Khmer Language OCR Training Data for Gemma Model

A multimodal dataset for training optical character recognition models on the Khmer language, created by KiteAether and hosted on Hugging Face. It was last updated on June 11, 2025. The dataset is categorized as containing between 1,000 and 10,000 samples.

Use Cases

• Fine-tune the Gemma model for Khmer text recognition from images using multimodal training data.
• Train an OCR model to extract Khmer text from documents, leveraging image-text pairs.
• Benchmark the performance of different OCR architectures on the Khmer language.
• Develop a pre-processing pipeline for digitizing historical or contemporary Khmer documents.

Strengths

• Dataset is specifically designed for the Khmer language, a low-resource script for OCR.
• Data is stored in the efficient Parquet file format, suitable for large-scale processing.
• Includes multimodal (image and text) data, which is essential for training OCR models.

Limitations

• The exact number of rows, columns, and sample data are unknown, making it difficult to assess scale and structure.
• Potential for class imbalance or label noise is unverified without access to sample data or statistics.
• Geographic and temporal coverage of the source documents is unspecified, which may limit generalizability.

Provenance

Source

Hugging Face, uploaded by author KiteAether.

Collection Method

Presumably collected and curated for training OCR models on Khmer text.

Freshness

Last updated on June 11, 2025.

Geography

Region is tagged as 'us', but the language focus is Khmer (Cambodia).