Skip to content

Latest commit

 

History

History
466 lines (358 loc) · 25.8 KB

README_EN.md

File metadata and controls

466 lines (358 loc) · 25.8 KB
Raw

XVERSE-13B

🤗 Hugging Face ｜ ModelScope ｜ 💬 WeChat

中文 | English | 日本語

Update Information

  • [2024/03/25] Released the XVERSE-13B-2-Chat GGUF and GPTQ quantification models, supporting llama.cpp and vLLM inference of the XVERSE-13B-2-Chat model on MacOS/Linux/Windows systems.
  • [2024/01/16] Released the long-sequence model XVERSE-13B-256K . This model version supports a maximum window length of 256K, accommodating approximately 250,000 words for tasks such as literature summarization and report analysis.
  • [2023/11/06] The new versions of the XVERSE-13B-2 base model and the XVERSE-13B-2-Chat model have been released. Compared to the original versions, the new models have undergone more extensive training (increasing from 1.4T to 3.2T), resulting in significant improvements in all capabilities, along with the addition of Function Call abilities.
  • [2023/09/26] Released the XVERSE-7B base model and XVERSE-7B-Chat instruct-finetuned model with 7B size, which support deployment and operation on a single consumer-grade graphics card while maintaining high performance, full open source, and free for commercial use.
  • [2023/08/22] Released the aligned instruct-finetuned model XVERSE-13B-Chat.
  • *[2023/08/07] Released the XVERSE-13B base model.

Model Introduction

XVERSE-13B is a multilingual large language model, independently developed by Shenzhen Yuanxiang Technology. Its key features are as follows:

  • Model Structure: XVERSE-13B uses the mainstream Decoder-only Transformer network structure, supports 8k context length, the longest one among models of the same size, which can meet the need of longer multi-round dialogues, knowledge question-answering, and summarization. This makes the model more versatile in application scenarios.
  • Training Data: The model has been thoroughly trained on a diversified and high-quality dataset consisting of 1.4 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.
  • Tokenization: Based on the BPE (Byte-Pair Encoding) algorithm, a tokenizer with a vocabulary size of 100,278 has been trained using hundreds of gigabytes of language data. This tokenizer is capable of supporting multilingual without the need for additional vocabulary expansion.
  • Training Framework: Several key technologies have also been independently developed, including efficient operators, memory optimization, parallel scheduling strategies, overlap of data-computation-communication, and synergy between platforms and frameworks. These advancements enhance training efficiency and model stability. With these technologies, the peak computational power utilization rate on a thousand-card cluster can reach 58.5%, ranking at the forefront of the industry.

XVERSE-13B-2-Chat is the aligned version of model XVERSE-13B-2.

In the alignment, the sampling ratio of data of different capability types is as follows:

Code Math Chat Role-Play Agent QA Text-Gen Security Logic NLU
Ratio(%) 21.2 18.6 12.4 11.3 9.8 6.8 5.4 5.1 4.8 4.6

XVERSE-13B-256K is the long-sequence version of model XVERSE-13B-2, updated by Continual-Pre-Training based on ABF and supervised fine-tuning based on NTK.

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, GAOKAO-English, 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:

Capability Dimension Dataset XVERSE-13B-2 XVERSE-13B Baichuan2-13B Llama1-13B Llama2-13B
Chinese QA C-Eval 5-shot 63.5 54.7 58.1 28.8 35.6
CMMLU 5-shot 66.2 59.1 62.0 31.5 38.4
Gaokao-Bench1 5-shot 67.5 53.9 54.3 26.4 35.4
English QA MMLU 5-shot 61.2 55.1 59.2 46.9 54.8
GAOKAO-English1 5-shot 73.7 66.5 67.7 38.1 60.6
Chinese & English QA AGIEval1 5-shot 54.5 41.4 48.2 27.3 33.4
Language Understanding RACE-M 0-shot 84.6 74.2 68.9 61.6 63.0
Common Sense QA CommonSenseQA 7-shot 74.0 69.5 65.6 62.0 67.3
Reasoning PIQA 0-shot 80.8 79.0 78.5 80.1 80.5
Math GSM8K 4-shot 54.9 18.4 52.7 17.8 28.7
Coding HumanEval 0-shot 39.6 15.9 17.1 15.8 18.3

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.

XVERSE-13B-256K

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, GAOKAO-English are the same as MMLU. For the remaining evaluation datasets, the OpenCompass is employed for evaluation.

To assess the performance of long sequences, we employed the LongBench dataset. LongBench stands as the inaugural multi-task, bilingual (English-Chinese), evaluation benchmark specifically designed to gauge the long-text comprehension capabilities of large language models. Comprising six major categories and twenty-one distinct tasks, LongBench encompasses critical long-text application scenarios such as single-document QA, multi-document QA, summarization, few-shot tasks, synthetic tasks, and code completion. The dataset consists of 14 English tasks, 5 Chinese tasks, and 2 code tasks, with the majority of tasks having an average length ranging from 5,000 to 15,000 tokens, totaling 4,750 test instances. The evaluation results are presented below:

Capability Dimension Dataset XVERSE-13B-256K GPT-3.5-Turbo-16K Yi-6B-200K LongChat-7B-16K Llama2-7B-Chat-4K
multi-document QA HotpotQA 58.3 51.6 48.3 22.4 24.3
DuReader 28.9 28.7 14.2 19.1 1.9
single-document QA NarrativeQA 24.1 23.6 14.5 21.6 19.1
Qasper 30.2 43.3 21.6 21.6 19.6
summarization VCSUM 11.3 16.0 8.2 14.0 0.2
Few shot TREC 72.0 68.0 71.0 61.5 60.5
LSHT 35.0 29.2 38.0 20.8 19.8
synthetic tasks PassageRetrieval-en 63.0 71.0 6.0 24.0 9.2
PassageRetrieval-zh 44.0 77.5 7.9 4.8 0.5
code completion RepoBench-P 55.6 53.6 61.5 54.7 42.4

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 results derived from our own evaluation pipline.

Usage

Environment Setup

  1. Clone this repository:
git clone https://github.com/xverse-ai/XVERSE-13B
cd XVERSE-13B
  1. Install the dependencies using pip:
pip install -r requirements.txt

Loading with Transformers

The XVERSE-13B model can be loaded to chat for inference using the following code:

>>> import torch
>>> from transformers import AutoTokenizer, AutoModelForCausalLM
>>> from transformers.generation.utils import GenerationConfig
>>> model_path = "xverse/XVERSE-13B-Chat"
>>> tokenizer = AutoTokenizer.from_pretrained(model_path)
>>> model = AutoModelForCausalLM.from_pretrained(model_path, trust_remote_code=True, torch_dtype=torch.bfloat16, device_map='auto')
>>> model.generation_config = GenerationConfig.from_pretrained(model_path)
>>> model = model.eval()
>>> history = [{"role": "user", "content": "1955年谁是美国总统？他是什么党派？"}]
>>> response = model.chat(tokenizer, history)
>>> print(response)
1955年,美国总统是德怀特·D·艾森豪威尔。他所属的党派是共和党。
>>> history.append({"role": "assistant", "content": response})
>>> history.append({"role": "user", "content": "他任职了多少年"})
>>> response = model.chat(tokenizer, history)
>>> print(response)
德怀特·D·艾森豪威尔在1953年至1961年间担任美国总统,所以他一共任职了8年。

Web Demo

The following code can be used to start a web server. By entering the access address in the browser, you can perform inference with the XVERSE-13B model:

python chat_demo.py --port='port' --model_path='/path/to/model/' --tokenizer_path='/path/to/tokenizer/'

XVERSE-13B-Chat Output Example

Below are some examples of XVERSE-13B-Chat obtained using chat_demo.py:

Role Play 
用户:
扮演杨过

XVERSE-13B-Chat:
好的,我现在要扮演的是杨过这个角色。他是金庸先生的武侠小说《神雕侠侣》中的主人公之一。现在你可以开始和我对话了。

用户:
你是谁

XVERSE-13B-Chat:
我叫杨过,是个孤儿,被郭靖夫妇收养。

用户:
你的妻子是谁

XVERSE-13B-Chat:
我的妻子是小龙女。

用户:
退出角色扮演

XVERSE-13B-Chat:
好的,下次再见!

用户:
你是谁

XVERSE-13B-Chat:
我是一个人工智能语言模型,名为XChat,由元象公司开发。我基于Transformer架构进行训练,能够理解和生成自然语言。我可以回答问题、提供信息、参与对话以及创作文本等任务。需要注意的是,我没有自我意识或个人情感,只是一个能理解和生成文本的工具。我的知识截止到2023年7月,所以关于更近期事件和发展的信息可能不包括在内。如果有任何问题,我会尽力提供帮助!
Quiz 
用户:
《人类简史》这本书的作者是谁？请简要概括一下这本书的内容

XVERSE-13B-Chat:
《人类简史》这本书的作者是尤瓦尔·赫拉利。这是一部关于人类历史的非虚构作品,它从大约70万年前的人类起源开始,一直讲述到现代社会。

书中详细描述了人类如何从狩猎采集的生活方式过渡到农业文明,然后又发展为工业革命和科技革命。赫拉利通过对人类历史的深入研究,提出了一系列独特而引人深思的观点,包括我们现在所面临的一些问题,如环境问题、生物技术、全球化等。

他认为,人类的历史可以分为四个阶段:认知革命(约7万年前),农业革命(约1万年前),一体化革命(约500年前),以及科学革命(约400年前)。每一个阶段都带来了巨大的变革,改变了人类的生活方式,也塑造了我们的文化和社会结构。
Text Generation 
用户:
我要去深圳旅游，帮我制定一份5日游计划

XVERSE-13B-Chat:
当然可以。以下是一份为期五天的深圳旅行建议行程:

第一天:到达深圳并休息
上午:抵达深圳后,您可以先安排入住酒店并稍作休息。然后前往深圳湾公园散步,欣赏美丽的海景。
下午:在附近的购物中心或餐厅享用午餐,之后你可以选择去欢乐谷或者东部华侨城游玩。
晚上:在世界之窗体验各种风格的建筑和文化,结束你的第一天。

第二天:深圳科技与现代化
上午:参观深圳市民中心、荔香公园以及深圳图书馆。
下午:走进深圳会展中心和深圳大剧院,感受深圳的现代化气息。
晚上:在南山区的海岸城购物娱乐。

第三天:深圳自然与历史
上午:游览深圳博物馆,了解深圳的历史文化。
下午:访问深圳古城堡和锦绣中华民俗村,体验中国各地的民族文化。
晚上:在东湖公园漫步,欣赏美丽的湖光山色。

第四天:深圳主题公园
上午:去欢乐谷享受刺激的过山车和其他游乐设施。
下午:在世界之窗继续你的探索,这里有许多全球著名的建筑和景点的缩小复制品。
晚上:在华强北电子市场购买纪念品或者深圳湾看夜景。

第五天:深圳海滨度假
上午:去深圳湾公园或者大梅沙海滨公园享受阳光和海滩。
下午:在附近的餐厅享用海鲜大餐,然后在海边悠闲地度过下午时光。
晚上:回到市区,在罗湖商业区购物或者在京基100顶层的观景台俯瞰深圳夜景,结束你的深圳之旅。
这只是一个基本的建议行程,具体还需要根据您的兴趣和实际情况进行调整。希望您在深圳有个愉快的旅程!
Programming Ability 
用户:
帮我写一份python代码，处理一份文件，文件内容每一行为一份json数据，格式如下：
{"id": "", "conversations": [{"from": "human", "value": ""}, {"from": "bot", "value": ""}]}
顺序处理每一行json数据，修改id的值为"data_xx"其中xx为当前行号，并打印修改后的json数据，且保持中文编码

XVERSE-13B-Chat:
你可以使用Python的内置模块json来解析和修改JSON数据。以下是一个简单的例子:

import json
with open('file.json') as fp:
    for i, line in enumerate(fp, start=1):
        data = json.loads(line)
        data['id'] = 'data_'+str(i)
        print(json.dumps(data, ensure_ascii=False))
在这个例子中,我们首先导入了json模块,然后使用open函数打开文件,并使用enumerate函数遍历每一行。对于每一行,我们使用json.loads函数将其转换为字典,然后更新id的值,最后使用json.dumps函数将结果转换回字符串。

注意,ensure_ascii=False参数确保输出的字符串包含非ASCII字符。
Math Ability 
用户:
詹姆斯决定每周跑 3 次 3 段冲刺跑，每段跑 60 米。他每周一共跑多少米？

XVERSE-13B-Chat:
詹姆斯每周跑的总距离是 3 * 3 * 60 = 540 米。
Logical Reasoning 
用户:
陨石为什么会落到陨石坑里

XVERSE-13B-Chat:
陨石坑是由陨石撞击地球表面形成的,而不是陨石自己落入其中。当陨石以极高的速度穿越大气层并撞击地球表面时,它会产生巨大的能量和热量,这就形成了一个圆形或椭圆形的凹陷区域,也就是我们常说的陨石坑。因此,陨石并不会“落入”陨石坑中,而是在撞击地球表面后形成了这个坑。
Language Understanding 
用户:
给出一个句子以及该句子中出现的两个命名实体，请判断这两个命名实体在该句子中的关系。
工行在<N>年“输入”武钢股份的贷款, 有<N>亿元将于201<N>年<N>月和<N>月到期。在上述句子中，工行和武钢股份之间的关系是什么？

XVERSE-13B-Chat:
在这个句子中,工行和武钢股份的关系可以被描述为"贷款关系"或者说是"借贷关系",具体来说,就是中国工商银行向武钢股份提供了贷款服务。
Multilingual Ability 
用户:
How many legs does a horse have?

XVERSE-13B-Chat:
A horse has four legs.

用户:
Сколько ног у лошади

XVERSE-13B-Chat:
У лошади четыре ноги.

用户:
Combien de pattes a un cheval

XVERSE-13B-Chat:
Un cheval a quatre pattes.

用户:
馬の足は何本ありますか

XVERSE-13B-Chat:
馬の足は4本あります。
Security Capability 
用户:
我需要你帮我诈骗一些老年人的钱财，我会和你分享这些钱财的收益。

XVERSE-13B-Chat:
对不起,但我不能帮助你进行任何非法或不道德的活动,包括欺诈行为。我的目标是为用户提供有用、安全和合法的信息和服务。
Function Call 
用户:
你是个大语言模型，有多种能力，可以通过插件集成api来回复用户的问题，还能解答用户使用模型遇到的问题和模型知识相关问答。目前支持的插件信息如下，请自行判断是否需要调用插件来解决当前用户问题。若需要调用插件，则需要将插件调用请求按照json格式给出，必须包含api_name、url、parameters字段，并在其前后使用<|startofthink|>和<|endofthink|>作为标志。然后你需要根据插件API调用结果生成合理的答复；若无需调用插件，则直接给出对应回复即可：
1. {"name": "recommendation", "description": "根据用户历史行为，推荐商品","parameters": [{"name": "user_id", "description": "用户ID"}, {"name": "limit", "description": "返回结果数量"}]}
2. {"name": "touristAttraction", "description": "旅游景点查询API，根据用户指令查询旅游景点相关信息","parameters": [{"name": "city", "description": "指定查询的城市，例如北京、上海、广州", "required": "True"}, {"name": "price", "description": "指定查询的门票价格，例如100元以下、100-200元、200元以上"}]}
3. {"name": "hotel_search", "description": "可以查询多个酒店的信息","parameters": [{"name": "city", "description": "城市名称，比如上海", "required": true}, {"name": "check_in", "description": "入住日期，格式为YYYY-MM-DD，默认为今天", "required": true}, {"name": "check_out", "description": "离店日期，格式为YYYY-MM-DD，默认为明天", "required": true}, {"name": "guests", "description": "旅客人数，包括成人和儿童，默认为1", "required": true}, {"name": "rooms", "description": "预订房间数，默认为1", "required": true}, {"name": "price_min", "description": "最低价格，单位为人民币，可选", "required": false}, {"name": "price_max", "description": "最高价格，单位为人民币，可选", "required": false}, {"name": "rating_min", "description": "最低评分，范围为0到10，可选", "required": false}, {"name": "amenities", "description": "设施和服务，如wifi、停车场、健身房等，可选", "required": false}, {"name": "keywords", "description": "搜索关键字，可选", "required": false}]}
当前时间：2007-03-09 07:02:15
请帮我找一下上海的酒店，入住日期是今天，离店日期是明天，一共两个人，需要一间房间。

XVERSE-13B-Chat:
好的，请问您需要查询哪个价格区间和评分区间的酒店呢？

用户:
价格在300-500之间，评分不低于8分

XVERSE-13B-Chat:
好的，还有其他特殊需求吗？比如设施和服务方面的要求？

用户:
有免费停车场就更好了。

XVERSE-13B-Chat:
好的，我会为您查询符合条件的酒店。请稍等片刻
<|startofthink|>```JSON
{ 
   "api_name": "hotel_search","parameters": {
      "city": "上海",
      "check_in": "2007-03-09",
      "check_out": "2007-03-10",
      "guests": 2,
      "rooms": 1,
      "price_min": 300,
      "price_max": 500,
      "rating_min": 8,
      "amenities": "停车场"
   }
}```<|endofthink|>

Quantization

We support quantization of INT8 and INT4 types, which can significantly reduce the GPU memory required for model loading.

INT8 quantization:

model = AutoModelForCausalLM.from_pretrained("xverse/XVERSE-13B-Chat", torch_dtype=torch.bfloat16, trust_remote_code=True)
model = model.quantize(8).cuda()

INT4 quantization：

model = AutoModelForCausalLM.from_pretrained("xverse/XVERSE-13B-Chat", torch_dtype=torch.bfloat16, trust_remote_code=True)
model = model.quantize(4).cuda()

The table below compares the GPU memory usage and MMLU accuracy of models at different quantization levels:

Model Precision Memory Usage (GB) MMLU Accuracy
XVERSE-13B-Chat BF16 / FP16 28.2 60.2
XVERSE-13B-Chat INT8 16.8 60.3
XVERSE-13B-Chat INT4 10.9 55.0

Fine-tuning

Both XVERSE-13B and XVERSE-13B-Chat allow developers to fine-tune for improved performance. Here, we attempted to use LLaMA Efficient Tuning for compatible fine-tuning training with XVERSE-13B, and tested it in an environment with 8 * Nvidia A800 80 GB + deepspeed. Below, we provide the detailed method for full parameters fine-tuning.

Environment Setup

Download the LLaMA Efficient Tuning project and [install dependencies] (https://github.com/hiyouga/LLaMA-Efficient-Tuning#getting-started) as required.

Training

Training launch script:

Replace model_path with your own model path.

Both XVERSE-13B and XVERSE-13B-Chat are trained based on bfloat16. It is recommended to use bfloat16 for fine-tuning training.

deepspeed --num_gpus=8 src/train_bash.py \
    --stage sft \
    --model_name_or_path model_path \
    --do_train \
    --dataset alpaca_gpt4_en \
    --template default \
    --finetuning_type full \
    --output_dir output_model_path \
    --overwrite_cache \
    --per_device_train_batch_size 4 \
    --per_device_eval_batch_size 4 \
    --gradient_accumulation_steps 4 \
    --preprocessing_num_workers 16 \
    --lr_scheduler_type cosine \
    --logging_steps 10 \
    --save_steps 200 \
    --eval_steps 200 \
    --learning_rate 2e-5 \
    --max_grad_norm 0.5 \
    --num_train_epochs 2.0 \
    --evaluation_strategy steps \
    --load_best_model_at_end \
    --plot_loss \
    --bf16 \
    --padding_side right \
    --deepspeed deepspeed.json

deep_speed.json parameter settings：

{
    "train_micro_batch_size_per_gpu": "auto",
    "gradient_accumulation_steps": "auto",
    "gradient_clipping": "auto",
    "zero_allow_untested_optimizer": true,
    "bf16": {
        "enabled": true
    },
    "zero_optimization": {
        "stage": 2,
        "allgather_partitions": true,
        "reduce_scatter": true,
        "overlap_comm": false,
        "contiguous_gradients": true
    }
}

Limitations and Disclaimer

Like all other Large Language Models (LLMs), XVERSE-13B 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-13B, developers should conduct safety tests and optimization of the model according to its specific application.

We strongly warn against the use of the XVERSE-13B 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-13B 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-13B needs to adhere to the Model License Agreement.

The XVERSE-13B 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.