<?xml version="1.0" encoding="utf-8" standalone="yes"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>Model Diagnostics on Nalar</title>
    <link>https://nalar.dev/tags/model-diagnostics/</link>
    <description>Recent content in Model Diagnostics on Nalar</description>
    <generator>Hugo</generator>
    <language>en-us</language>
    <lastBuildDate>Mon, 01 Jan 0001 00:00:00 +0000</lastBuildDate>
    <atom:link href="https://nalar.dev/tags/model-diagnostics/index.xml" rel="self" type="application/rss+xml" />
    <item>
      <title>Inspect Transformer Residual Streams with the Logit Lens</title>
      <link>https://nalar.dev/inspect-transformer-residual-streams-with-the-logit-lens/</link>
      <pubDate>Fri, 11 Sep 2026 00:00:00 +0700</pubDate>
      <guid>https://nalar.dev/inspect-transformer-residual-streams-with-the-logit-lens/</guid>
      <description>&lt;h1 id=&#34;inspect-transformer-residual-streams-with-the-logit-lens&#34;&gt;Inspect Transformer Residual Streams with the Logit Lens&lt;/h1&gt;&#xA;&lt;p&gt;A language model returns token probabilities only after its final transformer block, but developers often need to inspect what happens before that point. A prompt may produce an unexpected completion, a fine-tuned checkpoint may behave differently from its base model, or a model modification may alter predictions several blocks before the final output.&lt;/p&gt;&#xA;&lt;p&gt;The &lt;strong&gt;logit lens&lt;/strong&gt; is a simple diagnostic for these cases. It takes an intermediate residual-stream vector, maps it through the model&amp;rsquo;s output machinery, and ranks vocabulary tokens as if that intermediate state were already ready for prediction.&lt;/p&gt;</description>
    </item>
  </channel>
</rss>
