<?xml version="1.0" encoding="utf-8"?>
<feed xmlns="http://www.w3.org/2005/Atom" xmlns:media="http://search.yahoo.com/mrss/">
  <title>BrewSLM engineering blog</title>
  <subtitle>Engineering notes on small language model fine-tuning infrastructure.</subtitle>
  <link rel="self" type="application/atom+xml" href="https://brewslm.com/feed.xml" />
  <link rel="alternate" type="text/html" href="https://brewslm.com/blog.html" />
  <id>https://brewslm.com/feed.xml</id>
  <updated>2026-10-25T00:00:00Z</updated>
  <author><name>TensorGreed</name><uri>https://github.com/TensorGreed</uri></author>
  <entry>
    <title>Narrative in, JSON out: a 1.5B accident-report extractor that knows when to say null</title>
    <link rel="alternate" type="text/html" href="https://brewslm.com/blog-ntsb-accident-report-extractor.html" />
    <id>https://brewslm.com/blog-ntsb-accident-report-extractor.html</id>
    <published>2026-10-25T00:00:00Z</published>
    <updated>2026-10-25T00:00:00Z</updated>
    <summary>A 1.5B model fine-tuned to turn NTSB aviation accident narratives into a 13-field JSON record. 0.959 field accuracy on 1,234 pre-registered test narratives against 0.494 for the base model, and null when the text doesn't say. All three seeds passed. Every number, including where it still fails.</summary>
    <media:thumbnail url="https://brewslm.com/assets/blog/ntsb-extractor/01-verdict-three-seeds.png" />
  </entry>
  <entry>
    <title>Teaching a 1.5B model to read Pub 17: two misses, one pass</title>
    <link rel="alternate" type="text/html" href="https://brewslm.com/blog-irs-pub17-teaching-a-small-model-to-read.html" />
    <id>https://brewslm.com/blog-irs-pub17-teaching-a-small-model-to-read.html</id>
    <published>2026-10-22T00:00:00Z</published>
    <updated>2026-10-22T00:00:00Z</updated>
    <summary>Part 2 of the Pub 17 run. A 1.5B model fine-tuned to answer from retrieved passages beat the same base model reading the same passages: 0.688 vs 0.570 on 157 pre-registered questions, all three seeds better. Every number, including the two attempts that missed.</summary>
    <media:thumbnail url="https://brewslm.com/assets/blog/irs-pub17-reader/01-reader-verdict.png" />
  </entry>
  <entry>
    <title>The check said don't fine-tune. We did anyway. It was right.</title>
    <link rel="alternate" type="text/html" href="https://brewslm.com/blog-irs-pub17-fine-tune-lost-retrieval-won.html" />
    <id>https://brewslm.com/blog-irs-pub17-fine-tune-lost-retrieval-won.html</id>
    <published>2026-10-19T00:00:00Z</published>
    <updated>2026-10-22T00:00:00Z</updated>
    <summary>Same 41 questions, same judge: base model 0.17, base + retrieved passages 0.73, fine-tuned on 405 pairs 0.02. A pre-training check called it before training.</summary>
    <media:thumbnail url="https://brewslm.com/assets/blog/irs-pub17/01-passages-check.png" />
  </entry>
  <entry>
    <title>Build a sensitive-data detector from zero, with no labelled data</title>
    <link rel="alternate" type="text/html" href="https://brewslm.com/blog-sensitive-data-detector-from-zero.html" />
    <id>https://brewslm.com/blog-sensitive-data-detector-from-zero.html</id>
    <published>2026-10-14T00:00:00Z</published>
    <updated>2026-10-18T00:00:00Z</updated>
    <summary>Step by step: a model that finds PCI, GDPR, HIPAA and custom sensitive data with exact offsets, from an empty project and no labels. Every number real.</summary>
    <media:thumbnail url="https://brewslm.com/assets/blog/sensitive-data/02-synthetic-span-tools.png" />
  </entry>
  <entry>
    <title>Inside the task-aware eval handler dispatcher</title>
    <link rel="alternate" type="text/html" href="https://brewslm.com/blog-task-aware-eval-handler-dispatcher.html" />
    <id>https://brewslm.com/blog-task-aware-eval-handler-dispatcher.html</id>
    <published>2026-05-14T00:00:00Z</published>
    <updated>2026-05-31T00:00:00Z</updated>
    <summary>One generic 'pass rate' metric flatters classification and lies about NER. Here's how BrewSLM dispatches to the metric shape that matches the task.</summary>
  </entry>
  <entry>
    <title>Why the schema introspector beats hand-written converters</title>
    <link rel="alternate" type="text/html" href="https://brewslm.com/blog-schema-introspector-vs-hand-written-converters.html" />
    <id>https://brewslm.com/blog-schema-introspector-vs-hand-written-converters.html</id>
    <published>2026-05-14T00:00:00Z</published>
    <updated>2026-10-18T00:00:00Z</updated>
    <summary>We shipped a 175-line BIO-to-spans converter for the Kaggle PII dataset. Then we threw it away and built a column sniffer. Here's why the sniffer wins.</summary>
  </entry>
  <entry>
    <title>Gamifying a dev tool without making it a toy</title>
    <link rel="alternate" type="text/html" href="https://brewslm.com/blog-gamifying-a-dev-tool.html" />
    <id>https://brewslm.com/blog-gamifying-a-dev-tool.html</id>
    <published>2026-05-14T00:00:00Z</published>
    <updated>2026-10-18T00:00:00Z</updated>
    <summary>Most gamified dev tools feel infantilizing. We tried it anyway. Here are the design rules we landed on and what we explicitly refused to do.</summary>
  </entry>
</feed>
