AI Models
DiacTag 2.0
DiacNet family · open weights on Hugging Face.
Restores diacritics and tone marks across 10 languages by classifying each character rather than generating text, so the output is your input with marks added and nothing else. 37.6M parameters, CPU-native.
OnnxDiactagDiacriticsDiacritic RestorationTashkeelToken ClassificationLow ResourceAfrican LanguagesArabicYorubaIgboHausaViPlTrPtEsFrItArLicence: apache-2.0
0
Likes
6
Downloads
Oct 2026
Created
/ Model card
Built for production use
Open-weights repository on Hugging Face
Integrated ecosystem protocol tier.
Compatible with Transformers library
Integrated ecosystem protocol tier.
Optimised for low-latency inference
Integrated ecosystem protocol tier.
Community engagement: 0 likes
Integrated ecosystem protocol tier.
/ Quick start
Use DiacTag 2.0, straight from its model card.
example.pyfrom the model card
import sys, json
from huggingface_hub import snapshot_download
d = snapshot_download("olaverse/diactag-2.0", allow_patterns=["code/*", "*.json", "ckpt_final.pt"])
sys.path.insert(0, f"{d}/code")
from diactag.labels import LabelSpace
from diactag.model import DiacTagger
from diactag.infer import Diacritizer, InferConfig
ls = LabelSpace.load(f"{d}/labels.json")
model, _ = DiacTagger.load(f"{d}/ckpt_final.pt")
T = json.load(open(f"{d}/calibration.json"))["shared"]
dz = Diacritizer(model, ls, InferConfig(temperature=T))
dz.restore("se eranko naa si gbo o?", "yor") # 'ṣé ẹranko náà sì gbọ́ ọ?'
dz.restore("Toi khong biet tieng Viet", "vie") # 'Tôi không biết tiếng Việt'
dz.restore("ذهب الطالب إلى المدرسة في الصباح", "ara") # 'ذَهَبَ الطَّالِبُ إلَى الْمَدْرَسَةِ فِي الصَّبَاحِ'
dz.detect_language("Lodz jest piekna") # ('pol', 0.99999...)
/ Model card
Official repository README
Dynamically loaded from Hugging Face
Model card metadata is available on Hugging Face.
/ Built with
Transformers PyTorch Python
Ready to try DiacTag 2.0?
Restores diacritics and tone marks across 10 languages by classifying each character rather than generating text, so the output is your input with marks added and nothing else. 37.6M parameters, CPU-native.
/ Model ecosystem