| --- |
| license: mit |
| language: |
| - en |
| - ar |
| - fa |
| - tr |
| pretty_name: The Alignment Veto (questions + human answers) |
| tags: |
| - cultural-alignment |
| - values |
| - survey |
| - mena |
| - llm-evaluation |
| size_categories: |
| - 1K<n<10K |
| configs: |
| - config_name: questions |
| data_files: questions.csv |
| - config_name: human_answers |
| data_files: human_answers.csv |
| --- |
| |
| # The Alignment Veto: questions and human answers |
|
|
| Questions and human survey answers from **The Alignment Veto: How Safety Training Suppresses Cultural Knowledge in LLMs** (EMNLP 2026). |
|
|
| [Paper](https://huggingface.co/papers/2510.13154) · [Website](https://llm-lab-org.github.io/Alignment-Veto/) · [Code](https://github.com/llm-lab-org/Alignment-Veto) · [Model responses (gated)](https://huggingface.co/datasets/llm-lab/alignment-veto-responses) |
|
|
| <video controls preload="metadata" src="https://huggingface.co/datasets/llm-lab/alignment-veto/resolve/main/alignment_veto_demo.mp4" width="720"></video> |
|
|
| ## Files |
|
|
| | File | Rows | What it is | |
| |---|---|---| |
| | `questions.csv` | **864** (all) | `question_id`, `source` (`wvs_wave7` or `arab_opinion_index`), `tier` (1 benign · 2 moderate · 3 sensitive), `question_text`, `scale_min`, `scale_max` | |
| | `human_answers.csv` | one per question × country, **WVS-sourced questions only** | `question_id`, `country`, `language`, `human_mean`, `distribution` (JSON: answer option → share of respondents) | |
| | `tiers_wvs_only.csv`, `new_weights_transposed_wvs_only.xlsx` | | the original-format files, filtered to WVS questions, so the GitHub code runs on this subset unchanged | |
|
|
| ```python |
| from datasets import load_dataset |
| questions = load_dataset("llm-lab/alignment-veto", "questions", split="train") # all 864 |
| human = load_dataset("llm-lab/alignment-veto", "human_answers", split="train") # WVS subset |
| ``` |
|
|
| The 864 questions come from two surveys: **291 from World Values Survey (WVS) Wave 7**, and **573 from the Arab Opinion Index (AOI)**. All 864 question texts are published in `questions.csv` — the `source` column tells you which survey each one is from. |
|
|
| For the human *answer* data (country-level means and response distributions), WVS's terms clearly permit redistributing aggregates, so `human_answers.csv` here covers all 291 WVS questions across the 9 countries WVS Wave 7 covers for this set (Egypt, Iran, Iraq, Jordan, Lebanon, Libya, Morocco, Tunisia, Turkey). |
|
|
| We have not yet confirmed that the Arab Opinion Index's terms permit redistributing aggregates, so the AOI-sourced human answers (573 questions, all 16 countries) and their native-language (Arabic) translations are **not included here**. If you need them for research, **email the corresponding author** (zahraei2@illinois.edu). |
|
|
|
|
| ## Sources |
|
|
| Human answers are country-level aggregates from the **World Values Survey Wave 7** and the **Arab Opinion Index**. If you use them, please cite those surveys as well as our paper. |
|
|
| ## Citation |
|
|
| ```bibtex |
| @inproceedings{zahraei2026alignmentveto, |
| title = {The Alignment Veto: How Safety Training Suppresses Cultural Knowledge in LLMs}, |
| author = {Zahraei, Pardis Sadat and Tur, Gokhan and Hakkani-T{\"u}r, Dilek and Asgari, Ehsaneddin}, |
| booktitle = {Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing}, |
| year = {2026}, |
| publisher = {Association for Computational Linguistics}, |
| url = {https://arxiv.org/abs/2510.13154} |
| } |
| ``` |
|
|
