This model is the baseline system for the NAKBA NLP 2026: Arabic Manuscript Understanding Shared Task (Systems Track). It fine-tunes Qwen3-VL-8B-Instruct with LoRA to transcribe Arabic manuscript line images into text.

Model Details

Description

Model type: Vision-language OCR/HTR model (LoRA-adapted)
Finetuned from model: Qwen/Qwen3-VL-8B-Instruct

Sources

Repository: https://github.com/U4RASD/ar-ms-baseline
Shared Task: https://acrps.ai/nakba-nlp-manu-understanding-2026

Training Details

Training Data

NAKBA NLP 2026 Shared Task (Subtask 2) training split from the Omar Al-Saleh memoir collection.
Dataset includes line images with gold transcriptions.

Training Procedure

Supervised fine-tuning with LoRA adapters on Qwen/Qwen3-VL-8B-Instruct.

Training Hyperparameters

Config reference: Hyperparameters are listed in configs/default.json

Evaluation

Testing Data, Factors & Metrics

Testing Data

NAKBA NLP 2026 Shared Task (Subtask 2) released test set of line images.

Metrics

CER (Character Error Rate)
WER (Word Error Rate)

Results

On released test set:

CER: 0.2297
WER: 0.4998
Hardware: NVIDIA H100 SXM

Contact

ar-ms@dohainstitute.edu.qa

U4RASD/ar-ms-baseline

Model Card for ar-ms-baseline

Model Summary