fine-tuning-with-trl

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (high risk: 0.90). The skill's required workflows explicitly load public, user-contributed datasets (e.g., datasets.load_dataset("trl-lib/Capybara"), "trl-lib/ultrafeedback_binarized", and CLI dataset_name/argilla/Capybara-Preferences in SKILL.md and the Workflow steps) which the agent ingests to train SFT/reward/PPO/GRPO models—content that is untrusted and can materially alter model behavior, so it could enable indirect prompt injection.