<?xml version="1.0" encoding="utf-8" standalone="yes"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/">
  <channel>
    <title>Multilingual on Rahul Bhati</title>
    <link>https://therahulbhati.github.io/tags/multilingual/</link>
    <description>Recent content in Multilingual on Rahul Bhati</description>
    <generator>Hugo -- 0.151.0</generator>
    <language>en-us</language>
    <lastBuildDate>Fri, 10 Jul 2026 21:30:00 +0530</lastBuildDate>
    <atom:link href="https://therahulbhati.github.io/tags/multilingual/index.xml" rel="self" type="application/rss+xml" />
    <item>
      <title>One Vocabulary, Four Languages: A BPE Tokenizer From Scratch</title>
      <link>https://therahulbhati.github.io/posts/multilingual-bpe-tokenizer/</link>
      <pubDate>Fri, 10 Jul 2026 21:30:00 +0530</pubDate>
      <guid>https://therahulbhati.github.io/posts/multilingual-bpe-tokenizer/</guid>
      <description>&lt;p&gt;Commercial tokenizers quietly charge some languages more than others. I wanted to measure that myself. Then I wanted to see how far from-scratch engineering can push against it. So, an experiment. Take the Wikipedia &lt;strong&gt;India&lt;/strong&gt; page in English, Hindi, Telugu, and one more language. Build a BPE tokenizer from scratch with a single 10,000-token vocabulary shared across all four. Optimize one number: &lt;code&gt;1000 / (X_max − X_min)&lt;/code&gt;, where each &lt;code&gt;X&lt;/code&gt; is fertility, the tokens a language needs per word. The metric only rewards &lt;em&gt;balance&lt;/em&gt;. A tokenizer that&amp;rsquo;s great at English and bad at Telugu scores worse than one that&amp;rsquo;s mediocre at both.&lt;/p&gt;</description>
    </item>
  </channel>
</rss>
