diacnet-mini-2.0
diacnet-2.0’s features at half the size.
diacnet-mini-2.0 is the compact version of diacnet-2.0. It covers the same 11 languages, including Arabic, with the same meaning hints, partial-input support and alignment, in a 300M model that is cheaper to serve.
The problem it solves
Not every pipeline needs the most accurate model. If you are restoring marks on a large backlog, or serving restoration inside a busy product, half the parameters means half the hardware.
diacnet-mini-2.0 gives up very little to get there: it is close to the full model everywhere and even beats diactag-2.0 on five of the European languages.
The numbers
diacbench, 1,000 sentences per language
diacnet-mini-2.0 (orange) sits close to diacnet-2.0 on every language and well ahead of diacnet-1.1, at half the parameters.
Meaning hints
Hints work at the smaller size too: Vietnamese ambiguous-word accuracy rises from 0.53 to 0.78, Polish from 0.89 to 0.98.
Highlights
Half the size, most of the accuracy
Mean error rate 0.0141 against 0.0117 for diacnet-2.0, at 300M parameters instead of 582M.
Strong on European languages
Lower error than diactag-2.0 on Polish, Portuguese, Spanish, French and Italian, and level on Hausa.
Everything the full model does
Arabic in both styles, automatic language detection, meaning hints, and partial input where existing marks are kept.
Safe output
The same alignment step as diacnet-2.0 keeps every letter you typed and takes only the marks.
Run it in a few lines
from olaverse.nlp import Diacritizer
d = Diacritizer(model="diacnet-mini-2.0", lang="yor", device="auto")
d.restore("O so fun ara re pe oun ko ni isoro kankan.")
# → 'Ó sọ fún ara rẹ̀ pé òun kò ní ìṣòro kankan.'
# One call for a mixed batch
d.restore_batch(texts, lang=["yor", "vie", "auto"])Which DiacNet model?
Need the best accuracy? Use diacnet-2.0. Need CPU-only and letter-safe? Use diactag-2.0.
| diacnet-2.0 → | diacnet-mini-2.0THIS PAGE | diactag-2.0 → | |
|---|---|---|---|
| Parameters | 582M | 300M | 37.9M |
| How it works | Rewrites the text | Rewrites the text | Tags each character |
| Mean error rate, 10 languages | 0.0117 | 0.0141 | 0.0130 |
| Yorùbá error rate | 0.0698 | 0.0815 | 0.0787 |
| Strongest on | Yorùbá and European languages | Close to 2.0 at half the size | Igbo, Vietnamese, Modern Standard Arabic |
| Extras | Meaning hints, typo correction | Meaning hints, typo correction | CPU-ready ONNX, never alters letters |
Best for
- Large backlogs and busy products where serving cost matters.
- Polish, Portuguese, Spanish, French and Italian, where it outperforms diactag-2.0.
- Teams who want diacnet-2.0’s features on smaller GPUs.
More production models
All modelsPurpleMIST-Flash-1.0
typed-decisions 0.720 · African TD 0.668
PurpleMISTPurpleMIST-Mini-1.0
typed-decisions 0.601 · 1.9B
MIST-ExtractMIST-Extract-1.0
91.8% docs fully right · 4B
MIST-ExtractMIST-Extract-Flash-1.0
9.7 docs/s · 2B
LID-608lid-neural-608
0.867 acc · 140M
LID-608lid-lite-608
37 MB · 6,800 texts/s
MIST QGmist-qg-1.5b
0.955 keep-rate · 25 langs