book-sft-pipeline

Pass

Audited by Gen Agent Trust Hub on Sep 14, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process external literary works (ePub files) to generate fine-tuning datasets.
  • Ingestion points: Text extraction logic in SKILL.md using BeautifulSoup and paragraph-level segmentation in scripts/pipeline_example.py.
  • Boundary markers: The pipeline produces structured JSONL files with explicit role-based message formatting (system, user, assistant).
  • Capability inventory: Includes file system access for reading ePubs/JSONL files and network operations for model training via the Tinker API.
  • Sanitization: The extraction process includes paragraph-level filtering and removal of metadata to prevent noise in the training data.
  • [EXTERNAL_DOWNLOADS]: The skill references external datasets and research materials.
  • Evidence: Downloads a sample style-transfer dataset from Hugging Face (huggingface.co/datasets/MuratcanKoylan/gertrude-stein-style-sft) and references a research paper on ArXiv (arxiv.org/pdf/2510.13939). Both sources are well-known and consistent with the skill's stated purpose.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 14, 2026, 06:42 PM
Security Audit — agent-trust-hub — book-sft-pipeline