from transformers import AutoModelForCausalLM, AutoTokenizer

#تحميل النموذج
model_id = "Etherll/Tashkeel-350M"
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': 'system','content': "أنت نموذج لتشكيل النصوص العربية."},
      {"role": "user", "content": 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 (English)

A 350M parameter model for Arabic diacritization (Tashkeel). This model is a fine-tune of LiquidAI/LFM2-350M on the arbml/tashkeela dataset.

Base Model: LiquidAI/LFM2-350M
Dataset: arbml/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.