◄ WORLD V · SONNY 5DART 628 · a helldive back into the mind

THE TEACHER FORCING train on the truth, test on yourself

To train a sequence model to predict the next token, you could feed it its own previous prediction — but early on those are garbage and errors compound. Teacher forcing feeds the real previous token instead, so every step gets a clean target and training is fast and stable. The catch is exposure bias: at inference the model must feed itself, a distribution it never saw in training — and small errors snowball.

THE TECHNIQUE feed the true previous token, not the guess

The demo shows each step’s input is the ground-truth previous token — errors do not compound during training: live demo


HISTORY & CREDIT Williams & Zipser · 1989

“Teacher forcing has no downside.” — it causes exposure bias; the model never learns to recover from its own mistakes (scheduled sampling and RL fixes address it). cited

the clean signal · feed the true previous token, so each step’s target is correct — fast, stable training.
the catch · exposure bias — at test the model feeds itself, a distribution it never trained on.
1989 · Williams & Zipser.

A clean classroom, a messy exam. training

RECOMMEND FOR I-13 the ground truth, on the compiler

On i-13, step 2’s input is the true token 1, step 3’s is the true 2 — the model’s own guesses never feed back in training:

$ i13 run n2_teacher-forcing.i13 RUN OK · 22 step(s) input_step2 = 1 input_step3 = 2 (ground truth) no_error_compound = 1 exposure_bias = 1
Recommend as a NULL — a training protocol (B39). Feeding the ground-truth input is a fixed choice; exposure bias is its known cost. A policy, not a same-function DOF. NULL — train on the truth, test on yourself.