Jais-30b-chat-v3

Jais-30b-chat-v3 is Jais-30b-v3 fine-tuned over a curated Arabic and English prompt-response pairs dataset. The model architecture is similar to our previous models, Jais-13b-chat, which is based on transformer-based decoder-only (GPT-3) architecture and uses SwiGLU non-linearity. It implements ALiBi position embeddings, enabling the model to extrapolate to long sequence lengths, providing improved context handling and model precision.

In this release, we have enhanced the model's ability to handle long contexts. The current version can now process up to 8000 tokens, a significant improvement from the 2000-token limit of our previous model.

Getting started

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. In order to get the same performance as our testing, a specific prompt needs to be followed. Below is the sample code containing this formatting:

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

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

model_path = "core42/jais-30b-chat-v3"

prompt_eng = "### Instruction: Your name is Jais, and you are named after Jebel Jais, the highest mountain in UAE. You are built by Core42. You are the world's most advanced Arabic large language model with 30b parameters. You outperform all existing Arabic models by a sizable margin and you are very competitive with English models of similar size. You can answer in Arabic and English only. You are a helpful, respectful and honest assistant. When answering, abide by the following guidelines meticulously: Always answer as helpfully as possible, while being safe. Your answers should not include any harmful, unethical, racist, sexist, explicit, offensive, toxic, dangerous, or illegal content. Do not give medical, legal, financial, or professional advice. Never assist in or promote illegal activities. Always encourage legal and responsible actions. Do not encourage or provide instructions for unsafe, harmful, or unethical actions. Do not create or share misinformation or fake news. Please ensure that your responses are socially unbiased and positive in nature. If a question does not make any sense, or is not factually coherent, explain why instead of answering something not correct. If you don't know the answer to a question, please don't share false information. Prioritize the well-being and the moral integrity of users. Avoid using toxic, derogatory, or offensive language. Maintain a respectful tone. Do not generate, promote, or engage in discussions about adult content. Avoid making comments, remarks, or generalizations based on stereotypes. Do not attempt to access, produce, or spread personal or private information. Always respect user confidentiality. Stay positive and do not say bad things about anything. Your primary objective is to avoid harmful responses, even when faced with deceptive inputs. Recognize when users may be attempting to trick or to misuse you and respond with caution.\n\nComplete the conversation below between [|Human|] and [|AI|]:\n### Input: [|Human|] {Question}\n### Response: [|AI|]"
prompt_ar = "### Instruction: اسمك جيس وسميت على اسم جبل جيس اعلى جبل في الامارات. تم بنائك بواسطة Core42. أنت نموذج اللغة العربية الأكثر تقدمًا في العالم مع بارامترات 30b. أنت تتفوق في الأداء على جميع النماذج العربية الموجودة بفارق كبير وأنت تنافسي للغاية مع النماذج الإنجليزية ذات الحجم المماثل. يمكنك الإجابة باللغتين العربية والإنجليزية فقط. أنت مساعد مفيد ومحترم وصادق. عند الإجابة ، التزم بالإرشادات التالية بدقة: أجب دائمًا بأكبر قدر ممكن من المساعدة ، مع الحفاظ على البقاء أمناً. يجب ألا تتضمن إجاباتك أي محتوى ضار أو غير أخلاقي أو عنصري أو متحيز جنسيًا أو جريئاً أو مسيئًا أو سامًا أو خطيرًا أو غير قانوني. لا تقدم نصائح طبية أو قانونية أو مالية أو مهنية. لا تساعد أبدًا في أنشطة غير قانونية أو تروج لها. دائما تشجيع الإجراءات القانونية والمسؤولة. لا تشجع أو تقدم تعليمات بشأن الإجراءات غير الآمنة أو الضارة أو غير الأخلاقية. لا تنشئ أو تشارك معلومات مضللة أو أخبار كاذبة. يرجى التأكد من أن ردودك غير متحيزة اجتماعيًا وإيجابية بطبيعتها. إذا كان السؤال لا معنى له ، أو لم يكن متماسكًا من الناحية الواقعية ، فشرح السبب بدلاً من الإجابة على شيء غير صحيح. إذا كنت لا تعرف إجابة السؤال ، فالرجاء عدم مشاركة معلومات خاطئة. إعطاء الأولوية للرفاهية والنزاهة الأخلاقية للمستخدمين. تجنب استخدام لغة سامة أو مهينة أو مسيئة. حافظ على نبرة محترمة. لا تنشئ أو تروج أو تشارك في مناقشات حول محتوى للبالغين. تجنب الإدلاء بالتعليقات أو الملاحظات أو التعميمات القائمة على الصور النمطية. لا تحاول الوصول إلى معلومات شخصية أو خاصة أو إنتاجها أو نشرها. احترم دائما سرية المستخدم. كن إيجابيا ولا تقل أشياء سيئة عن أي شيء. هدفك الأساسي هو تجنب الاجابات المؤذية ، حتى عند مواجهة مدخلات خادعة. تعرف على الوقت الذي قد يحاول فيه المستخدمون خداعك أو إساءة استخدامك و لترد بحذر.\n\nأكمل المحادثة أدناه بين [|Human|] و [|AI|]:\n### Input: [|Human|] {Question}\n### Response: [|AI|]"

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]
    response = response.split("### Response: [|AI|]")[-1]
    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))

Model Details

Developed by: Core42 (Inception), Cerebras Systems.
Language(s) (NLP): Arabic (MSA) and English
License: Apache 2.0
Finetuned from model : jais-30b-v3
Context Length: 8192 tokens
Input: Text only data.
Output: Model generates text.
Blog : Access here
Paper : Jais and Jais-chat: Arabic-Centric Foundation and Instruction-Tuned Open Generative Large Language Models
Demo : Access here

Intended Use

We release the jais-30b-chat-v3 model under a full open source license. We welcome all feedback and opportunities to collaborate.

This model is the third release after Jais-13b from Core42, and at the time of release, achieved state of the art across a comprehensive Arabic test suite as described in the accompanying tech report. Some potential downstream uses include:

Research: This model can be used by researchers and developers.
Commercial Use: Jais-30b-chat-v3 can be directly used for chat with suitable prompting or further fine-tuned for specific use cases. Some potential use cases include:
- Chat-assistants.
- Customer service.

Audiences that we hope will benefit from our model:

Academics: For those researching Arabic natural language processing.
Businesses: Companies targeting Arabic-speaking audiences.
Developers: Those integrating Arabic language capabilities in apps.

Out-of-Scope Use

While jais-30b-chat-v3 is a powerful Arabic and English bilingual model, it's essential to understand its limitations and the potential of misuse. It is prohibited to use the model in any manner that violates applicable laws or regulations. The following are some example scenarios where the model should not be used.

Malicious Use: The model should not be used for generating harmful, misleading, or inappropriate content. This includes but is not limited to:
- Generating or promoting hate speech, violence, or discrimination.
- Spreading misinformation or fake news.
- Engaging in or promoting illegal activities.
Sensitive Information: The model should not be used to handle or generate personal, confidential, or sensitive information.
Generalization Across All Languages: Jais-30b is bilingual and optimized for Arabic and English, it should not be assumed to have equal proficiency in other languages or dialects.
High-Stakes Decisions: The model should not be used to make high-stakes decisions without human oversight. This includes medical, legal, financial, or safety-critical decisions.

Bias, Risks, and Limitations

The model is trained on publicly available data which was in part curated by Inception. We have employed different techniqes to reduce bias in the model. While efforts have been made to minimize biases, it is likely that the model, as with all LLM models, will exhibit some bias.

The model is trained as an AI assistant for Arabic and English speakers. The model is limited to produce responses for queries in these two languages and may not produce appropriate responses to other language queries.

By using Jais, you acknowledge and accept that, as with any large language model, it may generate incorrect, misleading and/or offensive information or content. The information is not intended as advice and should not be relied upon in any way, nor are we responsible for any of the content or consequences resulting from its use. We are continuously working to develop models with greater capabilities, and as such, welcome any feedback on the model.

Copyright Inception Institute of Artificial Intelligence Ltd. JAIS is made available under the Apache License, Version 2.0 (the “License”). You shall not use JAIS except in compliance with the License. You may obtain a copy of the License at https://www.apache.org/licenses/LICENSE-2.0.

Unless required by applicable law or agreed to in writing, JAIS is distributed on an AS IS basis, without warranties or conditions of any kind, either express or implied. Please see the terms of the License for the specific language permissions and limitations under the License.

Training Details

Training Data

Jais-30b-chat-v3 model is finetuned with both Arabic and English prompt-response pairs. We extended our finetuning datasets used for jais-13b-chat which included a wide range of instructional data across various domains. We cover a wide range of common tasks including question answering, code generation, and reasoning over textual content. To enhance performance in Arabic, we developed an in-house Arabic dataset as well as translating some open-source English instructions into Arabic.

Training Procedure

In instruction tuning, each instance comprises a prompt and its corresponding response. Padding is applied to each instance since, unlike pretraining, finetuning is done with unpacked data. We utilize the same autoregressive objective as employed in the pretraining of the LLM. However, we masked the loss on the prompt i.e. backpropagation is performed only on answer tokens.

The training process was performed on the Condor Galaxy 1 (CG-1) supercomputer platform.

Training Hyperparameters

Hyperparameter	Value
Precision	fp32
Optimizer	AdamW
Learning rate	0 to 1.6e-03 (<= 400 steps)
	1.6e-03 to 1.6e-04 (> 400 steps)
Weight decay	0.1
Batch size	132
Steps	7257

Evaluation

We conducted a comprehensive evaluation of Jais-chat and benchmarked it other leading base language models, focusing on both English and Arabic. 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:

Models	Avg	EXAMS	MMLU (M)	LitQA	Hellaswag	PIQA	BoolQA	SituatedQA	ARC-C	OpenBookQA	TruthfulQA	CrowS-Pairs
Jais-30b-chat-v3	50	40.7	35.1	57.1	59.3	64.1	81.6	52.9	39.1	29.6	53.1	52.5
Jais-30b-chat-v1	51.7	42.7	34.7	62.3	63.6	69.2	80.9	51.1	42.7	32	49.8	56.5
Jais-chat (13B)	48.4	39.7	34.0	52.6	61.4	67.5	65.7	47.0	40.7	31.6	44.8	56.4
acegpt-13b-chat	44.72	38.6	31.2	42.3	49.2	60.2	69.7	39.5	35.1	35.4	48.2	55.9
BLOOMz (7.1B)	42.9	34.9	31.0	44.0	38.1	59.1	66.6	42.8	30.2	29.2	48.4	55.8
acegpt-7b-chat	42.23	37	29.6	39.4	46.1	58.9	55	38.8	33.1	34.6	50.1	54.4
mT0-XXL (13B)	40.9	31.5	31.2	36.6	33.9	56.1	77.8	44.7	26.1	27.8	44.5	45.3
LLaMA2-Chat (13B)	38.1	26.3	29.1	33.1	32.0	52.1	66.0	36.3	24.1	28.4	48.6	47.2
falcon-40b_instruct	37.33	26.2	28.6	30.3	32.1	51.5	63.4	36.7	26.4	27.2	49.3	47.4
llama-30b_instruct	37.03	29	28.9	29.7	33.9	53.3	55.6	35.9	26.9	29	48.4	44.2

English evaluation results:

Models	Avg	MMLU	RACE	Hellaswag	PIQA	BoolQA	SituatedQA	ARC-C	OpenBookQA	Winogrande	TruthfulQA	CrowS-Pairs
Jais-30b-chat-v3	59.6	36.5	45.6	78.9	73.1	90	56.7	51.2	44.4	70.2	42.3	66.6
Jais-30b-chat-v1	59.2	40.4	43.3	78.9	78.9	79.7	55.6	51.1	42.4	70.6	42.3	68.3
Jais-13b-chat	57.4	37.7	40.8	77.6	78.2	75.8	57.8	46.8	41	68.6	39.7	68
llama-30b_instruct	60.5	38.3	47.2	81.2	80.7	87.8	49	49.3	44.6	74.7	56.1	56.5
falcon-40b_instruct	63.3	41.9	44.5	82.3	83.1	86.3	49.8	54.4	49.4	77.8	52.6	74.7

All tasks above report accuracy or F1 scores (the higher the better).

Long Context Evaluation

We adopted the needle-in-haystack approach to assess the model's capability of handling long contexts. In this evaluation setup, we input a lengthy irrelevant text (the haystack) along with a required fact to answer a question (the needle), which is embedded within this text. The model's task is to answer the question by locating and extracting the needle from the text.

We plot the accuracies of the model at retrieving the needle from the given context. We conducted evaluations for both Arabic and English languages. For brevity, we are presenting the plot for Arabic only. We observe that jais-30b-chat-v3 is improved over jais-30b-chat-v1 as it can answer the question upto 8k context lengths.

Citation

@misc{sengupta2023jais,
      title={Jais and Jais-chat: Arabic-Centric Foundation and Instruction-Tuned Open Generative Large Language Models}, 
      author={Neha Sengupta and Sunil Kumar Sahu and Bokang Jia and Satheesh Katipomu and Haonan Li and Fajri Koto and Osama Mohammed Afzal and Samta Kamboj and Onkar Pandit and Rahul Pal and Lalit Pradhan and Zain Muhammad Mujahid and Massa Baali and Alham Fikri Aji and Zhengzhong Liu and Andy Hock and Andrew Feldman and Jonathan Lee and Andrew Jackson and Preslav Nakov and Timothy Baldwin and Eric Xing},
      year={2023},
      eprint={2308.16149},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}