The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill provides instructions to download the official torchtitan repository and its dependencies from the PyTorch GitHub organization, which is a trusted source.
[CREDENTIALS_UNSAFE]: The documentation includes examples of using Hugging Face tokens for asset downloads, but correctly utilizes placeholders (e.g., YOUR_HF_TOKEN) and directs users to their official settings page, avoiding the exposure of actual secrets.
[COMMAND_EXECUTION]: The skill lists standard command-line instructions for launching distributed training using torchrun and srun (SLURM), which are routine operations for the described workflow.
[INDIRECT_PROMPT_INJECTION]: The skill defines ingestion points for external datasets (e.g., C4). While this represents a theoretical attack surface common to all training frameworks, the risk is mitigated as the data is consumed by the model training process rather than being used to influence agent instructions. This aligns with the skill's primary purpose.
[SAFE]: No obfuscation, persistence mechanisms, or unauthorized privilege escalation patterns were found. All referenced resources trace back to well-known and trusted infrastructure.

distributed-llm-pretraining-torchtitan