moe-training
Pass
Audited by Gen Agent Trust Hub on Oct 1, 2026
Risk Level: SAFE
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill includes instructions to clone the
Megatron-DeepSpeedrepository from Microsoft's official GitHub organization and install standard machine learning libraries (DeepSpeed, Transformers, Accelerate, Torch) from official package registries. These are well-known, trusted resources used in standard research and development workflows. - [COMMAND_EXECUTION]: Provides detailed templates for shell commands to initiate model training via the
deepspeedCLI. These commands utilize documented parameters for model configuration (e.g., hidden layers, expert parallel size) and are strictly aligned with the skill's stated purpose of training large-scale models. - [SAFE]: All code snippets are transparent PyTorch implementations for MoE layers, routing mechanisms, and load balancing. The skill adheres to best practices, such as suggesting lower learning rates for MoE training and providing clear documentation for auxiliary losses. No evidence of prompt injection, data exfiltration, or obfuscation was found.
Audit Metadata