The Agent Skills Directory

[COMMAND_EXECUTION]: The skill requires the agent to execute shell commands using torchrun and sbatch for launching distributed training jobs. These are standard operations for ML engineering workflows.
Evidence: Multiple shell script examples and torchrun commands are provided throughout SKILL.md and the reference files.
[EXTERNAL_DOWNLOADS]: The skill instructs the installation of standard machine learning libraries and the use of official NVIDIA Docker containers.
Evidence: pip install megatron-core and docker run nvcr.io/nvidia/pytorch:25.04-py3 in SKILL.md.
Context: These are well-known, official sources (NVIDIA) and represent standard practice.
[INDIRECT_PROMPT_INJECTION]: The skill defines a surface for processing large external datasets for training, which could theoretically contain malicious instructions.
Ingestion points: The --data-path parameter in training scripts (found in SKILL.md and references/training-recipes.md).
Boundary markers: None specified; standard ML training typically does not use NL delimiters for raw data tokens.
Capability inventory: The skill has the capability to write local shell scripts, execute system commands (torchrun), and initiate network communication for distributed training.
Sanitization: Not applicable as the data is consumed by low-level ML training loops rather than an LLM prompt processor.
[SAFE]: The skill is a high-quality educational and operational resource for ML training that adheres to safety and integrity rules. It uses trusted industry references such as NVIDIA, Meta, and Microsoft.

training-llms-megatron