The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: Fetches model weights, dataset samples, and technical documentation from established machine learning platforms including HuggingFace (huggingface.co) and GitHub (github.com/huggingface). These are well-known services and essential for machine learning workflows.
[COMMAND_EXECUTION]: Orchestrates Docker builds and training executions using system tools like docker, bash, and torchrun. These operations are standard for a containerized post-training environment and are executed with user-aware permissions (--user $(id -u):$(id -g)) to ensure host file system security.
[PROMPT_INJECTION]: Possesses an indirect prompt injection surface as part of its primary functionality. The skill reads external model cards and documentation to generate local Python scripts. This research-based generation is a core feature of the skill, and the risk is moderated by the focus on official model repositories and developer-controlled environments.
[DYNAMIC_EXECUTION]: Pseudocode logic in the compatibility registry references the use of eval() to process hardware and model detection rules. This evaluation is restricted to the skill's own internal reference data and serves the primary purpose of automating environment setup.

tao-finetune-huggingface-model