Datasets

Public Khmer data is the scarcest thing in this field, so the data we build is published, free to download on Hugging Face. Each one's page has the full card and the files.

  • bilingual-en-km

    Four English–Khmer corpora, khmer-english-pairs-raw among them, combined, deduplicated and tagged for both directions, plus 42,572 validation pairs

    Used to train Netra MT 1

    4,214,606 pairs

    Task
    Translation
    Kind
    Mixed
    Licence
    CC BY 4.0

  • khmer-english-parallel

    Khmer–English sentence pairs from news and government sources and the ALT corpus, each scored with COMET, with fixed validation and test splits of about 1,000 pairs each

    81,003 pairs

    Task
    Translation
    Kind
    Mixed

  • khmer-english-pairs-raw

    English–Khmer sentence pairs from legal, literary and governmental text

    Used to train Netra MT 1

    200,000 pairs

    Task
    Translation
    Kind
    Synthetic

  • khmer-scene-text-synthetic-contrast

    Khmer text lines set over photographs as scene text, each with its transcription

    Used to train Netra OCR 1

    102,500 images

    Task
    Text recognition
    Kind
    Synthetic

  • khmer-document-synthetic-low-res

    Khmer text lines rendered as low-resolution document text, each with its transcription

    Used to train Netra OCR 1

    100,000 images

    Task
    Text recognition
    Kind
    Synthetic

  • Textline-Detection-Dataset

    Boxes around every text line and embedded image, on real scene text, DocLayNet document pages and synthetic Khmer documents, in YOLO format

    33,422 images

    Task
    Text detection
    Kind
    Mixed
    Licence
    CC BY 4.0

  • khmer-textline-dataset_v2

    Khmer official documents, such as letters, memos and press releases, with boxes around each text line and each stamp, seal or logo

    4,243 images

    Task
    Text detection
    Kind
    Synthetic
    Licence
    MIT

  • khmer-id-detection

    Khmer ID cards with made-up details and a box around each of 12 fields, from the Khmer and English names to the three machine-readable lines

    10,000 images

    Task
    Field detection
    Kind
    Synthetic

  • khmer-historical-heritage-image-captioning

    Images of Khmer historical heritage, such as carved reliefs, each described in an English caption

    1,387 images

    Task
    Image captioning
    Kind
    Real

  • khmer-asr-bench

    Khmer speech from FLEURS and three of seanghay's corpora, with transcripts and a fixed train and test split, so speech encoders can be compared on identical audio. Each source keeps its own licence

    8,037 utterances

    Task
    Speech recognition
    Kind
    Real

  • Khmer–English code-switch speech

    Hosted by Panhapich

    Khmer sentences with English words mixed in, written by a language model and read aloud by VoxCPM2

    47 h

    Task
    Speech recognition
    Kind
    Synthetic

Have Khmer data that could help, even a small, well-labelled set? Here is how to contribute it.