DistilRoBERTa is a distilled version of RoBERTa. A smaller student model learns to reproduce the behaviour of the larger RoBERTa teacher.

Training process

  1. Give input text to the RoBERTa teacher
  2. Use the teacher's output probabilities as a learning signal
  3. Train the smaller student to produce similar outputs
RoBERTa teacher -> teacher output probabilities -> smaller student

Purpose

  • Reduce the size of the model
  • Reduce the resources needed to run the model
  • Retain useful language-understanding behaviour from the teacher

Distilled model comparison

StudentTeacher
DistilBERTBERT
DistilRoBERTaRoBERTa

The teacher determines which model behaviour the student learns. Each student requires its own evaluation of size, speed, and accuracy.