The Agent Skills Directory

[SAFE]: The skill is focused on providing legitimate PyTorch training recipes, including implementation of the standard five-step training loop, mixed precision (AMP), and optimizers like AdamW. All code snippets and instructions align with documented PyTorch 2.x best practices.
[COMMAND_EXECUTION]: The skill references the use of torchrun in references/distributed.md for launching distributed training processes across multiple GPUs. This is a standard command-line utility in the PyTorch ecosystem and is used here in a proper, non-malicious context.
[DATA_EXFILTRATION]: No unauthorized network operations or sensitive file access patterns were detected. The skill uses standard distributed training communication backends (NCCL/Gloo) for inter-GPU synchronization and only references official documentation at pytorch.org.
[PROMPT_INJECTION]: There are no patterns of prompt injection, role-play bypasses, or instructions to ignore safety guidelines. The content is strictly technical and educational.
[CREDENTIALS_UNSAFE]: The skill does not contain hardcoded API keys, tokens, or private credentials. It correctly references environment variables like LOCAL_RANK for distributed process identification, which is standard procedure.

deep-learning