fine-tuning-with-trl

Pass

Audited by Gen Agent Trust Hub on Sep 15, 2026

Risk Level: SAFEEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The skill provides instructions to install the trl, transformers, datasets, peft, and accelerate packages from official registries. It also includes examples that fetch datasets and model weights from Hugging Face's official repository (e.g., trl-lib/Capybara, Qwen/Qwen2.5-0.5B). These are well-known services and the downloads are standard for the described ML workflows.- [INDIRECT_PROMPT_INJECTION]: The skill facilitates the ingestion of external data for model training purposes, creating a surface for indirect prompt injection. Malicious instructions embedded in processed datasets could potentially influence the training process or the resulting model behavior.
  • Ingestion points: Code examples in SKILL.md (lines 72, 105, 233), references/reward-modeling.md (line 21), and references/sft-training.md (line 49) demonstrate importing data from external sources using load_dataset().
  • Boundary markers: The documentation does not implement or recommend the use of delimiters or 'ignore embedded instructions' markers for processing external data.
  • Capability inventory: The training scripts involve filesystem write capabilities through trainer.save_model() and subprocess execution via library-specific CLI tools.
  • Sanitization: No sanitization, validation, or filtering of the ingested dataset content is included in the documentation examples.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 15, 2026, 05:31 PM
Security Audit — agent-trust-hub — fine-tuning-with-trl