Model Card for TikZilla-8B

TikZilla-8B is a language model for generating TikZ/LaTeX figures from natural language descriptions.

It is based on Qwen3-8B-Base and was trained with supervised fine-tuning (SFT) on DaTikZ-V4 for scientific figure generation.

Installation

pip install torch==2.5.1 transformers==4.53.2 accelerate==1.8.1

Usage

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, GenerationConfig

model_id = "nllg/TikZilla-8B"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

eos_token_id = tokenizer.convert_tokens_to_ids("<|im_end|>")
pad_token_id = tokenizer.pad_token_id or tokenizer.eos_token_id or eos_token_id

gen_config = GenerationConfig(
    do_sample=True,
    temperature=1.0,
    top_p=0.9,
    max_new_tokens=2048,
    eos_token_id=eos_token_id,
    pad_token_id=pad_token_id,
)

your_input_description = "A scientific line plot showing two curves. The x-axis is labeled 'Time' ranging from 0 to 100, and the y-axis is labeled 'Value' ranging from 0 to 1. The first curve is a blue solid line that gradually increases from near 0 and levels off around 0.9. The second curve is a red dashed line that rises to a peak around the middle of the plot and then decreases. A legend in the upper right labels the blue line as 'Model A' and the red dashed line as 'Model B'. The background is white with light gray grid lines."

messages = [
    {
        "role": "user",
        "content": (
            "Generate a complete LaTeX document that contains a TikZ figure according to the following requirements:\n"
            + your_input_description +
            "\nWrap your code using \\documentclass[tikz]{standalone}, and include \\begin{document}...\\end{document}. "
            "Only output valid LaTeX code with no extra text."
        ),
    }
]

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)

inputs = tokenizer([text], return_tensors="pt").to(model.device)
output_ids = model.generate(**inputs, generation_config=gen_config)
response_ids = output_ids[0][len(inputs["input_ids"][0]):]
output = tokenizer.decode(response_ids, skip_special_tokens=True)

print(output)