2 min read

Evaluating five million strings

Table of Contents

Draft — this is an outline while I write the full piece. Check back soon.

Before any model, there was the data. This write-up covers the corpus evaluation that came ahead of the autocomplete prototype: how I sampled it, what I measured, and why that work mattered more than the modeling that followed.

What this will cover

  • How the 5M+ strings break down by locale and product
  • Measuring repetition, and where translation memory already helps
  • Cleaning decisions and the ones I got wrong the first time
  • What the evaluation changed about the model plan

More to come here — the notes above are the skeleton I am filling in.