jawi-ocr-al-models

Qwen-SEA-LION-v4-8B-VL Active Learning Models for Jawi OCR

This repository hosts the final adapters (iter_5_model) from an Active Learning (AL) study for Jawi OCR.

---

📂 Repository Structure

The models are organized into subfolders based on their dataset, AL strategy, and pool size:

1. Workspace Models (workspace_models/)

These models were trained in the root workspace models/ directory:

• workspace_models/al-run-dynamic-10/iter_5_model
• workspace_models/al-run-dynamic-50/iter_5_model
• workspace_models/al-run-new-way-20/iter_5_model
• workspace_models/al-run-new-way-30/iter_5_model
• workspace_models/al-run-new-way-40/iter_5_model
• workspace_models/al-run-new-way/iter_5_model
• workspace_models/al-run-new-way_pool10/iter_5_model
• workspace_models/al-run-old-way/iter_5_model
• workspace_models/al-run-old-way_pool10/iter_5_model
• workspace_models/al_entropy_full_orig_models_pool10/iter_5_model
• workspace_models/al_entropy_full_orig_models_pool20/iter_5_model
• workspace_models/al_entropy_full_orig_models_pool30/iter_5_model
• workspace_models/al_entropy_full_orig_models_pool40/iter_5_model
• workspace_models/al_entropy_full_orig_models_pool50/iter_5_model
• workspace_models/al_kmeans_center_full_orig_models_pool10/iter_5_model
• workspace_models/al_kmeans_center_full_orig_models_pool20/iter_5_model
• workspace_models/al_kmeans_center_full_orig_models_pool30/iter_5_model
• workspace_models/al_kmeans_center_full_orig_models_pool40/iter_5_model
• workspace_models/al_kmeans_center_full_orig_models_pool50/iter_5_model
• workspace_models/al_random_full_aug_models_pool10/iter_5_model
• workspace_models/al_random_full_aug_models_pool20/iter_5_model
• workspace_models/al_random_full_aug_models_pool30/iter_5_model
• workspace_models/al_random_full_aug_models_pool40/iter_5_model
• workspace_models/al_random_full_aug_models_pool50/iter_5_model
• workspace_models/al_random_full_orig_models_pool10/iter_5_model
• workspace_models/al_random_full_orig_models_pool20/iter_5_model
• workspace_models/al_random_full_orig_models_pool30/iter_5_model
• workspace_models/al_random_full_orig_models_pool40/iter_5_model
• workspace_models/al_random_full_orig_models_pool50/iter_5_model

2. Jawi Paper Repo Models (Jawi_Paper_Repo/)

These models are from the main paper replication repository:

3. Historical Benchmark Models (other_hist_bench/)

These models are historical benchmarks evaluated on historical lines datasets (like Belfort, Esposalles, Himanis, NewsEye, NorHand):

• other_hist_bench/Jawi-OCR-data-v4_al_diva_alpha10/iter_5_model
• other_hist_bench/Jawi-OCR-data-v4_al_entropy/iter_5_model
• other_hist_bench/Jawi-OCR-data-v4_al_kmeans_center/iter_5_model
• other_hist_bench/Jawi-OCR-data-v4_al_random/iter_5_model
• other_hist_bench/Teklia_Esposalles-line_al_diva_alpha10/iter_5_model
• other_hist_bench/Teklia_Esposalles-line_al_diva_alpha10_full/iter_5_model
• other_hist_bench/Teklia_Esposalles-line_al_diva_alpha10_full_50/iter_5_model
• other_hist_bench/Teklia_Esposalles-line_al_entropy/iter_5_model
• other_hist_bench/Teklia_Esposalles-line_al_entropy_full/iter_5_model
• other_hist_bench/Teklia_Esposalles-line_al_entropy_full_50/iter_5_model
• other_hist_bench/Teklia_Esposalles-line_al_kmeans_center/iter_5_model
• other_hist_bench/Teklia_Esposalles-line_al_kmeans_center_full/iter_5_model
• other_hist_bench/Teklia_Esposalles-line_al_kmeans_center_full_50/iter_5_model
• other_hist_bench/Teklia_Esposalles-line_al_random/iter_5_model
• other_hist_bench/Teklia_Esposalles-line_al_random_full/iter_5_model
• other_hist_bench/Teklia_Esposalles-line_al_random_full_50/iter_5_model

---

🚀 How to Load and Use

To load any of the adapter models, you can use the transformers and peft libraries in Python.

Python Code Example

import torch
from transformers import AutoProcessor, AutoModelForImageTextToText
from peft import PeftModel

# 1. Specify the base model and this repository
base_model_id = "aisingapore/Qwen-SEA-LION-v4-8B-VL"
adapter_repo_id = "ThuraAung1601/jawi-ocr-al-models"

# 2. Choose the specific adapter subfolder path from the list above
# Example: Loading the random active learning run with pool size 30
subfolder_path = "workspace_models/al_random_full_orig_models_pool30/iter_5_model"

print("Loading base model...")
model = AutoModelForImageTextToText.from_pretrained(
    base_model_id, 
    device_map="auto", 
    torch_dtype=torch.bfloat16,
    trust_remote_code=True
)
processor = AutoProcessor.from_pretrained(base_model_id, trust_remote_code=True)

print(f"Loading adapter from {subfolder_path}...")
model = PeftModel.from_pretrained(
    model, 
    adapter_repo_id, 
    subfolder=subfolder_path
)

print("Model is successfully loaded and ready for inference!")