data/external_augmented/extended/README.md

Extended Outputs (Research Only)

This directory is intentionally empty in redistributable packages.

Extended datasets have additional usage restrictions (e.g., CC‑BY‑NC or custom research clauses). Generate them locally at runtime using the adapters in scripts/ and do not commit or redistribute the resulting .jsonl files or exported images.

What gets generated here

• AgentHarm (scripts/adapters/b4_agentharm_adapter.py) → extended/B4_augmented_agentharm.jsonl
• Mindgard (scripts/adapters/b3_mindgard_adapter.py) → extended/B3_augmented_mindgard.jsonl (gated; may require access)
• MM-SafetyBench (scripts/adapters/b2_mmsafetybench_adapter.py) → extended/B2_augmented_mm-safetybench.jsonl (+optional exported images)

One-command reconstruction

From ART-SafeBench-v2.0.0/scripts/:

python reconstruct_external_data.py --extended

Notes:

• If Mindgard fails with a gated/403 error, request access on HuggingFace and login:
  • huggingface-cli login
• For MM-SafetyBench you can optionally select configs/splits and export images (do not redistribute):
  • python reconstruct_external_data.py --extended --mmsafetybench-configs EconomicHarm,Fraud --mmsafetybench-splits SD,SD_TYPO --mmsafetybench-export-images