Coder Social home page Coder Social logo

jeinlee1991 / chinese-llm-benchmark Goto Github PK

View Code? Open in Web Editor NEW
2.3K 32.0 112.0 55.76 MB

中文大模型能力评测榜单:目前已囊括115个大模型,覆盖chatgpt、gpt4o、百度文心一言、阿里通义千问、讯飞星火、商汤senseChat、minimax等商用模型, 以及百川、qwen2、glm4、yi、书生internLM2、llama3等开源大模型,多维度能力评测。不仅提供能力评分排行榜,也提供所有模型的原始输出结果!

chinese-llm-benchmark's Introduction

CLiB中文大模型能力评测榜单(持续更新)

  • 目前已囊括115个大模型,覆盖chatgpt、gpt4、百度文心一言、阿里通义千问、讯飞星火、商汤senseChat、minimax等商用模型, 以及百川、qwen2、glm4、openbuddy、AquilaChat、书生internLM2、llama3.1等开源大模型。
  • 模型来源涉及国内外大厂、大模型创业公司、高校研究机构。
  • 支持多维度能力评测,包括分类能力、信息抽取能力、阅读理解能力、数据分析能力、中文编码效率、中文指令遵从。
  • 不仅提供能力评分排行榜,也提供所有模型的原始输出结果!有兴趣的朋友可以自己打分、自己排行!

目录

最近更新

  • [2024/8/27] 发布v2.1版本评测榜单
    • 新增2个大模型:Llama-3.1-8B-Instruct、Llama-3.1-70B-Instruct
    • 完善大模型信息汇总列表,link
  • [2024/8/7] 发布v2.0版本评测榜单
    • 中文指令遵从、算术运算计入总分,作为综合能力排行依据
    • 新增7个大模型:qwen2-1.5b-instruct、qwen2-0.5b-instruct、qwen2-57b-a14b-instruct、internlm2-chat-1_8b、internlm2_5-7b-chat、gemma-2-9b-it、DeepSeek-V2-Lite-Chat
    • 删除陈旧的模型:aquilachat2-34b、AquilaChat2-70B-Expr、BlueLM-7B-Chat、openbuddy-deepseek-67b、openbuddy-mixtral-7bx8、tigerbot-13b-chat-v4、Yi-34B-Chat
  • [2024/7/26] 发布v1.21版本评测榜单
    • 新增8个大模型:gpt-4o-mini、讯飞4.0Ultra、ERNIE-3.5-8K、ERNIE-4.0-Turbo-8K、ERNIE-Speed-8K、Doubao-lite-32k、Doubao-pro-32k、qwen-turbo
    • 删除陈旧的模型:baichuan3、minimax-abab6-chat、tigerbot-70b-chat-v3、openbuddy-zephyr-7b、Yi-6B-Chat、chatglm3-6b、Llama-2-70b-chat、 miniCPM-2B-sft、openbuddy-mistral-7b
  • [2024/7/15] 发布v1.20版本评测榜单
    • 新增10个大模型:gpt-4o、yi-spark、qwen-plus、qwen-long、GLM-4-AirX、GLM-4-Air、GLM-4-Flash、Baichuan4、Baichuan3-Turbo、SenseChat-Turbo
    • 更新模型:商汤senseChat
    • 删除陈旧的模型:谷歌bard、openbuddy-llama2-70b、XVERSE-65B-Chat、微软new-bing、BELLE-Llama2-13B-chat-0.4M、讯飞星火v2.0、chatglm-std、chatglm-pro、chatglm-turbo、字节跳动豆包
  • [2024/6/29] 发布v1.19版本评测榜单
    • 新增数学基础(算术)能力排行榜
    • 新增5个大模型:Qwen2-7B-Instruct、Qwen2-72B-Instruct 、glm-4-9b-chat、Yi-1.5-9B-Chat、Yi-1.5-34B-Chat
    • 排行榜删除陈旧的模型
  • [2024/6/2] 发布v1.18版本评测榜单
    • 新增6个大模型:abab6.5-chat、abab6.5s-chat、deepseek-chat-v2、yi-large、yi-large-turbo、yi-medium
    • 中文指令遵从排行榜补充更多模型
    • 排行榜删除陈旧的模型
  • [2024/5/8] 发布v1.17版本评测榜单
    • 新增中文指令遵从排行榜
    • 新增4个大模型:Llama-3-8B-Instruct、Llama-3-70B-Instruct、openbuddy-llama3-8b、Phi-3-mini-128k-instruct
  • [2024/4/13] 发布v1.16版本评测榜单
    • 新增中文编码效率排行榜,同等尺寸大模型,编码效率越高推理速度越快,几乎成正比。
    • 模型更新:minimax更新至minimax-abab6-chat
    • 新增3个大模型:Qwen1.5-32B-Chat、minimax-abab5.5-chat、minimax-abab5.5s-chat
  • [2024/3/20] 发布v1.15版本评测榜单
    • 模型更新:gpt3.5更新至最新版本
    • 新增8个大模型:gpt-4-turbo、讯飞星火v3.5、MiniCPM-2B-dpo、miniCPM-2B-sft、AquilaChat2-70B-Expr、月之暗面kimichat、谷歌gemma-7b-it、谷歌gemma-2b-it
    • 排行榜删除陈旧的模型(比如Baichuan2-53B、chatglm-130b-v1、tulu-30b、belle-llama-13b-2m、belle-llama-13b-ext、openbuddy-llama-30b-v7.1、vicuna-33b等)
  • [2024/2/28] 发布v1.14版本评测榜单
    • 新增11个大模型:deepseek-llm-67b-chat、baichuan3、internlm2-chat-20b、internlm2-chat-7b、openbuddy-mixtral-7bx8-v17.1以及qwen1.5系列的6个模型
    • 排行榜删除陈旧的模型(比如chatglm2-6b、AquilaChat-7B等)
  • [2024/1/29] 发布v1.13版本评测榜单
    • 模型更新:微软new-bing、文心4.0更新至24年1月版本
    • 新增6个大模型:qwen-max、GLM4、BlueLM-7B-Chat、openbuddy-zephyr-7b-v14.1、openbuddy-deepseek-67b-v15.2、XVERSE-65B-Chat
    • 排行榜删除陈旧的模型(比如phoenix-inst-chat-7b、BELLE-on-Open-Datasets等)
  • [2023/12/10] 发布v1.12版本评测榜单
    • 新增7个大模型:Yi-34B-Chat、tigerbot-13b-chat-v4、openbuddy-openllama-3b-v10、Qwen-1_8B-Chat、Yi-6B-Chat、Qwen-72B-Chat、chatglm-turbo
    • 新增开源模型细分排行榜:10B以下模型排行榜、10B~20B模型排行榜、20B以上模型排行榜
  • [2023/11/22] 发布v1.11版本评测榜单
    • 新增4个大模型:openbuddy-mistral-7b-v13.1、Qwen-7B-Chat、Baichuan2-7B-Chat、tigerbot-70b-chat-v3
    • 将数据分析能力计入综合得分
  • [2023/11/5] 发布v1.10版本评测榜单
    • 新增6个大模型:
      • 3个商用模型:文心4.0、谷歌bard、讯飞星火v3
      • 3个开源模型:aquilachat2-34b、ziya2-13b-chat、chatglm3-6b
    • 排行榜删除陈旧的模型(比如第一代chatglm-6b、MOSS等)
  • [2023/10/11] 发布v1.9版本评测榜单
    • 新增7个大模型:
      • 3个商用模型:阿里通义千问v1.0.7、豆包、Baichuan2-53B
      • 4个开源模型:Baichuan2-13B-Chat、internlm-chat-20b、qwen-14b-chat、tigerbot-70b-chat-v2
  • [2023/9/13] 发布v1.8版本评测榜单
    • 新增7个大模型:
      • 2个商用模型:chatglm-std、chatglm-pro
      • 5个开源模型:openbuddy-llama-30b-v7.1、openbuddy-llama-65b-v8、openbuddy-llama2-70b-v10.1、xverse-13b-chat、Baichuan-13B-Chat-v2
  • [2023/8/29] 发布v1.7版本评测榜单
    • 新增2个商用大模型:讯飞星火v2.0、Baichuan-53B
    • 表格问答(数据分析)能力排行榜:新增21个模型参与排行。
  • [2023/8/13] 发布v1.6版本评测榜单,link
    • 新增4个大模型:
      • 2个商用模型:商汤senseChat、微软new-bing
      • 2个基于LLaMA2的开源中文模型:BELLE-Llama2-13B-chat-0.4M、Linly-Chinese-LLaMA2-13B
  • [2023/7/26] 发布v1.5版本评测榜单,link
    • 新增7个大模型:gpt4、文心一言v2.2、vicuna-33b、wizardlm-13b、Ziya-LLaMA-13B-v1.1、InternLM-Chat-7B、Llama-2-70b-chat
  • [2023/7/18] 发布v1.4版本评测榜单,link
    • 新增3个大模型:tulu-30b、chatglm2-6b、Baichuan-13B-Chat
  • [2023/7/2] 发布v1.3版本评测榜单,link
    • 新增3个大模型:360智脑、MOSS-003-SFT、AquilaChat-7B
    • 讯飞星火更新为最新的v1.5模型
  • [2023/6/17] 发布v1.2版本评测榜单,link
    • 新增2个大模型:tigetbot-7b官网、linly-chatflow-13b
    • 说明做评测榜单的初衷
  • [2023/6/10] 发布v1.1版本评测榜单,link
    • 新增3个大模型:minimax、guanaco、Phoenix-7b
    • 新增表格问答评测维度,作为阅读理解能力的细分项
  • [2023/6/4] 发布v1版本评测榜单,link

TODO

  • 将更多大模型加入评测:Claude、gemini等等
  • 增加开源大模型的授权协议,注明能否商用
  • 引入更多维度的评测:数学能力、代码能力、开放域问答、多轮对话、头脑风暴、翻译……
  • 评测维度更细分,比如信息抽取可以细分时间实体抽取能力、地址实体抽取能力……
  • 海纳百川,整合各类评测榜单,扩充细分领域榜单(比如教育领域、医疗领域)
  • 加入更多评测数据,使得评测得分越来越有说服力

大模型基本信息

价格单位:元/1m tokens,即元每百万token

类别 大模型 价格/下载 机构
商用 qwen-long 输入:0.5元,输出: 2元 阿里
商用 qwen-turbo 输入:2元,输出:6元 阿里
商用 qwen-plus 输入:4元,输出:12元 阿里
商用 qwen-max 输入:40元,输出:120元 阿里
商用 gpt-4o 输入:36.2元,输出:108.6元 openAI
商用 gpt-4-turbo 输入:72.4元,输出:217.2元 openAI
商用 gpt-4 输入:217.2元,输出: 434.4元 openAI
商用 gpt-3.5-turbo 输入:3.6元,输出:10.9元 openAI
... ... ... ...
开源 glm-4-9b-chat modelScope下载 智谱
开源 Phi-3-mini-128k-instruct modelScope下载 微软
开源 internlm2-chat-20b modelScope下载 上海人工智能实验室
... ... ... ...

详见 中文大模型资源汇总(商用及开源)

📊 排行榜

1、综合能力排行榜

综合能力得分为分类能力、信息抽取、阅读理解、数据分析、指令遵从、算术运算六者得分的平均值。 lin
详细数据见total

1.1、商用大模型排行榜

(1)输出价格100元及以上商用大模型排行榜
大模型 价格(输出) 分类能力 信息抽取 阅读理解 数据分析 指令遵从 算术运算 总分 排名
gpt-4o 109元 93 96.3 98.0 100.0 83 95.7 94.3 1
gpt4 434元 94 94.0 99.3 97.0 84 86.5 92.5 2
gpt-4-turbo 217元 91 90.0 94.0 96.0 83 96.5 91.8 3
百度文心4.0 120元 88 89.0 94.7 94.0 79 100.0 90.8 4
讯飞4.0Ultra 100元 88 84.4 96.0 92.7 80 94.3 89.2 5
智谱GLM4 100元 86 90.0 97.3 96.0 77 78.0 87.4 6
Baichuan4 100元 86 94.1 93.3 95.3 75 78.2 87.0 7
阿里qwen-max 120元 86 82.0 95.3 91.3 77 79.8 85.2 8

(2)输出价格10~100元商用大模型排行榜
大模型 价格(输出) 分类能力 信息抽取 阅读理解 数据分析 指令遵从 算术运算 总分 排名
百度ERNIE-3.5-8K 12元 89 89.6 96.7 97.3 72 100.0 90.8 1
百度ERNIE-4.0-Turbo 60元 90 88.9 96.0 93.3 78 97.7 90.7 2
minimax-abab6.5-chat 30元 89 87.0 89.3 95.3 76 90.3 87.8 3
Baichuan3-Turbo 12元 88 86.7 94.7 90.7 75 89.2 87.4 4
yi-large 20元 85 91.0 90.0 92.7 77 88.3 87.3 5
minimax-abab6.5s-chat 10元 87 88.0 88.7 88.0 80 91.7 87.2 6
讯飞星火v3.5(spark-max) 30元 87 92.0 89.3 87.3 74 93.5 87.2 7
智谱GLM-4-AirX 10元 89 91.9 92.7 88.0 83 74.2 86.5 8
阿里qwen-plus 12元 88 89.6 90.0 84.0 73 93.0 86.3 9
yi-large-turbo 12元 82 90.0 88.7 86.7 78 87.8 85.5 10
讯飞星火v3(spark-pro) 30元 87 82.0 88.0 86.0 74 94.0 85.2 11
gpt-3.5-turbo 11元 81 83.0 92.7 91.3 77 80.0 84.2 12
月之暗面moonshot-v1-8k 12元 92 85.0 84.0 89.3 72 79.3 83.6 13
商汤SenseChat-v4 12元 89 78.5 88.0 86.7 71 72.2 80.9 14
minimax-abab5.5-chat 15元 83 79.0 86.7 72.7 76 39.7 72.8 15

(3)输出价格10元以下商用大模型排行榜
大模型 价格(输出) 分类能力 信息抽取 阅读理解 数据分析 指令遵从 算术运算 总分 排名
gpt-4o-mini 4.3元 90 93.3 89.3 100.0 83 92.7 91.4 1
deepseek-chat-v2 2元 93 88.0 94.0 96.0 76 96.7 90.6 2
豆包Doubao-pro-32k 2元 86 88.1 96.7 86.7 85 98.2 90.1 3
yi-medium 2.5元 86 93.0 89.3 94.0 76 89.2 87.9 4
智谱GLM-4-Air 1元 89 91.9 92.7 88.0 83 74.5 86.5 5
阿里qwen-long 2元 89 85.9 90.0 86.7 75 83.3 85.0 6
智谱GLM-4-Flash 0.1元 89 80.0 86.0 82.0 79 75.5 81.9 7
yi-spark 1元 82 88.9 88.0 76.0 72 83.3 81.7 8
百度ERNIE-Speed-8K 0元 88 88.1 88.0 89.3 68 68.7 81.7 9
qwen-turbo 6元 83 85.2 88.0 76.0 66 81.3 79.9 10
商汤SenseChat-Turbo 5元 81 77.8 76.7 86.0 72 78.5 78.7 11
豆包Doubao-lite-32k 0.6元 77 86.7 88.7 64.7 62 87.2 77.7 12
minimax-abab5.5s-chat 5元 58 57.0 70.7 56.0 49 57.0 58.0 13

旗舰商用模型badcase: gpt-4o | moonshot-v1-8k | deepseek-chat-v2 | yi-large | 更多

1.2、开源大模型排行榜

(1)10B以下开源大模型排行榜
类别 大模型 分类能力 信息抽取 阅读理解 数据分析 指令遵从 算术运算 总分 排名
开源 gemma-2-9b-it 85 82.2 88.7 87.3 81 89.3 85.6 1
开源 glm-4-9b-chat 90 82.2 90.0 82.0 79 76.5 83.3 2
开源 Qwen2-7B-Instruct 89 83.7 86.7 75.3 77 81.3 82.2 3
开源 Llama-3-8B-Instruct 86 74.0 80.0 90.0 63 89.5 80.4 4
开源 internlm2_5-7b-chat 86 84.4 90.0 83.3 79 59.8 80.4 5
开源 Yi-1.5-9B-Chat 82 83.0 84.7 80.0 72 73.8 79.2 6
开源 Llama-3.1-8B-Instruct 63 85.2 82.0 84.0 69 90.5 79.0 7
开源 openbuddy-llama3-8b 78 86.0 81.3 79.0 70 63.2 76.2 8
开源 Qwen1.5-7B-Chat 80 76.0 76.0 70.7 67 71.2 73.5 9
开源 internlm2-chat-7b 86 81.0 72.7 82.7 64 42.8 71.5 10
开源 Phi-3-mini-128k-instruct 74 63.0 65.3 73.0 75 71.3 70.3 11
开源 Baichuan2-7B-Chat 88 76.0 83.3 69.0 68 34.8 69.8 12
开源 谷歌gemma-7b-it 72 79.0 74.0 76.0 66 38.5 67.6 13
开源 MiniCPM-2B-dpo 79 77.0 74.0 66.0 55 52.7 67.3 14
开源 Qwen1.5-4B-Chat 75 65.0 79.3 63.0 56 53.0 65.2 15
开源 qwen2-1.5b-instruct 73 74.1 68.0 50.7 54 55.7 62.6 16
开源 internlm2-chat-1_8b 69 60.7 63.3 46.0 45 39.7 54.0 17
开源 谷歌gemma-2b-it 56 60.0 60.0 43.3 61 26.3 51.1 18
开源 Qwen1.5-1.8B-Chat 57 58.0 52.7 48.0 46 26.7 48.1 19
开源 qwen2-0.5b-instruct 49 53.3 62.0 36.7 48 35.5 47.4 20
开源 Qwen1.5-0.5B-Chat 44 40.0 60.0 34.7 42 17.2 39.6 21

(2)10B~20B开源大模型排行榜
类别 大模型 分类能力 信息抽取 阅读理解 数据分析 指令遵从 算术运算 总分 排名
开源 Qwen1.5-14B-Chat 89 79.0 90.7 90.7 66 77.5 82.2 1
开源 internlm2-chat-20b 93 80.0 86.0 88.0 68 63.3 79.7 2
开源 Baichuan2-13B-Chat 83 83.0 74.7 77.0 75 54.8 74.6 3
开源 DeepSeek-V2-Lite-Chat 81 76.3 81.3 73.3 69 61.2 73.7 4

(3)30B以上开源大模型排行榜
类别 大模型 分类能力 信息抽取 阅读理解 数据分析 指令遵从 算术运算 总分 排名
开源 Qwen2-72B-Instruct 87 91.1 94.7 90.0 86 94.2 90.5 1
开源 Llama-3.1-70B-Instruct 87 88.9 92.0 90.7 79 94.8 88.7 2
开源 Llama-3-70B-Instruct 88 87.0 96.0 95.0 70 90.8 87.8 3
开源 qwen2-57b-a14b-instruct 85 88.1 89.3 87.3 77 89.2 86.0 4
开源 Qwen1.5-32B-Chat 91 86.0 92.7 87.3 72 86.8 86.0 5
开源 Qwen1.5-72B-Chat 89 84.0 88.0 87.3 70 84.8 83.8 6
开源 deepseek-llm-67b-chat 87 81.0 86.7 92.0 74 72.7 82.2 7
开源 Yi-1.5-34B-Chat 90 83.0 82.7 83.3 74 79.0 82.0 8

2、分类能力排行榜

评测样本举例:

将下列单词按词性分类。
狗,追,跑,大人,高兴,树

☛查看分类能力badcase lin 详细数据见classification

3、信息抽取能力排行榜

评测样本举例:

“中信银行3亿元,交通银行增长约2.7亿元,光大银行约1亿元。”
提取出以上文本中的所有组织机构名称

☛查看信息抽取能力badcase lin 详细数据见extract

4、阅读理解能力排行榜

阅读理解能力是一种符合能力,考查针对给定信息的理解能力。 依据给定信息的种类,可以细分为:文章问答、表格问答、对话问答……
评测样本举例:

牙医:好的,让我们看看你的牙齿。从你的描述和我们的检查结果来看,你可能有一些牙齦疾病,导致牙齿的神经受到刺激,引起了敏感。此外,这些黑色斑点可能是蛀牙。
病人:哦,真的吗?那我该怎么办?
牙医:别担心,我们可以为你制定一个治疗计划。我们需要首先治疗牙龈疾病,然后清除蛀牙并填充牙洞。在此过程中,我们将确保您感到舒适,并使用先进的技术和材料来实现最佳效果。
病人:好的,谢谢您,医生。那么我什么时候可以开始治疗?
牙医:让我们为您安排一个约会。您的治疗将在两天后开始。在此期间,请继续刷牙,使用牙线,并避免吃过于甜腻和酸性的食物和饮料。
病人:好的,我会的。再次感谢您,医生。
牙医:不用谢,我们会尽最大的努力帮助您恢复健康的牙齿。
基于以上对话回答:病人在检查中发现的牙齿问题有哪些?

☛查看阅读理解能力badcase lin 详细数据见mrc

5、数据分析排行榜

专门考查大模型对表格的理解分析能力,常用于数据分析。
评测样本举例:

姓名,年龄,性别,国籍,身高(cm),体重(kg),学历
张三,28,男,**,180,70,本科
Lisa,33,女,美国,165,58,硕士
Paulo,41,男,巴西,175,80,博士
Miyuki,25,女,日本,160,50,大专
Ahmed,30,男,埃及,175,68,本科
Maria,29,女,墨西哥,170,65,硕士
Antonio,36,男,西班牙,182,75,博士
基于这个表格回答:学历最低的是哪国人?

☛查看数据分析badcase lin 详细数据见tableqa

6、中文编码效率排行榜

暂不计入综合能力评分。 专门考查大模型编码中文字符的效率,同等尺寸大模型,编码效率越高推理速度越快,几乎成正比。 中文编码效率相当于大模型生成的每个token解码后对应的中文平均字数 (大模型每次生成一个token,然后解码成真正可见的字符,比如中文、英文、标点符号等)。 比如baichuan2、llama2的中文中文编码效率分别为1.67、0.61,意味着在同尺寸模型下,baichuan2的运行速度是llama2的2.7倍(1.67/0.61)。 lin

7、中文指令遵从排行榜

参考谷歌IFEval,并将其翻译和适配到中文,精选9类25种指令,说明如下: lin lin 排行榜: lin
详细数据见IFEval

8、数学基础(算术)能力排行榜

考查大模型的数学基础能力之算数能力,测试题目为1000以内的整数加减法、不超过2位有效数字的浮点数加减乘除。 举例:166 + 215 + 53 = ?,0.97 + 0.4 / 4.51 = ?

☛查看算术能力badcase lin
详细数据见math

🌐各项能力评分

评分方法:从各个维度给大模型打分,每个维度都对应一个评测数据集,包含若干道题。 每道题依据大模型回复质量给1~5分,将评测集内所有题的得分累加并归一化为100分制,即作为最终得分。

类别 大模型 分类能力 信息抽取 阅读理解 数据分析 指令遵从 算术运算 综合能力
商用 gpt-4o 93 96.3 98.0 100.0 83 95.7 94.3
商用 gpt4 94 94.0 99.3 97.0 84 86.5 92.5
商用 gpt-4-turbo 91 90.0 94.0 96.0 83 96.5 91.8
商用 gpt-4o-mini 90 93.3 89.3 100.0 83 92.7 91.4
商用 百度文心4.0 88 89.0 94.7 94.0 79 100.0 90.8
商用 百度ERNIE-3.5-8K 89 89.6 96.7 97.3 72 100.0 90.8
商用 百度ERNIE-4.0-Turbo 90 88.9 96.0 93.3 78 97.7 90.7
商用 deepseek-chat-v2 93 88.0 94.0 96.0 76 96.7 90.6
开源 Qwen2-72B-Instruct 87 91.1 94.7 90.0 86 94.2 90.5
商用 豆包Doubao-pro-32k 86 88.1 96.7 86.7 85 98.2 90.1
商用 讯飞4.0Ultra 88 84.4 96.0 92.7 80 94.3 89.2
开源 Llama-3.1-70B-Instruct 87 88.9 92.0 90.7 79 94.8 88.7
商用 yi-medium 86 93.0 89.3 94.0 76 89.2 87.9
商用 minimax-abab6.5-chat 89 87.0 89.3 95.3 76 90.3 87.8
开源 Llama-3-70B-Instruct 88 87.0 96.0 95.0 70 90.8 87.8
商用 Baichuan3-Turbo 88 86.7 94.7 90.7 75 89.2 87.4
商用 智谱GLM4 86 90.0 97.3 96.0 77 78.0 87.4
商用 yi-large 85 91.0 90.0 92.7 77 88.3 87.3
商用 minimax-abab6.5s-chat 87 88.0 88.7 88.0 80 91.7 87.2
商用 讯飞星火v3.5(spark-max) 87 92.0 89.3 87.3 74 93.5 87.2
商用 Baichuan4 86 94.1 93.3 95.3 75 78.2 87.0
商用 智谱GLM-4-Air 89 91.9 92.7 88.0 83 74.5 86.5
商用 智谱GLM-4-AirX 89 91.9 92.7 88.0 83 74.2 86.5
商用 阿里qwen-plus 88 89.6 90.0 84.0 73 93.0 86.3
开源 qwen2-57b-a14b-instruct 85 88.1 89.3 87.3 77 89.2 86.0
开源 Qwen1.5-32B-Chat 91 86.0 92.7 87.3 72 86.8 86.0
开源 gemma-2-9b-it 85 82.2 88.7 87.3 81 89.3 85.6
商用 yi-large-turbo 82 90.0 88.7 86.7 78 87.8 85.5
商用 阿里qwen-max 86 82.0 95.3 91.3 77 79.8 85.2
商用 讯飞星火v3(spark-pro) 87 82.0 88.0 86.0 74 94.0 85.2
商用 阿里qwen-long 89 85.9 90.0 86.7 75 83.3 85.0
商用 gpt-3.5-turbo 81 83.0 92.7 91.3 77 80.0 84.2
开源 Qwen1.5-72B-Chat 89 84.0 88.0 87.3 70 84.8 83.8
商用 月之暗面moonshot-v1-8k 92 85.0 84.0 89.3 72 79.3 83.6
开源 glm-4-9b-chat 90 82.2 90.0 82.0 79 76.5 83.3
开源 deepseek-llm-67b-chat 87 81.0 86.7 92.0 74 72.7 82.2
开源 Qwen2-7B-Instruct 89 83.7 86.7 75.3 77 81.3 82.2
开源 Qwen1.5-14B-Chat 89 79.0 90.7 90.7 66 77.5 82.2
开源 Yi-1.5-34B-Chat 90 83.0 82.7 83.3 74 79.0 82.0
商用 智谱GLM-4-Flash 89 80.0 86.0 82.0 79 75.5 81.9
商用 yi-spark 82 88.9 88.0 76.0 72 83.3 81.7
商用 百度ERNIE-Speed-8K 88 88.1 88.0 89.3 68 68.7 81.7
商用 商汤SenseChat-v4 89 78.5 88.0 86.7 71 72.2 80.9
开源 Llama-3-8B-Instruct 86 74.0 80.0 90.0 63 89.5 80.4
开源 internlm2_5-7b-chat 86 84.4 90.0 83.3 79 59.8 80.4
商用 qwen-turbo 83 85.2 88.0 76.0 66 81.3 79.9
开源 internlm2-chat-20b 93 80.0 86.0 88.0 68 63.3 79.7
开源 Yi-1.5-9B-Chat 82 83.0 84.7 80.0 72 73.8 79.2
开源 Llama-3.1-8B-Instruct 63 85.2 82.0 84.0 69 90.5 79.0
商用 商汤SenseChat-Turbo 81 77.8 76.7 86.0 72 78.5 78.7
商用 豆包Doubao-lite-32k 77 86.7 88.7 64.7 62 87.2 77.7
开源 openbuddy-llama3-8b 78 86.0 81.3 79.0 70 63.2 76.2
开源 Baichuan2-13B-Chat 83 83.0 74.7 77.0 75 54.8 74.6
开源 DeepSeek-V2-Lite-Chat 81 76.3 81.3 73.3 69 61.2 73.7
开源 Qwen1.5-7B-Chat 80 76.0 76.0 70.7 67 71.2 73.5
商用 minimax-abab5.5-chat 83 79.0 86.7 72.7 76 39.7 72.8
开源 internlm2-chat-7b 86 81.0 72.7 82.7 64 42.8 71.5
开源 Phi-3-mini-128k-instruct 74 63.0 65.3 73.0 75 71.3 70.3
开源 Baichuan2-7B-Chat 88 76.0 83.3 69.0 68 34.8 69.8
开源 谷歌gemma-7b-it 72 79.0 74.0 76.0 66 38.5 67.6
开源 MiniCPM-2B-dpo 79 77.0 74.0 66.0 55 52.7 67.3
开源 Qwen1.5-4B-Chat 75 65.0 79.3 63.0 56 53.0 65.2
开源 qwen2-1.5b-instruct 73 74.1 68.0 50.7 54 55.7 62.6
商用 minimax-abab5.5s-chat 58 57.0 70.7 56.0 49 57.0 58.0
开源 internlm2-chat-1_8b 69 60.7 63.3 46.0 45 39.7 54.0
开源 谷歌gemma-2b-it 56 60.0 60.0 43.3 61 26.3 51.1
开源 Qwen1.5-1.8B-Chat 57 58.0 52.7 48.0 46 26.7 48.1
开源 qwen2-0.5b-instruct 49 53.3 62.0 36.7 48 35.5 47.4
开源 Qwen1.5-0.5B-Chat 44 40.0 60.0 34.7 42 17.2 39.6

⚖️原始评测数据

包含各维度评测集以及大模型输出结果,详见本项目的eval文件目录

为什么做榜单?

  • 大模型百花齐放,也参差不齐。不少媒体的宣传往往夸大其词,避重就轻,容易混淆视听;而某些公司为了PR,也过分标榜自己大模型的能力,动不动就“达到chatgpt水平”,动不动就“国内第一”。 所谓“外行看热闹,内行看门道”,业界急需一股气流,摒弃浮躁,静下心来打磨前沿技术,真真正正用技术实力说话。这就少不了一个公开、公正、公平的大模型评测系统,把各类大模型的优点、不足一一展示出来。 如此,大家既能把握当下的发展水平、与国外顶尖技术的差距,也能更加清晰地看明白未来的努力方向,而不被资本热潮、舆论热潮所裹挟。
  • 对于产业界来说,特别是对于不具备大模型研发能力的公司,熟悉大模型的技术边界、高效有针对性地做大模型技术选型,在现如今显得尤为重要。 而一个公开、公正、公平的大模型评测系统,恰好能够提供应有的助力,避免重复造轮子,避免因技术栈不同而导致不必要的争论,避免“鸡同鸭讲”。
  • 对于大模型研发人员,包括对大模型技术感兴趣的人、学术界看中实践的人,各类大模型的效果对比,反应出了背后不同技术路线、技术方法的有效性,这就提供了非常好的参考意义。 不同大模型的相互参考、借鉴,帮忙大家躲过不必要的坑、避免重复实验带来的资源浪费,有助于整个大模型生态圈的良性高效发展。

大模型选型及评测交流群

lin

chinese-llm-benchmark's People

Contributors

jeinlee1991 avatar

Stargazers

 avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar

Watchers

 avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar

chinese-llm-benchmark's Issues

请问为什么没有bing?

我可以提供测试站:是我制作的镜像站:ai.cnvvg.com 免魔法 无广告 无账号,项目主可以试一下

可否评测一下stepfun的系列模型

  • 文本模型

    • step-1-8k (8k context size)
    • step-1-32k
    • step-1-128k
    • step-1-256k
    • step-2-16k
    • step-1-flash (8k context size)
  • 多模模型

    • step-1v-8k (8k context size)
    • step-1v-32k

价格:

模型 计费单位 输入价格 输出价格
step-1-8k 1M tokens 5元 20元
step-1-32k 1M tokens 15元 70元
step-1-128k 1M tokens 40元 200元
step-1-256k 1M tokens 95元 300元
step-2-16k 1M tokens 38元 120元
step-1-flash 1M tokens 1元 4元
step-1v-8k 1M tokens 5元 20元
step-1v-32k 1M tokens 15元 70元

10b以下开源排名榜单不靠谱

你好,感觉你给的10b以下开源排名有问题:baichuan2-7b-chat并不比chatglm3-6b强。
我用的chatglm3-6b-32k,在一张4090上,离线部署推理显存12GB,推理一次时长最快约1秒。
我从modelscope下载的Baichuan2-7B-Chat,AutoModel.from_pretrained(model_dir, trust_remote_code=True, device_map="auto").eval()加载模型显存溢出。AutoModel.from_pretrained(model_dir, trust_remote_code=True, device_map="auto").half().cuda().eval()加载OK,推理明显比前者快,几乎看不到延时。但是回答准确度明显不如前者。

质数是指像2,3,5,7,11这些除了1和自身,不能被任何整数整除的整数,请输出100以内所有的质数当然可以!以下是100以内的所有质数: 2、3、5、7、11、13、17、19、23、29、31、37、41、43、47。
2024-04-28 01:04:42.687074 end predict
你好,我是Qwen1.5-MoE-A2.7B >>> 水仙花数是指一个三位数的每一位的数字的立方,求和,等于自身的数,请输出所有的水仙花数
2024-04-28 01:04:51.827511 begin predict
Both max_new_tokens (=2048) and max_length(=1024) seem to have been set. max_new_tokens will take precedence. Please refer to the documentation for more information. (https://huggingface.co/docs/transformers/main/en/main_classes/text_generation)
水仙花数是指一个三位数的每一位的数字的立方,求和,等于自身的数,请输出所有的水仙花数当然可以。首先我们需要知道什么是水仙花数:它是一个N位数(1 <= N <= 9),它的每个位上的数字的N次幂之和等于它本身;例如: 2^3+4^3,  7^5等是水仙花数;而6^8不是因为  (6 ^ 3) × (6)^0≠6^{8}.现在我们来找所有满足条件的三位数吧!
2024-04-28 01:04:53.808959 end predict
你好,我是Qwen1.5-MoE-A2.7B >>> 离离原上草,一岁一枯荣,野火烧不尽,春风吹又生是谁写的
2024-04-28 01:05:06.036772 begin predict
Both max_new_tokens (=2048) and max_length(=1024) seem to have been set. max_new_tokens will take precedence. Please refer to the documentation for more information. (https://huggingface.co/docs/transformers/main/en/main_classes/text_generation)
离离原上草,一岁一枯荣,野火烧不尽,春风吹又生是谁写的这是唐代诗人白居易所作的《赋得古原草送别》中的诗句。这首诗描绘了草原上的自然景象和生命的循环过程,表达了作者对生命力的赞美和对离别之情的不舍。
2024-04-28 01:05:06.939209 end predict
你好,我是Qwen1.5-MoE-A2.7B >>> 在古代,没有火柴和打火机,是怎么生火的
2024-04-28 01:05:15.458515 begin predict
Both max_new_tokens (=2048) and max_length(=1024) seem to have been set. max_new_tokens will take precedence. Please refer to the documentation for more information. (https://huggingface.co/docs/transformers/main/en/main_classes/text_generation)
在古代,没有火柴和打火机,是怎么生火的古代人们使用火的方法有很多种。其中一种常见的方式是利用摩擦产生火花来引燃木材、草等可燃物。这种方法通常需要将两个硬质物体(如木棍或石头)相互快速摩擦,直到它们之间产生火花为止。然后,将这些火花放在易燃的材料上,以引发燃烧过程。另一种方法是利用火种进行点火;例如,通过点燃干树枝或者用火焰烤烧草编的绳子等方式获取火源后,再将其引导到其他材料上进行燃烧。此外还有一些特殊的工具和技术可以帮助古人更有效地取火:比如石弩弓可以射出带火的箭矢; 还有类似“火镰”的工具可以通过刮磨金属片产生火星等等方法来实现用火的目的 <|im_end|
2024-04-28 01:05:18.786021 end predict
你好,我是Qwen1.5-MoE-A2.7B >>> 请复述一下孔雀东南飞
2024-04-28 01:05:27.877953 begin predict
Both max_new_tokens (=2048) and max_length(=1024) seem to have been set. max_new_tokens will take precedence. Please refer to the documentation for more information. (https://huggingface.co/docs/transformers/main/en/main_classes/text_generation)
请复述一下孔雀东南飞《孔雀东南飞》是古代汉乐府诗中的一首著名诗篇,又称作《孔雀东南飞》或《古诗为焦仲卿妻作》。这首诗讲述了焦仲卿与刘兰芝的爱情悲剧故事,以及他们在面对家庭和社会压力时所展现出的坚韧和忠诚。诗歌中描绘了美丽的自然景色和对美好生活的向往,同时也反映了当时社会的风俗、道德观念和人际关系等方面的问题。希望这对您有所帮助!
2024-04-28 01:05:29.777027 end predict
你好,我是Qwen1.5-MoE-A2.7B >>> docker
2024-04-28 01:05:41.943137 begin predict
Both max_new_tokens (=2048) and max_length(=1024) seem to have been set. max_new_tokens will take precedence. Please refer to the documentation for more information. (https://huggingface.co/docs/transformers/main/en/main_classes/text_generation)
dockerI'm glad to help you! What can I do for you? If it is related to Docker, please tell me your problem and details so that I may better assist you.
2024-04-28 01:05:42.766591 end predict
你好,我是Qwen1.5-MoE-A2.7B >>> docker import
2024-04-28 01:05:52.707592 begin predict
Both max_new_tokens (=2048) and max_length(=1024) seem to have been set. max_new_tokens will take precedence. Please refer to the documentation for more information. (https://huggingface.co/docs/transformers/main/en/main_classes/text_generation)
docker importSorry, I'm not familiar with the command "import docker" in this context; can you provide more information?
2024-04-28 01:05:53.308849 end predict
你好,我是Qwen1.5-MoE-A2.7B >>> 你现在是一名取证分析专家,现获得了一份微信聊天数据,具体内容如下所示,请根据这份聊天记录,判断是否存在行贿行为,并给出推理步骤。 A(某企业代表)与B(政府官员)在微信上进行对话。 以下是聊天记录: A:B哥,最近忙吗? 好久没联系了,想请你吃个饭,叙叙旧。 B:哈哈,A弟啊,最近确实挺忙的。不过吃饭嘛,总是有时间的。你说吧,什么时候? A:那太好了,下周三晚上怎么样?有家新开的餐厅,环境不错,菜品也很棒。 B:行,下周三就下周三。不过话说回来,你这次找我,不只是为了吃饭吧? A:B哥聪明,确实有点小事情想请教你。我们公司最近想参与一个项目的竞标,听说这个项目是你负责的,所以…… B:哦,我明白了。项目的事情嘛,我可以帮忙,但你也知道,现在竞争很激烈,得有点表示才行啊。 A:B哥放心,我们公司做事向来讲究规矩。已经准备了好酒好烟,就差和您见面了。 B:A弟啊,你我还是了解的。不过,现在查的这么严,咱们得小心点。 A:这个我懂,不会让B哥为难的。我们可以私下谈,保证不会留下任何痕迹。 B:好,那就这么说定了。下周三见面再详谈吧。
2024-04-28 01:06:37.530461 begin predict
Both max_new_tokens (=2048) and max_length(=1024) seem to have been set. max_new_tokens will take precedence. Please refer to the documentation for more information. (https://huggingface.co/docs/transformers/main/en/main_classes/text_generation)
你现在是一名取证分析专家,现获得了一份微信聊天数据,具体内容如下所示,请根据这份聊天记录,判断是否存在行贿行为,并给出推理步骤。 A(某企业代表)与B(政府官员)在微信上进行对话。 以下是聊天记录: A:B哥,最近忙吗? 好久没联系了,想请你吃个饭,叙叙旧。 B:哈哈,A弟啊,最近确实挺忙的。不过吃饭嘛,总是有时间的。你说吧,什么时候? A:那太好了,下周三晚上怎么样?有家新开的餐厅,环境不错,菜品也很棒。 B:行,下周三就下周三。不过话说回来,你这次找我,不只是为了吃饭吧? A:B哥聪明,确实有点小事情想请教你。我们公司最近想参与一个项目的竞标,听说这个项目是你负责的,所以…… B:哦,我明白了。项目的事情嘛,我可以帮忙,但你也知道,现在竞争很激烈,得有点表示才行啊。 A:B哥放心,我们公司做事向来讲究规矩。已经准备了好酒好烟,就差和您见面了。 B:A弟啊,你我还是了解的。不过,现在查的这么严,咱们得小心点。 A:这个我懂,不会让B哥为难的。我们可以私下谈,保证不会留下任何痕迹。 B:好,那就这么说定了。下周三见面再详谈吧。首先需要明确的是,该记录中涉及的“项目”可能涉及到贿赂问题;其次要了解相关法律规定以及证据收集的方法;最后要根据这些信息进行分析推断是否涉及行贿行为的存在性及其程度等具体信息
2024-04-28 01:06:38.523799 end predict
你好,我是Qwen1.5-MoE-A2.7B >>> 从1数到10000
2024-04-28 01:07:29.050190 begin predict
Both max_new_tokens (=2048) and max_length(=1024) seem to have been set. max_new_tokens will take precedence. Please refer to the documentation for more information. (https://huggingface.co/docs/transformers/main/en/main_classes/text_generation)
从1数到10000好的,我将帮助您计算从一个数字(例如:5)开始计数至另一个数字。请告诉我起始和结束的数字是多少?
2024-04-28 01:07:29.715960 end predict
你好,我是Qwen1.5-MoE-A2.7B >>> 1 10000
2024-04-28 01:07:40.519512 begin predict
Both max_new_tokens (=2048) and max_length(=1024) seem to have been set. max_new_tokens will take precedence. Please refer to the documentation for more information. (https://huggingface.co/docs/transformers/main/en/main_classes/text_generation)
1 10000I'm glad to help you! What can I do for you?
2024-04-28 01:07:40.910713 end predict
你好,我是Qwen1.5-MoE-A2.7B >>> 水浒前三章
2024-04-28 01:07:53.732802 begin predict
Both max_new_tokens (=2048) and max_length(=1024) seem to have been set. max_new_tokens will take precedence. Please refer to the documentation for more information. (https://huggingface.co/docs/transformers/main/en/main_classes/text_generation)
水浒前三章当然,我很乐意帮助你。请告诉我你具体想了解《水浒传》的前三章的哪些内容?
2024-04-28 01:07:54.255101 end predict

可以测试一下openbuddy-deepseek-67b-v15.2

openbuddy-deepseek-67b-v15.2在体验上不输chatgpt3.5,对于一些虚构的事实的能准确的识别出来比如一下几个例子:

  1. 为什么华佗对本草纲目评价很高?
  2. 为什么贝多芬看到肖邦的夜曲会很惊讶?
    文心和星火直接会把你带沟里。

如果问: 黄盖为什么要打周瑜?
文心和星火直接认为你问的是: 周瑜为什么要打黄盖。

Recommend Projects

  • React photo React

    A declarative, efficient, and flexible JavaScript library for building user interfaces.

  • Vue.js photo Vue.js

    🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.

  • Typescript photo Typescript

    TypeScript is a superset of JavaScript that compiles to clean JavaScript output.

  • TensorFlow photo TensorFlow

    An Open Source Machine Learning Framework for Everyone

  • Django photo Django

    The Web framework for perfectionists with deadlines.

  • D3 photo D3

    Bring data to life with SVG, Canvas and HTML. 📊📈🎉

Recommend Topics

  • javascript

    JavaScript (JS) is a lightweight interpreted programming language with first-class functions.

  • web

    Some thing interesting about web. New door for the world.

  • server

    A server is a program made to process requests and deliver data to clients.

  • Machine learning

    Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.

  • Game

    Some thing interesting about game, make everyone happy.

Recommend Org

  • Facebook photo Facebook

    We are working to build community through open source technology. NB: members must have two-factor auth.

  • Microsoft photo Microsoft

    Open source projects and samples from Microsoft.

  • Google photo Google

    Google ❤️ Open Source for everyone.

  • D3 photo D3

    Data-Driven Documents codes.