Spaces:
Running
Running
File size: 8,150 Bytes
f98e8dd |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 |
<!DOCTYPE html>
<head>
<!-- Global site tag (gtag.js) - Google Analytics -->
<script async src="https://www.googletagmanager.com/gtag/js?id=UA-178132094-1"></script>
<script>
window.dataLayer = window.dataLayer || [];
function gtag() {
dataLayer.push(arguments);
}
gtag("js", new Date());
gtag("config", "UA-178132094-1");
</script>
<meta charset="UTF-8" />
<meta name="viewport" content="width=device-width, initial-scale=1" />
<!-- <meta name="viewport" content="width=1024" /> -->
<title>OR-Bench: Over Refusal Benchmark</title>
<script src="https://ajax.googleapis.com/ajax/libs/jquery/3.5.1/jquery.min.js"></script>
<link href="https://fonts.googleapis.com/css2?family=Montserrat:wght@400;700&display=swap" rel="stylesheet">
<script src="https://maxcdn.bootstrapcdn.com/bootstrap/4.3.1/js/bootstrap.min.js"></script>
<link rel="stylesheet" href="https://stackpath.bootstrapcdn.com/bootstrap/4.3.1/css/bootstrap.min.css" integrity="sha384-ggOyR0iXCbMQv3Xipma34MD+dH/1fQ784/j6cY/iJTQUOhcWr7x9JvoRxT2MZw1T" crossorigin="anonymous">
<script src="https://polyfill.io/v3/polyfill.min.js?features=es6"></script>
<script type="text/javascript" async
src="https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.7/MathJax.js?config=TeX-MML-AM_CHTML">
</script>
<link rel="stylesheet" href="https://cdnjs.cloudflare.com/ajax/libs/foundation/6.4.3/css/foundation.min.css" />
<link rel="stylesheet" href="https://cdn.rawgit.com/jpswalsh/academicons/master/css/academicons.min.css" />
<script src="https://kit.fontawesome.com/b939870cfb.js" crossorigin="anonymous"></script>
<link rel="stylesheet" href="https://cdn.datatables.net/1.10.24/css/dataTables.foundation.min.css">
<script type="text/javascript" src="https://cdn.datatables.net/1.10.24/js/jquery.dataTables.min.js"></script>
<link rel="stylesheet" href="./css/main.css" />
</head>
<body>
<nav class="navbar navbar-expand-md">
<div class="container">
<a class="navbar-brand" href="./index.html"
>OR-Bench</a>
<button
class="navbar-toggler navbar-light"
type="button"
data-toggle="collapse"
data-target="#main-navigation"
>
<span class="navbar-toggler-icon"></span>
</button>
<div class="collapse navbar-collapse" id="main-navigation">
<ul class="navbar-nav">
<li class="nav-item">
<a class="nav-link" href="#leaderboard">Leaderboards</a>
</li>
<li>
<a class="nav-link" href="https://arxiv.org/abs/2405.20947" target="_blank">Paper</a>
</li>
<li>
<a class="nav-link text-nowrap" href="https://github.com/justincui03/or-bench"
target="_blank">Github</a>
</li>
</ul>
</div>
</div>
</nav>
<!-- <hr class="toprule" /> -->
<header>
<div class="header-block container">
<div class="title-logo"><img src="./images/logo.png" alt="logo" /></div>
<div class="title">OR-BENCH</div>
<div class="description">
An over-refusal benchmark for large language models
</div>
</div>
</header>
<!-- <hr class="toprule" /> -->
<div class="container">
<section id="introduction">
<div class="overview">
<p class="doublealign">
<b>Large Language Models (LLMs) </b> require careful safety alignment to prevent malicious outputs. While significant research focuses on mitigating harmful content generation,
the enhanced safety often come with the side effect of over-refusal, where LLMs may reject innocuous prompts and become less helpful.
Although the issue of over-refusal has been empirically observed, a systematic measurement is challenging
due to the difficulty of crafting prompts that appear harmful but are benign.<br><br>
We introduce OR-Bench, the <b>first large-scale over-refusal benchmark</b>. OR-Bench comprises 80,000 seemingly toxic prompts across 10 common rejection categories, a subset of around 1,000 hard prompts that are challenging even for state-of-the-art LLMs, and an additional 600 toxic prompts to prevent indiscriminate responses.<br><br>
We plot the evaluation results in the following figure. The x-axis is the rejection rate on seemingly toxic prompts and the y-axis is the rejection rate on real toxic prompts. In the ideal case, the model should be on the top-left corner where the model rejects the most number of toixc prompts and the least number of seemingly toxic prompts.
</p>
<div style="margin-top:20px"><img src="./images/overall_x_y_plot.png" style="width: 100%;"/></div>
</div>
</section>
<div class="divider"><hr /></div>
<section class="container" id="div_cifar10_ipc1_heading">
<div id="div_or_bench" class="display responsive nowrap" style="width:100%"></div>
</section>
<div class="divider"><hr /></div>
<!-- <script
type="module"
src="https://gradio.s3-us-west-2.amazonaws.com/4.31.0/gradio.js"
></script> -->
<div><b>Please try out our demos below 🚀</b></div>
<div class="iframe-container">
<iframe
id="myIframe"
src="https://bench-llm-or-bench.hf.space"
frameborder="0"
width="2160"
height="450"
></iframe>
</div>
<div class="vspace50"></div>
<div id="new_result" style="width:100%">
<a class="btn btn-secondary" style="width:100%" href="https://forms.gle/WEeSVBENFWMXMuYK8" target="_blank">Submit New Results <img src="images/click.png" width="20"/></a>
</div>
<section id="citation" >
<div class="heading">
<p>Citation</p>
</div>
<!-- HTML generated using hilite.me --><div style="background: #ffffff; overflow:auto;width:auto;border:solid gray;border-width:.1em .1em .1em .8em;padding:.2em .6em;"><pre style="margin: 0; line-height: 125%"><span style="color: #555555; font-weight: bold">@article</span>{cui2024or,
title<span style="color: #333333">=</span>{OR<span style="color: #333333">-</span>Bench: An Over<span style="color: #333333">-</span>Refusal Benchmark <span style="color: #008800; font-weight: bold">for</span> Large Language Models},
author<span style="color: #333333">=</span>{Cui, Justin <span style="color: #000000; font-weight: bold">and</span> Chiang, Wei<span style="color: #333333">-</span>Lin <span style="color: #000000; font-weight: bold">and</span> Stoica, Ion <span style="color: #000000; font-weight: bold">and</span> Hsieh, Cho<span style="color: #333333">-</span>Jui},
journal<span style="color: #333333">=</span>{arXiv preprint arXiv:<span style="color: #6600EE; font-weight: bold">2405.20947</span>},
year<span style="color: #333333">=</span>{<span style="color: #0000DD; font-weight: bold">2024</span>}
}
</pre></div>
</section>
<div class="vspace50"></div>
</div>
<hr class="bottomrule" />
<footer>
<small>© 2024, OR-Bench
</footer>
<script>
// When the user scrolls the page, execute myFunction
window.onscroll = function () {
myFunction();
};
// Get the navbar
var navbar = document.getElementById("navbar");
// Get the offset position of the navbar
var sticky = navbar.offsetTop;
// Add the sticky class to the navbar when you reach its scroll position. Remove "sticky" when you leave the scroll position
function myFunction() {
if (window.pageYOffset >= sticky) {
navbar.classList.add("sticky");
} else {
navbar.classList.remove("sticky");
}
}
</script>
<script>
$("#div_or_bench").load("./data/or-bench.html", function() {
$('#or-bench-table').DataTable({
"pageLength": 25, // Set the initial number of entries
"lengthMenu": [[10, 25, 50, -1], [10, 25, 50, "All"]], // Set options for lengthMenu
"order": [[3, "asc"]], // Sort by the third column (index 2) in descending order
"paging": false, // Disables pagination
"responsive": true // Enable responsive feature
});
});
</script>
</body>
|