vaishnavi0901/kannada-ocr-dataset

Kannada OCR Fine-Tuning Dataset

Fine-tuning dataset for training vision-language models to read handwritten Kannada text from historical birth and death register images.

Dataset Description

This dataset pairs scanned register page images with structured JSON ground truth extracted from verified source data. The goal is to fine-tune a model that can look at a register page and output the correct field values for every row on that page.

Source

• Region: Haveri district, Karnataka, India
• Document Type: Historical birth and death registers
• Page Sizes: A3 and A5 scanned pages
• Language: Handwritten Kannada script

Dataset Statistics

Split	Images	Records
Train	119	404
Val	13	46
Total	132	450

Note: Birth register pages often contain multiple rows (~4.4 rows/page average), while death pages typically have 1 row each.

Data Format

Each sample contains:

• User message: Image + OCR prompt
• Assistant message: Ground-truth JSON

Birth Record Fields (8 fields)

• reg_no: Registration serial number
• b_dob: Date of birth (YYYY-MM-DD)
• gender: M or F
• b_baby_nme_kn: Baby's name in Kannada
• b_father_kn: Father's name in Kannada
• mother_namekn: Mother's name in Kannada
• info_name_kn: Informant's name in Kannada
• info_details: Informant relationship/address

Death Record Fields (10 fields)

• reg_no: Registration serial number
• d_dod: Date of death (YYYY-MM-DD)
• gender: M or F
• age: Age at death
• d_deceased_nme_kn: Deceased's name in Kannada
• d_father_kn: Father's name in Kannada
• d_h_w_name_kn: Husband/wife's name in Kannada
• mother_namekn: Mother's name in Kannada
• info_name_kn: Informant's name in Kannada
• info_details: Informant relationship/address

Usage

from datasets import load_dataset
from PIL import Image

# Load the dataset
dataset = load_dataset("YOUR_USERNAME/kannada-ocr-dataset")

# Access a sample
sample = dataset["train"][0]
print(sample["messages"])

For Vision Model Fine-tuning

import json
from pathlib import Path
from PIL import Image
from datasets import load_dataset

dataset = load_dataset("YOUR_USERNAME/kannada-ocr-dataset")

def load_sample_with_image(sample, dataset_path):
    """Load sample and convert image path to PIL Image."""
    for msg in sample["messages"]:
        if msg["role"] == "user":
            for part in msg["content"]:
                if part["type"] == "image":
                    img_path = Path(dataset_path) / part["image"]
                    part["image"] = Image.open(img_path).convert("RGB")
    return sample

Training Recommendations

• Use max_length=8192 (multi-record JSON responses are long)
• Use num_train_epochs=3-5
• Start with birth records only to validate pipeline
• A100 40GB recommended for Gemma 3 Vision fine-tuning

License

CC-BY-4.0