nllg/detikzify-cl-7b

Model Card for DeTikZify-CL_7b

DeTikZify-CL_7b is a language model that automatically converts sketches and existing scientific figures into editable, semantics-preserving TikZ graphics programs. It is based on CodeLlama 7b and was fine-tuned on DaTikZ_v2. Check out the DeTikZify project for more information and tips on how to best run the model.

Usage

from operator import itemgetter

from detikzify.model import load
from detikzify.infer import DetikzifyPipeline
import torch

image = "https://w.wiki/A7Cc"
pipeline = DetikzifyPipeline(*load(
    base_model="nllg/detikzify-cl-7b",
    device_map="auto",
    torch_dtype=torch.bfloat16,
))

# generate a single TikZ program
fig = pipeline.sample(image=image)

# if it compiles, rasterize it and show it
if fig.is_rasterizable:
    fig.rasterize().show()

# run MCTS for 10 minutes and generate multiple TikZ programs
figs = set()
for score, fig in pipeline.simulate(image=image, timeout=600):
    figs.add((score, fig))

# save the best TikZ program
best = sorted(figs, key=itemgetter(0))[-1][1]
best.save("fig.tex")