Skip to content

Commit

Permalink
update
Browse files Browse the repository at this point in the history
  • Loading branch information
@RZFan525
RZFan525 committed Feb 17, 2024
1 parent ce66583 commit 415a50a
Show file tree
Hide file tree
Showing 20 changed files with 3,434 additions and 1 deletion.
2 changes: 1 addition & 1 deletion README.md
View file
Original file line number Diff line number Diff line change
Expand Up @@ -34,7 +34,7 @@ Through this collaborative synergy, we expect the generated instruction data to


<div align=center><img src="./figs/Math_Results.jpg" style="zoom: 25%;" /></div>
<center>The accuracy of the GSM8K test set for LLaMA-2-13B and Mistral-7B models fine-tuned on the training set of GSM8K and MATH with and without \modelname. (a): Training on GSM8K. (b): Training on MATH and testing on GSM8K (Out-of-Distribution Setting).</center>
<center>The accuracy of the GSM8K test set for LLaMA-2-13B and Mistral-7B models fine-tuned on the training set of GSM8K and MATH with and without ReAlign. (a): Training on GSM8K. (b): Training on MATH and testing on GSM8K (Out-of-Distribution Setting).</center>


<div align=center><img src="./figs/overall_figs.jpg" style="zoom: 25%;" /></div>
Expand Down
Binary file added figs/alignment_results.jpg
View file
Loading
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
Binary file added figs/factuality_results.jpg
View file
Loading
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
Binary file modified figs/intro_graph.jpg
View file
Loading
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
Binary file added figs/math_reasoning.jpg
View file
Loading
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
Binary file added figs/readability_results.jpg
View file
Loading
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
Binary file added figs/scaling_law.jpg
View file
Loading
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
350 changes: 350 additions & 0 deletions index.html
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,350 @@
<!DOCTYPE html>
<html>
<head>
<meta charset="utf-8">
<meta name="description"
content="Reformatted Alignment: Distill LLM’s Generative Power instead of Generation">
<meta name="keywords" content="Alignment, Language Models, NLP">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>Reformatted Alignment: Distill LLM’s Generative Power instead of Generation</title>

<!-- Global site tag (gtag.js) - Google Analytics -->
<script async src="https://www.googletagmanager.com/gtag/js?id=G-PYVRSFMDRL"></script>
<script>
window.dataLayer = window.dataLayer || [];

function gtag() {
dataLayer.push(arguments);
}

gtag('js', new Date());

gtag('config', 'G-PYVRSFMDRL');
</script>

<link href="https://fonts.googleapis.com/css?family=Google+Sans|Noto+Sans|Castoro"
rel="stylesheet">

<link rel="stylesheet" href="./static/css/bulma.min.css">
<link rel="stylesheet" href="./static/css/bulma-carousel.min.css">
<link rel="stylesheet" href="./static/css/bulma-slider.min.css">
<link rel="stylesheet" href="./static/css/fontawesome.all.min.css">
<link rel="stylesheet"
href="https://cdn.jsdelivr.net/gh/jpswalsh/academicons@1/css/academicons.min.css">
<link rel="stylesheet" href="./static/css/index.css">
<link rel="icon" href="./static/images/mathpile-logo-1.png">

<script src="https://ajax.googleapis.com/ajax/libs/jquery/3.5.1/jquery.min.js"></script>
<script defer src="./static/js/fontawesome.all.min.js"></script>
<script src="./static/js/bulma-carousel.min.js"></script>
<script src="./static/js/bulma-slider.min.js"></script>
<script src="./static/js/index.js"></script>
</head>
<body>


<section class="hero">
<div class="hero-body">
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column has-text-centered">
<h1 class="title is-1 publication-title">Reformatted Alignment</h1>
<div class="is-size-5 publication-authors">
<span class="author-block">
<a href="https://scholar.google.com/citations?user=mhot7AUAAAAJ&hl=en" target="_blank">Run-Ze Fan</a><sup>1,6</sup>,</span>
<span class="author-block">
<a href="https://github.com/hongtangshui" target="_blank">Xuefeng Li</a><sup>1,6</sup>,</span>
<span class="author-block">
<a href="https://openreview.net/profile?id=~Haoyang_Zou1" target="_blank">Haoyang Zou</a><sup>3,6</sup>,</span>
<span class="author-block">
<a href="https://lockon-n.github.io/" target="_blank">Junlong Li</a><sup>1,6</sup>,</span>
<span class="author-block">
<a href="https://shwai-he.github.io/" target="_blank">Shwai He</a><sup>4</sup>,</span>
<span class="author-block">
<a href="https://ethanc111.github.io/" target="_blank">Ethan Chern</a><sup>1,6</sup>,</span>
<span class="author-block">
<a href="https://www.linkedin.com/in/jiewen-hu/" target="_blank">Jiewen Hu</a><sup>5,6</sup>,</span>
<span class="author-block">
<a href="http://pfliu.com/" target="_blank">Pengfei Liu</a><sup>1,2,6*</sup>
</span>
<!-- <span class="author-block">
<a href="http://sofienbouaziz.com">Sofien Bouaziz</a><sup>2</sup>,
</span>
<span class="author-block">
<a href="https://www.danbgoldman.com">Dan B Goldman</a><sup>2</sup>,
</span>
<span class="author-block">
<a href="https://homes.cs.washington.edu/~seitz/">Steven M. Seitz</a><sup>1,2</sup>,
</span>
<span class="author-block">
<a href="http://www.ricardomartinbrualla.com">Ricardo Martin-Brualla</a><sup>2</sup>
</span> -->
</div>

<div class="is-size-5 publication-authors">
<span class="author-block"><sup>1</sup>Shanghai Jiao Tong University,</span>
<span class="author-block"><sup>2</sup>Shanghai Artificial Intelligence Laboratory,</span>
<span class="author-block"><sup>3</sup>Fudan University,</span>
<span class="author-block"><sup>4</sup>University of Maryland, College Park,</span>
<span class="author-block"><sup>5</sup>CMU,</span>
<span class="author-block"><sup>6</sup>Generative AI Research Lab (GAIR)</span>
<span class="author-block"><sup>*</sup>Corresponding Author</span>
</div>

<div class="column has-text-centered">
<div class="publication-links">
<!-- PDF Link. -->
<!-- <span class="link-block">-->
<!-- <a href="https://huggingface.co/papers/2312.17120" target="_blank"-->
<!-- class="external-link button is-normal is-rounded is-dark">-->
<!-- <span class="icon">-->
<!-- <i class="fas fa-file-pdf"></i>-->
<!-- </span>-->
<!-- <span>Paper</span>-->
<!-- </a>-->
<!-- </span>-->
<span class="link-block">
<a href="https://arxiv.org/abs/xxx" target="_blank"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="ai ai-arxiv"></i>
</span>
<span>Paper</span>
</a>
</span>
<!-- Video Link. -->
<!-- <span class="link-block">
<a href="https://www.youtube.com/watch?v=MrKrnHhk8IA"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fab fa-youtube"></i>
</span>
<span>Video</span>
</a>
</span> -->
<!-- Code Link. -->
<span class="link-block">
<a href="https://github.com/GAIR-NLP/ReAlign" target="_blank"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fab fa-github"></i>
</span>
<span>Code</span>
</a>
</span>
<!-- Dataset Link. -->
<!-- <span class="link-block">-->
<!-- <a href="https://huggingface.co/datasets/GAIR/MathPile" target="_blank"-->
<!-- class="external-link button is-normal is-rounded is-dark">-->
<!-- <span class="icon">-->
<!-- <i class="far fa-images"></i>-->
<!-- </span>-->
<!-- <span>Datasets</span>-->
<!-- </a>-->
<!-- </span>-->
<!-- Twitter Link. -->
<!-- <span class="link-block">-->
<!-- <a href="https://twitter.com/_akhaliq/status/1740571256234057798" target="_blank"-->
<!-- class="external-link button is-normal is-rounded is-dark">-->
<!-- <span class="icon">-->
<!-- <i class="fab fa-twitter"></i>-->
<!-- </span>-->
<!-- <span>Twitter by AK</span>-->
<!-- </a>-->
<!-- </span>-->
</div>

</div>
</div>
</div>
</div>
</div>
</section>



<section class="section">
<div class="container is-max-desktop">
<!-- Abstract. -->
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<h2 class="title is-3">Introduction</h2>
<div class="content has-text-justified">
<p>
We explores elevating the quality of existing instruction data to better align with human values, introducing a simple and effective approach named <b>ReAlign</b> (<b>Re</b>formatted <b>Align</b>ment), which <i>reformats</i> the responses of instruction data into a format that better aligns with pre-established criteria and the collated evidence.
This approach minimizes human annotation, hallucination, and the difficulty in scaling, remaining orthogonal to existing alignment techniques.
Experimentally, ReAlign significantly boosts the general alignment ability, math reasoning, factuality, and readability of the LLMs.
</p>
<p>
Encouragingly, <i>without</i> introducing any additional data or advanced training techniques, and merely by reformatting the response, LLaMA-2-13B's mathematical reasoning ability on GSM8K can be improved <b>from 46.77% to 56.63%</b> in accuracy.
Additionally, a mere 5% of ReAlign data yields a 67% boost in general alignment ability measured by the Alpaca dataset.
This work highlights the need for further research into the <i>science</i> and <i>interpretability</i> of LLMs.
</p>
<p>
The underlying <i>philosophy</i> of ReAlign is to re-coordinate the roles of humans and LLMs in the alignment process, leveraging their complementary strengths -- humans articulate their preferences, and LLMs, in turn, reconstruct instructions based on their generative power (e.g., instruction-following ability), without directly using distilled LLM knowledge.
Through this collaborative synergy, we expect the generated instruction data to be not only more contextually precise but also more closely aligned with human preferences.
</p>
<figure>
<img src="figs/Math_Results.jpg" width=100% alt="Description of first image"/>
<figcaption>
The accuracy of the GSM8K test set for LLaMA-2-13B and Mistral-7B models fine-tuned on the training set of GSM8K and MATH with and without <b>ReAlign</b>. (a): Training on GSM8K. (b): Training on MATH and testing on GSM8K (Out-of-Distribution Setting).
</figcaption>
</figure>
</div>
</div>
</div>
</div>
</section>

<section class="section">
<div class="container is-max-desktop">
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<h2 class="title is-3">Methodology</h2>
<div class="content has-text-justified">


<figure>
<img src="figs/overall_figs.jpg" width=100% alt="Description of second image"/>
<figcaption>
An overview of our <strong>ReAlign</strong> including three steps. KILT denotes Knowledge Intensive Language Tasks.
</figcaption>
</figure>

<p>The ReAlign process unfolds in three main steps. </p>

<p>The first step involves <b>criteria definition</b>, where humans define their preferences (e.g., the preferred format of responses) in various scenarios in the form of natural language.
In this paper, we meticulously define criteria for 46 distinct scenarios. </p>

<p>The second step, <b>retrieval augmentation</b>, broadens the knowledge base for knowledge-intensive tasks like open-domain QA and fact verification. This is achieved by incorporating additional information, thereby improving the factuality and informativeness of responses. </p>

<p>The final step, <b>reformatting</b>, aims to re-align the responses with the pre-established criteria and the collated evidence, guaranteeing outputs that are both structured and substantiated.</p>
</div>
</div>
</div>
</div>
</section>

<section class="section">
<div class="container is-max-desktop">
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<h2 class="title is-3">Examples</h2>
<div class="content has-text-justified">

<figure>
<img src="figs/intro_graph.jpg" width=80% alt="Description of third image"/>
<figcaption>
<b>ReAlign</b> realigns the original response with the pre-defined criteria to be a better format.
</figcaption>
</figure>

<figure>
<img src="figs/model_example.jpg" width=100% alt="Description of fourth image"/>
<figcaption>
An example of the response from the original model and the response from the <b>ReAlign</b> Model.
</figcaption>
</figure>


</div>
</div>
</div>
</div>
</section>


<section class="section">
<div class="container is-max-desktop">
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<h2 class="title is-3">Results</h2>
<div class="content has-text-justified">

<figure>
<img src="figs/alignment_results.jpg" width=100% alt="Description of fifth image"/>
<figcaption>
The results of the general alignment ability on the original datasets and the ReAlign datasets.
</figcaption>
</figure>

<figure>
<img src="figs/math_reasoning.jpg" width=60% alt="Description of sixth image"/>
<figcaption>
The results of math reasoning on GSM8K, MATH and them + ReAlign based on LLaMA-2-13B and Mistral-7B. We test models on both GSM8K and MATH test sets. We report the accuracy by exact matching.
</figcaption>
</figure>

<figure>
<img src="figs/factuality_results.jpg" width=60% alt="Description of seventh image"/>
<figcaption>
The results of the factuality score.
</figcaption>
</figure>

<figure>
<img src="figs/readability_results.jpg" width=100% alt="Description of eighth image"/>
<figcaption>
The readability win-rate of the original dataset + ReAlign against the original dataset based on LLaMA-2-13B, judged by GPT-4 and human.
</figcaption>
</figure>

<figure>
<img src="figs/scaling_law.jpg" width=80% alt="Description of ninth image"/>
<figcaption>
The scaling trends in ReAlign data percentage, including general alignment ability and knowledge ability. We conduct the experiment in the Alpaca dataset based on LLaMA-2-13B.
</figcaption>
</figure>

</div>
</div>
</div>
</div>
</section>



<section class="section" id="BibTeX">
<div class="container is-max-desktop content">
<h2 class="title">BibTeX</h2>
<pre><code>

</code></pre>
</div>
</section>


<footer class="footer">
<div class="container">
<div class="content has-text-centered">
<a class="icon-link"
href="https://huggingface.co/papers/xxx">
<i class="fas fa-file-pdf"></i>
</a>
<a class="icon-link" href="https://github.com/GAIR-NLP/ReAlign" class="external-link" disabled>
<i class="fab fa-github"></i>
</a>
</div>
<div class="columns is-centered">
<div class="column is-8">
<div class="content">
<p>
This website is licensed under a <a rel="license"
href="http://creativecommons.org/licenses/by-sa/4.0/">Creative
Commons Attribution-ShareAlike 4.0 International License</a>.
</p>
<p>
This means you are free to borrow the <a
href="https://github.com/nerfies/nerfies.github.io">source code</a> of this website,
we just ask that you link back to this page in the footer.
Please remember to remove the analytics code included in the header of the website which
you do not want on your website.
</p>
</div>
</div>
</div>
</div>
</footer>

</body>
</html>
1 change: 1 addition & 0 deletions static/css/bulma-carousel.min.css
View file

Some generated files are not rendered by default. Learn more about how customized files appear on GitHub.

Loading

0 comments on commit 415a50a

Please sign in to comment.