modelopt-qad
Pass
Audited by Gen Agent Trust Hub on Nov 13, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [SAFE]: The skill provides detailed, legitimate instructions for Quantization-Aware Distillation (QAD) using the ModelOpt library. The functionality aligns with the stated purpose of model accuracy recovery.
- [COMMAND_EXECUTION]: The skill directs the agent to execute specific Python scripts such as
quantize.py,distill.py, andexport.py. It also includes instructions for managing Slurm jobs. These operations are essential to the distillation workflow and are intended for use in the user's controlled development environment. - [INDIRECT_PROMPT_INJECTION]: The skill involves processing external datasets, including a default Nvidia dataset (
nvidia/Nemotron-Cascade-2-SFT-Data) and user-provided sources. While data ingestion presents a theoretical surface for indirect prompt injection, the risk is inherent to the machine learning training task, and the skill includes technical guardrails such as validation checks and loss monitoring to detect anomalies. - Ingestion points: Dataset paths in
SKILL.md(e.g.,nvidia/Nemotron-Cascade-2-SFT-Data,examples/megatron_bridge/data/nemotron-cascade-2-blend.yaml). - Boundary markers: Absent in the prompt templates.
- Capability inventory: Execution of local Python scripts and Slurm job management.
- Sanitization: Not explicitly detailed for data content.
Audit Metadata