
  <rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
    <channel>
      <title>KShivendu</title>
      <link>https://www.kshivendu.dev/blog</link>
      <description>Kumar Shivendu&#39;s blog</description>
      <language>en-us</language>
      <managingEditor>mail@kshivendu.dev (Kumar Shivendu)</managingEditor>
      <webMaster>mail@kshivendu.dev (Kumar Shivendu)</webMaster>
      <lastBuildDate>Thu, 17 Sep 2026 00:00:00 GMT</lastBuildDate>
      <atom:link href="https://www.kshivendu.dev/tags/tokenization/feed.xml" rel="self" type="application/rss+xml"/>
      
  <item>
    <guid>https://www.kshivendu.dev/blog/improving-bpe</guid>
    <title>Improving BPE: Stop storing the space, store its absence</title>
    <link>https://www.kshivendu.dev/blog/improving-bpe</link>
    <description>BPE spends two vocabulary slots on every word, one with a leading space and one without. I mark the absence of a space instead, which frees ~14,000 slots and beats a properly-trained BPE baseline by 9.4% on code and 0.9% on English. Also: the baseline mistake that made my first results look 4x better than they were.</description>
    <pubDate>Thu, 17 Sep 2026 00:00:00 GMT</pubDate>
    <author>mail@kshivendu.dev (Kumar Shivendu)</author>
    <category>nlp</category><category>compression</category><category>tokenization</category><category>internals</category>
  </item>

    </channel>
  </rss>
