Report
@wightmanr reports an in-fusion boost for decoder-only OCR
Rebalancing image and text parameter widths improved decoder-only OCR experiments, @wightmanr says, but more trials are needed.
TLDR
Adding Meta’s in-fusion approach improved a decoder-only OCR model, @wightmanr reports, but more trials are needed. The post also describes slower convergence in a separate experiment using pretrained SmolLM2 135M. The revived project has a pastry counter’s worth of configurations: Cruller, Fritter, Cronut and Berliner, spanning encoder-decoder and decoder-only designs.
Combined views
164
3 Sources, first seen ago
Comments