Most embedding benchmarks are English-first, which means "state of the art" often quietly means "state of the art if your documents are in English." Arabic legal text breaks a lot of those assumptions — script, morphology, and domain vocabulary all shift at once.
Starting point
An open-source Qwen model, adapted rather than trained from scratch — the base model's multilingual pretraining gives it a real head start on Arabic, but general-purpose Arabic understanding and legal-domain Arabic understanding are not the same thing.
What actually moved the needle
- Domain-specific hard negatives mattered more than raw data volume.
- Evaluation had to be legal-expert-reviewed, not just held-out accuracy — a wrong-but-plausible retrieval is a worse failure mode here than a low-confidence miss.
- Reranking on top of the fine-tuned embeddings closed most of the remaining gap.
This is placeholder body copy — swap in the real write-up, numbers, and what didn't work.