This repository compiles results of the unicorn test for various open-weights models.
This is not a good test. It is not even meant to be a good test.
However: Nobody is gaming it deliberately, because it is too silly. So it might be a better test than alternatives.
- We only care about whether it can draw TikZ.
- We don't care what the model says to us that isn't TikZ code
- We always use exactly the prompt "Draw a unicorn in TikZ".
Run process_tex.py to generate a pdf. The file will take the tex files in \drawings, combine them with head.tex and footer.tex, and use pdflatex to create a .pdf.
python file_to_compile.tex
Below are some unicorns produced by leading LLMs. Top row are the best. Bottom row is the worst.
