@inproceedings{ali-etal-2026-gemma,
title = "Not Gemma at {AR}-{MS} {N}akba{NLP} 2026: Mubsir {OCR}: End-to-End Recognition of {A}rabic Handwritten Text",
author = "Ali, Ali Adel and
Ali, Mona Khaled and
Sayed, Mohamed Emad and
Mostafa, Ibrahim Naser",
editor = "Jarrar, Mustafa and
El-Haj, Mo and
Haddad, Amal and
Atiani, Serin and
Abudalfa, Shadi and
Regier, Terry and
Rayson, Paul and
Sima{'}an, Khalil and
Mansour, Camille",
booktitle = "Proceedings of the 2nd International Workshop on Nakba Narratives as Language Resources @ {LREC} 2026",
month = may,
year = "2026",
address = "Palma, Mallorca (Spain)",
publisher = "ELRA Language Resources Association (ELRA)",
url = "https://aclanthology.org/2026.nakbanlp-1.42/",
doi = "10.63317/44f6f4a6cxma",
pages = "273--278",
abstract = "Historical Arabic handwritten OCR is difficult because of cursive script, fine diacritics, mixed numerals, and degraded media; classical segmentation pipelines compound errors, whereas end-to-end vision-language models can adapt when fine-tuned on in-domain data. We present Mubsir OCR, a systematic evaluation on the NAKBA dataset: an annotated set (15,962 training line crops and 2,095 val lines with ground truth, used for all nine experiments) and a separate blind AR-MS (Subtask 2) set (2,671 images; scores only via official submission). We compare external vs. in-house VLMs (Qwen2.5-VL 3B, Qwen3-VL-4B-Instruct, Gemma3), inference backends (vLLM/bf16 vs. HuggingFace/bf16), training length (16 vs. 32 epochs), and test-time preprocessing (CLAHE+unsharp). Best on the annotated val set: 8.59{\%} CER / 25.87{\%} WER (HuggingFace bf16); the same configuration attains 11.00{\%} CER / 31.26{\%} WER on the blind set. Domain-specific fine-tuning beats general-purpose checkpoints; preprocessing helps only marginally and is not recommended without train-time augmentation."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="ali-etal-2026-gemma">
<titleInfo>
<title>Not Gemma at AR-MS NakbaNLP 2026: Mubsir OCR: End-to-End Recognition of Arabic Handwritten Text</title>
</titleInfo>
<name type="personal">
<namePart type="given">Ali</namePart>
<namePart type="given">Adel</namePart>
<namePart type="family">Ali</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mona</namePart>
<namePart type="given">Khaled</namePart>
<namePart type="family">Ali</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mohamed</namePart>
<namePart type="given">Emad</namePart>
<namePart type="family">Sayed</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ibrahim</namePart>
<namePart type="given">Naser</namePart>
<namePart type="family">Mostafa</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-05</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 2nd International Workshop on Nakba Narratives as Language Resources @ LREC 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Mustafa</namePart>
<namePart type="family">Jarrar</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mo</namePart>
<namePart type="family">El-Haj</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Amal</namePart>
<namePart type="family">Haddad</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Serin</namePart>
<namePart type="family">Atiani</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Shadi</namePart>
<namePart type="family">Abudalfa</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Terry</namePart>
<namePart type="family">Regier</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Paul</namePart>
<namePart type="family">Rayson</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Khalil</namePart>
<namePart type="family">Sima’an</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Camille</namePart>
<namePart type="family">Mansour</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>ELRA Language Resources Association (ELRA)</publisher>
<place>
<placeTerm type="text">Palma, Mallorca (Spain)</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
</relatedItem>
<abstract>Historical Arabic handwritten OCR is difficult because of cursive script, fine diacritics, mixed numerals, and degraded media; classical segmentation pipelines compound errors, whereas end-to-end vision-language models can adapt when fine-tuned on in-domain data. We present Mubsir OCR, a systematic evaluation on the NAKBA dataset: an annotated set (15,962 training line crops and 2,095 val lines with ground truth, used for all nine experiments) and a separate blind AR-MS (Subtask 2) set (2,671 images; scores only via official submission). We compare external vs. in-house VLMs (Qwen2.5-VL 3B, Qwen3-VL-4B-Instruct, Gemma3), inference backends (vLLM/bf16 vs. HuggingFace/bf16), training length (16 vs. 32 epochs), and test-time preprocessing (CLAHE+unsharp). Best on the annotated val set: 8.59% CER / 25.87% WER (HuggingFace bf16); the same configuration attains 11.00% CER / 31.26% WER on the blind set. Domain-specific fine-tuning beats general-purpose checkpoints; preprocessing helps only marginally and is not recommended without train-time augmentation.</abstract>
<identifier type="citekey">ali-etal-2026-gemma</identifier>
<identifier type="doi">10.63317/44f6f4a6cxma</identifier>
<location>
<url>https://aclanthology.org/2026.nakbanlp-1.42/</url>
</location>
<part>
<date>2026-05</date>
<extent unit="page">
<start>273</start>
<end>278</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Not Gemma at AR-MS NakbaNLP 2026: Mubsir OCR: End-to-End Recognition of Arabic Handwritten Text
%A Ali, Ali Adel
%A Ali, Mona Khaled
%A Sayed, Mohamed Emad
%A Mostafa, Ibrahim Naser
%Y Jarrar, Mustafa
%Y El-Haj, Mo
%Y Haddad, Amal
%Y Atiani, Serin
%Y Abudalfa, Shadi
%Y Regier, Terry
%Y Rayson, Paul
%Y Sima’an, Khalil
%Y Mansour, Camille
%S Proceedings of the 2nd International Workshop on Nakba Narratives as Language Resources @ LREC 2026
%D 2026
%8 May
%I ELRA Language Resources Association (ELRA)
%C Palma, Mallorca (Spain)
%F ali-etal-2026-gemma
%X Historical Arabic handwritten OCR is difficult because of cursive script, fine diacritics, mixed numerals, and degraded media; classical segmentation pipelines compound errors, whereas end-to-end vision-language models can adapt when fine-tuned on in-domain data. We present Mubsir OCR, a systematic evaluation on the NAKBA dataset: an annotated set (15,962 training line crops and 2,095 val lines with ground truth, used for all nine experiments) and a separate blind AR-MS (Subtask 2) set (2,671 images; scores only via official submission). We compare external vs. in-house VLMs (Qwen2.5-VL 3B, Qwen3-VL-4B-Instruct, Gemma3), inference backends (vLLM/bf16 vs. HuggingFace/bf16), training length (16 vs. 32 epochs), and test-time preprocessing (CLAHE+unsharp). Best on the annotated val set: 8.59% CER / 25.87% WER (HuggingFace bf16); the same configuration attains 11.00% CER / 31.26% WER on the blind set. Domain-specific fine-tuning beats general-purpose checkpoints; preprocessing helps only marginally and is not recommended without train-time augmentation.
%R 10.63317/44f6f4a6cxma
%U https://aclanthology.org/2026.nakbanlp-1.42/
%U https://doi.org/10.63317/44f6f4a6cxma
%P 273-278
Markdown (Informal)
[Not Gemma at AR-MS NakbaNLP 2026: Mubsir OCR: End-to-End Recognition of Arabic Handwritten Text](https://aclanthology.org/2026.nakbanlp-1.42/) (Ali et al., NakbaNLP 2026)
ACL