Tinker Diffusion language models

by Samy (Kondasamy Jayaraman) · drawn with Colorfigures by April Zhu

“ safety boots▢ high-visibility”

Chapter 2 of 7

Fill the blanks

The vendor’s model is not left to right. It starts from a reply of blanks and fills them, like Larkspur’s toy masked model.

This chapter: How does a model fill a blank it can see both sides of?

About this chapter

You will learn

  • How a masked model guesses a hidden token
  • Why it reads the words on both sides
  • How one pass guesses every blank at once

Words you’ll meet

Masked model
A model trained to guess hidden tokens, reading the words on both sides of each one.
Blank
A slot whose token is hidden. The model sees [MASK] there and guesses what belongs.
Confidence
The probability the model gives its best guess for a slot. Here, the glow on the slug.

Step 1 of 4

The frame holds a handbook phrase with one token hidden: a . Larkspur’s toy reads every slot at once.

Predict first.

Which words can the masked model use to guess the blank?