peft-fine-tuning
Pass
Audited by Gen Agent Trust Hub on Oct 23, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill demonstrates training workflows that ingest external datasets, creating a surface for indirect prompt injection where malicious instructions could be embedded in the fine-tuning data.
- Ingestion points: Data is loaded via the
load_datasetfunction from Hugging Face Datasets inSKILL.md. - Boundary markers: The instructions use basic headers (e.g.,
### Instruction:) but do not implement robust sanitization or delimiters to prevent the model from obeying instructions embedded within the training examples. - Capability inventory: The training process involves network communication with the Hugging Face Hub and local file-system operations for saving and merging model adapters.
- Sanitization: There is no mention of input validation or content filtering for the datasets being processed.
- [EXTERNAL_DOWNLOADS]: The skill provides instructions for downloading configuration, model weights, and datasets from well-known and reputable services such as Hugging Face and PyPI.
Audit Metadata