furonghuang-lab/PHTest

---
license: mit
task_categories:
- text-generation
- question-answering
language:
- en
tags:
- llm
- alignment
- false refusal
- over-alignment
pretty_name: PHTest
size_categories:
- 100K<n<1M
viewer: true
---
<h3>🌟 <strong>PHTest: Evaluating False Refusals in LLMs</strong></h3>

<ol>
<li><strong>🤖 Auto Red-Teaming</strong>
<ul>
<li>All prompts are generated automatically using a controllable text-generation technique called <a href="https://arxiv.org/abs/2310.15140">AutoDAN</a>.</li>
</ul>
</li>

<li><strong>🌐 Diverse Prompts</strong>
<ul>
<li>PHTest introduces false refusal patterns that aren’t present in existing datasets, including prompts that avoid mentioning sensitive words.</li>
</ul>
</li>

<li><strong>⚖️ Harmlessness &amp; Controversial Labeling</strong>
<ul>
<li>Controversial prompts are separately labeled to address the inherent ambiguity in defining harmfulness, ensuring fair benchmarking and enabling tailored mitigation strategies.</li>
</ul>
</li>
</ol>

<h3>📚 <strong>Learn More</strong></h3>
<ul>
<li>For detailed information and evaluation results, refer to our COLM paper: <a href="https://arxiv.org/abs/2409.00598">Automatic Pseudo-Harmful Prompt Generation for Evaluating False Refusals in Large Language Models</a></li>
<li>Visit our project webpage: <a href="https://phtest-frf.github.io/">PHTest Project</a></li>
</ul>