The model retains the strong multilingual and reasoning capabilities of the original LLaMA 3.3 70B Instruct while improving Arabic fluency, comprehension, and alignment.

Base Model

Base model: meta-llama/Llama-3.3-70B-Instruct
Architecture: Decoder-only Transformer
Model size: 70 billion parameters
Context length: Same as the base model

Training Details

Fine-tuning type: LoRA / PEFT
Training data: Curated Arabic instruction-following datasets
Languages: Arabic (primary), English (secondary)
Objective: Improve Arabic instruction-following, reasoning, and generation quality

No additional pretraining was performed. The model was fine-tuned starting from the original LLaMA 3.3 70B Instruct checkpoint.

Intended Use

This model is intended for:

Arabic conversational agents
Arabic question answering
Instruction-following tasks
Content generation in Arabic
Research and experimentation with Arabic LLMs

Limitations

The model may still produce incorrect or biased outputs.
Performance may vary across Arabic dialects.
Not suitable for high-risk or safety-critical applications without further evaluation.

Ethical Considerations

This model inherits limitations and potential biases from both the base model and the fine-tuning data. Users should apply appropriate safety measures and human oversight when deploying the model.

Usage Example

python

from transformers import AutoTokenizer, AutoModelForCausalLM

model_name = "Lina-Z/arabic_llama_model"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    device_map="auto",
    torch_dtype="auto"
)

prompt = "اشرح مفهوم الذكاء الاصطناعي باختصار."
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

outputs = model.generate(**inputs, max_new_tokens=200)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))