Participating in the NAKBA Competition through Misraj AI

Behind the Scenes: How We Built the Winning Model for the NAKBA NLP 2026 Competition at the LREC Conference.

This post walks through the model development process behind our winning entry — read the full NAKBA NLP 2026 research write-up — and explains the paper linked below. https://lnkd.in/eaGtpaQV

The Arabic handwriting recognition techniques behind this win also power Baseer OCR, our production Arabic document AI product.

Want to put this Arabic HTR expertise to work on your own documents? Talk to our team.

The competition

NAKBA NLP 2026, held with the LREC 2026 conference, ran a competition on Arabic handwritten text recognition: reading historical Arabic manuscripts and returning their text. The pages are highly cursive, written by many different hands and often degraded by age, which is exactly where general OCR tools struggle. Teams were ranked on character error rate (CER) and word error rate (WER) on a hidden test set hosted on Codabench.

The result: first place

Our model, entered as Misraj AI, took first place on the official leaderboard with a CER of 0.0790 and a WER of 0.2440. The top of the table:

Against the competition baseline, that is roughly a fifth of the character errors and about a third of the word errors.

Where we started: Baseer

We did not start from a blank model. The entry adapts Baseer, our 3-billion-parameter vision-language model for Arabic documents, described in arXiv:2509.18174. Baseer was fine-tuned from Qwen2.5-VL-3B-Instruct on 500,000 Arabic page and text pairs and has a word error rate of 0.25 on the Misraj-DocOCR benchmark for printed documents. Historical handwriting is a harder problem, so the work was about carrying that reading ability over to it.

How we built it

Training followed a multi-stage supervised fine-tuning curriculum:

1. Match the data. The Muharaf dataset, which is structurally similar to the competition pages, was converted to grayscale so its tones matched the Nakba data.
2. Train the decoder first. Only the text decoder was trained, on Muharaf, to condition the language-modelling head before the vision side was touched.
3. Tune the whole model. The vision encoder was then unfrozen and the full architecture trained on the Nakba data, with a lower learning rate for the encoder than for the decoder. This step improved WER by more than 5% over decoder-only tuning.
4. Merge the best checkpoints. The checkpoints from epochs 1 and 5 were merged with SLERP interpolation at equal weight, to stabilise predictions and generalise better.

The setup was modest: two NVIDIA H100 GPUs, five epochs, the AdamW optimiser with a cosine schedule, a batch size of 128, sequences of up to 1,200 tokens and input images of 644 by 644 pixels. The encoder learning rate was 9e-6, against 1e-4 for the decoder.

What is public

The weights are on Hugging Face as Misraj/Baseer__Nakba under a CC BY-NC-SA 4.0 licence, and the inference pipeline is on GitHub as misraj-ai/Nakba-pipeline. The model works best on pages like the Nakba data and on visually similar historical manuscripts.

Questions

How is Baseer Nakba different from Baseer? Baseer reads printed Arabic documents and powers Baseer OCR. Baseer Nakba is the same base adapted to historical handwritten manuscripts.

What do CER and WER measure? The share of characters and of words the model gets wrong against a reference transcription. Lower is better.

Can I use the model commercially? The published weights carry a non-commercial licence, CC BY-NC-SA 4.0. To try Baseer on your own documents, visit baseerocr.com.

Reading the full leaderboard

Eight teams were ranked on the hidden test set. Behind the top three came Latent Narratives (CER 0.1050, WER 0.3106), Al-Warraq (0.1142, 0.3780), Not Gemma (0.1217, 0.3063), NAMAA-Qari (0.1950, 0.5194) and Fahras (0.2269, 0.5223). The gap is widest on whole words: our WER of 0.2440 is the only one on the board below 0.29, and the next best, 3reeq's 0.2996, is more than five points higher.