Misraj/Baseer__Nakba

Baseer-Nakba HTR: A State-of-the-Art VLM for Arabic Handwritten Text Recognition

Overview

This repository contains the model weights and inference pipeline for our submission to the NAKBA NLP 2026 Arabic Handwritten Text Recognition (HTR) competition.

Our approach adapts the 3B-parameter Baseer Vision-Language Model (VLM) to effectively parse and recognize highly cursive, historical Arabic manuscripts. Through a progressive training pipeline, domain-matched data augmentation, and advanced checkpoint merging, this unified model mitigates the challenges of varying writer styles, age-related document degradation, and morphological complexity.

To try our Baseer model for document extraction, please visit: Baseer — Baseer is the SOTA model on Arabic Document Extraction.

---

🏆 Competition Results

Our final model (Misraj AI) secured 1st place on the official Nakba hidden test set leaderboard.

Rank	Team	CER	WER
🥇 1st	Misraj AI	0.0790	0.2440
🥈 2nd	Oblevit	0.0925	0.3268
🥉 3rd	3reeq	0.0938	0.2996
4th	Latent Narratives	0.1050	0.3106
5th	Al-Warraq	0.1142	0.3780
6th	Not Gemma	0.1217	0.3063
7th	NAMAA-Qari	0.1950	0.5194
8th	Fahras	0.2269	0.5223
—	Baseline	0.3683	0.6905

---

Training Methodology

Our model was trained using a multi-stage Supervised Fine-Tuning (SFT) curriculum.

1. Data Augmentation: The Muharaf enhancement dataset was converted to grayscale to match the visual complexity and tonal distribution of the Nakba competition data.
2. Decoder-Only SFT: We first trained the text decoder autoregressively on the structurally similar Muharaf dataset to condition the language modeling head.
3. Full Encoder-Decoder Tuning: We subsequently unfroze the vision encoder and trained the full architecture on the Nakba dataset using differential learning rates — a key step that yielded a >5% improvement in WER over decoder-only tuning.
4. Checkpoint Merging: To stabilize predictions and maximize generalization, we merged our top-performing checkpoints (Epoch 1 and Epoch 5) using SLERP interpolation.

---

Training Hyperparameters

All supervised experiments were conducted with standardized hyperparameters across configurations.

Parameter	Value
Hardware	2× NVIDIA H100 GPUs
Base Model	3B-parameter Baseer
Epochs	5
Optimizer	AdamW
Weight Decay	0.01
Learning Rate Schedule	Cosine
Batch Size	128
Max Sequence Length	1200 tokens
Input Image Resolution	644 × 644 pixels
Decoder-Only Learning Rate	1e-4
Encoder Learning Rate	9e-6
Decoder Learning Rate (Full Tuning)	1e-4

---

Image Examples

The model works reliably on images from the Nakba dataset and visually similar historical manuscripts.

---

Merge Method

This model was merged using the SLERP merge method.

Models Merged

• Baseer_Nakba_ep_1
• Baseer_Nakba_ep_5

Configuration

merge_method: slerp
base_model: Baseer_Nakba_ep_1
models:
  - model: Baseer_Nakba_ep_1
  - model: Baseer_Nakba_ep_5
parameters:
  t:
    - value: 0.50
dtype: bfloat16

---

Citation

If you use this model or find our work helpful, please consider citing our paper:

@inproceedings{misrajai2026nakba,
  title     = {Adapting Vision-Language Models for Historical Arabic Handwritten Text Recognition},
  author    = {Misraj AI},
  booktitle = {Nakba OCR Competition, NLP 2026},
  year      = {2026}
}

---

Links

• 🤗 Model weights: Misraj/Baseer__Nakba
• 💻 Inference pipeline: misraj-ai/Nakba-pipeline
• 🌐 Live demo: baseerocr.com
• 📄 Competition: Nakba Codabench