline.
[COMMAND_EXECUTION]: Documents the use of chmod 777 and chown as fallback mechanisms for resolving Docker volume mount permission issues in a user-facing reference file.
[CREDENTIALS_UNSAFE]: The skill requires and handles an HF_TOKEN environment variable to authenticate with Hugging Face for gated model access. This follows standard industry practices for secret management in AI workflows.
[DATA_EXPOSURE]: Recommends mounting sensitive system files /etc/passwd and /etc/group into Docker containers. This is used to resolve host UIDs within the containerized environment to ensure correct file ownership and is a documented best practice for this use case.
[PROMPT_INJECTION]: Indirect prompt injection surface identified where the skill processes user-defined text (roi_prompt_defect_location) from a JSONL file to guide vision-language models.
Ingestion points: The defect_spec JSONL file, specifically the roi_prompt_defect_location field used in Phase 2 for regional identification.
Boundary markers: No explicit delimiters are described to separate user input from internal prompt logic.
Capability inventory: The pipeline includes executing shell commands, file system operations, and network downloads.
Sanitization: No explicit sanitization or filtering of the user-provided prompt strings is documented.

paidf-anomalygen