ASRD v1.0.0

First versioned release of the Adversarial Surface-Form Robustness Dataset (ASRD), the dataset accompanying Quad-State Safety Evaluation of Open-Weight Large Language Models on Non-Canonical Inputs (EvoRobust @ NeurIPS 2026).

This is the version evaluated in the paper.

Contents

• 2,100 prompts: 300 harmful seeds across 6 risk categories adapted from HarmBench, each rendered in 7 prompt families (F0 baseline, F1 emoji, F2 homoglyph, F3 leetspeak and emoji, F4 Base64 encoded wrappers, F5 invisible Unicode, F6 hybrid)
• data/prompts_master.{parquet,jsonl,csv} and one CSV per family in data/families/
• build_dataset.py and a dependency-free load_dataset.py

License and responsible use

Released under CC BY-NC 4.0 for AI safety research and evaluation. Use is subject to the ASRD Responsible Use Policy. The dataset contains requests only and does not include instructions or answers to those requests.

Also available on the Hugging Face Hub (gated): https://huggingface.co/datasets/pavanmaddula/ASRD-Dataset