training-llms-megatron
Pass
Audited by Gen Agent Trust Hub on Sep 16, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill references official NVIDIA Docker containers on
nvcr.ioand the installation of themegatron-corepackage from the standard Python registry. These are well-known and trusted services. - [COMMAND_EXECUTION]: The skill provides shell script templates and command-line instructions for using
torchrun,bash, and SLURM (sbatch) to orchestrate distributed GPU training. These are standard practices for model training workflows. - [INDIRECT_PROMPT_INJECTION]: The skill defines workflows that ingest external datasets, creating a potential surface for indirect prompt injection if untrusted data is used.
- Ingestion points: Dataset files, vocabularies, and merge files loaded via
--data-path,--vocab-file, and--merge-fileparameters. - Boundary markers: Absent; training scripts are designed to process raw data inputs directly.
- Capability inventory: High-privilege command execution for distributed training and shell access.
- Sanitization: Absent; the skill expects users to provide pre-processed and trusted datasets.
Audit Metadata