The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill handles untrusted data as part of its core functionality for training machine learning models on external datasets and pre-trained weights.
Ingestion points: The skill ingests model assets, tokenizers, and datasets (such as C4) from Hugging Face and other external repositories as part of the training workflow.
Boundary markers: Boundary markers are not explicitly defined for dataset content, as the skill relies on standard PyTorch and TorchTitan data-loading abstractions.
Capability inventory: The skill is capable of executing training scripts (run_train.sh), distributed execution via torchrun, and package management through pip.
Sanitization: Training data and model weights are processed directly by the training engine without external sanitization.
[EXTERNAL_DOWNLOADS]: The skill facilitates the download of framework code and assets from well-known sources.
Fetches the TorchTitan library from the official PyTorch GitHub repository.
Downloads model checkpoints and tokenizers from Hugging Face for Llama 3.1 and other supported models.
Installs the torchao optimization library from its official repository to enable Float8 support.
[COMMAND_EXECUTION]: The skill uses standard command-line interfaces for training and setup.
Executes pip for dependency management and library installation.
Utilizes torchrun and SLURM commands (srun, sbatch) for orchestrating multi-node distributed training jobs.

distributed-llm-pretraining-torchtitan