Tashkeel-v3 (Gyoom copy)

Gyoom's copy of ahmedsamirtarjama/Tashkeel-v3 at revision 98b3985c, the one Hammel AI TTS is tested on. Upstream has replaced its weights since, so this copy keeps that version available.

Unchanged: every other file is byte for byte upstream's at that revision. Of its two weight files only model.safetensors is kept; pytorch_model.bin is left out.
Added: this header and LICENSE, the Apache-2.0 text upstream's card names.
Inference needs model.safetensors, tokenizer.json and tokenizer_config.json. Load the weights in fp32: transformers 5 defaults to bf16, which changes the output.

hf download gyoom-sa/Tashkeel-v3 model.safetensors tokenizer.json tokenizer_config.json \
  --local-dir models/tashkeel-v3

The model card below is upstream's, unchanged.

Tashkeel-v3: SOTA Arabic Diacritizer (Morpheme-Aware MARBERTv2 + Linear-Chain CRF)

Tashkeel-v3 is a state-of-the-art neural Arabic diacritization model achieving 2.85% DER / 7.91% WER on the Misraj/SadeedDiac-25 benchmark and 6.37% DER / 17.39% WER on Bisher/CATT_benchmark.

🌟 Key Features

Linear-Chain Conditional Random Field (CRF) & Viterbi Decoding:
- Integrates a global transition matrix $\mathbf{A}_{15 \times 15}$ that penalizes phonologically illegal transitions (e.g. consecutive Tanweens, improper geminations, illegal boundary vowels).
- Global Viterbi Decoding searches the sentence space for the globally optimal diacritization sequence.
Decoupled Multi-Task Heads (Lexical Stem vs. Terminal I'rab):
- Dedicated heads for word-interior morphological diacritics and terminal syntactic case endings.
Morpheme & Clitic Disambiguation:
- Character-level sub-token projection with proclitic/stem/enclitic segmentation embeddings.
Zero Hallucinations (0.00%):
- Every input consonant, digit, and punctuation mark is preserved 100% identically without any dropped tokens or runaway repetitions.
High Inference Speed:
- ~1,164 to 1,523 sentences per second on a single NVIDIA GPU.

📊 Benchmark Evaluation on `Misraj/SadeedDiac-25` (1,200 sentences)

Model / System	Total Scored	Hallucination Rate	Total DER (CE) ↓	Morph DER (w/o CE) ↓	Total WER (CE) ↓	Morph WER (w/o CE) ↓	Inference Speed
Claude-3.7-Sonnet	1,190 / 1,200	0.82%	1.39%	0.77%	4.67%	2.31%	~5 sent/s
Tashkeel-v3 (This Model)	1,200 / 1,200	0.00%	2.85%	2.13%	7.91% 🏆	5.54% 🏆	~1,164 sent/s ⚡
Tashkeel-v2	1,200 / 1,200	0.00%	2.85%	2.13%	8.02%	5.66%	~898 sent/s
Gemini-Flash-2.0	1,186 / 1,200	1.17%	3.19%	2.38%	7.99%	5.50%	~20 sent/s
Tashkeel-50M (v1)	12 / 1,200*	~99.0%*	3.08%*	2.26%*	9.56%*	6.77%*	~25 sent/s
GPT-4	1,188 / 1,200	1.02%	3.86%	3.86%	5.27%	10.93%	~10 sent/s
Sadeed (Published Paper)	1,114 / 1,200	7.19%	7.29%	5.26%	13.74%	9.92%	~35 sent/s
tash2 (50M Baseline)	1,199 / 1,200	0.08%	15.29%	11.41%	24.36%	20.40%	~25 sent/s

* Note on Causal LLMs: Scores calculated only on the subset of non-hallucinated generations.

📈 Out-of-Domain Generalization on `Bisher/CATT_benchmark` (742 sentences)

Model / System	Total Scored	Hallucination Rate	Total DER (CE) ↓	Morph DER (w/o CE) ↓	Total WER (CE) ↓	Morph WER (w/o CE) ↓	Inference Speed
Tashkeel-v3 (This Model)	742 / 742	0.00%	6.37% 🏆	5.33% 🏆	17.39% 🏆	14.84% 🏆	~1,523 sent/s ⚡
Tashkeel-v2	742 / 742	0.00%	6.41%	5.35%	17.66%	15.01%	~1,470 sent/s

🚀 Quickstart & Usage

1. Installation

pip install torch transformers pyarabic

2. Python Inference Code

import torch
from transformers import AutoModel, AutoTokenizer

# Load Tashkeel-v3 directly from Hugging Face Hub
model_id = "ahmedsamirtarjama/Tashkeel-v3"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModel.from_pretrained(model_id, trust_remote_code=True).cuda()

# Single sentence diacritization
text = "ذهب الطالب المجتهد إلى المدرسة صباحا"
output = model.diacritize(text, tokenizer=tokenizer)
print(output)
# Output: ذَهَبَ الطَّالِبُ الْمُجْتَهِدُ إِلَى الْمَدْرَسَةِ صَبَاحًا

# Batched inference
sentences = [
    "سطا لصوص على منزل وزيرة الدفاع التشيلية",
    "إن مع العسر يسرا",
    "العلم نور والجهل ظلام"
]
outputs = model.diacritize(sentences, tokenizer=tokenizer, batch_size=64)
for inp, out in zip(sentences, outputs):
    print(f"{inp} -> {out}")

🏗️ Model Architecture

                                [Input Arabic Text]
                                         │
                             ┌───────────┴───────────┐
                             ▼                       ▼
                    [Shared MARBERTv2]     [Morpheme + Position Embeddings]
                             │                       │
                             └───────────┬───────────┘
                                         ▼
                             [Bidirectional LSTM (384x2)]
                                         │
                             ┌───────────┴───────────┐
                             ▼                       ▼
                       [Stem Head]             [I'rab Head]
                       (Internal Diac)         (Terminal Case)
                             │                       │
                             └───────────┬───────────┘
                                         ▼
                           [Combined Emission Logits]
                                         │
                                         ▼
                       [Linear-Chain CRF Transition Matrix (A)]
                                         │
                                         ▼
                           [Global Viterbi Decoding (y*)]

📜 Citation

@misc{samir2026tashkeelv3,
  author = {Ahmed Samir},
  title = {Tashkeel-v3: SOTA Arabic Diacritization with Morpheme-Aware MARBERT and Linear-Chain CRF},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/ahmedsamirtarjama/Tashkeel-v3}}
}

Downloads last month: 33

Safetensors

Model size

0.2B params

Tensor type

F32