peft-fine-tuning

Pass

Audited by Gen Agent Trust Hub on Sep 17, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process external datasets for fine-tuning purposes, which introduces an attack surface for instructions embedded in data.
  • Ingestion points: External datasets are loaded using the datasets.load_dataset function in SKILL.md (e.g., "databricks/databricks-dolly-15k").
  • Boundary markers: The code utilizes delimited prompt templates (e.g., "### Instruction:", "### Response:") to separate fields, which helps mitigate instruction confusion.
  • Capability inventory: The skill utilizes transformers.Trainer and model.save_pretrained which perform write operations to the local file system.
  • Sanitization: No explicit sanitization or filtering of the dataset content is described before the tokens are processed by the training loop or saved to disk.
  • [EXTERNAL_DOWNLOADS]: The skill documentation includes commands to fetch code and models from external remote sources.
  • Fetches packages from official registries via pip install peft transformers bitsandbytes.
  • Downloads pre-trained models and datasets from the Hugging Face Hub (e.g., "meta-llama/Llama-3.1-8B").
  • Provides instructions in references/troubleshooting.md to download source code for manual compilation from github.com/TimDettmers/bitsandbytes to resolve CUDA environment issues.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 17, 2026, 07:53 PM
Security Audit — agent-trust-hub — peft-fine-tuning