Latest
Tashih v1
Our most accurate model. It fixes spelling, grammar and punctuation, turns dialect into standard Arabic, and rewrites text.
Try it now15.1
Word error rate on 78 test paragraphs
78.7
On the official QALB‑2014 test (F0.5)
1.5 s
For a 46-word paragraph on CPU
What's new
What changed.
Learner Arabic
It trained on thousands of texts written by learners of Arabic, and now removes about 86% of their errors.
Twenty-two dialects
From Moroccan darija to Omani, including words that look like standard Arabic but mean something else.
Rewriting
Two new modes reword your text while keeping its meaning, with a separate model that leaves correction untouched.
Measurement
It removes more than half the errors.
Word error rate on 78 test paragraphs; lower is better.
By error type
Strongest at spelling and grammar.
Share of errors left after correction, on held-out test text for each type; lower is better.
Against published work
Where it stands among other systems.
These are the best published results on the official QALB‑2014 benchmark, the test Arabic correction systems are compared on.
| System | Size | Dev | Test |
|---|---|---|---|
| Tashih v1 | 367M | 78.91 | 78.73 |
| Tashih v1 — QALB-only variant | 367M | 80.98 | 80.77 |
| ArbESC+ (9-model ensemble) | 9 models | — | 82.63 |
| 4-model ensemble with GPT‑4o | 900M+ | 81.8 | 81.7 |
| SWEET (two-model cascade) | 270M | 80.3 | 80.5 |
| AraT5 (single model) | 367M | — | 80.09 |
| Seq2Seq++ | 502M | 79.6 | 79.6 |
| GPT‑4o | — | — | 77.7 |
The metric is F0.5 (higher is better) with the official M2 scorer, punctuation included. Other systems' numbers come from their papers. We scored the test set once, after both systems were frozen. The QALB-only variant beats every published single system; Tashih v1 is trained on much broader text (learners, dialects, many error types), so it gives up two points here to be far better on real-world writing.
On our own test
And against an open model that does the same job.
We ran Gemma‑1B, an open one-billion-parameter Arabic correction model, on the same test paragraphs.
Word error rate on 78 test paragraphs; lower is better.
Specifications
Small enough to run where you are.
- Parameters
- 367 million
- Architecture
- Encoder–decoder (T5)
- Context window
- 256 tokens per segment
- Size on disk
- 370 MB (int8)
- Runs on
- CPU or GPU