Jais Family 2p7b

🚀 Jais Family Model Card

The Jais family of models is a comprehensive series of bilingual English - Arabic large language models (LLMs). These models are optimized for Arabic while maintaining strong English capabilities. This release includes two types of foundation models: models pre - trained from scratch (jais - family - *) and models pre - trained adaptively from Llama - 2 (jais - adapted - *). In total, 20 models across 8 sizes, with parameters ranging from 590M to 70B, are introduced, trained on up to 1.6T tokens of Arabic, English, and code data. All pre - trained models in this series are instruction fine - tuned (*-chat) for dialog using a curated mix of Arabic and English instruction data.

✨ Features

• Bilingual Excellence: Optimized for Arabic with strong English capabilities.
• Diverse Model Sizes: 20 models across 8 sizes, from 590M to 70B parameters.
• Extensive Training Data: Trained on up to 1.6T tokens of Arabic, English, and code data.
• Instruction Fine - Tuning: All pre - trained models are instruction fine - tuned for dialog.

📦 Installation

No installation steps are provided in the original document, so this section is skipped.

💻 Usage Examples

Basic Usage

# -*- coding: utf-8 -*-

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

model_path = "inceptionai/jais-family-2p7b"

device = "cuda" if torch.cuda.is_available() else "cpu"

tokenizer = AutoTokenizer.from_pretrained(model_path)
model = AutoModelForCausalLM.from_pretrained(model_path, device_map="auto", trust_remote_code=True)


def get_response(text, tokenizer=tokenizer, model=model):
    input_ids = tokenizer(text, return_tensors="pt").input_ids
    inputs = input_ids.to(device)
    input_len = inputs.shape[-1]
    generate_ids = model.generate(
        inputs,
        top_p=0.9,
        temperature=0.3,
        max_length=2048,
        min_length=input_len + 4,
        repetition_penalty=1.2,
        do_sample=True,
    )
    response = tokenizer.batch_decode(
        generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=True
    )[0]
    return response


text = "عاصمة دولة الإمارات العربية المتحدة ه"
print(get_response(text))

text = "The capital of UAE is"
print(get_response(text))

📚 Documentation

Jais Family Details

Pre - trained Models

Adapted Pre - trained Models

Model Architecture

All models in this family are auto - regressive language models that use a transformer - based, decoder - only architecture (GPT - 3). Jais models (jais - family - *) are trained from scratch, incorporating the SwiGLU non - linear activation function and ALiBi position encoding. Jais adapted models (jais - adapted - *) are built on top of Llama - 2, which employs RoPE position embedding and Grouped Query Attention. Tokenizer expansion with Arabic data is introduced, improving fertility and compute efficiency by over 3x.

Training Details

Pretraining Data

The Jais family of models are trained on up to 1.6 Trillion tokens of diverse English, Arabic, and Code data from web, code, books, scientific, and synthetic sources.

Finetuning data

All chat models in the Jais family are fine - tuned using Arabic and English prompt - response pairs in single - turn and multi - turn settings. Data sources include open - source fine - tuning datasets and internally curated human data, supplemented with synthetic content.

Training Procedure

• Pre - training (jais - family - *): Documents are packed into sequences separated by EOS tokens, and the model is trained autoregressively.
• Adapted pre - training (jais - adapted - *): A two - stage approach is used to train the new tokenizer and Arabic embeddings.
• Instruction tuning: Examples are packed together, and the loss is masked on the prompt tokens.

Training Hyperparameters: Jais - family - 2p7b

Compute Infrastructure

The training process was performed on the Condor Galaxy (CG) supercomputer platform, which contains 64 Cerebras CS - 2 Wafer - Scale Engines (WSE - 2) with 40 GB of SRAM and achieves a total of 960 PetaFLOP/s.

🔧 Technical Details

The models' architecture, training data, and training procedures are designed to optimize performance in both Arabic and English. The use of different activation functions, position encodings, and data sources contributes to the models' capabilities.

📄 License

The Jais family of models is released under the Apache 2.0 license.