Buckets:

hf-doc-build
/

doc-dev

Files

xet

hf-doc-build/doc-dev / course /pr_1021 /zh-CN /chapter6 /5.html

rtrm

about 2 months ago

download

raw

87.8 kB

	<meta charset="utf-8" /><meta name="hf:doc:metadata" content="{"title":"BPE tokenization 算法","local":"BPE tokenization算法","sections":[{"title":"BPE 训练","local":"BPE 训练","sections":[],"depth":2},{"title":"tokenization","local":"tokenization","sections":[],"depth":2},{"title":"实现 BPE 算法","local":"实现 BPE 算法","sections":[],"depth":2}],"depth":1}">
	<link href="/docs/course/pr_1021/zh-CN/_app/immutable/assets/0.e3b0c442.css" rel="modulepreload">
	<link rel="modulepreload" href="/docs/course/pr_1021/zh-CN/_app/immutable/entry/start.f3a1a511.js">
	<link rel="modulepreload" href="/docs/course/pr_1021/zh-CN/_app/immutable/chunks/scheduler.37c15a92.js">
	<link rel="modulepreload" href="/docs/course/pr_1021/zh-CN/_app/immutable/chunks/singletons.9bf55235.js">
	<link rel="modulepreload" href="/docs/course/pr_1021/zh-CN/_app/immutable/chunks/index.18351ede.js">
	<link rel="modulepreload" href="/docs/course/pr_1021/zh-CN/_app/immutable/chunks/paths.0ba10750.js">
	<link rel="modulepreload" href="/docs/course/pr_1021/zh-CN/_app/immutable/entry/app.c39e37cf.js">
	<link rel="modulepreload" href="/docs/course/pr_1021/zh-CN/_app/immutable/chunks/index.2bf4358c.js">
	<link rel="modulepreload" href="/docs/course/pr_1021/zh-CN/_app/immutable/nodes/0.dad18ce3.js">
	<link rel="modulepreload" href="/docs/course/pr_1021/zh-CN/_app/immutable/chunks/each.e59479a4.js">
	<link rel="modulepreload" href="/docs/course/pr_1021/zh-CN/_app/immutable/nodes/48.a4455964.js">
	<link rel="modulepreload" href="/docs/course/pr_1021/zh-CN/_app/immutable/chunks/Tip.363c041f.js">
	<link rel="modulepreload" href="/docs/course/pr_1021/zh-CN/_app/immutable/chunks/Youtube.1e50a667.js">
	<link rel="modulepreload" href="/docs/course/pr_1021/zh-CN/_app/immutable/chunks/CodeBlock.4e987730.js">
	<link rel="modulepreload" href="/docs/course/pr_1021/zh-CN/_app/immutable/chunks/CourseFloatingBanner.9ff4c771.js">
	<link rel="modulepreload" href="/docs/course/pr_1021/zh-CN/_app/immutable/chunks/DocNotebookDropdown.efc1fb7c.js">
	<link rel="modulepreload" href="/docs/course/pr_1021/zh-CN/_app/immutable/chunks/getInferenceSnippets.ebf8be91.js"><!-- HEAD_svelte-u9bgzb_START --><meta name="hf:doc:metadata" content="{"title":"BPE tokenization 算法","local":"BPE tokenization算法","sections":[{"title":"BPE 训练","local":"BPE 训练","sections":[],"depth":2},{"title":"tokenization","local":"tokenization","sections":[],"depth":2},{"title":"实现 BPE 算法","local":"实现 BPE 算法","sections":[],"depth":2}],"depth":1}"><!-- HEAD_svelte-u9bgzb_END --> <p></p> <h1 class="relative group"><a id="BPE tokenization算法" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#BPE tokenization算法"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>BPE tokenization 算法</span></h1> <div class="flex space-x-1 absolute z-10 right-0 top-0"><a href="https://discuss.huggingface.co/t/chapter-6-questions" target="_blank"><img alt="Ask a Question" class="!m-0" src="https://img.shields.io/badge/Ask%20a%20question-ffcb4c.svg?logo=data:image/svg+xml;base64,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"></a> <a href="https://colab.research.google.com/github/huggingface/notebooks/blob/master/course/zh-CN/chapter6/section5.ipynb" target="_blank"><img alt="Open In Colab" class="!m-0" src="https://colab.research.google.com/assets/colab-badge.svg"></a> <a href="https://studiolab.sagemaker.aws/import/github/huggingface/notebooks/blob/master/course/zh-CN/chapter6/section5.ipynb" target="_blank"><img alt="Open In Studio Lab" class="!m-0" src="https://studiolab.sagemaker.aws/studiolab.svg"></a></div> <p data-svelte-h="svelte-cafj1v">字节对编码（BPE）最初被开发为一种压缩文本的算法，然后在预训练 GPT 模型时被 OpenAI 用于 tokenization。许多 Transformer 模型都使用它，包括 GPT、GPT-2、RoBERTa、BART 和 DeBERTa。</p> <iframe class="w-full xl:w-4/6 h-80" src="https://www.youtube-nocookie.com/embed/HEikzVL-lZU" title="YouTube video player" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture" allowfullscreen></iframe> <div class="course-tip bg-gradient-to-br dark:bg-gradient-to-r before:border-green-500 dark:before:border-green-800 from-green-50 dark:from-gray-900 to-white dark:to-gray-950 border border-green-50 text-green-700 dark:text-gray-400"><p data-svelte-h="svelte-1waktky">💡 本节深入介绍了 BPE，甚至展示了一个完整的实现。如果你只想大致了解 tokenization 算法，可以跳到最后。</p></div> <h2 class="relative group"><a id="BPE 训练" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#BPE 训练"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>BPE 训练</span></h2> <p data-svelte-h="svelte-1xtlx6q">BPE 训练首先计算语料库中使用的唯一单词集合（在完成标准化和预分词步骤之后），然后取出用来编写这些词的所有符号来构建词汇表。举一个非常简单的例子，假设我们的语料库使用了这五个词：</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-string">"hug"</span>, <span class="hljs-string">"pug"</span>, <span class="hljs-string">"pun"</span>, <span class="hljs-string">"bun"</span>, <span class="hljs-string">"hugs"</span><!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-qrzm6i">基础单词集合将是 <code>["b", "g", "h", "n", "p", "s", "u"]</code> 。在实际应用中，基本词汇表将至少包含所有 ASCII 字符，可能还包含一些 Unicode 字符。如果你正在 tokenization 不在训练语料库中的字符，则该字符将转换为未知 tokens，这就是为什么许多 NLP 模型在分析带有表情符号的内容的结果非常糟糕的原因之一。</p> <div class="course-tip bg-gradient-to-br dark:bg-gradient-to-r before:border-green-500 dark:before:border-green-800 from-green-50 dark:from-gray-900 to-white dark:to-gray-950 border border-green-50 text-green-700 dark:text-gray-400"><p data-svelte-h="svelte-ec0vh8">GPT-2 和 RoBERTa （这两者非常相似）的 tokenizer 有一个巧妙的方法来处理这个问题：他们不把单词看成是用 Unicode 字符编写的，而是用字节编写的。这样，基本词汇表的大小很小（256），但是能包含几乎所有你能想象的字符，而不会最终转换为未知 tokens 这个技巧被称为 <code>字节级（byte-level） BPE</code> 。</p></div> <p data-svelte-h="svelte-1pf74eh">获得这个基础单词集合后，我们通过学习 <code>合并（merges）</code> 来添加新的 tokens 直到达到期望的词汇表大小。合并是将现有词汇表中的两个元素合并为一个新元素的规则。所以，一开始会创建出含有两个字符的 tokens 然后，随着训练的进展，会产生更长的子词。</p> <p data-svelte-h="svelte-fgjr6p">在分词器训练期间的任何一步，BPE 算法都会搜索最常见的现有 tokens 对（在这里，“对”是指一个词中的两个连续 tokens ）。最常见的这一对会被合并，然后我们重复这个过程。</p> <p data-svelte-h="svelte-1kuwrk5">回到我们之前的例子，让我们假设单词具有以下频率：</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->(<span class="hljs-string">"hug"</span><span class="hljs-punctuation">,</span> <span class="hljs-number">10</span>)<span class="hljs-punctuation">,</span> (<span class="hljs-string">"pug"</span><span class="hljs-punctuation">,</span> <span class="hljs-number">5</span>)<span class="hljs-punctuation">,</span> (<span class="hljs-string">"pun"</span><span class="hljs-punctuation">,</span> <span class="hljs-number">12</span>)<span class="hljs-punctuation">,</span> (<span class="hljs-string">"bun"</span><span class="hljs-punctuation">,</span> <span class="hljs-number">4</span>)<span class="hljs-punctuation">,</span> (<span class="hljs-string">"hugs"</span><span class="hljs-punctuation">,</span> <span class="hljs-number">5</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1jexcr3">意思是 <code>"hug"</code> 在语料库中出现了 10 次， <code>"pug"</code> 出现了 5 次， <code>"pun"</code> 出现了 12 次， <code>"bun"</code> 出现了 4 次， <code>"hugs"</code> 出现了 5 次。我们通过将每个单词拆分为字符（形成我们初始词汇表的字符）来开始训练，这样我们就可以将每个单词视为一个 tokens 列表：</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->(<span class="hljs-string">"h"</span> <span class="hljs-string">"u"</span> <span class="hljs-string">"g"</span><span class="hljs-punctuation">,</span> <span class="hljs-number">10</span>)<span class="hljs-punctuation">,</span> (<span class="hljs-string">"p"</span> <span class="hljs-string">"u"</span> <span class="hljs-string">"g"</span><span class="hljs-punctuation">,</span> <span class="hljs-number">5</span>)<span class="hljs-punctuation">,</span> (<span class="hljs-string">"p"</span> <span class="hljs-string">"u"</span> <span class="hljs-string">"n"</span><span class="hljs-punctuation">,</span> <span class="hljs-number">12</span>)<span class="hljs-punctuation">,</span> (<span class="hljs-string">"b"</span> <span class="hljs-string">"u"</span> <span class="hljs-string">"n"</span><span class="hljs-punctuation">,</span> <span class="hljs-number">4</span>)<span class="hljs-punctuation">,</span> (<span class="hljs-string">"h"</span> <span class="hljs-string">"u"</span> <span class="hljs-string">"g"</span> <span class="hljs-string">"s"</span><span class="hljs-punctuation">,</span> <span class="hljs-number">5</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-cu00o7">然后我们看看相邻的字符对。 <code>("h", "u")</code> 在词 <code>"hug"</code> 和 <code>"hugs"</code> 中出现，所以在语料库中总共出现了 15 次。然而，最常见的对属于 <code>("u", "g")</code> ，它在 <code>"hug"</code> 、 <code>"pug"</code> 和 <code>"hugs"</code> 中出现，总共在词汇表中出现了 20 次。</p> <p data-svelte-h="svelte-1pkk3x3">因此，tokenizer 学习的第一个合并规则是 <code>("u", "g") -> "ug"</code> ，意思就是 <code>"ug"</code> 将被添加到词汇表中，且应在语料库的所有词中合并这一对。在这个阶段结束时，词汇表和语料库看起来像这样：</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->词汇表: <span class="hljs-selector-attr">[<span class="hljs-string">"b"</span>, <span class="hljs-string">"g"</span>, <span class="hljs-string">"h"</span>, <span class="hljs-string">"n"</span>, <span class="hljs-string">"p"</span>, <span class="hljs-string">"s"</span>, <span class="hljs-string">"u"</span>, <span class="hljs-string">"ug"</span>]</span>
	语料库: (<span class="hljs-string">"h"</span> <span class="hljs-string">"ug"</span>, <span class="hljs-number">10</span>), (<span class="hljs-string">"p"</span> <span class="hljs-string">"ug"</span>, <span class="hljs-number">5</span>), (<span class="hljs-string">"p"</span> <span class="hljs-string">"u"</span> <span class="hljs-string">"n"</span>, <span class="hljs-number">12</span>), (<span class="hljs-string">"b"</span> <span class="hljs-string">"u"</span> <span class="hljs-string">"n"</span>, <span class="hljs-number">4</span>), (<span class="hljs-string">"h"</span> <span class="hljs-string">"ug"</span> <span class="hljs-string">"s"</span>, <span class="hljs-number">5</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-19iv4fv">现在我们有一些对，继续合并的话会产生一个比两个字符长的 tokens 例如 <code>("h", "ug")</code> ，在语料库中出现 15 次。然而，这个阶段出现频率最高的对是 <code>("u", "n")</code> ，在语料库中出现 16 次，所以学到的第二个合并规则是 <code>("u", "n") -> "un"</code> 。将其添加到词汇表并合并所有现有的这个对，将出现：</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->词汇表: <span class="hljs-selector-attr">[<span class="hljs-string">"b"</span>, <span class="hljs-string">"g"</span>, <span class="hljs-string">"h"</span>, <span class="hljs-string">"n"</span>, <span class="hljs-string">"p"</span>, <span class="hljs-string">"s"</span>, <span class="hljs-string">"u"</span>, <span class="hljs-string">"ug"</span>, <span class="hljs-string">"un"</span>]</span>
	语料库: (<span class="hljs-string">"h"</span> <span class="hljs-string">"ug"</span>, <span class="hljs-number">10</span>), (<span class="hljs-string">"p"</span> <span class="hljs-string">"ug"</span>, <span class="hljs-number">5</span>), (<span class="hljs-string">"p"</span> <span class="hljs-string">"un"</span>, <span class="hljs-number">12</span>), (<span class="hljs-string">"b"</span> <span class="hljs-string">"un"</span>, <span class="hljs-number">4</span>), (<span class="hljs-string">"h"</span> <span class="hljs-string">"ug"</span> <span class="hljs-string">"s"</span>, <span class="hljs-number">5</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-scrtmz">现在最频繁的一对是 <code>("h", "ug")</code> ，所以我们学习了合并规则 <code>("h", "ug") -> "hug"</code> ，这形成了我们第一个三个字母的 tokens 合并后，语料库如下所示：</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->词汇表: <span class="hljs-selector-attr">[<span class="hljs-string">"b"</span>, <span class="hljs-string">"g"</span>, <span class="hljs-string">"h"</span>, <span class="hljs-string">"n"</span>, <span class="hljs-string">"p"</span>, <span class="hljs-string">"s"</span>, <span class="hljs-string">"u"</span>, <span class="hljs-string">"ug"</span>, <span class="hljs-string">"un"</span>, <span class="hljs-string">"hug"</span>]</span>
	语料库: (<span class="hljs-string">"hug"</span>, <span class="hljs-number">10</span>), (<span class="hljs-string">"p"</span> <span class="hljs-string">"ug"</span>, <span class="hljs-number">5</span>), (<span class="hljs-string">"p"</span> <span class="hljs-string">"un"</span>, <span class="hljs-number">12</span>), (<span class="hljs-string">"b"</span> <span class="hljs-string">"un"</span>, <span class="hljs-number">4</span>), (<span class="hljs-string">"hug"</span> <span class="hljs-string">"s"</span>, <span class="hljs-number">5</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-4ek1mp">我们继续这样合并，直到达到我们所需的词汇量。</p> <div class="course-tip bg-gradient-to-br dark:bg-gradient-to-r before:border-green-500 dark:before:border-green-800 from-green-50 dark:from-gray-900 to-white dark:to-gray-950 border border-green-50 text-green-700 dark:text-gray-400"><p data-svelte-h="svelte-1abbvo6">✏️ <strong>现在轮到你了！</strong> 你认为下一个合并规则是什么？</p></div> <h2 class="relative group"><a id="tokenization" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#tokenization"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>tokenization</span></h2> <p data-svelte-h="svelte-vwd30d">完成训练之后就可以对新的输入 tokenization 了，从某种意义上说，新的输入会依照以下步骤对新输入进行 tokenization：</p> <ol data-svelte-h="svelte-m3fmz7"><li>标准化</li> <li>预分词</li> <li>将单词拆分为单个字符</li> <li>根据学习的合并规则，按顺序合并拆分的字符</li></ol> <p data-svelte-h="svelte-1gfifff">让我们以我们在训练期间使用的示例为例，Tokenizer 学习到了三个合并规则：</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-function"><span class="hljs-params">(<span class="hljs-string">"u"</span>, <span class="hljs-string">"g"</span>)</span> -></span> <span class="hljs-string">"ug"</span>
	<span class="hljs-function"><span class="hljs-params">(<span class="hljs-string">"u"</span>, <span class="hljs-string">"n"</span>)</span> -></span> <span class="hljs-string">"un"</span>
	<span class="hljs-function"><span class="hljs-params">(<span class="hljs-string">"h"</span>, <span class="hljs-string">"ug"</span>)</span> -></span> <span class="hljs-string">"hug"</span><!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-19k5wls">在这种情况下，单词 <code>"bug"</code> 将被转化为 <code>["b", "ug"]</code> 。然而 <code>"mug"</code> ，将被转换为 <code>["[UNK]", "ug"]</code> ，因为字母 <code>"m"</code> 不再基本词汇表中。同样，单词 <code>"thug"</code> 会被转换为 <code>["[UNK]", "hug"]</code> ：字母 <code>"t"</code> 不在基本词汇表中，使用合并规则首先会将 <code>"u"</code> 和 <code>"g"</code> 合并，然后将 <code>"h"</code> 和 <code>"ug"</code> 合并。</p> <div class="course-tip bg-gradient-to-br dark:bg-gradient-to-r before:border-green-500 dark:before:border-green-800 from-green-50 dark:from-gray-900 to-white dark:to-gray-950 border border-green-50 text-green-700 dark:text-gray-400"><p data-svelte-h="svelte-18fkeee">✏️ <strong>现在轮到你了！</strong> 你认为这个词 <code>"unhug"</code> 将如何被 tokenization？</p></div> <h2 class="relative group"><a id="实现 BPE 算法" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#实现 BPE 算法"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>实现 BPE 算法</span></h2> <p data-svelte-h="svelte-92kbvp">现在，让我们看一下 BPE 算法的实现。这并不是在大型语料库上实际使用的经过优化的版本；我们只是想向你展示代码，以便你可以更好地理解算法</p> <p data-svelte-h="svelte-19oi16y">首先，我们需要一个语料库，让我们创建一个含有几句话的简单语料库：</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->corpus = [
	<span class="hljs-string">"This is the Hugging Face Course."</span>,
	<span class="hljs-string">"This chapter is about tokenization."</span>,
	<span class="hljs-string">"This section shows several tokenizer algorithms."</span>,
	<span class="hljs-string">"Hopefully, you will be able to understand how they are trained and generate tokens."</span>,
	]<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-193w4gh">接下来，我们需要将该语料库预分词为单词。由于我们正在复现一个 BPE tokenizer （例如 GPT-2），我们将使用 <code>gpt2</code> 分词器进行预分词：</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> AutoTokenizer

	tokenizer = AutoTokenizer.from_pretrained(<span class="hljs-string">"gpt2"</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-vwpm4m">然后，我们在进行预分词的同时计算语料库中每个单词的频率：</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-keyword">from</span> collections <span class="hljs-keyword">import</span> defaultdict

	word_freqs = defaultdict(<span class="hljs-built_in">int</span>)

	<span class="hljs-keyword">for</span> text <span class="hljs-keyword">in</span> corpus:
	words_with_offsets = tokenizer.backend_tokenizer.pre_tokenizer.pre_tokenize_str(text)
	new_words = [word <span class="hljs-keyword">for</span> word, offset <span class="hljs-keyword">in</span> words_with_offsets]
	<span class="hljs-keyword">for</span> word <span class="hljs-keyword">in</span> new_words:
	word_freqs[word] += <span class="hljs-number">1</span>

	<span class="hljs-built_in">print</span>(word_freqs)<!-- HTML_TAG_END --></pre></div> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->defaultdict(<span class="hljs-built_in">int</span>, {<span class="hljs-string">'This'</span>: <span class="hljs-number">3</span>, <span class="hljs-string">'Ġis'</span>: <span class="hljs-number">2</span>, <span class="hljs-string">'Ġthe'</span>: <span class="hljs-number">1</span>, <span class="hljs-string">'ĠHugging'</span>: <span class="hljs-number">1</span>, <span class="hljs-string">'ĠFace'</span>: <span class="hljs-number">1</span>, <span class="hljs-string">'ĠCourse'</span>: <span class="hljs-number">1</span>, <span class="hljs-string">'.'</span>: <span class="hljs-number">4</span>, <span class="hljs-string">'Ġchapter'</span>: <span class="hljs-number">1</span>,
	<span class="hljs-string">'Ġabout'</span>: <span class="hljs-number">1</span>, <span class="hljs-string">'Ġtokenization'</span>: <span class="hljs-number">1</span>, <span class="hljs-string">'Ġsection'</span>: <span class="hljs-number">1</span>, <span class="hljs-string">'Ġshows'</span>: <span class="hljs-number">1</span>, <span class="hljs-string">'Ġseveral'</span>: <span class="hljs-number">1</span>, <span class="hljs-string">'Ġtokenizer'</span>: <span class="hljs-number">1</span>, <span class="hljs-string">'Ġalgorithms'</span>: <span class="hljs-number">1</span>,
	<span class="hljs-string">'Hopefully'</span>: <span class="hljs-number">1</span>, <span class="hljs-string">','</span>: <span class="hljs-number">1</span>, <span class="hljs-string">'Ġyou'</span>: <span class="hljs-number">1</span>, <span class="hljs-string">'Ġwill'</span>: <span class="hljs-number">1</span>, <span class="hljs-string">'Ġbe'</span>: <span class="hljs-number">1</span>, <span class="hljs-string">'Ġable'</span>: <span class="hljs-number">1</span>, <span class="hljs-string">'Ġto'</span>: <span class="hljs-number">1</span>, <span class="hljs-string">'Ġunderstand'</span>: <span class="hljs-number">1</span>, <span class="hljs-string">'Ġhow'</span>: <span class="hljs-number">1</span>,
	<span class="hljs-string">'Ġthey'</span>: <span class="hljs-number">1</span>, <span class="hljs-string">'Ġare'</span>: <span class="hljs-number">1</span>, <span class="hljs-string">'Ġtrained'</span>: <span class="hljs-number">1</span>, <span class="hljs-string">'Ġand'</span>: <span class="hljs-number">1</span>, <span class="hljs-string">'Ġgenerate'</span>: <span class="hljs-number">1</span>, <span class="hljs-string">'Ġtokens'</span>: <span class="hljs-number">1</span>})<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1fa8ke2">下一步是计算基础词汇表，这由语料库中使用的所有字符组成：</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->alphabet = []

	<span class="hljs-keyword">for</span> word <span class="hljs-keyword">in</span> word_freqs.keys():
	<span class="hljs-keyword">for</span> letter <span class="hljs-keyword">in</span> word:
	<span class="hljs-keyword">if</span> letter <span class="hljs-keyword">not</span> <span class="hljs-keyword">in</span> alphabet:
	alphabet.append(letter)
	alphabet.sort()

	<span class="hljs-built_in">print</span>(alphabet)<!-- HTML_TAG_END --></pre></div> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->[ <span class="hljs-string">','</span>, <span class="hljs-string">'.'</span>, <span class="hljs-string">'C'</span>, <span class="hljs-string">'F'</span>, <span class="hljs-string">'H'</span>, <span class="hljs-string">'T'</span>, <span class="hljs-string">'a'</span>, <span class="hljs-string">'b'</span>, <span class="hljs-string">'c'</span>, <span class="hljs-string">'d'</span>, <span class="hljs-string">'e'</span>, <span class="hljs-string">'f'</span>, <span class="hljs-string">'g'</span>, <span class="hljs-string">'h'</span>, <span class="hljs-string">'i'</span>, <span class="hljs-string">'k'</span>, <span class="hljs-string">'l'</span>, <span class="hljs-string">'m'</span>, <span class="hljs-string">'n'</span>, <span class="hljs-string">'o'</span>, <span class="hljs-string">'p'</span>, <span class="hljs-string">'r'</span>, <span class="hljs-string">'s'</span>,
	<span class="hljs-string">'t'</span>, <span class="hljs-string">'u'</span>, <span class="hljs-string">'v'</span>, <span class="hljs-string">'w'</span>, <span class="hljs-string">'y'</span>, <span class="hljs-string">'z'</span>, <span class="hljs-string">'Ġ'</span>]<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1tg2x2o">我们还在该词汇表的开头添加了模型使用的特殊 tokens 对于 GPT-2，唯一的特殊 tokens 是 <code>"<\|endoftext\|>"</code> ：</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->vocab = [<span class="hljs-string">"<\|endoftext\|>"</span>] + alphabet.copy()<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1m50py">我们现在需要将每个单词拆分为单独的字符，以便能够开始训练：</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->splits = {word: [c <span class="hljs-keyword">for</span> c <span class="hljs-keyword">in</span> word] <span class="hljs-keyword">for</span> word <span class="hljs-keyword">in</span> word_freqs.keys()}<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1kbibkn">现在我们已准备好进行训练，让我们编写一个函数来计算每对字符的频率。我们需要在训练的每个步骤中使用它：</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-keyword">def</span> <span class="hljs-title function_">compute_pair_freqs</span>(<span class="hljs-params">splits</span>):
	pair_freqs = defaultdict(<span class="hljs-built_in">int</span>)
	<span class="hljs-keyword">for</span> word, freq <span class="hljs-keyword">in</span> word_freqs.items():
	split = splits[word]
	<span class="hljs-keyword">if</span> <span class="hljs-built_in">len</span>(split) == <span class="hljs-number">1</span>:
	<span class="hljs-keyword">continue</span>
	<span class="hljs-keyword">for</span> i <span class="hljs-keyword">in</span> <span class="hljs-built_in">range</span>(<span class="hljs-built_in">len</span>(split) - <span class="hljs-number">1</span>):
	pair = (split[i], split[i + <span class="hljs-number">1</span>])
	pair_freqs[pair] += freq
	<span class="hljs-keyword">return</span> pair_freqs<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1ec489l">让我们来看看这个字典在初始合并后的一些结果：</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->pair_freqs = compute_pair_freqs(splits)

	<span class="hljs-keyword">for</span> i, key <span class="hljs-keyword">in</span> <span class="hljs-built_in">enumerate</span>(pair_freqs.keys()):
	<span class="hljs-built_in">print</span>(<span class="hljs-string">f"<span class="hljs-subst">{key}</span>: <span class="hljs-subst">{pair_freqs[key]}</span>"</span>)
	<span class="hljs-keyword">if</span> i >= <span class="hljs-number">5</span>:
	<span class="hljs-keyword">break</span><!-- HTML_TAG_END --></pre></div> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->(<span class="hljs-string">'T'</span>, <span class="hljs-string">'h'</span>): <span class="hljs-number">3</span>
	(<span class="hljs-string">'h'</span>, <span class="hljs-string">'i'</span>): <span class="hljs-number">3</span>
	(<span class="hljs-string">'i'</span>, <span class="hljs-string">'s'</span>): <span class="hljs-number">5</span>
	(<span class="hljs-string">'Ġ'</span>, <span class="hljs-string">'i'</span>): <span class="hljs-number">2</span>
	(<span class="hljs-string">'Ġ'</span>, <span class="hljs-string">'t'</span>): <span class="hljs-number">7</span>
	(<span class="hljs-string">'t'</span>, <span class="hljs-string">'h'</span>): <span class="hljs-number">3</span><!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-11ofgxl">现在，只需要一个简单的循环就可以找到出现频率最高的对：</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->best_pair = <span class="hljs-string">""</span>
	max_freq = <span class="hljs-literal">None</span>

	<span class="hljs-keyword">for</span> pair, freq <span class="hljs-keyword">in</span> pair_freqs.items():
	<span class="hljs-keyword">if</span> max_freq <span class="hljs-keyword">is</span> <span class="hljs-literal">None</span> <span class="hljs-keyword">or</span> max_freq < freq:
	best_pair = pair
	max_freq = freq

	<span class="hljs-built_in">print</span>(best_pair, max_freq)<!-- HTML_TAG_END --></pre></div> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->(<span class="hljs-string">'Ġ'</span>, <span class="hljs-string">'t'</span>) <span class="hljs-number">7</span><!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-2ohqia">所以第一个要学习的合并规则是 <code>('Ġ', 't') -> 'Ġt'</code> ，我们将 <code>'Ġt'</code> 添加到词汇表：</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->merges = {(<span class="hljs-string">"Ġ"</span>, <span class="hljs-string">"t"</span>): <span class="hljs-string">"Ġt"</span>}
	vocab.append(<span class="hljs-string">"Ġt"</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-8v6wgo">接下来，我们需要在我们的 <code>splits</code> 字典中进行这个合并。让我们为此编写另一个函数：</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-keyword">def</span> <span class="hljs-title function_">merge_pair</span>(<span class="hljs-params">a, b, splits</span>):
	<span class="hljs-keyword">for</span> word <span class="hljs-keyword">in</span> word_freqs:
	split = splits[word]
	<span class="hljs-keyword">if</span> <span class="hljs-built_in">len</span>(split) == <span class="hljs-number">1</span>:
	<span class="hljs-keyword">continue</span>

	i = <span class="hljs-number">0</span>
	<span class="hljs-keyword">while</span> i < <span class="hljs-built_in">len</span>(split) - <span class="hljs-number">1</span>:
	<span class="hljs-keyword">if</span> split[i] == a <span class="hljs-keyword">and</span> split[i + <span class="hljs-number">1</span>] == b:
	split = split[:i] + [a + b] + split[i + <span class="hljs-number">2</span> :]
	<span class="hljs-keyword">else</span>:
	i += <span class="hljs-number">1</span>
	splits[word] = split
	<span class="hljs-keyword">return</span> splits<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1atajuj">我们可以观察一下第一次合并的结果：</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->splits = merge_pair(<span class="hljs-string">"Ġ"</span>, <span class="hljs-string">"t"</span>, splits)
	<span class="hljs-built_in">print</span>(splits[<span class="hljs-string">"Ġtrained"</span>])<!-- HTML_TAG_END --></pre></div> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->[<span class="hljs-string">'Ġt'</span>, <span class="hljs-string">'r'</span>, <span class="hljs-string">'a'</span>, <span class="hljs-string">'i'</span>, <span class="hljs-string">'n'</span>, <span class="hljs-string">'e'</span>, <span class="hljs-string">'d'</span>]<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-wrfq07">现在我们有了我们需要的所有代码，可以循环直到我们学习到我们想要的所有合并。让我们把目标词汇表的大小设定为 50：</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->vocab_size = <span class="hljs-number">50</span>

	<span class="hljs-keyword">while</span> <span class="hljs-built_in">len</span>(vocab) < vocab_size:
	pair_freqs = compute_pair_freqs(splits)
	best_pair = <span class="hljs-string">""</span>
	max_freq = <span class="hljs-literal">None</span>
	<span class="hljs-keyword">for</span> pair, freq <span class="hljs-keyword">in</span> pair_freqs.items():
	<span class="hljs-keyword">if</span> max_freq <span class="hljs-keyword">is</span> <span class="hljs-literal">None</span> <span class="hljs-keyword">or</span> max_freq < freq:
	best_pair = pair
	max_freq = freq
	splits = merge_pair(*best_pair, splits)
	merges[best_pair] = best_pair[<span class="hljs-number">0</span>] + best_pair[<span class="hljs-number">1</span>]
	vocab.append(best_pair[<span class="hljs-number">0</span>] + best_pair[<span class="hljs-number">1</span>])<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-12dmu1y">最终，我们学习了 19 条合并规则（初始词汇量为 31 —— 字母表中的 30 个字符，加上特殊 token ）：</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-built_in">print</span>(merges)<!-- HTML_TAG_END --></pre></div> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->{(<span class="hljs-string">'Ġ'</span>, <span class="hljs-string">'t'</span>): <span class="hljs-string">'Ġt'</span>, (<span class="hljs-string">'i'</span>, <span class="hljs-string">'s'</span>): <span class="hljs-string">'is'</span>, (<span class="hljs-string">'e'</span>, <span class="hljs-string">'r'</span>): <span class="hljs-string">'er'</span>, (<span class="hljs-string">'Ġ'</span>, <span class="hljs-string">'a'</span>): <span class="hljs-string">'Ġa'</span>, (<span class="hljs-string">'Ġt'</span>, <span class="hljs-string">'o'</span>): <span class="hljs-string">'Ġto'</span>, (<span class="hljs-string">'e'</span>, <span class="hljs-string">'n'</span>): <span class="hljs-string">'en'</span>,
	(<span class="hljs-string">'T'</span>, <span class="hljs-string">'h'</span>): <span class="hljs-string">'Th'</span>, (<span class="hljs-string">'Th'</span>, <span class="hljs-string">'is'</span>): <span class="hljs-string">'This'</span>, (<span class="hljs-string">'o'</span>, <span class="hljs-string">'u'</span>): <span class="hljs-string">'ou'</span>, (<span class="hljs-string">'s'</span>, <span class="hljs-string">'e'</span>): <span class="hljs-string">'se'</span>, (<span class="hljs-string">'Ġto'</span>, <span class="hljs-string">'k'</span>): <span class="hljs-string">'Ġtok'</span>,
	(<span class="hljs-string">'Ġtok'</span>, <span class="hljs-string">'en'</span>): <span class="hljs-string">'Ġtoken'</span>, (<span class="hljs-string">'n'</span>, <span class="hljs-string">'d'</span>): <span class="hljs-string">'nd'</span>, (<span class="hljs-string">'Ġ'</span>, <span class="hljs-string">'is'</span>): <span class="hljs-string">'Ġis'</span>, (<span class="hljs-string">'Ġt'</span>, <span class="hljs-string">'h'</span>): <span class="hljs-string">'Ġth'</span>, (<span class="hljs-string">'Ġth'</span>, <span class="hljs-string">'e'</span>): <span class="hljs-string">'Ġthe'</span>,
	(<span class="hljs-string">'i'</span>, <span class="hljs-string">'n'</span>): <span class="hljs-string">'in'</span>, (<span class="hljs-string">'Ġa'</span>, <span class="hljs-string">'b'</span>): <span class="hljs-string">'Ġab'</span>, (<span class="hljs-string">'Ġtoken'</span>, <span class="hljs-string">'i'</span>): <span class="hljs-string">'Ġtokeni'</span>}<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1spn0nn">词汇表由特殊 token 初始字母和所有合并结果组成：</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-built_in">print</span>(vocab)<!-- HTML_TAG_END --></pre></div> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->[<span class="hljs-string">'<\|endoftext\|>'</span>, <span class="hljs-string">','</span>, <span class="hljs-string">'.'</span>, <span class="hljs-string">'C'</span>, <span class="hljs-string">'F'</span>, <span class="hljs-string">'H'</span>, <span class="hljs-string">'T'</span>, <span class="hljs-string">'a'</span>, <span class="hljs-string">'b'</span>, <span class="hljs-string">'c'</span>, <span class="hljs-string">'d'</span>, <span class="hljs-string">'e'</span>, <span class="hljs-string">'f'</span>, <span class="hljs-string">'g'</span>, <span class="hljs-string">'h'</span>, <span class="hljs-string">'i'</span>, <span class="hljs-string">'k'</span>, <span class="hljs-string">'l'</span>, <span class="hljs-string">'m'</span>, <span class="hljs-string">'n'</span>, <span class="hljs-string">'o'</span>,
	<span class="hljs-string">'p'</span>, <span class="hljs-string">'r'</span>, <span class="hljs-string">'s'</span>, <span class="hljs-string">'t'</span>, <span class="hljs-string">'u'</span>, <span class="hljs-string">'v'</span>, <span class="hljs-string">'w'</span>, <span class="hljs-string">'y'</span>, <span class="hljs-string">'z'</span>, <span class="hljs-string">'Ġ'</span>, <span class="hljs-string">'Ġt'</span>, <span class="hljs-string">'is'</span>, <span class="hljs-string">'er'</span>, <span class="hljs-string">'Ġa'</span>, <span class="hljs-string">'Ġto'</span>, <span class="hljs-string">'en'</span>, <span class="hljs-string">'Th'</span>, <span class="hljs-string">'This'</span>, <span class="hljs-string">'ou'</span>, <span class="hljs-string">'se'</span>,
	<span class="hljs-string">'Ġtok'</span>, <span class="hljs-string">'Ġtoken'</span>, <span class="hljs-string">'nd'</span>, <span class="hljs-string">'Ġis'</span>, <span class="hljs-string">'Ġth'</span>, <span class="hljs-string">'Ġthe'</span>, <span class="hljs-string">'in'</span>, <span class="hljs-string">'Ġab'</span>, <span class="hljs-string">'Ġtokeni'</span>]<!-- HTML_TAG_END --></pre></div> <div class="course-tip bg-gradient-to-br dark:bg-gradient-to-r before:border-green-500 dark:before:border-green-800 from-green-50 dark:from-gray-900 to-white dark:to-gray-950 border border-green-50 text-green-700 dark:text-gray-400"><p data-svelte-h="svelte-x4k8ma">💡 在同一语料库上使用 <code>train_new_from_iterator()</code> 可能不会产生完全相同的词汇表。这是因为当有多个出现频率最高的对时，我们选择遇到的第一个，而 🤗 Tokenizers 库根据内部 ID 选择第一个。</p></div> <p data-svelte-h="svelte-1r7uxle">为了对新文本进行分词，我们对其进行预分词、拆分，然后使用学到的所有合并规则：</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-keyword">def</span> <span class="hljs-title function_">tokenize</span>(<span class="hljs-params">text</span>):
	pre_tokenize_result = tokenizer._tokenizer.pre_tokenizer.pre_tokenize_str(text)
	pre_tokenized_text = [word <span class="hljs-keyword">for</span> word, offset <span class="hljs-keyword">in</span> pre_tokenize_result]
	splits = [[l <span class="hljs-keyword">for</span> l <span class="hljs-keyword">in</span> word] <span class="hljs-keyword">for</span> word <span class="hljs-keyword">in</span> pre_tokenized_text]
	<span class="hljs-keyword">for</span> pair, merge <span class="hljs-keyword">in</span> merges.items():
	<span class="hljs-keyword">for</span> idx, split <span class="hljs-keyword">in</span> <span class="hljs-built_in">enumerate</span>(splits):
	i = <span class="hljs-number">0</span>
	<span class="hljs-keyword">while</span> i < <span class="hljs-built_in">len</span>(split) - <span class="hljs-number">1</span>:
	<span class="hljs-keyword">if</span> split[i] == pair[<span class="hljs-number">0</span>] <span class="hljs-keyword">and</span> split[i + <span class="hljs-number">1</span>] == pair[<span class="hljs-number">1</span>]:
	split = split[:i] + [merge] + split[i + <span class="hljs-number">2</span> :]
	<span class="hljs-keyword">else</span>:
	i += <span class="hljs-number">1</span>
	splits[idx] = split

	<span class="hljs-keyword">return</span> <span class="hljs-built_in">sum</span>(splits, [])<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-18srrkp">我们可以尝试在任何由字母表中的字符组成的文本上进行此操作：</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->tokenize(<span class="hljs-string">"This is not a token."</span>)<!-- HTML_TAG_END --></pre></div> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->[<span class="hljs-string">'This'</span>, <span class="hljs-string">'Ġis'</span>, <span class="hljs-string">'Ġ'</span>, <span class="hljs-string">'n'</span>, <span class="hljs-string">'o'</span>, <span class="hljs-string">'t'</span>, <span class="hljs-string">'Ġa'</span>, <span class="hljs-string">'Ġtoken'</span>, <span class="hljs-string">'.'</span>]<!-- HTML_TAG_END --></pre></div> <div class="course-tip course-tip-orange bg-gradient-to-br dark:bg-gradient-to-r before:border-orange-500 dark:before:border-orange-800 from-orange-50 dark:from-gray-900 to-white dark:to-gray-950 border border-orange-50 text-orange-700 dark:text-gray-400"><p data-svelte-h="svelte-ezw7bq">⚠️ 如果存在未知字符，我们的实现将抛出错误，因为我们没有做任何处理它们。GPT-2 实际上没有未知 tokens （使用字节级 BPE 时不可能得到未知字符），但这里的代码可能会出现这个错误，因为我们并未在初始词汇中包含所有可能的字节。BPE 的这一部分已超出了本节的范围，因此我们省略了一些细节。</p></div> <p data-svelte-h="svelte-1isqaa6">至此，BPE 算法的介绍就到此结束！接下来，我们将研究 WordPiece 算法。</p> <a class="!text-gray-400 !no-underline text-sm flex items-center not-prose mt-4" href="https://github.com/huggingface/course/blob/main/chapters/zh-CN/chapter6/5.mdx" target="_blank"><span data-svelte-h="svelte-1kd6by1"><</span> <span data-svelte-h="svelte-x0xyl0">></span> <span data-svelte-h="svelte-1dajgef"><span class="underline ml-1.5">Update</span> on GitHub</span></a> <p></p>

	<script>
	{
	__sveltekit_9aawfw = {
	assets: "/docs/course/pr_1021/zh-CN",
	base: "/docs/course/pr_1021/zh-CN",
	env: {}
	};

	const element = document.currentScript.parentElement;

	const data = [null,null];

	Promise.all([
	import("/docs/course/pr_1021/zh-CN/_app/immutable/entry/start.f3a1a511.js"),
	import("/docs/course/pr_1021/zh-CN/_app/immutable/entry/app.c39e37cf.js")
	]).then(([kit, app]) => {
	kit.start(app, element, {
	node_ids: [0, 48],
	data,
	form: null,
	error: null
	});
	});
	}
	</script>

Xet Storage Details

Size:: 87.8 kB
Xet hash:: 14f7fafde0c238104266bf3abec8ad4817ca0ec3a862b6c7eb0273f9677f8b26

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.