The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill provides instructions to download and install official libraries and models from trusted sources, specifically the PyTorch and Meta-Llama GitHub organizations. These references are standard for the described machine learning workflows.- [COMMAND_EXECUTION]: Instructions include standard operations for distributed training, such as torchrun, srun for SLURM job submission, and pip for dependency management. These commands are localized to the training environment and follow established developer patterns.- [CREDENTIALS_UNSAFE]: The skill correctly handles sensitive information by using placeholders (e.g., YOUR_HF_TOKEN) and directing users to official settings pages for credential generation, rather than hardcoding actual secrets.- [INDIRECT_PROMPT_INJECTION]: The skill involves processing model configurations (TOML) and training checkpoints. While these represent a data ingestion surface common to ML tools, there are no malicious patterns or unsafe interpolation of untrusted data identified.

distributed-llm-pretraining-torchtitan