The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external user-supplied datasets and validation files which enter the agent's context during the training preparation phase.\n- Ingestion points: scripts/prepare_finetune_recipe.py and scripts/validate_nn_improvement.py read user-provided JSONL and CSV files.\n- Boundary markers: The skill does not use explicit delimiters in prompts but enforces strict structural validation on all ingested data.\n- Capability inventory: The skill can write configuration files (recipes) and execute shell commands via scripts/finetune_anomalygennext.sh.\n- Sanitization: Implements robust regex validation (SAFE_NAME) for dataset and job names, strictly whitelists image extensions to supported formats (.jpg, .jpeg, .png), and verifies mask integrity by ensuring pixel values are exactly 0 and 255.\n- [COMMAND_EXECUTION]: The skill executes training workloads and validation logic using shell commands. This includes the use of torchrun and Python-based validation scripts. Test scripts (scripts/tests/) appropriately use subprocess.run for functional verification in a controlled environment.

tao-finetune-anomalygennext