The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the user to clone the Megatron-DeepSpeed repository from Microsoft's official GitHub account. This is a reference to a well-known and trusted technology organization.
Evidence: git clone https://github.com/microsoft/Megatron-DeepSpeed in SKILL.md and references/training.md.
[INDIRECT_PROMPT_INJECTION]: The skill defines a workflow for training large-scale models that ingest external data. This creates an attack surface where maliciously crafted training data or vocabulary files could theoretically influence model behavior.
Ingestion points: Untrusted training data, vocabulary files, and merge files are ingested via command-line arguments such as --data-path, --vocab-file, and --merge-file in SKILL.md and references/training.md.
Boundary markers: The provided training script templates do not specify boundary markers or input validation for the ingested data.
Capability inventory: The skill utilizes subprocess execution for training via the deepspeed CLI and performs file system reads on the provided data paths.
Sanitization: No sanitization or filtering of the input training data is described in the provided instructions.

moe-training