One arm of a study on teaching a language model to notice when its own conversation has collapsed into repetition, emit a special <end> token, and change the subject.

Version 6.1 at 3 turns: v6 with canned openers and stray formatting cleaned out of the training data. Reading the transcripts, the uncleaned original came out ahead 11 to 4.

Two things that will silently break this. <end> is a special token (id 151665), so decode with skip_special_tokens=False or the model fires and you see nothing. And feed it about 5 turns: it was trained on 5-turn windows, and a 16-turn window dropped firing to 0 of 17 conversations in our tests.

Full write-up, training recipe and limitations: center-for-humans-and-machines/end_token. Datasets: Farrous05/end_token_data_full.

Built at the Center for Humans and Machines, Max Planck Institute for Human Development.