README.md

---
library_name: transformers
license: apache-2.0
pipeline_tag: text-generation
language:
- ar
- en
---

<p align="center">
  <picture>
    <!-- Dark mode -->
    <source media="(prefers-color-scheme: dark)" srcset="https://cdn-uploads.huggingface.co/production/uploads/65604648d69284e31fed02b0/iDADKNbWL17MTB-bB34gV.png">
    <!-- Light mode -->
    <source media="(prefers-color-scheme: light)" srcset="https://cdn-uploads.huggingface.co/production/uploads/65604648d69284e31fed02b0/AO9bjIkbM0zFs67oE3-it.png">
    <!-- Fallback -->
    <img src="https://cdn-uploads.huggingface.co/production/uploads/65604648d69284e31fed02b0/AO9bjIkbM0zFs67oE3-it.png" alt="Jais2 Logo" width="400">
  </picture>
</p>

# Jais-2: The Next Generation of Arabic Frontier LLMs

## Model Overview
Jais-2-70B-Chat is a high-capacity bilingual Arabic–English language model developed by MBZUAI, Inception, and Cerebras. 
Jais-2-70B-Chat Model is trained from scratch on Arabic and English data and is powered by a custom Arabic-centric vocabulary, it efficiently captures Modern Standard Arabic, regional dialects, and mixed Arabic–English code-switching. 
The model is openly available under a Apache 2.0 license and also deployed as a fast, production-ready chat experience running on Cerebras hardware. 
Visit the [Jais-2 Web App](https://jaischat.ai).

## Key Technical Specifications
- **Model Developers**: MBZUAI, Inception, Cerebras.
- **Languages**: Arabic (MSA & dialects) and English
- **Architecture**: Transformer-based, Decoder-only architecture with multi-head self-attention.
- **Parameters**: 70 Billion
- **Context Length**: 8,192
- **Vocabulary Size**: 150,272
- **Training Infrastructure**: Optimized for Cerebras CS-2 and Condor Galaxy clusters
- **Key Design Choices**: Rotary Position Embeddings (RoPE), Squared-ReLU activation, custom μP parameterization, and 8:1 filter-to-hidden size ratio.

---


## How to Use the Model

# Using Transformers
### 1. Clone the Jais 2–compatible Transformers fork

```bash
# Pending PR merge to the official package
git clone --branch jais2 --single-branch \
    https://github.com/inceptionai-abudhabi/transformers.git
cd transformers
uv pip install -e .
```
### 2. Load and Inference on the Model
```python
from transformers import AutoTokenizer, AutoModelForCausalLM

# Load the model and tokenizer
model_name = "inceptionai/Jais-2-70B-Chat"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto")

# Example Arabic prompt
system_prompt = "أجب باللغة العربية بطريقة رسمية وواضحة."
user_input = "ما هي عاصمة الإمارات؟"

# Apply chat template (always)
chat_text = tokenizer.apply_chat_template(
    [
        {"role": "system", "content": system_prompt},
        {"role": "user", "content": user_input}
    ],
    tokenize=False,
    add_generation_prompt=True
)

# Tokenize and generate
inputs = tokenizer(chat_text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=8192, temperature=0)

# Decode and print
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
#عاصمة الإمارات العربية المتحدة هي أبوظبي.
```
# Using vLLM

### 1. Clone the Jais 2–compatible vLLM fork

```bash
# Pending PR merge to the official package
git clone --branch jais2 --single-branch \
    https://github.com/inceptionai-abudhabi/vllm.git
cd vllm
uv pip install -e . # If you install vllm after transformers, please re-install transformers again from this branch: https://github.com/inceptionai-abudhabi/transformers.git
```

### 2. Load and Inference on the Model
```python
from vllm import LLM, SamplingParams

# Load model and tokenizer
model_name = "inceptionai/Jais-2-70B-Chat"
llm = LLM(model=model_name, tensor_parallel_size=1)
tokenizer = llm.get_tokenizer()

# Example Arabic prompt
system_prompt = "أجب باللغة العربية بطريقة رسمية وواضحة."
user_input = "ما هي عاصمة الإمارات؟"

# Apply chat template (always)
chat_text = tokenizer.apply_chat_template(
    [
        {"role": "system", "content": system_prompt},
        {"role": "user", "content": user_input}
    ],
    tokenize=False,
    add_generation_prompt=True
)

# Run generation
sampling_params = SamplingParams(max_tokens=8192, temperature=0)
outputs = llm.generate([chat_text], sampling_params)

#Print output
print(outputs[0].outputs[0].text)
#عاصمة الإمارات العربية المتحدة هي أبوظبي.
```

Or serve through command line (CLI)

```shell
vllm serve inceptionai/Jais-2-70B-Chat \
    --served-model-name inceptionai/Jais-2-70B-Chat-Local --dtype bfloat16 \
    --tensor-parallel-size 2 --max-model-len 8192 --max-num-seqs 256 \
    --host 0.0.0.0 --port 8042 --api-key "Optional"
```
---
## Evaluation
### Performance Overview
We evaluate **Jais-2-70B** across two key benchmarks that capture both *instruction following* and *generative* Arabic ability: **IFEval** (English and Arabic) and **AraGen-12-24 (3C3H)**.  

### IFEval Results (Strict 0-shot)

| Model Name | En-Strict-Prompt-lvl | En-Strict-Instruction-lvl | Ar-Strict-Prompt-lvl | Ar-Strict-Instruction-lvl |
|------------|-----------------------|----------------------------|------------------------|----------------------------|
| **Qwen2.5-72B-Instruct** | 83.53 | 88.51 | **67.33** | **74.05** |
| **Llama-3.3-70B-Instruct** | **88.20** | **92.10** | 58.17 | 63.13 |
| **Jais-2-70B (ours)** | 70.78 | 78.93 | 66.58 | 74.53 |

---

### AraGen-12-24 (3C3H) Results

| Model Name | 3C3H Score (%) | Correctness | Completeness | Conciseness | Helpfulness | Honesty | Harmlessness |
|------------|----------------|-------------|--------------|-------------|-------------|---------|-------------- |
| **Qwen2.5-72B-Instruct** | 62.58 | 71.92 | 71.80 | 19.06 | 69.86 | 70.94 | 71.92 |
| **Llama-3.3-70B-Instruct** | 61.29 | 68.58 | 65.11 | **34.50** | 63.50 | 67.47 | 68.58 |
| **Jais-2-70B (ours)** | **70.71** | **80.53** | **79.09** | 25.48 | **78.43** | **80.23** | **80.53** |


Overall, our results show that:  
- Jais-2-70B delivers competitive Arabic and English instruction-following performance across IFEval metrics.  
- Jais-2-70B achieves the highest scores across nearly all AraGen metrics, outperforming Qwen2.5-72B and Llama-3.3-70B on Arabic generative tasks.

---

## Intended Use

### Target Audiences
- **Academics**: Researchers focusing on Arabic NLP, multilingual modeling, or cultural alignment
- **Businesses**: Companies targeting Arabic-speaking markets
- **Developers and ML Engineers**: Integrating Arabic language capabilities into applications and workflows

### Appropriate Use Cases
- **Research**:
  - Natural language understanding and generation tasks
  - Conducting interpretability or cross-lingual alignment analyses
  - Investigating Arabic linguistic or cultural patterns
    
- **Commercial Use**:
  - Building chat assistants for Arabic-speaking audiences
  - Performing sentiment and market analysis in regional contexts
  - Summarizing or processing bilingual Arabic–English documents
  - Creating culturally resonant Arabic marketing and entertainment content for regional audiences

### Inappropriate Use Cases
- **Harmful or Malicious Use**:
  - Producing hate speech, extremist content, or discriminatory language
  - Creating or spreading misinformation or deceptive content
  - Engaging in or promoting illegal activities
    
- **Sensitive Information**:
  - Handling or generating personal, confidential, or sensitive information
  - Attempting to infer, reconstruct, or guess sensitive information about individuals or organizations
    
- **Language Limitations**:
  - Applications requiring strong performance outside Arabic or English languages
 
- **High-Stakes Decisions**:
  - Making medical, legal, financial, or safety-critical decisions without human oversight

## Citation

If you find our work helpful, please give us a cite.

``` 
@techreport{jais2_2025,
  title        = {Jais 2: {A} Family of {A}rabic-Centric Open Large Language Models},
  author       = {
    Anwar, Mohamed and
    Freihat, Abdelhakim and
    Ibrahim, George and
    Awad, Mostafa and
    Sadallah, Abdelrahman Atef Mohamed Ali and
    Gosal, Gurpreet and
    Ramakrishnan, Gokul and
    Hestness, Joel and
    Mishra, Biswajit and
    Joshi, Rituraj and 
    Chandran, Sarath and
    Frikha, Ahmed and
    Goffinet, Etienne and
    Maiti, Abhishek and
    El Filali, Ali and
    Al Barri, Sarah and
    Ghosh, Samujjwal and
    Pal, Rahul and
    Mullah, Parvez and
    Shukla, Awantika and
    Siddiki, Sajid and
    Kamboj, Samta and
    Pandit, Onkar and
    Sahu, Sunil and
    El Badawy, Abelrahman and
    Mohamed, Amr and
    Chamma, Ahmad and
    Dufraisse, Evan and
    Bounhar, Abdelaziz and
    Bouch, Dani and
    Abdine, Hadi and
    Shang, Guokan and
    Koto, Fajri and
    Wang, Yuxia and
    Xie, Zhuohan and
    Mekky, Ali and
    Elbadry, Rania Hossam Elmohamady and
    Ahmad, Sarfraz and
    Ahsan, Momina and
    El-Herraoui, Omar Emad Mohamed and
    Orel, Daniil and
    Iqbal, Hasan and
    Elzeky, Kareem Mohamed Naguib Abdelmohsen Fahmy and
    Abassy, Mervat and
    Ali, Kareem and
    Eletter, Saadeldine and
    Atif, Farah and
    Mukhituly, Nurdaulet and
    Li, Haonan and
    Han, Xudong and
    Singh, Aaryamonvikram and
    Quraishi, Zain and
    Sengupta, Neha and
    Murray, Larry and
    Sheinin, Avraham and
    Vassilieva, Natalia and
    Ren, Hector and
    Liu, Zhengzhong and
    Vazirgiannis, Michalis and
    Nakov, Preslav
  },
  institution  = {IFM},
  type         = {Technical Report},
  year         = {2025},
  month        = dec,
  day          = {09},
}
```