XVERSE-MoE-A4.2B

更新信息

• [2024/04/28] 发布 MoE 架构的 XVERSE-MoE-A4.2B-Chat 对话模型。
• [2024/04/02] 发布 MoE 架构的 XVERSE-MoE-A4.2B 底座模型，Chat 对齐模型将在后续发布。

Update Information

• [2024/04/28] Released XVERSE-MoE-A4.2B-Chat MoE chat model.
• [2024/04/02] Released XVERSE-MoE-A4.2B MoE base model, the Chat version model will be released later.

模型介绍

XVERSE-MoE-A4.2B 是由深圳元象科技自主研发的支持多语言的大语言模型（Large Language Model），使用混合专家模型（MoE，Mixture-of-experts）架构，模型的总参数规模为 258 亿，实际激活的参数量为 42 亿，本次开源的模型为底座模型 XVERSE-MoE-A4.2B，主要特点如下：

• 模型结构：XVERSE-MoE-A4.2B 为 Decoder-only 的 Transformer 架构，将密集模型的 FFN 层扩展为专家层，不同于传统 MoE 中每个专家的大小与标准 FFN 相同（如Mixtral 8x7B ），使用了更细粒度的专家，每个专家是标准 FFN 大小的 1/4，并设置了共享专家（Shared Expert）和非共享专家（Non-shared Expert）两类，共享专家在计算时始终被激活，非共享专家通过 Router 选择性激活。
• 训练数据：构建了 2.7 万亿 token 的高质量、多样化的数据对模型进行充分训练，包含中、英、俄、西等 40 多种语言，通过精细化设置不同类型数据的采样比例，使得中英两种语言表现优异，也能兼顾其他语言效果；模型使用 8K 长度的训练样本进行训练。
• 训练框架：针对 MoE 模型中独有的专家路由和权重计算逻辑，进行了深入定制优化，开发出一套高效的融合算子，以提升计算效率。同时，为解决 MoE 模型显存占用和通信量大的挑战，设计了计算、通信和 CPU-Offload 的 Overlap 处理方式，从而提高整体吞吐量。

XVERSE-MoE-A4.2B 的模型大小、架构和学习率如下：

total params	activated params	n_layers	d_model	n_heads	d_ff	n_non_shared_experts	n_shared_experts	top_k	lr
25.8B	4.2B	28	2560	32	1728	64	2	6	3.5e−4

Model Introduction

XVERSE-MoE-A4.2B is a multilingual large language model, independently developed by Shenzhen Yuanxiang Technology which is using Mixture-of-experts (MoE) architecture. The total parameter scale of the model is 25.8 billion, with an actual number of activated parameters being 4.2 billion. The models released this time is the base model XVERSE-MoE-A4.2B. Its key features are as follows:

• Model Structure: XVERSE-MoE-A4.2B uses the mainstream Decoder-only Transformer network structure that extends the FFN layer of dense models to expert layers. Unlike traditional MoE model where each expert has the same size as standard FFN (such as Mixtral 8x7B), it uses more fine-grained experts, with each expert being 1/4 the size of a standard FFN. It includes shared experts and non-shared experts, where shared experts are always activated during computation, and non-shared experts are selectively activated through a Router.
• Training Data: The model has been thoroughly trained on a diversified and high-quality dataset consisting of 3.2 trillion of tokens, including more than 40 languages such as Chinese, English, Russian, and Spanish. The sampling ratio of different types of data is finely set, which makes the performance of Chinese and English excellent, and also takes into account the effect of other languages; The model is trained using training samples of length 8k.
• Training Framework: We conducted in-depth customized optimization for the unique expert routing and weight calculation logic in the MoE model, developed an efficient fusion operator to improve computational efficiency. At the same time, to address the challenges of high memory consumption and communication volume in the MoE model, we designed a processing method for overlapping computation, communication, and CPU-Offload to increase overall throughput.

The models sizes, architectures and learning rate of XVERSE-MoE-A4.2B are showed as follows:

total params	activated params	n_layers	d_model	n_heads	d_ff	n_non_shared_experts	n_shared_experts	top_k	lr
25.8B	4.2B	28	2560	32	1728	64	2	6	3.5e−4

评测结果

为了综合评估模型的性能，我们在一系列标准数据集上进行了全面测试，包括C-Eval、CMMLU、Gaokao-Bench、MMLU、AGIEval、RACE-M、CommonSenseQA、PIQA、GSM8K和HumanEval。这些评估覆盖了模型在多个领域的能力，具体包括中文问答、英文问答、语言理解、常识问答、逻辑推理、数学问题解答以及编程能力。评估结果如下：

数据集	XVERSE-MoE-A4.2B	XVERSE-13B-2	Baichuan2-13B	Llama2-13B	Llama1-65B	XVERSE-7B	DeepSeek-7B	Mistral-7B	Gemma-7B	DeepSeek-MoE-16B
C-Eval	60.5	62.0	58.1	35.6	38.8	57.1	45.0	45.1	50.0	40.6
CMMLU	64.5	65.4	62.0	38.4	40.6	61.3	47.2	44.9	50.5	42.5
Gaokao-Bench1	60.3	65.3	54.3	35.4	38.9	61.7	35.4	40.2	42.3	29.1
MMLU	60.2	60.0	59.2	54.8	63.4	56.6	48.2	62.5	64.3	45
AGIEval1	48.0	52.4	48.2	33.4	42.4	46.9	26.4	41.2	41.7	31.7
RACE-M	75.4	82.4	68.9	63.0	67.9	79.0	63.2	67.5	80.2	61.9
CommonSenseQA	70.0	68.0	65.6	67.3	74.0	64.1	56.4	68.8	74.0	54.8
PIQA	81.4	79.8	78.5	80.5	82.8	76.7	79.2	82.2	81.2	80.2
GSM8K	51.2	52.7	52.7	28.7	50.9	19.3	17.4	35.4	46.4	18.8
HumanEval	29.9	32.3	17.1	18.3	23.7	10.4	26.2	26.2	32.3	26.8

^{1：只针对其中的单项选择题进行测试，即排除了填空题、开放性问题和多项选择题}

对于上述所有比较模型，我们优先汇报其官方公布的结果。在缺少官方结果的情况下，我们采用了 OpenCompass 榜单的报告结果。其他结果则来自于我们自行执行的评估流程所获得的数据。
对于 MMLU ，我们采用作者提供的评测工具，C-Eval、AGIEval、GAOKAO-Bench 与 MMLU 的评测方式相同，其余评测数据集使用 OpenCompass 评估框架进行评估。

Model Evaluation

To comprehensively assess the performance of the model, we conducted extensive testing across a range of standard datasets, including C-Eval, CMMLU, Gaokao-Bench, MMLU, AGIEval, RACE-M, CommonSenseQA, PIQA, GSM8K and HumanEval. These evaluations spanned multiple capabilities of the model, specifically including Chinese question answering, English question answering, language comprehension, common sense questioning, logical reasoning, mathematical problem-solving, and coding ability. The results of the evaluations are as follows:

Dataset	XVERSE-MoE-A4.2B	XVERSE-13B-2	Baichuan2-13B	Llama2-13B	Llama1-65B	XVERSE-7B	DeepSeek-7B	Mistral-7B	Gemma-7B	DeepSeek-MoE-16B
C-Eval	60.5	62.0	58.1	35.6	38.8	57.1	45.0	45.1	50.0	40.6
CMMLU	64.5	65.4	62.0	38.4	40.6	61.3	47.2	44.9	50.5	42.5
Gaokao-Bench1	60.3	65.3	54.3	35.4	38.9	61.7	35.4	40.2	42.3	29.1
MMLU	60.2	60.0	59.2	54.8	63.4	56.6	48.2	62.5	64.3	45
AGIEval1	48.0	52.4	48.2	33.4	42.4	46.9	26.4	41.2	41.7	31.7
RACE-M	75.4	82.4	68.9	63.0	67.9	79.0	63.2	67.5	80.2	61.9
CommonSenseQA	70.0	68.0	65.6	67.3	74.0	64.1	56.4	68.8	74.0	54.8
PIQA	81.4	79.8	78.5	80.5	82.8	76.7	79.2	82.2	81.2	80.2
GSM8K	51.2	52.7	52.7	28.7	50.9	19.3	17.4	35.4	46.4	18.8
HumanEval	29.9	32.3	17.1	18.3	23.7	10.4	26.2	26.2	32.3	26.8

^{1: Tests are conducted only on single-answer multiple-choice questions, thus excluding fill-in-the-blanks, open-ended questions, and multiple-answer multiple-choice questions.}

For all the comparison models mentioned above, we prioritize the disclosure of their officially published results. In the absence of official data, we refer to the reported outcomes from OpenCompass Leaderboard. Results not covered by the aforementioned sources are derived from our own evaluation pipline.
For MMLU, we adopt the evaluation tools provided by the authors, C-Eval, AGIEval, GAOKAO-Bench are the same as MMLU. For the remaining evaluation datasets, the OpenCompass is employed for evaluation.

使用方法

Transformers 加载方式

可通过以下代码加载 XVERSE-MoE-A4.2B 模型来进行推理：

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
tokenizer = AutoTokenizer.from_pretrained("xverse/XVERSE-MoE-A4.2B")
model = AutoModelForCausalLM.from_pretrained("xverse/XVERSE-MoE-A4.2B", trust_remote_code=True, torch_dtype=torch.bfloat16, device_map='auto')
model = model.eval()
inputs = tokenizer('北京的景点：故宫、天坛、万里长城等。\n深圳的景点：', return_tensors='pt').input_ids
inputs = inputs.cuda()
generated_ids = model.generate(inputs, max_new_tokens=64, eos_token_id=tokenizer.eos_token_id, repetition_penalty=1.1)
print(tokenizer.batch_decode(generated_ids, skip_special_tokens=True))

Usage

Loading with Transformers

The XVERSE-MoE-A4.2B model can be loaded for inference using the following code:

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
tokenizer = AutoTokenizer.from_pretrained("xverse/XVERSE-MoE-A4.2B")
model = AutoModelForCausalLM.from_pretrained("xverse/XVERSE-MoE-A4.2B", trust_remote_code=True, torch_dtype=torch.bfloat16, device_map='auto')
model = model.eval()
inputs = tokenizer('北京的景点：故宫、天坛、万里长城等。\n深圳的景点：', return_tensors='pt').input_ids
inputs = inputs.cuda()
generated_ids = model.generate(inputs, max_new_tokens=64, eos_token_id=tokenizer.eos_token_id, repetition_penalty=1.1)
print(tokenizer.batch_decode(generated_ids, skip_special_tokens=True))

局限性与免责申明

XVERSE-MoE-A4.2B 与其他所有 LLM 一样，在某些情况下可能会产生不准确、有偏见或其他令人反感的内容。因此，请谨慎使用模型生成的内容，请勿将生成的有害内容进行传播，在部署任何 XVERSE-MoE-A4.2B 的应用之前，开发人员应根据其具体应用对模型进行安全测试和调优。

我们强烈警告不要将 XVERSE-MoE-A4.2B 模型用于制造或传播有害信息，或进行任何可能损害公众、国家、社会安全或违反法规的活动。如果使用 XVERSE-MoE-A4.2B 模型产生任何问题，无论是数据安全问题、公共舆论风险，还是模型被误解、滥用、传播或不合规使用所引发的任何风险和问题，我们将不承担任何责任。

模型开源协议

使用本仓库的源码需要遵循 Apache-2.0 开源协议，使用 XVERSE-MoE-A4.2B 的模型权重则需要遵循模型许可协议。

XVERSE-MoE-A4.2B 模型权重对学术研究完全开放，并且支持免费商用。如需申请商业许可证，请填写【申请表】，如有其他问题或合作，请联系 opensource@xverse.cn。

Limitations and Disclaimer

Like all other Large Language Models (LLMs), XVERSE-MoE-A4.2B may produce inaccurate, biased, or otherwise offensive content under certain circumstances. Therefore, please use the content generated by the model with caution and refrain from disseminating harmful content. Before deploying any application of XVERSE-MoE-A4.2B, developers should conduct safety tests and optimization of the model according to its specific application.

We strongly warn against the use of the XVERSE-MoE-A4.2B model for producing or spreading harmful information, or conducting any activities that might harm the public, national, or social security, or violate regulations. We assume no responsibility for any problems arising from the use of the XVERSE-MoE-A4.2B model, whether it be data security issues, public opinion risks, or any risks and issues caused by misunderstanding, misuse, dissemination, or non-compliance with the model.

Open Source License

The use of the source code in this repository must follow the Apache-2.0 open-source license, while the use of the model weights of XVERSE-MoE-A4.2B needs to adhere to the Model License Agreement.

The XVERSE-MoE-A4.2B model weights are fully open to academic research and support free commercial use. To apply for a commercial license, please fill in the application form. For other questions or collaborations, please contact opensource@xverse.cn.