from transformers import AutoModelForCausalLM, AutoTokenizer

#تحميل النموذج
model_id = "Etherll/Tashkeel-350M-v2"
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
    torch_dtype="bfloat16",
)
tokenizer = AutoTokenizer.from_pretrained(model_id)

# إضافة التشكيل
prompt = "السلام عليكم" 
input_ids = tokenizer.apply_chat_template(
    [{"role": "user", "content": "قم بتشكيل هذا النص "+ ":\n"+prompt}],
    add_generation_prompt=True,
    return_tensors="pt",
    tokenize=True,
).to(model.device)

output = model.generate(
    input_ids,
    do_sample=False,  
)

print(tokenizer.decode(output[0, input_ids.shape[-1]:], skip_special_tokens=True))

مثال

النص المدخل: السلام عليكم
الناتج: اَلسَلَامُ عَلَيْكُمْ

Tashkeel-350M-v2 (English)

A 350M parameter model for Arabic diacritization (Tashkeel). This model is a fine-tune of ibm-granite/granite-4.0-h-350m on the Misraj/Sadeed_Tashkeela dataset.

Base Model: ibm-granite/granite-4.0-h-350m
Dataset: Misraj/Sadeed_Tashkeela

How to Use

The Python code for usage is the same as listed in the Arabic section above.

Example

Input: السلام عليكم
Output: اَلسَلَامُ عَلَيْكُمْ

This lfm2 model was trained 2x faster with Unsloth and Huggingface's TRL library.