Teaching machines to think

Netra Lab is an independent, non-commercial research lab. We build small models that read, hear, translate and write Khmer, and every one is free to download, test and build on.

Look ahead

Explore the research
Footage and photographs from Wikimedia Commons · Credits
  1. Ocean Moon Glint and City Night Lights in 4K · NASA · Public domain
  2. Artemis I Slow Motion Liftoff (KSC-20221116--MH-FMX-01-0001-Artemis 1 slowmo liftoff-3314595) · NASA Kennedy Space Center · Public domain
  3. Apollo 11 launch, video of engines at 500 fps (camera E-8) · NASA/Kennedy Space Center · Public domain
  4. Time-lapse Earth Flyover from NASA Astronaut in Space · NASA · Public domain
  5. Aurora Borealis from Space (4K) · NASA · Public domain
  6. Angkor - Zentrum des Königreichs der Khmer (CC BY-SA 4.0) · ZDF/Terra X/interscience film/Faber Courtial/Gero von Boehm/Alexander Hein/Andreas Tiletzek, Jörg Courtial · CC BY-SA 4.0
  7. The monolithic rock-hewn churches of Lalibela, Terra X (English redub) · ZDF/Terra X/interscience film/Faber Courtial/Gero von Boehm/Alexander Hein/Andreas Tiletzek, Jörg Courtial · CC BY 4.0
  8. Die Hagia Sophia in Istanbul · ZDF/Terra X/Bilderfest/M. Richter-Jähnig/H. Neuhäuser/K. Weiss/M. Heß · CC BY-SA 4.0
  9. Alhambra-Kuppel 2025 Zeitraffer, Museum für Islamische Kunst, Staatliche Museen zu Berlin · Maximilian Heiden · CC BY-SA 4.0
  10. Nagasaki lantern festival - 2023 1 28 · Nesnad · CC BY-SA 4.0
  11. Ultra High Definition Video from the International Space Station (Reel 1) · NASA Johnson · Public domain
  12. NASA-Apollo8-Dec24-Earthrise · NASA/Bill Anders · Public domain
  13. Saturn from Cassini Orbiter (2004-10-06) · NASA/JPL/Space Science Institute · Public domain
  14. Jupiter Showcases Auroras, Hazes (NIRCam Widefield View) · NASA, ESA, CSA, Jupiter ERS Team; Judy Schmidt, Ricardo Hueso · Public domain
  15. NASA’s Webb Reveals Cosmic Cliffs, Glittering Landscape of Star Birth · NASA, ESA, CSA, STScI · Public domain
  16. Webb's First Deep Field · NASA, ESA, CSA, STScI · Public domain
  17. Hubble and Webb Showcase the Pillars of Creation (Side by Side) (weic2216d) · NASA, ESA, CSA, STScI; J. DePasquale, A. Koekemoer, A. Pagan (STScI) · CC BY 4.0
  18. Cuevamanos3399 · Fjturban · Public domain
  19. Das Lächeln von Angkor · Hans Stieglitz · CC BY-SA 3.0
  20. Tsunami by hokusai 19th century · Katsushika Hokusai · Public domain
  21. Vincent van Gogh - Starry Night - Google Art Project · Vincent van Gogh · Public domain
  22. Katsushika Hokusai - Fine Wind, Clear Morning (Gaifū kaisei) - Google Art Project · Katsushika Hokusai · Public domain
  23. Johannes Vermeer - Girl with a Pearl Earring - 670 - Mauritshuis · Johannes Vermeer · Public domain
  24. Court of Gayumars Persian miniature · Attributed to Sultan Muhammad · Public domain
  25. Claude Monet - Reflections of Clouds on the Water-Lily Pond · Claude Monet · Public domain
  26. Michelangelo - Creation of Adam (cropped) · Michelangelo · Public domain
  27. Kuo Hsi - Early Spring - Google Art Project · Guo Xi · Public domain
  28. Janneau-G-Khom-script-1870.16 · Gustave Janneau · Public domain

The film is shared under CC BY-SA 4.0.

What we work on

  • Models

    Small open models for Khmer, sized to run on an ordinary laptop, with the weights free to download and fine-tune.

    See the models

  • Datasets

    The Khmer data we build to train and test them, published so the next project doesn't have to start from zero.

    See the datasets

  • Benchmarks

    Test sets and scores that show where each model stands against the strongest alternative, including where it falls short.

    See the benchmarks

  • Tools

    Apps that put the models to work on your own computer, offline, with no account and nothing uploaded.

    See the tools

Latest research

All research

Paper/

Sota

A 204M-parameter speech recogniser for Khmer and English, and a 73M-parameter one, made by structurally pruning Meta's Omnilingual ASR to 80% encoder sparsity. The larger has a character error rate of 1.77% on the DDD Khmer test set, below its 975M-parameter teacher's 2.47%, at a quarter of the compute.

Parameters
73M / 204M
CER
1.77%

Paper/

Netra OCR 1

A 19M-parameter text-line recognizer trained only on synthetic images. It has the lowest character error rate of five systems on every test set, including vision-language models about 150 times its size.

Paper/

Netra MT 1

A 77M-parameter translation model trained from scratch on 220 million tokens of English–Khmer text. It beats open models up to 40 times its size on English→Khmer overlap metrics and runs 16 times faster than NLLB-600M on a CPU, but still trails them on Khmer→English.

Most AI research can’t be checked. Ours can

When data, code and weights stay private, no one outside can see what a model has learned, or build on it. We think research on how machines think should be something anyone can read, test and question.

Teach machines to think, and show all the work. We publish every model with its code, data, tests and an account of where it fails, so anyone can check the research and take it further. Our work is in Khmer, a language most AI has barely learned. About the lab

How we work

Our core values in full

  • You don't need an account, API key or credit card to use anything we make. Code, weights and reports are released under licences that allow commercial use, changes and redistribution. We won't lock the work away later.

Get involved

How to reach us

  • Contribute data

    Annotated Khmer text, scanned documents, or transcriptions are the scarcest resource in this whole space. Even small, well-labelled sets are useful.

  • Report where models fail

    Real documents that the models get wrong are worth more than another benchmark. Failure cases directly shape the next training run.

  • Build on the work

    Build something on top of the models and say what broke. Nothing surfaces gaps faster than someone trying to ship with them.