Al-Khwarizmi

AI Math Tutor

Visit huggingface.co Discuss it

LiveAI & productivityWeb, iOS, Android, Desktop, Bot / APIArabic, English, Other languages

What it does

A small language model trained with full fine-tuning, LoRA, and DPO, to solve simple math problems, supports 9 languages and can easily be downloaded and used offline (1.9-6 GB variants). Training reached 87.21% mean token accuracy and cut validation loss by about 19%. Named after Muhammad ibn Musa al-Khwarizmi, the 9th-century mathematician whose name is the direct origin of the word "algorithm".

Build story

How AI helped

First working version

Around 2 weeks

From idea to something people could use.

What AI did

Helped shape the training configurations I chose into code, translated an EN math dataset for multilingual training, helped draft the model card, and created the online demo.

What the maker did

Chose the training methods and configurations for a number of iterations aimed at better results with step-by-step reasoning and more efficient local usage with smaller sizes.

Biggest lesson

“Small AI models can perform well while being usable locally when trained on a specific task or field.”