Hugging Peers

phi-4-eval-result

☆ Save
by easyllm text-generation
2 downloads 0 likes on ModelScope updated

Model card

模型信息

Model Summary
Developers Microsoft Research
Description phi-4 is a state-of-the-art open model built upon a blend of synthetic datasets, data from filtered public domain websites, and acquired academic books and Q&A datasets. The goal of this approach was to ensure that small capable models were trained with data focused on high quality and advanced reasoning.

phi-4 underwent a rigorous enhancement and alignment process, incorporating both supervised fine-tuning and direct preference optimization to ensure precise instruction adherence and robust safety measures
Architecture 14B parameters, dense decoder-only Transformer model
Inputs Text, best suited for prompts in the chat format
Context length 16K tokens
GPUs 1920 H100-80G
Training time 21 days
Training data 9.8T tokens
Outputs Generated text in response to input
Dates October 2024 – November 2024
Status Static model trained on an offline dataset with cutoff dates of June 2024 and earlier for publicly available data
Release date December 12, 2024
License MIT

Intended Use
Primary Use Cases Our model is designed to accelerate research on language models, for use as a building block for generative AI powered features. It provides uses for general purpose AI systems and applications (primarily in English) which require:

1. Memory/compute constrained environments.
2. Latency bound scenarios.
3. Reasoning and logic.
Out-of-Scope Use Cases Our models is not specifically designed or evaluated for all downstream purposes, thus:

1. Developers should consider common limitations of language models as they select use cases, and evaluate and mitigate for accuracy, safety, and fairness before using within a specific downstream use case, particularly for high-risk scenarios.
2. Developers should be aware of and adhere to applicable laws or regulations (including privacy, trade compliance laws, etc.) that are relevant to their use case, including the model’s focus on English.
3. Nothing contained in this Model Card should be interpreted as or deemed a restriction or modification to the license the model is released under.

Data Overview

Training Datasets

Our training data is an extension of the data used for Phi-3 and includes a wide variety of sources from:

  1. Publicly available documents filtered rigorously for quality, selected high-quality educational data, and code.

  2. Newly created synthetic, “textbook-like” data for the purpose of teaching math, coding, common sense reasoning, general knowledge of the world (science, daily activities, theory of mind, etc.).

  3. Acquired academic books and Q&A datasets.

  4. High quality chat format supervised data covering various topics to reflect human preferences on different aspects such as instruct-following, truthfulness, honesty and helpfulness.

Multilingual data constitutes about 8% of our overall data. We are focusing on the quality of data that could potentially improve the reasoning ability for the model, and we filter the publicly available documents to contain the correct level of knowledge.

模型下载

点击下载

评测结果

开源模型对比

排名 大模型 机构 size 总分 医疗 教育 金融 法律 行政公务 心理健康 推理与数学计算 语言与指令遵从
1 DeepSeek-R1 深度求索 671B 80.8 81.9 89.2 82.9 73.8 83.0 61.5 89.2 84.8
2 hunyuan-large 腾讯 389B 80.2 88.6 85.5 84.6 82.4 70.4 73.2 76.8 80.1
3 deepseek-chat-v3-0324(new) 深度求索 671B 77.7 78.0 87.0 76.1 61.7 81.2 64.6 87.6 85.4
4 qwq-32b 阿里巴巴 32B 77.5 76.5 87.2 78.3 60.9 82.2 63.0 87.2 85.2
5 GLM-Z1-32B-0414(new) 智谱AI 32B 75.6 72.6 81.9 74.8 62.2 80.0 63.3 88.5 81.3
6 DeepSeek-R1-Distill-Qwen-32B 深度求索 32B 71.8 69.4 85.1 73.8 51.8 76.0 53.8 83.6 81.4
7 GLM-4-32B-0414(new) 智谱AI 32B 71.8 71.7 76.5 72.3 54.5 76.0 60.9 81.2 81.3
8 Llama-4-Maverick-17B-128E-Instruct-FP8(new) Meta 400B 71.3 75.8 83.0 71.1 48.1 69.0 59.0 82.7 81.6
9 qwen2.5-72b-instruct 阿里巴巴 72B 71.2 74.1 80.0 75.6 51.6 67.0 59.5 79.7 81.7
10 MiniMax-Text-01 minimax 456B 69.7 70.6 79.8 70.0 50.7 71.0 57.8 76.0 81.3
11 qwen2.5-32b-instruct 阿里巴巴 32B 69.1 71.6 82.0 71.8 50.8 64.5 57.8 73.5 80.7
12 DeepSeek-R1-Distill-Qwen-14B 深度求索 14B 66.5 62.8 79.3 67.5 40.2 66.5 55.6 81.0 78.7
13 qwen2.5-14b-instruct 阿里巴巴 14B 65.9 67.7 78.6 68.0 45.4 61.0 56.1 70.4 79.9
14 GLM-Z1-9B-0414(new) 智谱AI 9B 63.2 58.0 66.9 63.8 31.5 75.5 48.8 84.8 76.2
15 qwen2.5-7b-instruct 阿里巴巴 7B 62.5 64.5 74.8 66.5 41.8 53.0 56.0 67.6 76.1
16 Llama-4-Scout-17B-16E-Instruct(new) Meta 109B 62.4 65.4 78.0 62.1 31.8 55.5 54.0 76.0 76.2
17 DeepSeek-R1-Distill-Llama-70B 深度求索 70B 62.2 54.1 72.9 60.6 34.7 70.0 46.2 81.8 77.0
18 internlm2_5-20b-chat 上海人工智能实验室 20B 61.6 65.4 73.9 62.9 44.7 53.5 52.7 62.4 77.2
19 Meta-Llama-3.1-405B-Instruct Meta 405B 61.0 66.1 64.1 59.3 34.7 59.0 53.9 73.2 77.9
20 internlm2_5-7b-chat 上海人工智能实验室 7B 60.6 64.8 66.1 64.1 43.1 55.5 51.0 64.3 76.3
21 GLM-4-9B-0414(new) 智谱AI 9B 60.3 59.3 72.0 62.1 40.8 53.5 49.2 68.3 76.8
22 Llama-3.3-70B-Instruct Meta 70B 59.0 59.2 63.8 56.9 29.9 60.5 49.6 74.2 78.0
23 Mistral-Small-3.1-24B-Instruct-2503(new) Mistral 24B 58.7 53.1 68.0 57.6 33.3 57.0 47.1 76.3 77.1
24 Llama-3.3-70B-Instruct-fp8 Meta 70B 58.4 58.7 63.0 56.0 29.2 59.0 48.5 74.7 78.1
25 GLM-Z1-Rumination-32B-0414(new) 智谱AI 32B 57.5 47.9 69.6 54.3 38.9 56.7 44.1 78.3 70.2
26 Hermes-3-Llama-3.1-405B NousResearch 405B 57.4 55.6 62.3 59.2 29.8 52.5 48.9 72.7 78.0
27 glm-4-9b-chat 智谱AI 9B 56.8 56.0 73.4 57.3 36.5 52.0 47.1 60.1 71.7
28 gemma-3-27b-it Google 27B 56.4 48.2 64.3 56.1 21.3 66.5 44.5 77.4 73.1
29 phi-4 微软 14B 55.3 43.4 59.7 56.0 24.9 64.5 43.2 77.3 73.7
30 Mistral-Small-24B-Instruct-2501 Mistral 24B 53.2 45.8 60.5 53.1 26.3 53.5 40.2 71.4 74.4
31 gemma-3-12b-it Google 12B 52.2 43.6 58.2 46.6 20.4 59.0 41.1 77.3 71.6
32 qwen2.5-3b-instruct 阿里巴巴 3B 50.6 49.8 62.4 53.8 26.9 42.0 43.8 59.1 67.6
33 DeepSeek-R1-Distill-Qwen-7B 深度求索 7B 45.4 28.8 52.8 45.8 19.2 54.0 30.4 70.9 61.2
34 DeepSeek-R1-Distill-Llama-8B 深度求索 8B 45.2 29.7 51.4 42.6 20.3 52.0 31.9 68.5 64.8
35 Llama-3.1-8B-Instruct Meta 8B 44.0 39.9 44.3 40.9 21.3 43.0 37.2 59.9 65.8
36 Meta-Llama-3.1-8B-Instruct-fp8 Meta 8B 42.9 39.2 43.0 40.8 19.5 37.5 34.0 63.1 65.7
37 qwen2.5-1.5b-instruct 阿里巴巴 2B 42.3 43.8 56.0 48.6 25.6 35.0 39.6 38.5 51.5
38 gemma-3-4b-it Google 4B 41.1 31.8 46.6 39.7 16.6 39.5 29.2 67.6 58.0
39 Llama-3.2-3B-Instruct Meta 3B 38.5 34.6 37.5 35.0 17.1 34.0 29.6 57.6 62.7
40 DeepSeek-R1-Distill-Qwen-1.5B 深度求索 2B 35.5 23.8 42.8 34.5 15.4 34.5 23.9 61.8 47.1
41 Llama-3.2-1B-Instruct Meta 1B 30.1 25.4 28.3 28.8 14.6 29.0 21.1 41.6 51.9
42 qwen2.5-0.5b-instruct 阿里巴巴 1B 30.0 29.2 36.8 33.9 18.4 22.5 24.5 35.2 39.2
43 gemma-3-1b-it Google 1B 28.6 23.6 31.9 28.7 15.4 29.0 20.6 31.5 48.2

精益评测：更细分的维度

  • 医疗
    • 医师
    • 护理
    • 药师
    • 医技
    • 医学基础知识
    • 医学考研
  • 教育
    • 小学学科
    • 初中学科
    • 中考
    • 高中学科
    • 高考
    • 高等教育
    • 考研
    • 教师资格
  • 金融排行榜
    • 财务
    • 银行
    • 保险
    • 证券
    • 其他金融资格考试
    • 金融基础知识
    • 金融应用
  • 法律排行榜
    • 律师资格考试
  • 行政公务排行榜
    • 公务员考试
  • 心理健康排行榜
  • 推理与数学计算排行榜
    • 演绎推理
    • 常识推理
    • 符号推理BBH
    • 算术能力
    • 表格问答
    • 高中奥数
    • 初中奥数
    • 小学奥数
  • 语言与指令遵从排行榜
    • 成语理解
    • 情感分析
    • 文本蕴含
    • 分类能力
    • 信息抽取
    • 阅读理解
    • C3中文阅读理解
    • 代词理解CLUEWSC
    • 诗词匹配CCPM
    • 中文指令遵从

以上细分能力评测结果，详见github