Tensor by Tensor

02.28 · UNIT 08 · Train, evaluate, and serve your transformer · Lab

Lab: train complete transformer models

A complete transformer lab joins data, masks, loss, updates, and held-out generation.

PLAIN-LANGUAGE INTRODUCTION

What is this?

A complete transformer lab joins data, masks, loss, updates, and held-out generation.

One simple example

For reversal, source [2,5,7] has target [7,5,2]. Evaluation must generate all three target tokens.

What goes in?

Training batches, shifted targets, masks, and a fixed model configuration.

What comes out?

Updated parameters and free-running held-out predictions.

Why does it matter?

The lab proves that separate components form one learning system.

What is it not?

Low teacher-forced loss does not prove correct free generation.

WORK THROUGH THE IDEA

See the idea in more detail

  1. This task maps source [2,5,7] to target [7,5,2]. Keep some sequences held out.
  2. Training uses shifted targets. The loss compares each predicted next token with its correct label.
  3. An optimizer updates parameters after backpropagation. Record the configuration and random seed.
  4. During evaluation, generate [7,5,2] without supplying correct earlier target tokens.
  5. Common mistake: a separately sampled test tensor can still repeat a generated training sequence. Check overlap.
Open the detailed notes ↗