fine-tuning-with-trl

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

• Third-party content exposure detected (high risk: 0.80). The skill instructs the agent to load and train on public datasets (e.g., load_dataset("trl-lib/Capybara"), load_dataset("trl-lib/ultrafeedback_binarized"), --dataset_name argilla/Capybara-Preferences, or load_dataset("json", data_files="preferences.json")), which are open third‑party/user‑generated sources that the agent ingests as part of its workflow and could contain untrusted instructions.