Re:Mlx-community/OLMo-2-0325-32B-Instruct-4bit
by simonw on Monday March 17, 2025 @10:48PM
(#43395283)
The models are mostly terrible at SVG output, at least if you ask for something that's hard (or impossible?) to draw like a pelican riding a bicycle. That's why I use it as a benchmark, I think it's amusing: https://simonwillison.net/tags/pelican-riding-a-bicycle/
Some of them can do good SVGs for things that make sense, like simple diagrams.
Re:Mlx-community/OLMo-2-0325-32B-Instruct-4bit
by showmexyz on Monday March 17, 2025 @10:53PM
(#43395308)
These works well for some svg that are simple and already in training data but doesn't work for harder svgs, even simple one if they are out of distribution of training data.
In simon's example whole purpose is to make it draw something that it has not seen before but can easily infer from geometry, spatial arrangement. I think it makes a fun problem.
Am I doing something wrong? Everyone seems to say how well models work in producing SVGs but I get shapes in all sorts of the wrong places. SVG documents are quite low level (verses editing them in Inkscape or Illustrator) so its tricky to modify, beyond very simple shapes.