About
We are building an Arabic checker that is small and accurate, and we tell the truth about its numbers.
Tashih is a research project. It began with a simple question: does correcting Arabic text need a giant model? Our answer so far: no.
Principles
Three things we do not give up.
Small on purpose
We believe one well-defined task does not need a model with billions of parameters. Our models run on an ordinary machine.
Honest measurement
When we found test text inside the training data, we removed it and retrained. A number that cannot be reproduced is a number we do not publish.
Careful with edits
A good editor knows when to stay quiet. We measure what the model breaks as carefully as what it fixes.
What we are working on
Reused layers
We are testing an encoder that reuses its own layers. In a first experiment it came close to the full model on QALB text with about 39% of the encoder's parameters, and still trails on dialect and grammar.
Safer rewriting
Rewriting changes the meaning in about one sentence in eight. We are testing ways to check the meaning before showing a result.
Rare dialect words
We are collecting examples of the words the model gets wrong, such as words that look like standard Arabic but mean something else.