# JAIS Intiative: Atlas-Chat Models ## Model Overview Atlas-Chat is a family of open models instruction-tuned for Darija, the colloquial Arabic of Morocco, developed as part of the [Jais](https://arxiv.org/abs/2308.16149) project for standard Arabic and its extentions to dialectal Arabic. These models are designed for language generation and excel in various applications such as question answering, summarization, and translation. Thanks to their compact size, Atlas-Chat models can be deployed in resource-constrained environments like laptops, desktops, or personal cloud setups, making advanced AI accessible to Darija speakers and promoting widespread innovation. Three sizes are available: * [Atlas-Chat-2B](https://huggingface.co/MBZUAI-Paris/Atlas-Chat-2B): A small-sized version with 2 billion parameters, capable of generating fluent Moroccan Darija text while maintaining efficiency. * [Atlas-Chat-9B](https://huggingface.co/MBZUAI-Paris/Atlas-Chat-9B): A medium-sized with 9 billion parameters, providing more nuanced, contextually rich language generation for complex tasks. * [Atlas-Chat-27B](https://huggingface.co/MBZUAI-Paris/Atlas-Chat-27B): A large-sized version with 27 billion parameters, offering even more advanced capabilities for complex tasks and nuanced language generation compared to the 2B and 9B versions. The models are designed to assist with: * Conversational agents and chatbots that operate in Darija. * Translation, summarization, and content generation in informal dialect. * Cultural research related to Morocco and its language. **Paper:** [Atlas-Chat: Adapting Large Language Models for Low-Resource Moroccan Arabic Dialect](https://arxiv.org/abs/2409.17912) ## 👥 Our Team The model is developed by MBZUAI France Lab, an AI research center in Paris affiliated with the [Mohamed bin Zayed University of Artificial Intelligence (MBZUAI)](https://mbzuai.ac.ae/) headquartered in Abu Dhabi. ## Usage Below we share some code snippets on how to get quickly started with running the model. First, install the Transformers library with: ```sh pip install -U transformers sentencepiece ``` Then, copy the snippet from the section that is relevant for your use case. #### Running with the `pipeline` API ```python import torch from transformers import pipeline pipe = pipeline( "text-generation", model="MBZUAI-Paris/Atlas-Chat-2B", model_kwargs={"torch_dtype": torch.bfloat16}, device="cuda" # replace with "mps" to run on a Mac device ) messages = [ {"role": "user", "content": 'شكون لي صنعك؟'}, ] outputs = pipe(messages, max_new_tokens=256, temperature=0.0) assistant_response = outputs[0]["generated_text"][-1]["content"].strip() print(assistant_response) ``` - Response: >قادّوني الباحثين والمهندسين ديال جامعة محمد بن زايد للذكاء الاصطناعي. جامعة محمد بن زايد للذكاء الاصطناعي هي جامعة ديال البحت والدراسات العليا، كتّخصّص فتعزيز الذكاء الاصطناعي والاستعمال ديالو لمصلحة الإنسانية. يمكن ليك تزور https://mbzuai.ac.ae/ar/about/ باش تعرف كثر على جامعة محمد بن زايد للذكاء الاصطناعي والمهمة ديالها! #### Running the model on a single / multi GPU ```sh pip install accelerate ``` ```python from transformers import AutoTokenizer, AutoModelForCausalLM import torch model_id = "MBZUAI-Paris/Atlas-Chat-2B" tokenizer = AutoTokenizer.from_pretrained(model_id) model = AutoModelForCausalLM.from_pretrained( model_id, device_map="auto", torch_dtype=torch.bfloat16, ) messages = [ {"role": "user", "content": "شنو كيتسمى المنتخب المغربي ؟"}, ] input_ids = tokenizer.apply_chat_template(messages, return_tensors="pt", return_dict=True, , add_generation_prompt=True) outputs = model.generate(**input_ids, max_new_tokens=256) print(tokenizer.decode(outputs[0])) ``` - Response: >المنتخب المغربي كيتسمى "أسود الاطلس". #### Quantized Versions through `bitsandbytes`

Using 8-bit precision (int8)

```sh pip install bitsandbytes accelerate ``` ```python # pip install bitsandbytes accelerate from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig model_id = "MBZUAI-Paris/Atlas-Chat-2B" quantization_config = BitsAndBytesConfig(load_in_8bit=True) tokenizer = AutoTokenizer.from_pretrained(model_id) model = AutoModelForCausalLM.from_pretrained( model_id, quantization_config=quantization_config, ) text = f""" شرح ليا هاد الهضرة: في القرن 19 لقاو الذّهب في كاليفورنيا، ناضو لّي كيبيعو العتلة والفاس كيقنعو الناس بلي غيديرو لاباس يلا قلبو على الذهب... فالأخير اغتنى تجار أدوات التنقيب والحفر. وحاليا كاين لّي كيقنع الأخرين بلي هو مليونير، وعندو الوقت يورّي للآخرين كيفاش يديرو لاباس. """ messages = [ {"role": "user", "content": text}, ] input_ids = tokenizer.apply_chat_template(messages, return_tensors="pt", return_dict=True).to("cuda") outputs = model.generate(**input_ids, max_new_tokens=256) print(tokenizer.decode(outputs[0]).split("model")[-1]) ``` - Response: >ف القرن 19، لقاو الذهب ف كاليفورنيا، وهاد الشي جاب بزاف ديال الناس باش يمشيو ليه. هاد الناس كانو كيبيعو العتلة والفاس وكيتظاهرو بلي إلا قلبو على الذهب غادي يلقاو ليه. ف الآخر، هاد التجار ديال التنقيب والحفر كانو كيتغلبو على الناس اللي بغاو يقلبو على الذهب. > >دابا، كاينين ناس اللي كيتظاهرو بلي هوما مليونيرين وكيتظاهرو بلي عندهم الوقت يورّيو للناس كيفاش يلقاو الذهب. هاد الناس كيتظاهرو بلي عندهم الخبرة والخبرة باش يلقاو الذهب، ولكن ف الحقيقة، هاد الشي ماشي صحيح.

Using 4-bit precision

```python # pip install bitsandbytes accelerate from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig model_id = "MBZUAI-Paris/Atlas-Chat-2B" quantization_config = BitsAndBytesConfig(load_in_4bit=True) tokenizer = AutoTokenizer.from_pretrained(model_id) model = AutoModelForCausalLM.from_pretrained( model_id, quantization_config=quantization_config, ) text = f"""ترجم للدارجة: Atlas Chat is the first open source large language model that talks in Darija. """ messages = [ {"role": "user", "content": text}, ] input_ids = tokenizer.apply_chat_template(messages, return_tensors="pt", return_dict=True, add_generation_prompt=True) outputs = model.generate(**input_ids, max_new_tokens=256, temperature=0.0) print(tokenizer.decode(outputs[0]).split("model")[-1]) ``` - Response: >إن أطلاس شات هو أول نموذج لغة كبير مفتوح المصدر كيهضر بالدارجة.

### Chat Template The models use a chat template that must be adhered to conversational use. The easiest way to apply it is using the tokenizer's built-in chat template, as shown in the following snippet. Let's load the model and apply the chat template to a conversation. In this example, we'll start with a single user interaction: ```python from transformers import AutoTokenizer, AutoModelForCausalLM import transformers import torch model_id = "MBZUAI-Paris/Atlas-Chat-2B" dtype = torch.bfloat16 tokenizer = AutoTokenizer.from_pretrained(model_id) model = AutoModelForCausalLM.from_pretrained( model_id, device_map="cuda", torch_dtype=dtype,) chat = [ { "role": "user", "content": "اشنو هو الطاجين ؟"}, ] prompt = tokenizer.apply_chat_template(chat, tokenize=False, add_generation_prompt=True) ``` At this point, the prompt contains the following text: ``` user اشنو هو الطاجين ؟ model ``` As you can see, each turn is preceded by a `` delimiter and then the role of the entity (either `user`, for content supplied by the user, or `model` for LLM responses). Turns finish with the `` token. You can follow this format to build the prompt manually, if you need to do it without the tokenizer's chat template. After the prompt is ready, generation can be performed like this: ```python inputs = tokenizer.encode(prompt, add_special_tokens=False, return_tensors="pt") outputs = model.generate(input_ids=inputs.to(model.device), max_new_tokens=512) print(tokenizer.decode(outputs[0])) ``` - Response: >الطاجين هو طبق تقليدي مغربي كيتصاوب من اللحم ولا الدجاج ولا الخضرة، مع الخضرة، والبهارات، والصلصة. كيتطيب فالمقلاة ولا فالمقلى على نار هادية لمدة طويلة، وهاد الشي كيخلي اللحم يطيب بشوية ويبدا يذوب. الطاجين معروف بعمق النكهة ديالو والريحة ديالو، وغالبا كيتقدم مع الرز ولا الخبز. ### Inputs and outputs * **Input:** Text string, such as a question, a prompt, or a document to be summarized. * **Output:** Generated Darija text in response to the input, such as an answer to a question, or a summary of a document. ### Chatbot interface using Ollama You can also use Ollama and chatbot-ollama to create a chatbot user-interface to better test the model. First you need to install Ollama on your machine from [here](https://github.com/ollama/ollama) and have node.js installed as well. Then, download and prepare the model as follows: ```bash huggingface-cli download MBZUAI-Paris/Atlas-Chat-2B --local-dir Atlas-Chat-2B/ ollama create Atlas-Chat-2B -f Atlas-Chat-2B/modelfile ollama serve ``` Finally, in a new terminal clone chatbot-ollama repository from Github and run it: ```bash git clone https://github.com/ivanfioravanti/chatbot-ollama.git cd chatbot-ollama npm ci npm run dev ``` You can start chatting with the model by visiting http://localhost:3000. ### Citation If you use Atlas-Chat in your research, please cite our paper: ```none @article{shang2024atlaschatadaptinglargelanguage, title={Atlas-Chat: Adapting Large Language Models for Low-Resource Moroccan Arabic Dialect}, author={Guokan Shang and Hadi Abdine and Yousef Khoubrane and Amr Mohamed and Yassine Abbahaddou and Sofiane Ennadir and Imane Momayiz and Xuguang Ren and Eric Moulines and Preslav Nakov and Michalis Vazirgiannis and Eric Xing}, year={2024}, eprint={2409.17912}, archivePrefix={arXiv}, primaryClass={cs.CL}, url={https://arxiv.org/abs/2409.17912}, } ``` ## Training Data The model was trained on diverse datasets focusing on Darija consisting for approximatley 450k instructions of a maximum length of 2048 tokens, including: * Synthetic instructions created to guide the model in processing various types of language tasks tailord towards Moroccan culture. * Instruction samples created from publicly available Moroccan Arabic datasets including translation, summarization and sentiment analysis. * Translated English and multi-lingual instruction-tuning datasets. Our training dataset [Darija-SFT-Mixture](https://huggingface.co/datasets/MBZUAI-Paris/Darija-SFT-Mixture) is publicly available. ## Implementation Information Atlas-Chat models are based on Gemma 2 models. The Atlas-Chat models were trained using 8 Nvidia's A100 80 GB GPUs in parallel using FSDP on AWS Sagemaker. The model is trained using HuggingFace transformers and parameter-efficient fine-tuning with LoRA rank of 256. ## Evaluation The Atlas-Chat models were evaluated on a comprehensive suite of tasks using various datasets and benchmarks to assess their performance across multiple dimensions. These included tasks such as: * **DarijaMMLU:** A Darija version of ArabicMMLU and MMLU benchmarks translated from MSA and English respectively. * **DarijaHellaSwag:** A Darija version of HellaSwag. * **Belebele Ary_Arab:** Belebele is a multiple-choice machine reading comprehension dataset published by Facebook spanning 122 language variants. The Evaluation is done on the Ary_Arab part of Belebele that refers to Darija. * **DarijaAlpacaEval:** A Darija version of AlpacaEval translated to Darija and adapted to the Moroccan culture. * **Sentiment Analysis.** * **Translation:** Including six directions and four languages: Darija, MSA, English and French. * **Transliteration:** Transforming a sentence from Darija (written in Arabic characters) to Arabizi (Written in Latin characters) and vice-versa. * **Summarization.** The models were compared against a collection of existing open-source Arabic models to gauge their effectiveness, with a particular focus on performance in Darija. All scores are based on zero-shot performance. The prompts are written mainly in Darija. The metric used for DarijaMMLU, DarijaHellaSwag, Belebele Ary and Sentiment Analysis is the normalized accuracy. We used [Language Model Evaluation Harness](https://github.com/MBZUAI-Paris/lm-evaluation-harness-atlas-chat) to conduct these evaluations. **LLMs Benchmarks:**

Model	DarijaMMLU	DarijaHellaSwag	Belebele Ary	DarijaAlpacaEval
jais-family-1p3b-chat	35.39	27.71	38.33	35.56
jais-family-2p7b-chat	37.44	29.10	44.11	52.97
gemma-2-2b-it	28.58	32.42	25.22	58.67
Llama-3.2-1B-Instruct	27.66	26.88	28.89	23.57
Llama-3.2-3B-Instruct	32.60	28.33	38.00	47.62
Atlas-Chat-2B	44.97	35.08	53.89	92.31
jais-family-6p7b-chat	39.96	32.64	51.22	65.18
jais-adapted-7b-chat	39.30	29.55	43.67	61.84
jais-family-13b-chat	45.11	33.98	58.67	69.93
jais-adapted-13b-chat	45.20	32.84	49.67	77.52
AceGPT-7b-chat	35.98	30.33	30.11	47.31
AceGPT-13b-chat	41.09	38.35	33.11	52.79
gemma-2-9b-it	35.91	32.19	31.00	90.86
Llama-3.1-8B-Instruct	44.13	31.40	47.00	78.08
Atlas-Chat-9B	58.23	43.65	74.56	95.62
jais-family-30b-8k-chat	51.88	35.61	65.67	24.64
gemma-2-27b-it	36.47	37.04	35.78	95.07
Atlas-Chat-27B	61.95	48.37	75.67	96.58

**Standard NLP Tasks:**

Model	DODa-10k (Translation)		MADAR (Translation)		FLORES+ (Translation)		NLLB-Seed (Translation)		DODa-10k (Transliteration)		MArSum (Summarization) (LLM as a judge)	Sentiment Analysis
Model	BLEU	chrF	BLEU	chrF	BLEU	chrF	BLEU	chrF	BLEU	chrF	MArSum (Summarization) (LLM as a judge)	Sentiment Analysis
jais-family-1p3b-chat	00.13	06.18	00.50	15.43	02.44	19.14	01.99	12.60	00.01	03.01	00.50	45.29
jais-family-2p7b-chat	00.25	07.46	00.62	16.36	04.25	18.22	03.10	08.19	00.01	03.27	00.90	51.56
gemma-2-2b-it	00.10	04.96	00.12	06.66	01.55	18.59	02.78	23.69	00.01	02.08	06.80	53.36
Llama-3.2-1B-Instruct	00.07	05.95	00.80	18.71	04.53	18.39	04.52	17.06	00.02	03.74	08.23	46.27
Llama-3.2-3B-Instruct	00.62	13.67	01.18	22.12	08.59	35.21	13.75	43.63	00.21	09.68	08.23	49.20
Atlas-Chat-2B	22.76	44.86	16.67	41.64	14.92	43.03	23.88	52.19	08.18	21.54	55.22	73.99
jais-family-6p7b-chat	00.73	11.85	01.88	23.22	04.25	18.22	04.62	20.22	00.02	03.79	03.02	56.78
jais-adapted-7b-chat	00.60	09.43	03.45	25.88	07.25	23.21	01.25	02.22	00.04	03.24	02.82	52.72
jais-family-13b-chat	00.92	11.71	04.01	28.48	05.70	27.24	04.50	22.56	00.03	03.57	01.77	41.73
jais-adapted-13b-chat	00.87	10.52	04.02	25.29	06.66	23.46	20.14	47.87	0.04	04.77	01.92	66.68
AceGPT-7b-chat	00.44	11.33	01.05	19.24	06.92	36.03	11.05	44.55	00.06	04.74	02.28	40.23
AceGPT-13b-chat	00.98	16.70	00.81	20.23	08.73	40.76	14.02	48.28	00.12	06.32	02.80	59.58
gemma-2-9b-it	03.10	19.16	01.72	24.35	05.18	36.96	08.23	43.57	00.17	09.14	13.81	59.87
Llama-3.1-8B-Instruct	00.92	14.19	01.46	23.82	08.89	33.08	11.85	35.51	00.11	06.02	16.14	44.08
Atlas-Chat-9B	28.08	50.48	18.16	43.91	18.63	47.53	29.98	58.26	22.08	34.17	59.76	81.89
jais-family-30b-8k-chat	01.10	14.40	01.67	23.37	08.52	35.41	13.71	41.33	00.05	04.48	00.46	56.73
gemma-2-27b-it	00.67	13.04	01.74	24.63	05.17	37.08	07.36	42.49	00.03	04.94	11.10	57.59
Atlas-Chat-27B	29.55	51.74	19.66	45.65	20.34	49.19	31.61	59.37	33.03	40.95	60.70	73.00

## Usage and Limitations These models have certain limitations that users should be aware of.

Intended Usage

Open Large Language Models (LLMs) have a wide range of applications across various industries and domains. The following list of potential uses is not comprehensive. The purpose of this list is to provide contextual information about the possible use-cases that the model creators considered as part of model training and development. * Content Creation and Communication * Text Generation: These models can be used to generate creative text formats such as poems, scripts, code, marketing copy, and email drafts. * Chatbots and Conversational AI: Power conversational interfaces for customer service, virtual assistants, or interactive applications. * Text Summarization: Generate concise summaries of a text corpus, research papers, or reports. * Research and Education * Natural Language Processing (NLP) Research: These models can serve as a foundation for researchers to experiment with NLP techniques, develop algorithms, and contribute to the advancement of the field. * Language Learning Tools: Support interactive language learning experiences, aiding in grammar correction or providing writing practice. * Knowledge Exploration: Assist researchers in exploring large bodies of text by generating summaries or answering questions about specific topics.

Limitations

* Training Data * The quality and diversity of the training data significantly influence the model's capabilities. Biases or gaps in the training data can lead to limitations in the model's responses. * The scope of the training dataset determines the subject areas the model can handle effectively. * Context and Task Complexity * LLMs are better at tasks that can be framed with clear prompts and instructions. Open-ended or highly complex tasks might be challenging. * A model's performance can be influenced by the amount of context provided (longer context generally leads to better outputs, up to a certain point). * Language Ambiguity and Nuance * Natural language is inherently complex. LLMs might struggle to grasp subtle nuances, sarcasm, or figurative language. * Factual Accuracy * LLMs generate responses based on information they learned from their training datasets, but they are not knowledge bases. They may generate incorrect or outdated factual statements. * Common Sense * LLMs rely on statistical patterns in language. They might lack the ability to apply common sense reasoning in certain situations.

Ethical Considerations and Risks

The development of large language models (LLMs) raises several ethical concerns. In creating an open model, we have carefully considered the following: * Bias and Fairness * LLMs trained on large-scale, real-world text data can reflect socio-cultural biases embedded in the training material. * Misinformation and Misuse * LLMs can be misused to generate text that is false, misleading, or harmful. * Guidelines are provided for responsible use with the model, see the [Responsible Generative AI Toolkit][rai-toolkit]. * Transparency and Accountability: * This model card summarizes details on the models' architecture, capabilities, limitations, and evaluation processes. * A responsibly developed open model offers the opportunity to share innovation by making LLM technology accessible to developers and researchers across the AI ecosystem. Risks identified and mitigations: * Perpetuation of biases: It's encouraged to perform continuous monitoring (using evaluation metrics, human review) and the exploration of de-biasing techniques during model training, fine-tuning, and other use cases. * Generation of harmful content: Mechanisms and guidelines for content safety are essential. Developers are encouraged to exercise caution and implement appropriate content safety safeguards based on their specific product policies and application use cases. * Privacy violations: Models were trained on data filtered for removal of PII (Personally Identifiable Information). Developers are encouraged to adhere to privacy regulations with privacy-preserving techniques.

## Acknowledgement We would like to express our gratitude to the following institutions for their contributions to this work: École Polytechnique, LINAGORA and KTH Royal Institute of Technology. Additionally, we extend our thanks to the AtlasIA community.