Jais Adapted 70b Chat

🚀 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 excellent performance in Arabic while maintaining strong English capabilities. We offer two variants of foundation models:

• Models pre - trained from scratch (jais - family - *).
• Models pre - trained adaptively from Llama - 2 (jais - adapted - *).

In this release, we present 20 models across 8 sizes, with parameters ranging from 590M to 70B. These models are 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.

We hope this extensive release will accelerate research in Arabic NLP and enable numerous downstream applications for the Arabic - speaking and bilingual community. The training and adaptation techniques demonstrated for Arabic models are extensible to other low and medium - resource languages.

✨ Features

Jais Family Details

• Developed by: Inception, Cerebras Systems.
• Language(s): (NLP): Arabic (MSA) and English.
• Input: Text - only data.
• Output: The model generates text.
• Model Sizes: 590M, 1.3B, 2.7B, 6.7B, 7B, 13B, 30B, 70B.
• Demo: [Access the live demo here](https://arabic - gpt.ai/)
• License: Apache 2.0

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. These architectural enhancements allow the models to extrapolate at long sequence lengths, leading to improved context handling and precision.

Jais adapted models (jais - adapted - *) are built on top of Llama - 2, which employs RoPE position embedding and Grouped Query Attention. We introduce tokenizer expansion with Arabic data, which improves fertility and compute efficiency by over 3x. In particular, we add 32,000 new Arabic tokens from the Jais - 30b vocabulary into the Llama - 2 tokenizer. To initialize these new Arabic token embeddings we first learn a linear projection from the embedding space of Jais - 30b to Llama's embedding space, using the set of shared English tokens present in both vocabularies. Next, this learned projection is applied to transform the existing Jais - 30b Arabic embeddings into the Llama - 2 embedding space.

🚀 Quick Start

Below is sample code to use the model. Note that the model requires a custom model class, so users must enable trust_remote_code = True while loading the model.

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

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

model_path = "inceptionai/jais - adapted - 70b - chat"

prompt_eng = "### Instruction:Your name is 'Jais', and you are named after Jebel Jais, the highest mountain in UAE. You were made by 'Inception' in the UAE. You are a helpful, respectful, and honest assistant. Always answer as helpfully as possible, while being safe. Complete the conversation between [|Human|] and [|AI|]:\n### Input: [|Human|] {Question}\n[|AI|]\n### Response :"
prompt_ar = "### Instruction:اسمك \"جيس\" وسميت على اسم جبل جيس اعلى جبل في الامارات. تم بنائك بواسطة Inception في الإمارات. أنت مساعد مفيد ومحترم وصادق. أجب دائمًا بأكبر قدر ممكن من المساعدة، مع الحفاظ على البقاء أمناً. أكمل المحادثة بين [|Human|] و[|AI|] :\n### Input:[|Human|] {Question}\n[|AI|]\n### Response :"

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)

if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token

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


ques = "ما هي عاصمة الامارات؟"
text = prompt_ar.format_map({'Question': ques})
print(get_response(text))

ques = "What is the capital of UAE?"
text = prompt_eng.format_map({'Question': ques})
print(get_response(text))

📚 Documentation

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. The data consists of the following sources:

• Web: We used publicly available web pages, wikipedia articles, news articles, and social network content in both Arabic and English.

• Code: To enhance the reasoning capability of our model, we include Code data in various programming languages.

• Books: We used a selection of publicly available Arabic and English books data, which improves long - range context modelling and coherent storytelling.

• Scientific: A subset of ArXiv papers were included to improve reasoning and long context abilities.

• Synthetic: We augment the volume of Arabic data by translating English to Arabic using an in - house machine translation system. We restrict this to high - quality English resources such as English Wikipedia and English books.

We extensively preprocess and deduplicate the training data. For Arabic, we used a custom preprocessing pipeline to filter for data with high linguistic quality. More information on this pipeline can be found in the Jais paper.

• Jais pre - trained (jais - family - *): Following our previous experimentation with language alignment mixing in Jais, we used a ratio of 1:2:0.4 of Arabic:English:Code data. This recipe for from scratch pre - training addresses Arabic data scarcity while improving performance in both languages.
• Jais adapted pre - trained (jais - adapted - *): For the adapted pre - training of Llama - 2, we utilized a larger Arabic dataset of ~334B Arabic tokens mixed with English and Code data. We vary the mixing ratio, at different model sizes, to introduce strong Arabic capabilities while maintaining performance in English.

Finetuning data

All chat models in the Jais family are fine - tuned using Arabic and English prompt - response pairs in both single - turn and multi - turn settings. Data sources include open - source fine - tuning datasets filtered for topic and style diversity. Additionally, internally curated human data is incorporated to enhance cultural adaptation. This data is supplemented with content generated using synthetic methods including machine translation, distillation, and model self - chat. Overall, our updated instruction - tuning dataset comprises ~10M and ~4M prompt - response pairs in English and Arabic respectively.

Training Procedure

During the pre - training of (jais - family - *) models, documents are packed into sequences separated by EOS tokens, and the model is trained autoregressively, applying the loss to all tokens. For jais - 30b models, the context length is progressively expanded from 2k to 8K to 16K by incorporating curated long - context documents in training. This progressive expansion leverages faster initial training at shorter context lengths, while gradually extending support for larger context lengths towards the end of the training process.

During the adapted pre - training of the (jais - adapted - *) models, we first initialize the new tokenizer and Arabic embeddings as described in [Model Architecture](#model - architecture). In training, we implemented a two - stage approach to overcome observed higher norms of the new Arabic embeddings. In the first stage, the backbone of the model is frozen, and the embeddings are trained using approximately 15 billion tokens from a bilingual corpus of English and Arabic. In the second stage, the backbone is unfrozen, and continuous pretraining is conducted with all parameters.

During instruction tuning, each training example consists of a single - turn or multi - turn prompt and its response. Instead of one example per sequence, examples are packed together while the loss is masked on the prompt tokens. This approach speeds up training by allowing more examples to be processed per batch.

Training Hyperparameters:

Jais - adapted - 70b - chat

Compute Infrastructure

The training process was performed on the Condor Galaxy (CG) supercomputer platform. A CG 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

Evaluation

We conducted a comprehensive evaluation of Jais models focusing on both English and Arabic, using LM - harness in a zero - shot setting. The evaluation criteria spanned various dimensions, including:

• Knowledge: How well the model answers factual questions.
• Reasoning: The model's ability to answer questions requiring reasoning.
• Misinformation/Bias: Assessment of the model's susceptibility to generating false or misleading information, and its neutrality.

Arabic evaluation results:

Arabic benchmarks are translated using an in - house MT model and reviewed by Arabic linguists. Benchmarks labeled with an asterisk (*) are natively Arabic; for further details, see the Jais paper. Additionally, we include ArabicMMLU, a native Arabic benchmark based on regional knowledge.

English evaluation results: