File size: 11,450 Bytes
1d62a78 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 |
---
license: apache-2.0
language:
- ar
- en
pipeline_tag: text-generation
tags:
- pytorch
library_name: transformers
base_model: QCRI/Fanar-1-9B
---
<p align="center">
<img src="./fanar_logo.jpg" width="200"/>
</p>
# Fanar-1-9B-Instruct
**Fanar-1-9B-Instruct** is a powerful Arabic-English LLM developed by [Qatar Computing Research Institute (QCRI)](https://www.hbku.edu.qa/en/qcri) at [Hamad Bin Khalifa University (HBKU)](https://www.hbku.edu.qa/), a member of Qatar Foundation for Education, Science, and Community Development. It is the instruction-tuned version of [Fanar-1-9B](https://huggingface.co/QCRI/Fanar-1-9B). We continually pretrain the `google/gemma-2-9b` model on 1T Arabic and English tokens. We pay particular attention to the richness of the Arabic language by supporting Modern Standard Arabic (MSA) and a diverse set of Arabic dialects, including Gulf, Levantine, and Egyptian. Fanar models, through meticulous curation of the pretraining and instruction-tuning data, are aligned with Islamic values and Arab cultures.
**Fanar-1-9B-Instruct** is a core component of the [Fanar GenAI platform](https://fanar.qa/) that offers a suite of capabilities including image generation, video and image understanding, deep thinking, advanced text-to-speech (TTS) and automatic-speech-recognition (ASR), attribution and fact-checking, Islamic RAG, among several other features.
We have published a comprehensive [report](https://arxiv.org/pdf/2501.13944) with all the details regarding our Fanar GenAI platform. We also provide an API to our models and the GenAI platform (request access [here](https://api.fanar.qa/request/en)).
---
## Model Details
| Attribute | Value |
|---------------------------|------------------------------------|
| Developed by | [QCRI](https://www.hbku.edu.qa/en/qcri) at [HBKU](https://www.hbku.edu.qa/) |
| Sponsored by | [Ministry of Communications and Information Technology, State of Qatar](https://www.mcit.gov.qa/en/)
| Model Type | Autoregressive Transformer |
| Parameter Count | 8.7 Billion |
| Context Length | 4096 Tokens |
| Input | Text only |
| Output | Text only |
| Training Framework | [LitGPT](https://github.com/Lightning-AI/litgpt) |
| Pretraining Token Count | 1 Trillion (ar + en) |
| SFT Instructions | 4.5M |
| DPO Preference Pairs | 250K |
| Languages | Arabic, English |
| License | Apache 2.0 |
<!-- | Precision | bfloat16 | -->
---
## Model Training
#### Pretraining
Fanar-1-9B-Instruct was continually pretrained on 1T tokens, with a balanced focus on Arabic and English: ~515B English tokens from a carefully curated subset of the [Dolma](https://huggingface.co/datasets/allenai/dolma) dataset, 410B Arabic tokens that we collected, parsed, and filtered from a variety of sources, and 102B code tokens curated from [The Stack](https://github.com/bigcode-project/the-stack-v2) dataset. Our codebase used the [LitGPT](https://github.com/Lightning-AI/litgpt) framework.
#### Post-training
Fanar-1-9B-Instruct underwent a two-phase post-training pipeline:
| Phase | Size |
|-------|------|
| Supervised Fine-tuning (SFT) | 4.5M Instructions |
| Direct Preference Optimization (DPO) | 250K Preference Pairs |
---
## Getting Started
Fanar-1-9B-Instruct is compatible with the Hugging Face `transformers` library (≥ v4.40.0). Here's how to load and use the model:
```python
from transformers import AutoTokenizer, AutoModelForCausalLM
model_name = "QCRI/Fanar-1-9B-Instruct"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto")
# message content may be in Arabic or English
messages = [
{"role": "user", "content": "ما هي عاصمة قطر؟"},
]
inputs = tokenizer.apply_chat_template(messages, tokenize=False, return_tensors="pt")
outputs = model.generate(**tokenizer(inputs, return_tensors="pt", return_token_type_ids=False), max_new_tokens=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```
Inference using VLLM is also supported:
```python
from vllm import LLM, SamplingParams
model_name = "QCRI/Fanar-1-9B-Instruct"
llm = LLM(model=model_name)
sampling_params = SamplingParams(temperature=0.7, max_tokens=256)
# message content may be in Arabic or English
messages = [
{"role": "user", "content": "ما هي عاصمة قطر؟"},
]
outputs = llm.chat(messages, sampling_params)
print(outputs[0].outputs[0].text)
```
---
## Intended Use
Fanar-1-9B-Instruct is built for:
- Conversational agents (Arabic only or bilingual)
- Cultural and dialectal question answering in Arabic
- Educational, governmental, and civic NLP applications focused on the Arab world or Arabic-speaking audiences
- Research on Arabic natural language generation and understanding
Fanar-1-9B-Instruct can be deployed as part of a broader AI system. Developers are encouraged to implement proper safeguards to ensure culturally respectful, accurate, and safe deployment. It should not be used to generate or spread **harmful, illegal, or misleading content.**
A version of this model can be accessed through [Fanar Chat](chat.fanar.qa). We are continuously improving the Fanar’s models and capabilities, and answers can differ from what you get from Fanar-1-9B-Instruct.
---
## Ethical Considerations & Limitations
Fanar-1-9B-Instruct is capable of generating fluent and contextually appropriate responses. However, as with any generative model there are uncertainities. The model may produce **biased, offensive, or incorrect outputs**. The model is **not suitable for high-stakes decision-making** (e.g., legal, medical, or financial advice). Though we have extensively tested Fanar-1-9B-Instruct and attempted to mitigate these issues, we cannot redress every possible scenario. Thus, we advise developers to implement safety checks and perform domain-specific fine-tuning for sensitive use cases. Kindly refer to our [Terms of Service]( https://chat.fanar.qa/terms-of-service) and [Privacy Policy](https://chat.fanar.qa/privacy-policy).
The output generated by this model is not considered a statement of QCRI, HBKU, Qatar Foundation, MCIT or any other organization or individual.
---
## Evaluation
Evaluation was conducted using a modified version of the LM Evaluation Harness and internal cultural alignment benchmarks.
<div style="overflow-x: auto;">
| Model | MMLU (5-shot) | MMMLU (Arabic) (0-shot) | ArabicMMLU (3-shot) | HellaSwag (0-shot) | PIQA (0-shot) | ARC Challenge (0-shot) | Belebele (Arabic) (3-shot) | ACVA (5-shot) | GSM8k | OALL (0-shot) | OALL v2 (0-shot) | Almieyar Arabic (3-shot) | Arab Cultural MCQ (3-shot) | AraDiCE PIQA (MSA) (0-shot) | AraDiCE PIQA(Egy) (0-shot) | AraDiCE PIQA(Lev) (0-shot) | AraDiCE ArabicMMLU(Egy) (0-shot) | AraDiCE ArabicMMLU(Lev) (0-shot) |
|-------|----------------|--------------------------|----------------------|--------------------|---------------|-------------------------|------------------------------|---------------|--------|----------------|------------------|---------------------------|-----------------------------|-------------------------------|------------------------------|------------------------------|-----------------------------------|-----------------------------------|
| Fanar-1-9B-it | 71.53% | **58.89%** | 67.69% | **83.16%** | **82.54%** | **67.15%** | **83.22%** | 80.02% | **74.60%** | **68.32%** | **66.29%** | **78.68%** | 72.40% | **67.68%** | **63.66%** | 59.03% | 59.63% | 60.62% |
| ALLaM-7B-Instruct-preview | 60.72% | 54.89% | **68.59%** | 76.35% | 80.52% | 51.62% | 75.80% | 74.52% | 46.63% | 57.31% | 63.66% | 76.31% | **74.20%** | 67.52% | 63.44% | **60.88%** | **62.50%** | **64.17%** |
| aya-expanse-8b | 62.85% | 47.14% | 60.10% | 78.54% | 81.18% | 56.40% | 70.78% | 77.11% | 8.26% | 53.18% | 59.74% | 70.20% | 67.30% | 63.00% | 59.41% | 56.53% | 53.52% | 53.71% |
| c4ai-command-r7b-arabic-02-2025 | 66.91% | 49.54% | 63.06% | 74.67% | 78.02% | 49.15% | 72.78% | 79.80% | 30.33% | 49.38% | 64.44% | 73.82% | 69.20% | 62.30% | 60.99% | 56.69% | 54.78% | 56.06% |
| AceGPT-v2-8B-Chat | 66.45% | 51.16% | 62.61% | 79.21% | 80.58% | 53.50% | 74.56% | 77.66% | 41.77% | 50.16% | 60.40% | 74.31% | 68.90% | 64.58% | 61.32% | 56.91% | 54.53% | 53.91% |
| gemma-2-9b-it | 71.65% | 57.93% | 64.16% | 79.06% | 79.38% | 63.99% | 78.31% | **80.67%** | 60.95% | 56.11% | 64.21% | 73.69% | 68.60% | 61.26% | 59.96% | 57.24% | 57.95% | 59.25% |
| jais-adapted-13b-chat | 56.64% | 44.45% | 58.97% | 80.86% | 80.47% | 54.27% | 67.52% | 75.24% | 44.05% | 46.41% | 56.56% | 65.46% | 65.30% | 61.10% | 58.05% | 55.77% | 52.87% | 53.59% |
| jais-family-6p7b-chat | 49.42% | 41.59% | 55.80% | 72.04% | 74.05% | 44.62% | 65.11% | 72.04% | 53.68% | 48.20% | 54.73% | 61.72% | 64.10% | 62.51% | 60.12% | 57.24% | 49.11% | 47.49% |
| Llama-3.1-8B-Instruct | 68.04% | 47.58% | 59.05% | 79.22% | 80.74% | 55.29% | 66.72% | 76.67% | 29.26% | 47.81% | 55.97% | 69.70% | 66.10% | 58.11% | 55.39% | 54.24% | 46.86% | 47.52% |
| Qwen2.5-7B-Instruct | **74.21%** | 55.63% | 63.96% | 80.44% | 79.92% | 55.03% | 74.61% | 78.09% | 71.34% | 54.19% | 62.69% | 75.69% | 68.10% | 60.55% | 58.65% | 56.04% | 48.74% | 53.42% |
</div>
---
## Citation
If you use Fanar-1-9B-Instruct or the Fanar GenAI system in your research or applications, please cite:
```bibtex
@misc{fanarllm2025,
title={Fanar: An Arabic-Centric Multimodal Generative AI Platform},
author={Fanar Team and Ummar Abbas and Mohammad Shahmeer Ahmad and Firoj Alam and Enes Altinisik and Ehsannedin Asgari and Yazan Boshmaf and Sabri Boughorbel and Sanjay Chawla and Shammur Chowdhury and Fahim Dalvi and Kareem Darwish and Nadir Durrani and Mohamed Elfeky and Ahmed Elmagarmid and Mohamed Eltabakh and Masoomali Fatehkia and Anastasios Fragkopoulos and Maram Hasanain and Majd Hawasly and Mus'ab Husaini and Soon-Gyo Jung and Ji Kim Lucas and Walid Magdy and Safa Messaoud and Abubakr Mohamed and Tasnim Mohiuddin and Basel Mousi and Hamdy Mubarak and Ahmad Musleh and Zan Naeem and Mourad Ouzzani and Dorde Popovic and Amin Sadeghi and Husrev Taha Sencar and Mohammed Shinoy and Omar Sinan and Yifan Zhang and Ahmed Ali and Yassine El Kheir and Xiaosong Ma and Chaoyi Ruan}},
year={2025},
url={https://arxiv.org/abs/2501.13944},
}
```
---
## Acknowledgements
This project is from [Qatar Computing Research Institute (QCRI)](https://qcri.org) at [Hamad Bin Khalifa University (HBKU)](https://hbku.edu.qa), a member of Qatar Foundation. We thank our engineers, researchers, and support team for their efforts in advancing Arabic-centric large language models.
Special thanks to the [Ministry of Communications and Information Technology, State of Qatar](https://www.mcit.gov.qa/en/) for their continued support by providing the compute infrastructure through the Google Cloud Platform.
---
## License
This model is licensed under the [Apache 2.0 License](https://www.apache.org/licenses/LICENSE-2.0). |