fine-tuning-with-trl
Pass
Audited by Gen Agent Trust Hub on Sep 9, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [EXTERNAL_DOWNLOADS]: Fetches pre-trained models and datasets from Hugging Face Hub. The skill references models such as Qwen2.5 and datasets from the TRL library and Argilla.
- [INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to ingest and process external datasets for training and alignment tasks.
- Ingestion points: The skill uses
datasets.load_dataset()to pull content from external repositories (e.g.,trl-lib/Capybara,argilla/Capybara-Preferences) into the training pipeline in SKILL.md, references/reward-modeling.md, and references/sft-training.md. - Boundary markers: Boundary markers or specific instructions to ignore embedded prompts in the training data are absent in the provided example scripts.
- Capability inventory: The skill demonstrates training capabilities, including file writing (saving models to
output_dir) and shell command execution via TRL CLI tools. - Sanitization: Example code does not include explicit sanitization or validation of the ingested dataset content before processing.
Audit Metadata