The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The documentation in references/getting_started.md provides instructions for system-level administrative tasks, such as using sudo to uninstall CUDA drivers (sudo /usr/local/cuda-12.1/bin/cuda-uninstaller).
[COMMAND_EXECUTION]: The skill guides the user and agent through executing various command-line operations using llamafactory-cli, torchrun, and accelerate to manage model training and inference.
[INDIRECT_PROMPT_INJECTION]: The skill defines workflows for ingesting and processing untrusted external datasets for fine-tuning, which presents an attack surface where malicious instructions could be embedded in the training data.
Ingestion points: User-provided datasets in Alpaca, ShareGPT, and KTO formats as described in references/getting_started.md.
Boundary markers: Absent; there are no specific instructions or delimiters provided to prevent the agent from being influenced by instructions contained within the datasets.
Capability inventory: The skill instructions facilitate shell execution via the llamafactory-cli and script execution for training and evaluation.
Sanitization: Absent; the documentation does not mention validation or filtering of input dataset content.

llama-factory