How to fine-tuning DeTikZify?

[Peter233666]

Hello!
Unlike large models such as DeepSeek, I wasn’t able to find any online tutorials for fine-tuning DeTikZify, so I’m reaching out for your help. I’d like to fine-tune the latest version of DeTikZify to enable it to understand certain symbols found on engineering drawings and then generate corresponding TikZ code to recreate the diagrams. Additionally, DeTikZify does not support Chinese content recognition. How can I make it capable of recognizing and constructing Chinese fonts? Please provide a somewhat detailed step-by-step guide. I really appreciate your help—I'm a beginner in this area! This is the first time I try fine-tuning.Thanks again.

[potamides]

Hi! We have example scripts for pretraining, fine-tuning, and post-training DeTikZify that can help you get started with training your own DeTikZify variants.

To make DeTikZify_v2 work better for Chinese, I can think of several approaches. If you have Chinese TikZ data available, you can use it to fine-tune DeTikZify directly. Alternatively, if you have Chinese figures without TikZ code, you could try post-training the model on this data, since our post-training setup doesn't require TikZ data. Another option is to take our DaTikZ_v3 dataset and generate synthetic Chinese data by machine translating all the text into Chinese. It's worth noting that the DeTikZify_v2 models are based on LLaMA_3.1, which doesn't perform particularly well on Chinese text in general. For better results, you might consider training your own DeTikZify model from scratch using a base model that works better on Chinese, such as Qwen2.5-VL (7b). This approach would likely yield better zero-shot performance on Chinese after training.

[Peter233666]

Thanks for your patient reply.I will keep going.And I've send an email to the address listed on your github profile,please check it when you are available. :)