Org card: models by what they do, with numbers and limits

README.md

@@ -7,11 +7,38 @@ sdk: static
 pinned: false
 ---
 
-**Netra Lab** is an independent, non-commercial research lab building open AI for the Khmer language.
+# Netra Lab
 
-More than 16 million people speak Khmer, and most AI still treats it as an afterthought. There is little public data to train on, and tools made for English stumble on a script where consonants stack and words run together without spaces.
+Open AI for the Khmer language, built in the open.
 
-We build small models that run on the laptops and phones people already own, the datasets to train and test them, and a written report for every release, including where the model falls short. Code, weights and data are free, under licenses that allow commercial use. If someone trains something better on top of them, that is the point.
+Netra Lab is an independent, non-commercial research lab. More than 16 million people speak Khmer, and most AI still treats it as an afterthought: there is little public data to train on, and tools made for English stumble on a script where consonants stack and words run together without spaces. We build small models for it that run on the laptops and phones people already own, and publish the code, weights, data and a written report for each one, including where it falls short.
 
+## Models
 
-The reports are on [our website](https://netra-lab.netlify.app/research), the code is on [GitHub](https://github.com/netra-ai-lab).
+### Reading
+
+**[Netra OCR 1](https://huggingface.co/Netra-Lab/Netra-OCR)** reads printed Khmer and English. A 19M-parameter text-line recognizer trained only on synthetic images, it has a 0.96% character error rate on the KHOB benchmark, against 5.12% for Tesseract 5. It comes with line and layout detection, PDF input and an editor that runs in the browser.<br>
+`pip install netra-ocr` · [GitHub](https://github.com/netra-ai-lab/Netra-OCR) · [Report](https://netra-lab.netlify.app/research/netra-ocr)
+
+### Translation
+
+**[Netra MT 1](https://huggingface.co/Netra-Lab/Netra-NMT)** translates English↔Khmer, both directions in one 77M-parameter model trained from scratch on 220M tokens. On English→Khmer it scores 25.8 spBLEU on ALT, second only to Google Translate and ahead of NLLB-200 1.3B. On Khmer→English it still trails the larger models. About 40 ms per sentence on a desktop CPU.<br>
+`pip install netra-nmt` · [GitHub](https://github.com/netra-ai-lab/Netra-NMT) · [Demo](https://huggingface.co/spaces/Darayut/Netra-NMT) · [Report](https://netra-lab.netlify.app/research/netra-mt)
+
+### Speech
+
+**Sota** transcribes Khmer and English speech, in two sizes pruned from Meta's 975M-parameter Omnilingual ASR. The 204M model has a 1.77% character error rate on the DDD Khmer test set, below its teacher's 2.47%, at a quarter of the compute.<br>
+[sota-large](https://huggingface.co/Darayut/sota-large) · [sota-base](https://huggingface.co/Darayut/sota-base) · [GitHub](https://github.com/netra-ai-lab/SOTA) · [Report](https://netra-lab.netlify.app/research/sota)
+
+### Typing
+
+**AkaraAlpha** turns Khmer typed in the Latin alphabet back into Khmer script. With 78.7K parameters, it has a lower character error rate than a Transformer three times its size, and with a dictionary check its first suggestion is right 74% of the time.<br>
+[GitHub](https://github.com/netra-ai-lab/Romanized-Khmer-to-Khmer-Transliteration) · [Report](https://netra-lab.netlify.app/research/akara-alpha)
+
+The training corpora and test sets we built for these models are listed below with the datasets.
+
+## The name
+
+នេត្រ (*netra*) is Khmer for "eye". It comes from the Cat's Eye Nebula, which resolves into the shape of an eye, and our symbol is that eye, drawn pixel by pixel, the way an OCR model reads a page.
+
+[Website](https://netra-lab.netlify.app/) · [GitHub](https://github.com/netra-ai-lab) · [Get in touch](https://netra-lab.netlify.app/contact)