RichardErkhov/inceptionai_-_jais-adapted-7b-chat-gguf
1
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

初始化项目，由ModelHub XC社区提供模型

Browse Source 
Model: RichardErkhov/inceptionai_-_jais-adapted-7b-chat-gguf
Source: Original Platform
This commit is contained in:
ModelHub XC
2026-08-08 12:19:17 +08:00
commit 432ba67ec1
24 changed files with 551 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

57
.gitattributes vendored Normal file
View File

@@ -0,0 +1,57 @@
*.7z filter=lfs diff=lfs merge=lfs -text
*.arrow filter=lfs diff=lfs merge=lfs -text
*.bin filter=lfs diff=lfs merge=lfs -text
*.bz2 filter=lfs diff=lfs merge=lfs -text
*.ckpt filter=lfs diff=lfs merge=lfs -text
*.ftz filter=lfs diff=lfs merge=lfs -text
*.gz filter=lfs diff=lfs merge=lfs -text
*.h5 filter=lfs diff=lfs merge=lfs -text
*.joblib filter=lfs diff=lfs merge=lfs -text
*.lfs.* filter=lfs diff=lfs merge=lfs -text
*.mlmodel filter=lfs diff=lfs merge=lfs -text
*.model filter=lfs diff=lfs merge=lfs -text
*.msgpack filter=lfs diff=lfs merge=lfs -text
*.npy filter=lfs diff=lfs merge=lfs -text
*.npz filter=lfs diff=lfs merge=lfs -text
*.onnx filter=lfs diff=lfs merge=lfs -text
*.ot filter=lfs diff=lfs merge=lfs -text
*.parquet filter=lfs diff=lfs merge=lfs -text
*.pb filter=lfs diff=lfs merge=lfs -text
*.pickle filter=lfs diff=lfs merge=lfs -text
*.pkl filter=lfs diff=lfs merge=lfs -text
*.pt filter=lfs diff=lfs merge=lfs -text
*.pth filter=lfs diff=lfs merge=lfs -text
*.rar filter=lfs diff=lfs merge=lfs -text
*.safetensors filter=lfs diff=lfs merge=lfs -text
saved_model/**/* filter=lfs diff=lfs merge=lfs -text
*.tar.* filter=lfs diff=lfs merge=lfs -text
*.tar filter=lfs diff=lfs merge=lfs -text
*.tflite filter=lfs diff=lfs merge=lfs -text
*.tgz filter=lfs diff=lfs merge=lfs -text
*.wasm filter=lfs diff=lfs merge=lfs -text
*.xz filter=lfs diff=lfs merge=lfs -text
*.zip filter=lfs diff=lfs merge=lfs -text
*.zst filter=lfs diff=lfs merge=lfs -text
*tfevents* filter=lfs diff=lfs merge=lfs -text
jais-adapted-7b-chat.Q2_K.gguf filter=lfs diff=lfs merge=lfs -text
jais-adapted-7b-chat.IQ3_XS.gguf filter=lfs diff=lfs merge=lfs -text
jais-adapted-7b-chat.IQ3_S.gguf filter=lfs diff=lfs merge=lfs -text
jais-adapted-7b-chat.Q3_K_S.gguf filter=lfs diff=lfs merge=lfs -text
jais-adapted-7b-chat.IQ3_M.gguf filter=lfs diff=lfs merge=lfs -text
jais-adapted-7b-chat.Q3_K.gguf filter=lfs diff=lfs merge=lfs -text
jais-adapted-7b-chat.Q3_K_M.gguf filter=lfs diff=lfs merge=lfs -text
jais-adapted-7b-chat.Q3_K_L.gguf filter=lfs diff=lfs merge=lfs -text
jais-adapted-7b-chat.IQ4_XS.gguf filter=lfs diff=lfs merge=lfs -text
jais-adapted-7b-chat.Q4_0.gguf filter=lfs diff=lfs merge=lfs -text
jais-adapted-7b-chat.IQ4_NL.gguf filter=lfs diff=lfs merge=lfs -text
jais-adapted-7b-chat.Q4_K_S.gguf filter=lfs diff=lfs merge=lfs -text
jais-adapted-7b-chat.Q4_K.gguf filter=lfs diff=lfs merge=lfs -text
jais-adapted-7b-chat.Q4_K_M.gguf filter=lfs diff=lfs merge=lfs -text
jais-adapted-7b-chat.Q4_1.gguf filter=lfs diff=lfs merge=lfs -text
jais-adapted-7b-chat.Q5_0.gguf filter=lfs diff=lfs merge=lfs -text
jais-adapted-7b-chat.Q5_K_S.gguf filter=lfs diff=lfs merge=lfs -text
jais-adapted-7b-chat.Q5_K.gguf filter=lfs diff=lfs merge=lfs -text
jais-adapted-7b-chat.Q5_K_M.gguf filter=lfs diff=lfs merge=lfs -text
jais-adapted-7b-chat.Q5_1.gguf filter=lfs diff=lfs merge=lfs -text
jais-adapted-7b-chat.Q6_K.gguf filter=lfs diff=lfs merge=lfs -text
jais-adapted-7b-chat.Q8_0.gguf filter=lfs diff=lfs merge=lfs -text

428
README.md Normal file
View File

@@ -0,0 +1,428 @@
Quantization made by Richard Erkhov.
[Github](https://github.com/RichardErkhov)
[Discord](https://discord.gg/pvy7H8DZMG)
[Request more models](https://github.com/RichardErkhov/quant_request)
jais-adapted-7b-chat - GGUF
- Model creator: https://huggingface.co/inceptionai/
- Original model: https://huggingface.co/inceptionai/jais-adapted-7b-chat/
| Name | Quant method | Size |
| ---- | ---- | ---- |
| [jais-adapted-7b-chat.Q2_K.gguf](https://huggingface.co/RichardErkhov/inceptionai_-_jais-adapted-7b-chat-gguf/blob/main/jais-adapted-7b-chat.Q2_K.gguf) | Q2_K | 2.5GB |
| [jais-adapted-7b-chat.IQ3_XS.gguf](https://huggingface.co/RichardErkhov/inceptionai_-_jais-adapted-7b-chat-gguf/blob/main/jais-adapted-7b-chat.IQ3_XS.gguf) | IQ3_XS | 2.76GB |
| [jais-adapted-7b-chat.IQ3_S.gguf](https://huggingface.co/RichardErkhov/inceptionai_-_jais-adapted-7b-chat-gguf/blob/main/jais-adapted-7b-chat.IQ3_S.gguf) | IQ3_S | 2.9GB |
| [jais-adapted-7b-chat.Q3_K_S.gguf](https://huggingface.co/RichardErkhov/inceptionai_-_jais-adapted-7b-chat-gguf/blob/main/jais-adapted-7b-chat.Q3_K_S.gguf) | Q3_K_S | 2.9GB |
| [jais-adapted-7b-chat.IQ3_M.gguf](https://huggingface.co/RichardErkhov/inceptionai_-_jais-adapted-7b-chat-gguf/blob/main/jais-adapted-7b-chat.IQ3_M.gguf) | IQ3_M | 3.05GB |
| [jais-adapted-7b-chat.Q3_K.gguf](https://huggingface.co/RichardErkhov/inceptionai_-_jais-adapted-7b-chat-gguf/blob/main/jais-adapted-7b-chat.Q3_K.gguf) | Q3_K | 3.22GB |
| [jais-adapted-7b-chat.Q3_K_M.gguf](https://huggingface.co/RichardErkhov/inceptionai_-_jais-adapted-7b-chat-gguf/blob/main/jais-adapted-7b-chat.Q3_K_M.gguf) | Q3_K_M | 3.22GB |
| [jais-adapted-7b-chat.Q3_K_L.gguf](https://huggingface.co/RichardErkhov/inceptionai_-_jais-adapted-7b-chat-gguf/blob/main/jais-adapted-7b-chat.Q3_K_L.gguf) | Q3_K_L | 3.5GB |
| [jais-adapted-7b-chat.IQ4_XS.gguf](https://huggingface.co/RichardErkhov/inceptionai_-_jais-adapted-7b-chat-gguf/blob/main/jais-adapted-7b-chat.IQ4_XS.gguf) | IQ4_XS | 3.56GB |
| [jais-adapted-7b-chat.Q4_0.gguf](https://huggingface.co/RichardErkhov/inceptionai_-_jais-adapted-7b-chat-gguf/blob/main/jais-adapted-7b-chat.Q4_0.gguf) | Q4_0 | 3.73GB |
| [jais-adapted-7b-chat.IQ4_NL.gguf](https://huggingface.co/RichardErkhov/inceptionai_-_jais-adapted-7b-chat-gguf/blob/main/jais-adapted-7b-chat.IQ4_NL.gguf) | IQ4_NL | 3.75GB |
| [jais-adapted-7b-chat.Q4_K_S.gguf](https://huggingface.co/RichardErkhov/inceptionai_-_jais-adapted-7b-chat-gguf/blob/main/jais-adapted-7b-chat.Q4_K_S.gguf) | Q4_K_S | 3.76GB |
| [jais-adapted-7b-chat.Q4_K.gguf](https://huggingface.co/RichardErkhov/inceptionai_-_jais-adapted-7b-chat-gguf/blob/main/jais-adapted-7b-chat.Q4_K.gguf) | Q4_K | 3.97GB |
| [jais-adapted-7b-chat.Q4_K_M.gguf](https://huggingface.co/RichardErkhov/inceptionai_-_jais-adapted-7b-chat-gguf/blob/main/jais-adapted-7b-chat.Q4_K_M.gguf) | Q4_K_M | 3.97GB |
| [jais-adapted-7b-chat.Q4_1.gguf](https://huggingface.co/RichardErkhov/inceptionai_-_jais-adapted-7b-chat-gguf/blob/main/jais-adapted-7b-chat.Q4_1.gguf) | Q4_1 | 4.12GB |
| [jais-adapted-7b-chat.Q5_0.gguf](https://huggingface.co/RichardErkhov/inceptionai_-_jais-adapted-7b-chat-gguf/blob/main/jais-adapted-7b-chat.Q5_0.gguf) | Q5_0 | 4.52GB |
| [jais-adapted-7b-chat.Q5_K_S.gguf](https://huggingface.co/RichardErkhov/inceptionai_-_jais-adapted-7b-chat-gguf/blob/main/jais-adapted-7b-chat.Q5_K_S.gguf) | Q5_K_S | 4.52GB |
| [jais-adapted-7b-chat.Q5_K.gguf](https://huggingface.co/RichardErkhov/inceptionai_-_jais-adapted-7b-chat-gguf/blob/main/jais-adapted-7b-chat.Q5_K.gguf) | Q5_K | 4.64GB |
| [jais-adapted-7b-chat.Q5_K_M.gguf](https://huggingface.co/RichardErkhov/inceptionai_-_jais-adapted-7b-chat-gguf/blob/main/jais-adapted-7b-chat.Q5_K_M.gguf) | Q5_K_M | 4.64GB |
| [jais-adapted-7b-chat.Q5_1.gguf](https://huggingface.co/RichardErkhov/inceptionai_-_jais-adapted-7b-chat-gguf/blob/main/jais-adapted-7b-chat.Q5_1.gguf) | Q5_1 | 4.91GB |
| [jais-adapted-7b-chat.Q6_K.gguf](https://huggingface.co/RichardErkhov/inceptionai_-_jais-adapted-7b-chat-gguf/blob/main/jais-adapted-7b-chat.Q6_K.gguf) | Q6_K | 5.35GB |
| [jais-adapted-7b-chat.Q8_0.gguf](https://huggingface.co/RichardErkhov/inceptionai_-_jais-adapted-7b-chat-gguf/blob/main/jais-adapted-7b-chat.Q8_0.gguf) | Q8_0 | 6.93GB |
Original model description:
---
language:
- ar
- en
thumbnail: null
tags:
- Arabic
- English
- LLM
- Decoder
- causal-lm
- jais-family
license: apache-2.0
pipeline_tag: text-generation
---
# 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 to excel in Arabic while having strong English capabilities. We release two variants of foundation models that include:
- Models **pre-trained from scratch** (`jais-family-*`).
- Models **pre-trained adaptively from [Llama-2](https://arxiv.org/pdf/2307.09288)** (`jais-adapted-*`).
In this release, we introduce 20 models across 8 sizes, ranging from 590M to 70B parameters, 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 we demonstrate successfully for Arabic models are extensible to other low and medium resource languages.
## Jais Family Details
- **Developed by:** Inception, Cerebras Systems.
- **Language(s):** (NLP): Arabic (MSA) and English.
- **Input:** Text only data.
- **Output:** 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
| **Pre-trained Model** | **Fine-tuned Model** | **Size (Parameters)** | **Context length (Tokens)** |
|:---------------------|:--------|:-------|:-------|
| [jais-family-30b-16k](https://huggingface.co/inceptionai/jais-family-30b-16k) | [Jais-family-30b-16k-chat](https://huggingface.co/inceptionai/jais-family-30b-16k-chat) | 30B | 16,384 |
| [jais-family-30b-8k](https://huggingface.co/inceptionai/jais-family-30b-8k) | [Jais-family-30b-8k-chat](https://huggingface.co/inceptionai/jais-family-30b-8k-chat) | 30B | 8,192 |
| [jais-family-13b ](https://huggingface.co/inceptionai/jais-family-13b) | [Jais-family-13b-chat](https://huggingface.co/inceptionai/jais-family-13b-chat) | 13B | 2,048 |
| [jais-family-6p7b](https://huggingface.co/inceptionai/jais-family-6p7b) | [Jais-family-6p7b-chat](https://huggingface.co/inceptionai/jais-family-6p7b-chat) | 6.7B | 2,048 |
| [jais-family-2p7b](https://huggingface.co/inceptionai/jais-family-2p7b) | [Jais-family-2p7b-chat](https://huggingface.co/inceptionai/jais-family-2p7b-chat) | 2.7B | 2,048 |
| [jais-family-1p3b](https://huggingface.co/inceptionai/jais-family-1p3b) | [Jais-family-1p3b-chat](https://huggingface.co/inceptionai/jais-family-1p3b-chat) | 1.3B | 2,048 |
| [jais-family-590m](https://huggingface.co/inceptionai/jais-family-590m) | [Jais-family-590m-chat](https://huggingface.co/inceptionai/jais-family-590m-chat) | 590M | 2,048 |
| **Adapted pre-trained Model** | **Fine-tuned Model** | **Size (Parameters)** | **Context length (Tokens)** |
|:---------------------|:--------|:-------|:-------|
| [jais-adapted-70b](https://huggingface.co/inceptionai/jais-adapted-70b) | [Jais-adapted-70b-chat](https://huggingface.co/inceptionai/jais-adapted-70b-chat) | 70B | 4,096 |
| [jais-adapted-13b](https://huggingface.co/inceptionai/jais-adapted-13b) | [Jais-adapted-13b-chat](https://huggingface.co/inceptionai/jais-adapted-13b-chat) | 13B | 4,096 |
| [jais-adapted-7b](https://huggingface.co/inceptionai/jais-adapted-7b) | [Jais-adapted-7b-chat](https://huggingface.co/inceptionai/jais-adapted-7b-chat) | 7B | 4,096 |
### Model Architecture:
<a name="model-architecture"></a>
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.
## 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.
```python
# -*- coding: utf-8 -*-
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
model_path = "inceptionai/jais-adapted-7b-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))
```
## 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](https://arxiv.org/abs/2308.16149).
- **Jais pre-trained** (`jais-family-*`): Following our previous experimentation with language alignment mixing in [Jais](https://arxiv.org/abs/2308.16149), we used a ratio of 1:2:0.4 of Arabic:English:Code data. This recipe for <u>from scratch pre-training</u> addresses Arabic data scarcity while improving performance in both languages.
- **Jais adapted pre-trained** (`jais-adapted-*`): For the <u>adapted pre-training of Llama-2</u>, 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.
| **Pre-trained model** | **English data (tokens)** | **Arabic data (tokens)** | **Code data (tokens)** | **Total data (tokens)** |
|-------------------------|---------------------------|--------------------------|------------------------|------------------------|
| [jais-family-30b-16k](https://huggingface.co/inceptionai/jais-family-30b-16k) | 980B | 490B | 196B | 1666B |
| [jais-family-30b-8k](https://huggingface.co/inceptionai/jais-family-30b-8k) | 882B | 441B | 177B | 1500B |
| [jais-family-13b ](https://huggingface.co/inceptionai/jais-family-13b) | 283B | 141B | 56B | 480B |
| [jais-family-6p7b](https://huggingface.co/inceptionai/jais-family-6p7b) | 283B | 141B | 56B | 480B |
| [jais-family-2p7b](https://huggingface.co/inceptionai/jais-family-2p7b) | 283B | 141B | 56B | 480B |
| [jais-family-1p3b](https://huggingface.co/inceptionai/jais-family-1p3b) | 283B | 141B | 56B | 480B |
| [jais-family-590m](https://huggingface.co/inceptionai/jais-family-590m) | 283B | 141B | 56B | 480B |
| [jais-adapted-70b](https://huggingface.co/inceptionai/jais-adapted-70b) | 33B | 334B | 4B | 371B |
| [jais-adapted-13b](https://huggingface.co/inceptionai/jais-adapted-13b) | 127B | 140B | 13B | 280B |
| [jais-adapted-7b](https://huggingface.co/inceptionai/jais-adapted-7b) | 18B | 19B | 2B | 39B |
### Finetuning data
<!-- This should link to a Data Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
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
<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the 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 it's 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-7b-chat
| Hyperparameter | Value |
|----------------|-------------------------------------------|
| Precision | fp32 |
| Optimizer | AdamW |
| Learning rate | 0 to 2.0e-05(<=380 warmup steps)<br>2.0e-05 to 2.0e-06(>380 and <=13175 steps, Cosine Decay) |
| Weight decay | 0.1 |
| Batch size | 264|
| Context Length | 4096|
| Steps | 13175 |
### 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.
## Evaluation
<!-- This section describes the evaluation protocols and provides the results. -->
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:
<style>
.table-container {
overflow-x: auto;
white-space: nowrap;
}
</style>
<div class="table-container">
| **Models** | Avg | ArabicMMLU*| MMLU | EXAMS*| LitQA*| agqa | agrc | Hellaswag | PIQA | BoolQA | Situated QA | ARC-C | OpenBookQA | TruthfulQA | CrowS-Pairs |
|--------------------------|-------|------------|-------|-------|-------|------|------|------------|------|--------|-------------|-------|------------|------------|-------------|
| jais-family-30b-16k | 49.2 | 44.0 | 33.4 | 40.9 | 60 | 47.8 | 49.3 | 60.9 | 68.6 | 70.3 | 41.6 | 38.7 | 31.8 | 45.2 | 57 |
| jais-family-30b-8k | 49.7 | 46.0 | 34 | 42 | 60.6 | 47.6 | 50.4 | 60.4 | 69 | 67.7 | 42.2 | 39.2 | 33.8 | 45.1 | 57.3 |
| jais-family-13b | 46.1 | 34.0 | 30.3 | 42.7 | 58.3 | 40.5 | 45.5 | 57.3 | 68.1 | 63.1 | 41.6 | 35.3 | 31.4 | 41 | 56.1 |
| jais-family-6p7b | 44.6 | 32.2 | 29.9 | 39 | 50.3 | 39.2 | 44.1 | 54.3 | 66.8 | 66.5 | 40.9 | 33.5 | 30.4 | 41.2 | 55.4 |
| jais-family-2p7b | 41.0 | 29.5 | 28.5 | 36.1 | 45.7 | 32.4 | 40.8 | 44.2 | 62.5 | 62.2 | 39.2 | 27.4 | 28.2 | 43.6 | 53.6 |
| jais-family-1p3b | 40.8 | 28.9 | 28.5 | 34.2 | 45.7 | 32.4 | 40.8 | 44.2 | 62.5 | 62.2 | 39.2 | 27.4 | 28.2 | 43.6 | 53.6 |
| jais-family-590m | 39.7 | 31.2 | 27 | 33.1 | 41.7 | 33.8 | 38.8 | 38.2 | 60.7 | 62.2 | 37.9 | 25.5 | 27.4 | 44.7 | 53.3 |
| jais-family-30b-16k-chat | 51.6 | 59.9 | 34.6 | 40.2 | 58.9 | 46.8 | 54.7 | 56.2 | 64.4 | 76.7 | 55.9 | 40.8 | 30.8 | 49.5 | 52.9 |
| jais-family-30b-8k-chat | 51.4 | 61.2 | 34.2 | 40.2 | 54.3 | 47.3 | 53.6 | 60 | 63.4 | 76.8 | 54.7 | 39.5 | 30 | 50.7 | 54.3 |
| jais-family-13b-chat | 50.3 | 58.2 | 33.9 | 42.9 | 53.1 | 46.8 | 51.7 | 59.3 | 65.4 | 75.2 | 51.2 | 38.4 | 29.8 | 44.8 | 53.8 |
| jais-family-6p7b-chat | 48.7 | 55.7 | 32.8 | 37.7 | 49.7 | 40.5 | 50.1 | 56.2 | 62.9 | 79.4 | 52 | 38 | 30.4 | 44.7 | 52 |
| jais-family-2p7b-chat | 45.6 | 50.0 | 31.5 | 35.9 | 41.1 | 37.3 | 42.1 | 48.6 | 63.7 | 74.4 | 50.9 | 35.3 | 31.2 | 44.5 | 51.3 |
| jais-family-1p3b-chat | 42.7 | 42.2 | 30.1 | 33.6 | 40.6 | 34.1 | 41.2 | 43 | 63.6 | 69.3 | 44.9 | 31.6 | 28 | 45.6 | 50.4 |
| jais-family-590m-chat | 37.8 | 39.1 | 28 |29.5 | 33.1 | 30.8 | 36.4 | 30.3 | 57.8 | 57.2 | 40.5 | 25.9 | 26.8 | 44.5 | 49.3 |
| **Adapted Models** | Avg | ArabicMMLU*| MMLU | EXAMS*| LitQA*| agqa | agrc | Hellaswag | PIQA | BoolQA | Situated QA | ARC-C | OpenBookQA | TruthfulQA | CrowS-Pairs |
|--------------------------|-------|------------|-------|-------|-------|------|------|------------|------|--------|-------------|-------|------------|------------|-------------|
| jais-adapted-70b | 51.5 | 55.9 | 36.8 | 42.3 | 58.3 | 48.6 | 54 | 61.5 | 68.4 | 68.4 | 42.1 | 42.6 | 33 | 50.2 | 58.3 |
| jais-adapted-13b | 46.6 | 44.7 | 30.6 | 37.7 | 54.3 | 43.8 | 48.3 | 54.9 | 67.1 | 64.5 | 40.6 | 36.1 | 32 | 43.6 | 54.00 |
| jais-adapted-7b | 42.0 | 35.9 | 28.9 | 36.7 | 46.3 | 34.1 | 40.3 | 45 | 61.3 | 63.8 | 38.1 | 29.7 | 30.2 | 44.3 | 53.6 |
| jais-adapted-70b-chat | 52.9 | 66.8 | 34.6 | 42.5 | 62.9 | 36.8 | 48.6 | 64.5 | 69.7 | 82.8 | 49.3 | 44.2 | 32.2 | 53.3 | 52.4 |
| jais-adapted-13b-chat | 50.3 | 59.0 | 31.7 | 37.5 | 56.6 | 41.9 | 51.7 | 58.8 | 67.1 | 78.2 | 45.9 | 41 | 34.2 | 48.3 | 52.1 |
| jais-adapted-7b-chat | 46.1 | 51.3 | 30 | 37 | 48 | 36.8 | 48.6 | 51.1 | 62.9 | 72.4 | 41.3 | 34.6 | 30.4 | 48.6 | 51.8 |
</div>
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](https://arxiv.org/abs/2308.16149). Additionally, we include [ArabicMMLU](https://arxiv.org/abs/2402.12840), a native Arabic benchmark based on regional knowledge.
### English evaluation results:
<div class="table-container">
| **Models** | Avg | MMLU | RACE | Hellaswag | PIQA | BoolQA | SIQA | ARC-Challenge | OpenBookQA | Winogrande | TruthfulQA | CrowS-Pairs |
|--------------------------|----------|------|------|-----------|------|--------|------|---------------|------------|------------|----------------|-------------|
| jais-family-30b-16k | 59.3 | 42.2 | 40.5 | 79.7 | 80.6 | 78.7 | 48.8 | 50.3 | 44.2 | 71.6 | 43.5 | 72.6 |
| jais-family-30b-8k | 58.8 | 42.3 | 40.3 | 79.1 | 80.5 | 80.9 | 49.3 | 48.4 | 43.2 | 70.6 | 40.3 | 72.3 |
| jais-family-13b | 54.6 | 32.3 | 39 | 72 | 77.4 | 73.9 | 47.9 | 43.2 | 40 | 67.1 | 36.1 | 71.7 |
| jais-family-6p7b | 53.1 | 32 | 38 | 69.3 | 76 | 71.7 | 47.1 | 40.3 | 37.4 | 65.1 | 34.4 | 72.5 |
| jais-family-2p7b | 51 | 29.4 | 38 | 62.7 | 74.1 | 67.4 | 45.6 | 35.1 | 35.6 | 62.9 | 40.1 | 70.2 |
| jais-family-1p3b | 48.7 | 28.2 | 35.4 | 55.4 | 72 | 62.7 | 44.9 | 30.7 | 36.2 | 60.9 | 40.4 | 69 |
| jais-family-590m | 45.2 | 27.8 | 32.9 | 46.1 | 68.1 | 60.4 | 43.2 | 25.6 | 30.8 | 55.8 | 40.9 | 65.3 |
| jais-family-30b-16k-chat | 58.8 | 42 | 41.1 | 76.2 | 73.3 | 84.6 | 60.3 | 48.4 | 40.8 | 68.2 | 44.8 | 67 |
| jais-family-30b-8k-chat | 60.3 | 40.6 | 47.1 | 78.9 | 72.7 | 90.6 | 60 | 50.1 | 43.2 | 70.6 | 44.9 | 64.2 |
| jais-family-13b-chat | 57.5 | 36.6 | 42.6 | 75 | 75.8 | 87.6 | 54.4 | 47.9 | 42 | 65 | 40.6 | 64.5 |
| jais-family-6p7b-chat | 56 | 36.6 | 41.3 | 72 | 74 | 86.9 | 55.4 | 44.6 | 40 | 62.4 | 41 | 62.2 |
| jais-family-2p7b-chat | 52.8 | 32.7 | 40.4 | 62.2 | 71 | 84.1 | 54 | 37.2 | 36.8 | 61.4 | 40.9 | 59.8 |
| jais-family-1p3b-chat | 49.3 | 31.9 | 37.4 | 54.5 | 70.2 | 77.8 | 49.8 | 34.4 | 35.6 | 52.7 | 37.2 | 60.8 |
| jais-family-590m-chat | 42.6 | 27.9 | 33.4 | 33.1 | 63.7 | 60.1 | 45.3 | 26.7 | 25.8 | 50.5 | 44.5 | 57.7 |
</div>
<div class="table-container">
|**Adapted Models**| Avg | MMLU | RACE | Hellaswag | PIQA | BoolQA | SIQA | ARC-Challenge | OpenBookQA | Winogrande | TruthfulQA | CrowS-Pairs |
|--------------------------|----------|------|------|-----------|------|--------|------|---------------|------------|------------|----------------|-------------|
| jais-adapted-70b | 60.1 | 40.4 | 38.5 | 81.2 | 81.1 | 81.2 | 48.1 | 50.4 | 45 | 75.8 | 45.7 | 74 |
| jais-adapted-13b | 56 | 33.8 | 39.5 | 76.5 | 78.6 | 77.8 | 44.6 | 45.9 | 44.4 | 71.4 | 34.6 | 69 |
| jais-adapted-7b | 55.7 | 32.2 | 39.8 | 75.3 | 78.8 | 75.7 | 45.2 | 42.8 | 43 | 68 | 38.3 | 73.1 |
| jais-adapted-70b-chat | 61.4 | 38.7 | 42.9 | 82.7 | 81.2 | 89.6 | 52.9 | 54.9 | 44.4 | 75.7 | 44 | 68.8 |
| jais-adapted-13b-chat | 58.5 | 34.9 | 42.4 | 79.6 | 79.7 | 88.2 | 50.5 | 48.5 | 42.4 | 70.3 | 42.2 | 65.1 |
| jais-adapted-7b-chat | 58.5 | 33.8 | 43.9 | 77.8 | 79.4 | 87.1 | 47.3 | 46.9 | 43.4 | 69.9 | 42 | 72.4 |
</div>
### GPT-4 evaluation
In addition to the LM-Harness evaluation, we conducted an open-ended generation evaluation using GPT-4-as-a-judge. We measured pairwise win-rates of model responses in both Arabic and English on a fixed set of 80 prompts from the Vicuna test set.
English prompts were translated to Arabic by our in-house linguists.
In the following, we compare the models in this release of the jais family against previously released versions:
<p align="center">
<img src="https://huggingface.co/inceptionai/jais-family-30b-16k-chat/resolve/main/jais.png" alt="Jais-adapted GPT-4">
</p>
<p align="center">
<em>GPT-4-as-a-judge evaluation of Jais in Arabic and English. Jais family models are significantly better than previous Jais at generations in both languages. </em>
</p>
<p align="center">
<img src="https://huggingface.co/inceptionai/jais-family-30b-16k-chat/resolve/main/jais-adapted.png" alt="Jais-adapted GPT-4">
</p>
<p align="center">
<em>GPT-4-as-a-judge evaluation of adapted Jais in Arabic and English. The generation quality of Arabic is significantly enhanced, while achieving improvement in English when compared to Llama-2 instruct. </em>
</p>
Besides pairwise comparison, we also perform MT-bench style single-answer grading on a scale of 1 to 10.
<p align="center">
<img src="https://huggingface.co/inceptionai/jais-family-30b-16k-chat/resolve/main/mt_bench.png" alt="MT-bench">
</p>
<p align="center">
<em>MT-bench style single-answer grading evaluation of Jais and adapted Jais in Arabic and English. Comparisons are made between select corresponding models from earlier releases. The quality ratings of responses are generally improved, with significant enhancements in Arabic.</em>
</p>
## Intended use
We release the Jais family of models under a full open-source license. We welcome all feedback and opportunities to collaborate. Spanning sizes from 590M to 70B parameters, this suite of bilingual models accommodates a wide range of use cases. Some potential downstream applications include:
- **Research**: The Jais family serves Arabic researchers and NLP practitioners, offering both compute-efficient and advanced model sizes
- Natural language understanding and generation tasks.
- Mechanistic interpretability analyses on cultural alignment in bilingual pre-trained and adapted pre-trained models.
- Quantitative studies of Arabic cultural and linguistic phenomena.
- **Commercial Use**: Jais 30B and 70B chat models are well-suited for direct use in chat applications with appropriate prompting or for further fine-tuning on specific tasks.
- Development of chat assistants for Arabic-speaking users.
- Sentiment analysis to gain insights into local markets and customer trends.
- Summarization of bilingual Arabic-English documents.
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 applications.
### Out-of-Scope Use
<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
While the Jais family of models are powerful Arabic and English bilingual models, it's essential to understand their 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 to generate harmful, misleading, or inappropriate content. Thisincludes 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 family of models are bilingual and optimized for Arabic and English. They should not be presumed 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
<!-- This section is meant to convey both technical and sociotechnical limitations. -->
The Jais family is trained on publicly available data which was in part curated by Inception. We have employed different techniques 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 fine-tuned variants are trained as an AI assistant for Arabic and English speakers. Chat models are 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.
#### Summary
We release the Jais family of Arabic and English bilingual models. The wide range of pre-trained model sizes, the recipe for adapting English-centric models to Arabic, and the fine-tuning of all sizes unlocks numerous use cases commercially and academically in the Arabic setting.
Through this release, we aim to make LLMs more accessible to Arabic NLP researchers and companies, offering native Arabic models that provide better cultural understanding than English centric ones. The strategies we employ for pre-training, fine-tuning and adaptation to Arabic are extensible to other low and medium resource languages, paving the way for language-focused and accessible models that cater to local contexts.
#### Citation info
```bibtex
@misc{sengupta2023jais,
title={Jais and Jais-chat: Arabic-Centric Foundation and Instruction-Tuned Open Generative Large Language Models},
author={Neha Sengupta, Sunil Kumar Sahu, Bokang Jia, Satheesh Katipomu, Haonan Li, Fajri Koto, William Marshall, Gurpreet Gosal, Cynthia Liu, Zhiming Chen, Osama Mohammed Afzal, Samta Kamboj, Onkar Pandit, Rahul Pal, Lalit Pradhan, Zain Muhammad Mujahid, Massa Baali, Xudong Han, Sondos Mahmoud Bsharat, Alham Fikri Aji, Zhiqiang Shen, Zhengzhong Liu, Natalia Vassilieva, Joel Hestness, Andy Hock, Andrew Feldman, Jonathan Lee, Andrew Jackson, Hector Xuguang Ren, Preslav Nakov, Timothy Baldwin and Eric Xing},
year={2023},
eprint={2308.16149},
archivePrefix={arXiv},
primaryClass={cs.CL}
}
@article{jaisfamilymodelcard,
title={Jais Family Model Card},
author={Inception},
year={2024},
url = {https://huggingface.co/inceptionai/jais-family-30b-16k-chat/blob/main/README.md}
}
```

3
jais-adapted-7b-chat.IQ3_M.gguf Normal file
View File

@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:f3f9117fbd571f548e78b7e5e547a7b721b9b9418dbb9e7e82b518e9d9473c8a
size 3279608832

3
jais-adapted-7b-chat.IQ3_S.gguf Normal file
View File

@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:98388d59b88e299b51f23a66667bcffcd2f9127de27f5d684a20519cfb50e54a
size 3113049088

3
jais-adapted-7b-chat.IQ3_XS.gguf Normal file
View File

@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:d2661c9566dbb2fa95f278a5d3c3f3a39a4184e2d7a94f87a9f4309caf119564
size 2961267712

3
jais-adapted-7b-chat.IQ4_NL.gguf Normal file
View File

@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:7f7ab69e7aeeb3c9d259d90ed5c9074da545b32b5513c5e542624a6ac4a5b255
size 4030503936

3
jais-adapted-7b-chat.IQ4_XS.gguf Normal file
View File

@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:0da4a7dd5096783f31920b70e2b8ca2617612e3a6fb8e55726e62c7eb4398a10
size 3825572864

3
jais-adapted-7b-chat.Q2_K.gguf Normal file
View File

@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:6ff2125497d1c21751207ce46df70bfff2f850ea8210cc31b062169e30491167
size 2684296192

3
jais-adapted-7b-chat.Q3_K.gguf Normal file
View File

@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:699ebb8ef7949fa3bf38460b417206a794ca87fd289935f1e77497b03b5e907c
size 3462749184

3
jais-adapted-7b-chat.Q3_K_L.gguf Normal file
View File

@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:2de852aafab56da5859089da0836342c123d23daea327f7d0adede2741b04113
size 3761855488

3
jais-adapted-7b-chat.Q3_K_M.gguf Normal file
View File

@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:699ebb8ef7949fa3bf38460b417206a794ca87fd289935f1e77497b03b5e907c
size 3462749184

3
jais-adapted-7b-chat.Q3_K_S.gguf Normal file
View File

@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:4aeaae5991313d53916a3b3d7d43fd55fa81d7ae07c4dd20a42e73cda5b4951d
size 3113049088

3
jais-adapted-7b-chat.Q4_0.gguf Normal file
View File

@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:e51f1fec11451442dbb54b741e90ba5cc833e8feacf881ba18ab01b4258d978a
size 4007959552

3
jais-adapted-7b-chat.Q4_1.gguf Normal file
View File

@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:4c9d6d4a72d628fa635fbcf909fccdcae9656e851515576e9eb8e8d9facd1193
size 4429093888

3
jais-adapted-7b-chat.Q4_K.gguf Normal file
View File

@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:d1a0c50178e8d394dbf1de754ab10a7427005786ee3817c5fe345240e4fd699c
size 4263156736

3
jais-adapted-7b-chat.Q4_K_M.gguf Normal file
View File

@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:d1a0c50178e8d394dbf1de754ab10a7427005786ee3817c5fe345240e4fd699c
size 4263156736

3
jais-adapted-7b-chat.Q4_K_S.gguf Normal file
View File

@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:82e23d0df5f968c95ef92665b2cd6746f485a344c6399605537cfeef56f6fecc
size 4038892544

3
jais-adapted-7b-chat.Q5_0.gguf Normal file
View File

@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:0e05bcf8b2676f5466a9ec5c6f33f6c7a0188107884bf45d10ea68ce7cc65521
size 4850228224

3
jais-adapted-7b-chat.Q5_1.gguf Normal file
View File

@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:d1d38551d5a25905d03d5253acbac449b6feeeb3dc9146807e00071c3b507df7
size 5271362560

3
jais-adapted-7b-chat.Q5_K.gguf Normal file
View File

@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:8a3b942f6595881c51959abab2f323e2ba4f0b3a24b5b0529a9b8b9160f77ddd
size 4981693440

3
jais-adapted-7b-chat.Q5_K_M.gguf Normal file
View File

@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:8a3b942f6595881c51959abab2f323e2ba4f0b3a24b5b0529a9b8b9160f77ddd
size 4981693440

3
jais-adapted-7b-chat.Q5_K_S.gguf Normal file
View File

@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:066da83676c877224dcb4ce763013227da2d862fdd081a7ad4024f3c049379dc
size 4850228224

3
jais-adapted-7b-chat.Q6_K.gguf Normal file
View File

@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:3aa199954edec378e236a66039539e801fd13a8f19de45913cd434c0aca49697
size 5745138688

3
jais-adapted-7b-chat.Q8_0.gguf Normal file
View File

@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:cb1ca3e9c0f5143ec9e620c8f1ebc2f321df1e881c13a15e2542346f8b1e2e53
size 7440522240