Buckets:
| <meta charset="utf-8" /><meta name="hf:doc:metadata" content="{"title":"Токенизация Byte-Pair Encoding","local":"byte-pair-encoding-tokenization","sections":[{"title":"Алгоритм обучения","local":"training-algorithm","sections":[],"depth":2},{"title":"Алгоритм токенизации","local":"tokenization-algorithm","sections":[],"depth":2},{"title":"Реализация BPE","local":"implementing-bpe","sections":[],"depth":2}],"depth":1}"/> | |
| <link href="/docs/course/pr_1349/ru/_app/immutable/entry/start.BsLZCO1o.js" rel="modulepreload"> | |
| <link href="/docs/course/pr_1349/ru/_app/immutable/chunks/BUjtLMej.js" rel="modulepreload"> | |
| <link href="/docs/course/pr_1349/ru/_app/immutable/chunks/C1g_JXwL.js" rel="modulepreload"> | |
| <link href="/docs/course/pr_1349/ru/_app/immutable/entry/app.Cc2HB5tW.js" rel="modulepreload"> | |
| <link href="/docs/course/pr_1349/ru/_app/immutable/chunks/CI14Xrrj.js" rel="modulepreload"> | |
| <link href="/docs/course/pr_1349/ru/_app/immutable/chunks/CtQcHPv9.js" rel="modulepreload"> | |
| <link href="/docs/course/pr_1349/ru/_app/immutable/chunks/DGU_2Sfd.js" rel="modulepreload"> | |
| <link href="/docs/course/pr_1349/ru/_app/immutable/chunks/DsnmJJEf.js" rel="modulepreload"> | |
| <link href="/docs/course/pr_1349/ru/_app/immutable/nodes/0.B6NbQ79Y.js" rel="modulepreload"> | |
| <link href="/docs/course/pr_1349/ru/_app/immutable/chunks/ChX01CHW.js" rel="modulepreload"> | |
| <link href="/docs/course/pr_1349/ru/_app/immutable/nodes/47.CzB8h1lF.js" rel="modulepreload"> | |
| <link href="/docs/course/pr_1349/ru/_app/immutable/chunks/D_OGVccK.js" rel="modulepreload"> | |
| <link href="/docs/course/pr_1349/ru/_app/immutable/chunks/Dv56shkE.js" rel="modulepreload"> | |
| <link href="/docs/course/pr_1349/ru/_app/immutable/chunks/C2e3vw1h.js" rel="modulepreload"> | |
| <link href="/docs/course/pr_1349/ru/_app/immutable/chunks/Bt8BkiRx.js" rel="modulepreload"> | |
| <!--b0b49a--><meta name="hf:doc:metadata" content="{"title":"Токенизация Byte-Pair Encoding","local":"byte-pair-encoding-tokenization","sections":[{"title":"Алгоритм обучения","local":"training-algorithm","sections":[],"depth":2},{"title":"Алгоритм токенизации","local":"tokenization-algorithm","sections":[],"depth":2},{"title":"Реализация BPE","local":"implementing-bpe","sections":[],"depth":2}],"depth":1}"/><!----> | |
| <link href="/docs/course/pr_1349/ru/_app/immutable/assets/0.tn0RQdqM.css" rel="modulepreload"> <!--[--><!--[0--><!--[--><!--[0--><!--[--><p></p> <div class="items-center shrink-0 min-w-[100px] max-sm:min-w-[50px] justify-end ml-auto flex" style="float: right; margin-left: 10px; display: inline-flex; position: relative; z-index: 10;"><div class="inline-flex rounded-md max-sm:rounded-sm"><button class="inline-flex items-center gap-1 h-7 max-sm:h-7 px-2 max-sm:px-1.5 text-sm font-medium text-gray-800 border border-r-0 rounded-l-md max-sm:rounded-l-sm border-gray-200 bg-white hover:shadow-inner dark:border-gray-850 dark:bg-gray-950 dark:text-gray-200 dark:hover:bg-gray-800" aria-live="polite"><span class="inline-flex items-center justify-center rounded-md p-0.5 max-sm:p-0 hover:text-gray-800 dark:hover:text-gray-200"><svg class="sm:size-3.5 size-3" 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><!----></span> <span>Copy page</span></button> <button class="inline-flex items-center justify-center w-6 max-sm:w-5 h-7 max-sm:h-7 disabled:pointer-events-none text-sm text-gray-500 hover:text-gray-700 dark:hover:text-white rounded-r-md max-sm:rounded-r-sm border border-l transition border-gray-200 bg-white hover:shadow-inner dark:border-gray-850 dark:bg-gray-950 dark:text-gray-200 dark:hover:bg-gray-800" aria-haspopup="menu" aria-expanded="false" aria-label="Open copy menu"><svg class="transition-transform text-gray-400 overflow-visible sm:size-3.5 size-3 rotate-0" width="1em" height="1em" viewBox="0 0 12 7" fill="none" xmlns="http://www.w3.org/2000/svg"><path d="M1 1L6 6L11 1" stroke="currentColor"></path></svg><!----></button></div> <!--[-1--><!--]--></div><!----> <!--[0--><h1 class="relative group"><a id="byte-pair-encoding-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="#byte-pair-encoding-tokenization"><span><svg 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>Токенизация Byte-Pair Encoding</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><!----> <!--[0--><a href="https://colab.research.google.com/github/huggingface/notebooks/blob/master/course/en/chapter6/section5.ipynb" target="_blank"><img alt="Open In Colab" class="!m-0" src="https://colab.research.google.com/assets/colab-badge.svg"/></a><!--]--> <!--[0--><a href="https://studiolab.sagemaker.aws/import/github/huggingface/notebooks/blob/master/course/en/chapter6/section5.ipynb" target="_blank"><img alt="Open In Studio Lab" class="!m-0" src="https://studiolab.sagemaker.aws/studiolab.svg"/></a><!--]--></div><!----> <p>Byte-Pair Encoding (BPE) изначально была разработана как алгоритм для сжатия текстов, а затем использовалась OpenAI для токенизации при предварительном обучении модели GPT. Она используется во многих моделях трансформеров, включая GPT, GPT-2, RoBERTa, BART и 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><!----> <blockquote class="tip"><p>💡 В этом разделе подробно рассматривается BPE, вплоть до демонстрации полной реализации. Вы можете пропустить этот раздел, если вам нужен только общий обзор алгоритма токенизации.</p></blockquote> <!--[1--><h2 class="relative group"><a id="training-algorithm" 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="#training-algorithm"><span><svg 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>Алгоритм обучения</span></h2><!--]--><!----> <p>Обучение BPE начинается с вычисления уникального набора слов, используемых в корпусе (после завершения этапов нормализации и предварительной токенизации), затем создается словарь, в который заносятся все символы, используемые для записи этих слов. В качестве очень простого примера предположим, что в нашем корпусе используются следующие пять слов:</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 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=" "><!----><span class="hljs-string">"hug"</span><span class="hljs-punctuation">,</span> <span class="hljs-string">"pug"</span><span class="hljs-punctuation">,</span> <span class="hljs-string">"pun"</span><span class="hljs-punctuation">,</span> <span class="hljs-string">"bun"</span><span class="hljs-punctuation">,</span> <span class="hljs-string">"hugs"</span><!----></pre></div><!----> <p>Тогда базовым словарем будет <code>["b", "g", "h", "n", "p", "s", "u"]</code>. В реальном мире этот базовый словарь будет содержать, как минимум, все символы ASCII, а возможно, и некоторые символы Unicode. Если в примере, который вы обрабатываете, используется символ, которого нет в обучающем корпусе, этот символ будет преобразован в неизвестный токен. Это одна из причин, по которой многие модели NLP очень плохо анализируют контент с эмоджи, например.</p> <blockquote class="tip"><p>Токенизаторы GPT-2 и RoBERTa (которые довольно похожи) имеют умный способ решения этой проблемы: они рассматривают слова не как символы Unicode, а как байты. Таким образом, базовый словарь имеет небольшой размер (256), но все символы, которые вы можете придумать, все равно будут включены и не будут преобразованы в неизвестный токен. Этот трюк называется <em>byte-level BPE</em>.</p></blockquote> <p>После получения базового словаря мы добавляем новые токены, пока не достигнем желаемого объема словаря, обучаясь <em>слияниям</em>, которые представляют собой правила слияния двух элементов существующего словаря в новый. Таким образом, в начале эти слияния будут создавать токены с двумя символами, а затем, по мере обучения, более длинные подслова.</p> <p>На любом шаге обучения токенизатора алгоритм BPE будет искать наиболее частую пару существующих токенов (под “парой” здесь понимаются два последовательных токена в слове). Эта наиболее часто встречающаяся пара и будет объединена, после чего все повторяется для следующего шага.</p> <p>Возвращаясь к нашему предыдущему примеру, предположим, что слова имеют следующую частоту:</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 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=" "><!---->(<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>)<!----></pre></div><!----> <p>значение <code>" hug"</code> встречалось в корпусе 10 раз, <code>"pug"</code> - 5 раз, <code>"pun"</code> - 12 раз, <code>"bun"</code> - 4 раза, и <code>"hugs"</code> - 5 раз. Мы начинаем обучение с разбиения каждого слова на части символов (те, которые формируют наш начальный словарь), чтобы мы могли рассматривать каждое слово как список токенов:</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 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=" "><!---->(<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>)<!----></pre></div><!----> <p>Затем мы посмотрим на пары. Пара <code>("h", "u")</code> присутствует в словах <code>"hug"</code> и <code>"hugs"</code>, всего 15 раз в корпусе. Однако это не самая частая пара: эта честь принадлежит <code>("u", "g")</code>, которая присутствует в словах <code>"hug"</code>, <code>"pug"</code> и <code>"hugs"</code>, в общей сложности 20 раз в словаре.</p> <p>Таким образом, первое правило слияния, выученное токенизатором, - <code>("u", "g") -> "ug"</code>, что означает, что <code>"ug"</code> будет добавлено в словарь, и эта пара должна быть объединена во всех словах корпуса. В конце этого этапа словарь и корпус выглядят следующим образом:</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 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=" "><!----><span class="hljs-symbol">Vocabulary:</span> [<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-symbol">Corpus:</span> (<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>)<!----></pre></div><!----> <p>Теперь у нас есть несколько пар, в результате которых получается токен длиннее двух символов: например, пара <code>("h", "ug")</code> (встречается в корпусе 15 раз). Самая частая пара на этом этапе - <code>("u", "n")</code>, однако она встречается в корпусе 16 раз, поэтому второе выученное правило слияния - <code>("u", "n") -> "un"</code>. Добавив это в словарь и объединив все существующие вхождения, мы получаем:</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 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=" "><!----><span class="hljs-symbol">Vocabulary:</span> [<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-symbol">Corpus:</span> (<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>)<!----></pre></div><!----> <p>Теперь наиболее частой парой является <code>("h", "ug")</code>, поэтому мы изучаем правило слияния <code>("h", "ug") -> "hug"</code>, что дает нам первый трехбуквенный токен. После слияния корпус выглядит следующим образом:</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 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=" "><!----><span class="hljs-symbol">Vocabulary:</span> [<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 class="hljs-symbol">Corpus:</span> (<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>)<!----></pre></div><!----> <p>И продолжаем в том же духе, пока не достигнем желаемого размера словаря.</p> <blockquote class="tip"><p>✏️ <strong>Теперь ваша очередь!</strong> Как вы думаете, каким будет следующее правило слияния?</p></blockquote> <!--[1--><h2 class="relative group"><a id="tokenization-algorithm" 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-algorithm"><span><svg 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>Алгоритм токенизации</span></h2><!--]--><!----> <p>Токенизация следует за процессом обучения в том смысле, что новые входные данные подвергаются токенизации путем применения следующих шагов:</p> <ol><li>Нормализация</li> <li>Предварительная токенизация</li> <li>Разделение слов на отдельные символы</li> <li>Применение правил слияния, изученных по порядку, к этим частям</li></ol> <p>Возьмем пример, который мы использовали во время обучения, с тремя выученными правилами слияния:</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 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=" "><!----><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><!----></pre></div><!----> <p>Слово <code>"bug"</code> будет токенизировано как <code>["b", "ug"]</code>. Слово <code>"mug"</code>, однако, будет токенизировано как <code>["[UNK]", "ug"]</code>, поскольку буква <code>"m"</code> отсутствует в базовом словаре. Аналогично, слово <code>"thug" будет токенизировано как</code>[”[UNK]”, “hug”]<code>: буква</code>“t” отсутствует в базовом словаре, и применение правил слияния приводит сначала к слиянию <code>"u"</code> и <code>"g"</code>, а затем к слиянию <code>"h"</code> и <code>"ug"</code>.</p> <blockquote class="tip"><p>✏️ <strong>Теперь ваша очередь!</strong> Как вы думаете, как будет токенизировано слово <code>'unhug'</code>?</p></blockquote> <!--[1--><h2 class="relative group"><a id="implementing-bpe" 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="#implementing-bpe"><span><svg 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>Реализация BPE</span></h2><!--]--><!----> <p>Теперь давайте посмотрим на реализацию алгоритма BPE. Это не будет оптимизированная версия, которую вы сможете использовать на большом корпусе; мы просто хотим показать вам код, чтобы вы могли лучше понять алгоритм.</p> <p>Для начала нам нужен корпус текста, поэтому давайте создадим простой корпус с несколькими предложениями:</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 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="language-python "><!---->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>, | |
| ]<!----></pre></div><!----> <p>Далее нам нужно предварительно токенизировать корпус в слова. Поскольку мы воспроизводим токенизатор BPE (например, GPT-2), для предварительной токенизации мы будем использовать токенизатор <code>gpt2</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 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="language-python "><!----><span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> AutoTokenizer | |
| tokenizer = AutoTokenizer.from_pretrained(<span class="hljs-string">"gpt2"</span>)<!----></pre></div><!----> <p>Затем мы вычисляем частоту каждого слова в корпусе, как и при предварительной токенизации:</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 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="language-python "><!----><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)<!----></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 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="language-python "><!---->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>})<!----></pre></div><!----> <p>Следующий шаг - составление базового словаря, состоящего из всех символов, используемых в корпусе:</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 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="language-python "><!---->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)<!----></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 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="language-python "><!---->[ <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>]<!----></pre></div><!----> <p>Мы также добавляем специальные токены, используемые моделью, в начало этого словаря. В случае GPT-2 единственным специальным токеном является <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 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="language-python "><!---->vocab = [<span class="hljs-string">"<|endoftext|>"</span>] + alphabet.copy()<!----></pre></div><!----> <p>Теперь нам нужно разделить каждое слово на отдельные символы, чтобы можно было начать обучение:</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 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="language-python "><!---->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()}<!----></pre></div><!----> <p>Теперь, когда мы готовы к обучению, давайте напишем функцию, которая вычисляет частоту каждой пары. Нам нужно будет использовать ее на каждом шаге обучения:</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 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="language-python "><!----><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<!----></pre></div><!----> <p>Давайте посмотрим на часть этого словаря после первых разделений:</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 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="language-python "><!---->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><!----></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 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="language-python "><!---->(<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><!----></pre></div><!----> <p>Теперь, чтобы найти наиболее часто встречающуюся пару, нужно всего лишь сделать быстрый цикл:</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 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="language-python "><!---->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)<!----></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 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="language-python "><!---->(<span class="hljs-string">'Ġ'</span>, <span class="hljs-string">'t'</span>) <span class="hljs-number">7</span><!----></pre></div><!----> <p>Итак, первое слияние, которое нужно выучить, это <code>('Ġ', 't') -> 'Ġt'</code>, и мы добавляем <code>'Ġt'</code> в словарь:</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 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="language-python "><!---->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>)<!----></pre></div><!----> <p>Чтобы продолжить, нам нужно применить это объединение в нашем экземпляре <code>splits</code> словаря. Давайте напишем для этого еще одну функцию:</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 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="language-python "><!----><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<!----></pre></div><!----> <p>И мы можем посмотреть на результат первого слияния:</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 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="language-py "><!---->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>])<!----></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 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="language-python "><!---->[<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>]<!----></pre></div><!----> <p>Теперь у нас есть все, что нужно, чтобы проитерироваться до тех пор, пока мы не выучим все слияния, которые нам нужны. Пусть размер словаря будет 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 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="language-python "><!---->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>])<!----></pre></div><!----> <p>В результате мы выучили 19 правил слияния (исходный словарь имел размер 31 - 30 символов в алфавите плюс специальный токен):</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 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="language-py "><!----><span class="hljs-built_in">print</span>(merges)<!----></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 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="language-python "><!---->{(<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>}<!----></pre></div><!----> <p>А словарь состоит из специального токена, начального алфавита и всех результатов слияния:</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 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="language-py "><!----><span class="hljs-built_in">print</span>(vocab)<!----></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 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="language-python "><!---->[<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>]<!----></pre></div><!----> <blockquote class="tip"><p>💡 Использование <code>train_new_from_iterator()</code> на том же корпусе не приведет к созданию точно такого же словаря. Это связано с тем, что при выборе наиболее частотной пары мы выбираем первую попавшуюся, в то время как библиотека 🤗 Tokenizers выбирает первую пару, основываясь на ее внутренних ID.</p></blockquote> <p>Чтобы токенизировать новый текст, мы предварительно токенизируем его, разбиваем на части, а затем применяем все изученные правила слияния:</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 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="language-python "><!----><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, [])<!----></pre></div><!----> <p>Мы можем попробовать это на любом тексте, состоящем из символов алфавита:</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 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="language-py "><!---->tokenize(<span class="hljs-string">"This is not a token."</span>)<!----></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 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="language-python "><!---->[<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>]<!----></pre></div><!----> <blockquote class="warning"><p>⚠️ Наша реализация будет выбрасывать ошибку при наличии неизвестного символа, поскольку мы ничего не сделали для их обработки. На самом деле в GPT-2 нет неизвестного токена (невозможно получить неизвестный символ при использовании BPE на уровне байтов), но здесь это может произойти, поскольку мы не включили все возможные байты в начальный словарь. Этот аспект BPE выходит за рамки данного раздела, поэтому мы опустили подробности.</p></blockquote> <p>Вот и все об алгоритме BPE! Далее мы рассмотрим WordPiece.</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/ru/chapter6/5.mdx" target="_blank"><svg class="mr-1" 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="M31,16l-7,7l-1.41-1.41L28.17,16l-5.58-5.59L24,9l7,7z"></path><path d="M1,16l7-7l1.41,1.41L3.83,16l5.58,5.59L8,23l-7-7z"></path><path d="M12.419,25.484L17.639,6.552l1.932,0.518L14.351,26.002z"></path></svg><!----> <span><span class="underline">Update</span> on GitHub</span></a><!----> <p></p><!--]--><!----><!--]--><!--]--><!--]--> <!--[-1--><!--]--><!--]--> | |
| <script> | |
| { | |
| __sveltekit_1scm2e0 = { | |
| base: "/docs/course/pr_1349/ru", | |
| assets: "/docs/course/pr_1349/ru" | |
| }; | |
| const element = document.currentScript.parentElement; | |
| Promise.all([ | |
| import("/docs/course/pr_1349/ru/_app/immutable/entry/start.BsLZCO1o.js"), | |
| import("/docs/course/pr_1349/ru/_app/immutable/entry/app.Cc2HB5tW.js") | |
| ]).then(([kit, app]) => { | |
| kit.start(app, element, { | |
| node_ids: [0, 47], | |
| data: [null,null], | |
| form: null, | |
| error: null | |
| }); | |
| }); | |
| } | |
| </script> | |
Xet Storage Details
- Size:
- 93.7 kB
- Xet hash:
- 62dd58bc9c024627df45d0797e3657891464844968e88dbdedf7446d54d051fd
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.