peft-fine-tuning
Pass
Audited by Gen Agent Trust Hub on Apr 4, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADS
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill references and downloads pre-trained model weights and datasets from Hugging Face's official repositories (e.g., meta-llama/Llama-3.1-8B, databricks/databricks-dolly-15k). Hugging Face is an industry-recognized trusted organization. Troubleshooting guides also mention installing libraries from official source repositories (e.g., bitsandbytes from TimDettmers on GitHub).
- [SAFE]: No malicious patterns, obfuscation, persistence mechanisms, or unauthorized privilege escalations were detected. The content consists of legitimate technical documentation for the PEFT ecosystem.
- [PROMPT_INJECTION]: The skill presents an indirect prompt injection surface by documenting the ingestion of external datasets for training. 1. Ingestion points: load_dataset calls in SKILL.md and training examples in references/advanced-usage.md. 2. Boundary markers: Not present in example training scripts, which is typical for fine-tuning tutorials. 3. Capability inventory: model.save_pretrained (file-write), push_to_hub (network), and trainer.train (compute execution). 4. Sanitization: No explicit data sanitization or input validation is shown in the training examples.
Audit Metadata