Irth | إرث

Chat with a voice from the 19th-century Arab world.

What happens when you ask a language model to speak from another century?

Irth explores the language, writing styles and perspectives of the 1800s Arab world through an experimental Arabic conversational model, drawing on a curated corpus of approximately one million words from historical books, newspapers and magazines.

Try Irth · Model · Dataset

Behind the voice

This repository contains the historical text processing pipeline and training code, including:

• Source-specific cleaning and corpus preparation.
• Training and held-out data preparation.
• Supervised fine-tuning with loss masking and hand-written refusal examples.

A simulation of the past

Irth generates interpretations, not authentic historical testimony. It can invent details, introduce modern ideas and reproduce prejudices present in its sources. Its outputs do not represent every nineteenth-century Arabic-speaking community.

Learning objectives

• Historical text collection and preprocessing
• Continued pretraining and supervised fine-tuning
• Model evaluation

Research purpose and limitations

This project studies historical language and perspectives; it does not endorse prejudices or discriminatory views found in the source material or generated outputs.

The model is an experimental historical simulation. Its responses are generated interpretations, not authentic historical testimony. It may invent details, introduce modern ideas, or overgeneralize from a limited collection of authors. It should not be treated as representative of all nineteenth-century Arabic-speaking communities.

The project aims to document source provenance, reuse permissions, dataset selection, and preprocessing decisions. Any public demonstration will clearly identify the model as a simulation and explain its limitations.

Initial corpus

• Book One (Published 1834)

  • Approximately 62,000 cleaned words

• Book Two (Published 1895)

  • Approximately 47,873 cleaned words

• Book Three (Published 1857)

  • Approximately 102,506 cleaned words

• Newspaper One (Published 1884)

  • Approximately 95,822 cleaned words

• Book Four (Published 1891)

  • Approximately 242,874 cleaned words

• Book Four (Published 1891)

  • Approximately 242,874 cleaned words

• Book Five (Published 1881)

  • Approximately 23,586 cleaned words

• Book Six (Published 1866)

  • Approximately 102,211 cleaned words

• Magazine One (Published 1892)

  • Approximately 221,395 cleaned words

• Book Eight (Published 1865)

  • Approximately 25,784 cleaned words

• Book Nine (Published 1897)

  • Approximately 109,480 cleaned words

Total: approximately 1,033,531 cleaned words across nine books and one newspaper collection.

Long-term goal

Expand the corpus and train a language-model from scratch as a separate phase.