clean_repeated_substrings is a dirty hack
#9
by
PartyParrot
- opened
Your model is suffering from the repetition bug, which is difficult to fix after training. You might want to consider penalizing repetitions during training, see for example section 4.4 in this paper: https://arxiv.org/pdf/2503.08525
If that does not work, you can also predict coordinates and font size in addition to tokens to draw a kind of heatmap and then penalize if the heatmap gets too crowded or the characters go out of bounds of the document.
Is there a way to fix post training on vLLM?