SSL-HWD is a large-scale handwritten text dataset introduced in the paper "Learning Beyond Labels: Self-Supervised Handwritten Text Recognition" (WACV 2026). The dataset comprises 10 million word-level handwritten images from 852 writers across diverse domains including Physics, Computer Science, Biology, Mathematics, and more.

The dataset is specifically designed to support self-supervised learning approaches for Handwritten Text Recognition (HTR), addressing the critical challenge of reducing dependence on large volumes of labeled data.

Dataset Composition

Total Words: 10 million word-level images
Writers: 852 unique contributors
Pages: 81,280 scanned document pages
Domains: 20+ domains including sciences, literature, and mathematics
Labeled Subset: 2.08M words (20.8%) with ground truth transcriptions
Unlabeled Subset: 7.92M words (79.2%) for self-supervised pretraining
Unique Vocabulary: 107,813 unique words

Key Features

Diversity and Complexity: The dataset includes challenging real-world scenarios:

Texts with different font colors (varying ink and pen usage)
Texts with difficult backgrounds (lines, noise interference)
Texts with distorted characters (irregular strokes, structural inconsistencies)
Texts with blurring effects (motion or focus issues)
Texts with highlighted backgrounds (color markings obscuring content)

Quality Assurance:

All samples automatically annotated using Amazon Textract with confidence ≥99%
Labeled subset manually verified by language experts
High-quality word segmentation with precise bounding boxes

Comparison with Existing Datasets

Dataset	Pages	Writers	Words	Unique Words
IAM	1.5K	657	115K	10.5K
GNHK	687	-	39K	12.3K
IIIT-HW-English-Word	20.8K	1,215	757K	174K
SSL-HWD (Ours)	81.2K	852	10M	107K

Vocabulary Distribution

Category	SSL-HWD
Alphabetic Words	61,088
Numeric Words	4,981
Stop-words	457
Other Words	41,287
Total Unique	107,813

Dataset Structure

Data Organization

SSL-HWD/
├── labeled/              # 2.08M labeled samples
│   ├── writer1/
│   │   ├── writer1.csv   # Ground truth transcriptions
│   │   ├── writer1_1.png
│   │   ├── writer1_2.png
│   │   └── ...
│   ├── writer2/
│   └── ...
└── unlabeled/            # 7.92M unlabeled samples
    ├── writer1/
    │   ├── writer1_1.png
    │   ├── writer1_2.png
    │   └── ...
    └── ...

Data Fields

CSV Format (for labeled data):

image_filename (string): Name of the word image file
transcription (string): Ground truth text transcription

Example:

image_filename,transcription
writer1_1.png,handwritten
writer1_2.png,recognition
writer1_3.png,dataset

Data Splits

The labeled subset (2.08M samples) is divided as follows:

Training: 60% (1.25M samples)
Testing: 40% (0.83M samples)

The unlabeled subset (7.92M samples) is used for self-supervised pretraining.

Supported Tasks

1. Handwriting Text Recognition (HTR)

Train models to recognize handwritten text from word images.

2. Self-Supervised Learning

Use the large unlabeled subset for pretraining with methods like contrastive learning.

3. Writer Identification

Identify writers based on handwriting characteristics (852 unique writers).

4. Domain Adaptation

Transfer learning across different handwriting styles and domains.

5. Semi-Supervised Learning

Combine small labeled and large unlabeled subsets for improved performance.

Usage

Loading the Dataset

from datasets import load_dataset

# Load the full dataset
dataset = load_dataset("your-username/ssl-hwd")

# Access labeled data
labeled = dataset['labeled']

# Access unlabeled data for self-supervised learning
unlabeled = dataset['unlabeled']

Basic Example

from PIL import Image
import pandas as pd
from pathlib import Path

# Load a writer's data
writer_folder = Path("labeled/writer1")
df = pd.read_csv(writer_folder / "writer1.csv")

# Load first image and transcription
img_name = df.iloc[0]['image_filename']
transcription = df.iloc[0]['transcription']

image = Image.open(writer_folder / img_name)
print(f"Transcription: {transcription}")

PyTorch DataLoader

import torch
from torch.utils.data import Dataset, DataLoader
from PIL import Image
import pandas as pd

class SSLHWDDataset(Dataset):
    def __init__(self, writer_folder, transform=None):
        self.folder = Path(writer_folder)
        csv_file = list(self.folder.glob("*.csv"))[0]
        self.data = pd.read_csv(csv_file)
        self.transform = transform
        
    def __len__(self):
        return len(self.data)
    
    def __getitem__(self, idx):
        img_name = self.data.iloc[idx]['image_filename']
        transcription = self.data.iloc[idx]['transcription']
        
        img_path = self.folder / img_name
        image = Image.open(img_path).convert('RGB')
        
        if self.transform:
            image = self.transform(image)
            
        return image, transcription

# Usage
dataset = SSLHWDDataset('labeled/writer1')
dataloader = DataLoader(dataset, batch_size=32, shuffle=True)

Benchmark Results

State-of-the-Art Performance (from LoGo-HTR paper)

IAM Dataset:

With 80% labeled data: WER 11.93%, CER 2.31%
With 100% labeled data: WER 10.27%, CER 2.01%

GNHK Dataset:

With 80% labeled data: WER 19.69%, CER 9.05%
With 100% labeled data: WER 12.07%, CER 7.20%

RIMES Dataset:

With 80% labeled data: WER 6.15%, CER 1.89%
With 100% labeled data: WER 5.50%, CER 1.78%

LAM Dataset (line-level):

With 80% labeled data: WER 7.2%, CER 3.2%
With 100% labeled data: WER 6.3%, CER 2.39%

Cross-Dataset Generalization

The dataset demonstrates strong cross-dataset transfer capabilities:

SSL-HWD → IAM: WER 13.2%, CER 2.9%
SSL-HWD → GNHK: WER 10.1%, CER 6.8%
SSL-HWD → RIMES: WER 11.2%, CER 3.5%
SSL-HWD → LAM: WER 16.4%, CER 7.2%

Dataset Creation

Source Data

The dataset was curated from publicly available digitized manuscripts from web sources, selected for being fully or substantially handwritten. Documents span:

Personal diaries
Academic notes
Historical correspondence
Scientific manuscripts
Mathematical writings
Literature and more

Data Quality

Diverse Sources: 852 unique writers across 20+ domains
Real-world Challenges: Includes blur, noise, distortions, and background interference

Applications

Self-Supervised Learning (Primary Use)

Use the 7.92M unlabeled samples for pretraining with methods like:

Contrastive learning (SimCLR, MoCo)
Masked image modeling
Local-global objectives (as in LoGo-HTR)

Semi-Supervised Learning

Combine labeled and unlabeled subsets for improved performance with limited annotations.

Few-Shot Learning

Train models with minimal labeled data by leveraging pretrained representations.

Transfer Learning

Pretrain on SSL-HWD and fine-tune on domain-specific datasets.

Limitations and Considerations

Known Limitations

Language: Primarily English handwritten text
Geographic Bias: Predominantly Western handwriting styles
Historical Period: Concentrated in specific time periods
Domain Coverage: While diverse, may not represent all handwriting variations

Ethical Considerations

Dataset contains historical documents and handwritten materials
Personal information may be present in some samples
Users should be aware of privacy considerations when using this data

Citation

If you use the SSL-HWD dataset in your research, please cite:

@inproceedings{mitra2026learning,
  title={Learning Beyond Labels: Self-Supervised Handwritten Text Recognition},
  author={Mitra, Shree and Mondal, Ajoy and Jawahar, C. V.},
  booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
  year={2026}
}

Additional Resources

Project Website: https://logo-ssl.github.io/

License

This dataset is released under the Apache License 2.0.

Acknowledgments

This work is supported by MeitY, Government of India, through the NLTM-Bhashini project.

Contact

For questions or issues regarding the dataset:

Authors: Shree Mitra, Ajoy Mondal, C.V. Jawahar
Institution: IIIT Hyderabad
Email: [email protected]

Dataset Version: 1.0
Last Updated: January 2026
Status: Active

Downloads last month: 14