quantizing-models-bitsandbytes
Pass
Audited by Gen Agent Trust Hub on Sep 9, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill describes workflows for fine-tuning models on external datasets, creating a surface where malicious instructions embedded in the data could potentially influence the agent's behavior during training or evaluation.\n
- Ingestion points: Loading datasets from external sources (e.g., HuggingFace Datasets) in
references/qlora-training.md.\n - Boundary markers: The provided code snippets do not include explicit delimiters or instructions to ignore potential commands within the dataset content.\n
- Capability inventory: The workflows include file system write operations (
save_pretrained,output_dir) and the use of theaccelerate launchcommand for distributed training.\n - Sanitization: The instructions do not specify validation or sanitization steps for the ingested dataset content.
Audit Metadata