The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill references a pretrained backbone model hosted on GitHub (SwinTransformer repository). This is a standard operation in machine learning for initializing architecture weights from established research sources.
[COMMAND_EXECUTION]: Utilizes the Bash tool to execute oneformer CLI commands for model training, quantization, and inference within a containerized TAO Toolkit environment.
[INDIRECT_PROMPT_INJECTION]: The skill ingests external data including annotation files and label maps which are processed by the underlying model tools, presenting a surface for indirect instructions.
Ingestion points: Dataset inputs defined in references/skill_info.yaml, such as annotations.json and label_map.json, are loaded into the agent's context and model environment.
Boundary markers: Absent. There are no explicit instructions or delimiters used to warn the agent to ignore potentially malicious content embedded within the dataset files.
Capability inventory: The skill possesses the capability to execute shell commands via the Bash tool to facilitate model operations.
Sanitization: Absent. There is no evidence of specific validation or sanitization of the content within the ingested dataset files before processing.
[DYNAMIC_EXECUTION]: The documentation mentions using the TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=1 environment variable to allow loading of full PyTorch checkpoints. The author correctly includes a safety warning against using this setting with untrusted checkpoints to mitigate the risk of arbitrary code execution via unsafe deserialization.

tao-train-oneformer