vichetkao/wild_khmer

Wild Khmer Grounding Dataset (OCR + Visual Grounding)

This dataset is a high-quality collection of Khmer text images "in the wild." It is a derivative work specifically enhanced for Visual Grounding and High-Accuracy OCR fine-tuning.

📌 Provenance & Credits

• Original Source: Khmer Word Dataset (Kaggle)
• Original Owner: Saly KEO
• Modification: The original dataset (which provided single text labels per image) has been re-annotated and processed to include normalized bounding boxes for multi-region text detection and grounding tasks.

📄 License

This dataset is distributed under the Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license.

• Attribution: You must give appropriate credit to Saly KEO (original source) and the current maintainer.
• Non-Commercial: You may not use the material for commercial purposes.

---

Dataset Summary

This version of the dataset provides localized bounding boxes for every text region within an image. This enables Vision-Language Models (VLMs) to learn spatial reasoning—understanding exactly where a piece of text is located before reading it.

• Primary Language: Khmer (UTF-8)
• Format: Hugging Face Parquet (Embedded Images)
• Normalization: Coordinates scaled to 0-1000.
• Task Type: Visual Grounding, Document Intelligence, Khmer OCR.

Dataset Structure

Column	Type	Description
image	Image	The source image (embedded bytes).
text	String	A JSON string containing a list of text regions and their normalized bounding boxes.

JSON Label Example

{
  "regions": [
    {
      "bbox_2d": [ymin, xmin, ymax, xmax],
      "label": "សោភ័ណ ស៊ីថា"
    }
  ]
}

How to Use

Loading the Dataset

from datasets import load_dataset

dataset = load_dataset("vichetkao/wild_khmer")
print(dataset['train'][0])

Formatting for Qwen3-VL / Unsloth

import json

def convert_to_conversation(sample):
    grounding_data = json.loads(sample["text"])
    instruction = "Detect all Khmer text regions in this image and return the labels with their bounding boxes in JSON format."

    return {
        "messages": [
            {"role": "user", "content": [{"type": "text", "text": instruction}, {"type": "image", "image": sample["image"]}]},
            {"role": "assistant", "content": [{"type": "text", "text": sample["text"]}]}
        ]
    }

Supported Models

• Qwen3-VL-8B / 32B
• Qwen2-VL Series
• InternVL2
• Llama-3.2-Vision

Citation

If you use this dataset, please credit the original author:

Original Creator: Saly KEO (Kaggle)
Derivative Work: Kao Vichet
License: CC BY-NC 4.0