Paper summary: Mutarjim: Advancing Bidirectional Arabic-English Translation with a Small Language Model

Problem

Large LLMs translate well but are expensive. Can a small, specialised model compete for Arabic–English? Existing benchmarks also have issues: narrow domains, short sentences and an English-source bias.

Method

Mutarjim builds on Kuwain-1.5B, a model tailored for Arabic and English, with an optimised two-phase training approach and a carefully curated corpus. The authors also introduce Tarjama-25, 5,000 expert-reviewed sentence pairs across a wide range of domains.

Key results

From the abstract: Mutarjim outperforms much larger models on several established benchmarks and rivals models up to 20× larger at much lower cost. On Tarjama-25 it reaches state of the art for English→Arabic, surpassing larger and proprietary models such as GPT-4o mini.

Limitations

There’s no dedicated limitations section in the extracted text. The headline English→Arabic result is on the authors’ own new benchmark.