vichetkao/khmer_word_dataset

Khmer OCR Dataset for Qwen3-VL Fine-Tuning

This dataset contains Khmer text images paired with their corresponding text labels, prepared for OCR and vision-language fine-tuning, specifically targeting models like Qwen3-VL.

All images and annotations were created by the authors based on Khmer text from Khmer Wikipedia. The dataset is fully self-contained in Parquet files with embedded images for easy loading via the Hugging Face datasets library.

---

Dataset Features

Column	Type	Description
id	string	Unique identifier for each example (e.g., khmer_0)
image	Image	The raw image data (embedded bytes, JPEG/PNG format)
conversations	list	Conversation-style format for Qwen3-VL fine-tuning
conversations.from	string	"user" or "assistant"
conversations.value	string	The input instruction or target text

Example Entry:

{
  "id": "khmer_0",
  "image": "<embedded image bytes>",
  "conversations": [
    {
      "from": "user",
      "value": "<image>\nExtract the Khmer text from this image."
    },
    {
      "from": "assistant",
      "value": "ខ្មែរ text here"
    }
  ]
}

---

Dataset Splits

• Train: 90% of examples
• Test: 10% of examples

Files:

train-00000-of-00001.parquet
test-00000-of-00001.parquet

---

Dataset Size

• Number of images: 26,873
• Total raw image size: ~298 MB
• Parquet size (with embedded images): ~200 MB

---

Usage

Load the dataset using Hugging Face datasets library:

from datasets import load_dataset

dataset = load_dataset("vichetkao/khmer_word_dataset")
train_ds = dataset['train']
test_ds = dataset['test']

# Access an image and show it
sample = train_ds[0]["image"]
sample.show()

# Access the conversation text
print(train_ds[0]["conversations"])

---

Purpose

This dataset is intended for:

• Khmer OCR model training
• Vision-Language fine-tuning (e.g., Qwen3-VL, LLaVA, Pixtral)
• Document recognition and extraction

---

License

This dataset is released under Apache-2.0 license.

---

Source / Attribution

The Khmer text used to generate this dataset was collected from Kaggle. All images and annotations were created by the authors and do not contain copyrighted images.

---

Citation

@misc{khmer_ocr_dataset,
  title = {Khmer OCR Dataset for Qwen3-VL},
  author = {Keat Chakravuth},
  year = {2026},
  howpublished = {\\url{https://huggingface.co/vichetkao/khmer_word_dataset}}
}