OpenCompass 大模型评测
AI-摘要
切换
Tianli GPT
AI初始化中...
介绍自己
生成本文简介
推荐相关文章
前往主页
前往tianli博客
OpenCompass 大模型评测
项目地址: tutorial/opencompass/opencompass_tutorial.md at main · InternLM/tutorial (github.com)
视频地址: OpenCompass 大模型评测
✍课程笔记
关于评测的三个问题?
我们为什么需要评测?
我们需要评测什么?
如何进行模型的评测
通过不同模型的来进行评测
基座模型：一般是经过海量的文本数据以自监督学习的方式进行训练获得的模型（如OpenAI的GPT-3，Meta的LLaMA），往往具有强大的文字续写能力。
对话模型：一般是在的基座模型的基础上，经过指令微调或人类偏好对齐获得的模型（如OpenAI的ChatGPT、上海人工智能实验室的书生·浦语），能理解人类指令，具有较强的对话能力。
工具架构
模型层：大模型评测所涉及的主要模型种类，OpenCompass以基座模型和对话模型作为重点评测对象。
能力层：OpenCompass从本方案从通用能力和特色能力两个方面来进行评测维度设计。在模型通用能力方面，从语言、知识、理解、推理、安全等多个能力维度进行评测。在特色能力方面，从长文本、代码、工具、知识增强等维度进行评测。
方法层：OpenCompass采用客观评测与主观评测两种评测方式。客观评测能便捷地评估模型在具有确定答案（如选择，填空，封闭式问答等）的任务上的能力，主观评测能评估用户对模型回复的真实满意度，OpenCompass采用基于模型辅助的主观评测和基于人类反馈的主观评测两种方式。
工具层：OpenCompass提供丰富的功能支持自动化地开展大语言模型的高效评测。包括分布式评测技术，提示词工程，对接评测数据库，评测榜单发布，评测报告生成等诸多功能。
能力维度
设计思路
为准确、全面、系统化地评估大语言模型的能力，OpenCompass从通用人工智能的角度出发，结合学术界的前沿进展和工业界的最佳实践，提出一套面向实际应用的模型能力评价体系。OpenCompass能力维度体系涵盖通用能力和特色能力两大部分。
通用能力涵盖学科综合能力、知识能力、语言能力、理解能力、推理能力、安全能力，共计六大维度构造立体全面的模型能力评价体系。
特色能力
评测方法
OpenCompass采取客观评测与主观评测相结合的方法。针对具有确定性答案的能力维度和场景，通过构造丰富完善的评测集，对模型能力进行综合评价。针对体现模型能力的开放式或半开放式的问题、模型安全问题等，采用主客观相结合的评测方式。
客观评测
针对具有标准答案的客观问题，我们可以我们可以通过使用定量指标比较模型的输出与标准答案的差异，并根据结果衡量模型的性能。同时，由于大语言模型输出自由度较高，在评测阶段，我们需要对其输入和输出作一定的规范和设计，尽可能减少噪声输出在评测阶段的影响，才能对模型的能力有更加完整和客观的评价。
为了更好地激发出模型在题目测试领域的能力，并引导模型按照一定的模板输出答案，OpenCompass采用提示词工程 （prompt engineering）和语境学习（in-context learning）进行客观评测。
在客观评测的具体实践中，我们通常采用下列两种方式进行模型输出结果的评测：
判别式评测：该评测方式基于将问题与候选答案组合在一起，计算模型在所有组合上的困惑度（perplexity），并选择困惑度最小的答案作为模型的最终输出。例如，若模型在
问题? 答案1上的困惑度为 0.1，在问题? 答案2上的困惑度为 0.2，最终我们会选择答案1作为模型的输出。生成式评测：该评测方式主要用于生成类任务，如语言翻译、程序生成、逻辑分析题等。具体实践时，使用问题作为模型的原始输入，并留白答案区域待模型进行后续补全。我们通常还需要对其输出进行后处理，以保证输出满足数据集的要求。
主观评测
语言表达生动精彩，变化丰富，大量的场景和能力无法凭借客观指标进行评测。针对如模型安全和模型语言能力的评测，以人的主观感受为主的评测更能体现模型的真实能力，并更符合大模型的实际使用场景。
OpenCompass采取的主观评测方案是指借助受试者的主观判断对具有对话能力的大语言模型进行能力评测。在具体实践中，我们提前基于模型的能力维度构建主观测试问题集合，并将不同模型对于同一问题的不同回复展现给受试者，收集受试者基于主观感受的评分。由于主观测试成本高昂，本方案同时也采用使用性能优异的大语言模拟人类进行主观打分。在实际评测中，本文将采用真实人类专家的主观评测与基于模型打分的主观评测相结合的方式开展模型能力评估。
在具体开展主观评测时，OpenComapss采用单模型回复满意度统计和多模型满意度比较两种方式开展具体的评测工作。
主流大模型评测框架
OpenCompass评测体系介绍
🖊课程作业
基础作业
使用 OpenCompass 评测 InternLM2-Chat-7B 模型在 C-Eval 数据集上的性能
安装
面向GPU的环境安装
conda create --name opencompass --clone=/root/share/conda_envs/internlm-base
source activate opencompass
git clone https://github.com/open-compass/opencompass
cd opencompass
pip install -e .有部分第三方功能,如代码能力基准测试 Humaneval 以及 Llama格式的模型评测,可能需要额外步骤才能正常运行，如需评测，详细步骤请参考安装指南。
数据准备
# 解压评测数据集到 data/ 处
cp /share/temp/datasets/OpenCompassData-core-20231110.zip /root/opencompass/
unzip OpenCompassData-core-20231110.zip
# 将会在opencompass下看到data文件夹查看支持的数据集和模型
# 列出所有跟 internlm 及 ceval 相关的配置
python tools/list_configs.py internlm ceval将会看到
+--------------------------+--------------------------------------------------------+
| Model | Config Path |
|--------------------------+--------------------------------------------------------|
| hf_internlm_20b | configs/models/hf_internlm/hf_internlm_20b.py |
| hf_internlm_7b | configs/models/hf_internlm/hf_internlm_7b.py |
| hf_internlm_chat_20b | configs/models/hf_internlm/hf_internlm_chat_20b.py |
| hf_internlm_chat_7b | configs/models/hf_internlm/hf_internlm_chat_7b.py |
| hf_internlm_chat_7b_8k | configs/models/hf_internlm/hf_internlm_chat_7b_8k.py |
| hf_internlm_chat_7b_v1_1 | configs/models/hf_internlm/hf_internlm_chat_7b_v1_1.py |
| internlm_7b | configs/models/internlm/internlm_7b.py |
| ms_internlm_chat_7b_8k | configs/models/ms_internlm/ms_internlm_chat_7b_8k.py |
+--------------------------+--------------------------------------------------------+
+----------------------------+------------------------------------------------------+
| Dataset | Config Path |
|----------------------------+------------------------------------------------------|
| ceval_clean_ppl | configs/datasets/ceval/ceval_clean_ppl.py |
| ceval_gen | configs/datasets/ceval/ceval_gen.py |
| ceval_gen_2daf24 | configs/datasets/ceval/ceval_gen_2daf24.py |
| ceval_gen_5f30c7 | configs/datasets/ceval/ceval_gen_5f30c7.py |
| ceval_ppl | configs/datasets/ceval/ceval_ppl.py |
| ceval_ppl_578f8d | configs/datasets/ceval/ceval_ppl_578f8d.py |
| ceval_ppl_93e5ce | configs/datasets/ceval/ceval_ppl_93e5ce.py |
| ceval_zero_shot_gen_bd40ef | configs/datasets/ceval/ceval_zero_shot_gen_bd40ef.py |
+----------------------------+------------------------------------------------------+
启动评测
确保按照上述步骤正确安装 OpenCompass 并准备好数据集后，可以通过以下命令评测 InternLM-Chat-7B 模型在 C-Eval 数据集上的性能。由于 OpenCompass 默认并行启动评估过程，我们可以在第一次运行时以 --debug 模式启动评估，并检查是否存在问题。在 --debug 模式下，任务将按顺序执行，并实时打印输出。
python run.py --datasets ceval_gen --hf-path /share/temp/model_repos/internlm-chat-7b/ --tokenizer-path /share/temp/model_repos/internlm-chat-7b/ --tokenizer-kwargs padding_side='left' truncation='left' trust_remote_code=True --model-kwargs trust_remote_code=True device_map='auto' --max-seq-len 2048 --max-out-len 16 --batch-size 4 --num-gpus 1 --debug命令解析
--datasets ceval_gen \
--hf-path /share/temp/model_repos/internlm-chat-7b/ \ # HuggingFace 模型路径
--tokenizer-path /share/temp/model_repos/internlm-chat-7b/ \ # HuggingFace tokenizer 路径（如果与模型路径相同，可以省略）
--tokenizer-kwargs padding_side='left' truncation='left' trust_remote_code=True \ # 构建 tokenizer 的参数
--model-kwargs device_map='auto' trust_remote_code=True \ # 构建模型的参数
--max-seq-len 2048 \ # 模型可以接受的最大序列长度
--max-out-len 16 \ # 生成的最大 token 数
--batch-size 2 \ # 批量大小
--num-gpus 1 # 运行模型所需的 GPU 数量
--debug如果一切正常，您应该看到屏幕上显示 “Starting inference process”：
[2024-01-12 18:23:55,076] [opencompass.openicl.icl_inferencer.icl_gen_inferencer] [INFO] Starting inference process...如果中途遇到如下所示的报错
ERROR - /root/opencompass/opencompass/tasks/openicl_eval.py - _score - 236 - Task lopencompass.models.huggingface.HuggingFace_model_repos_internlm2-chat-7b/ceval-marxism]: No predictions found.将命令改成，将--batch-size调小一点调成2或1
python run.py --datasets ceval_gen --hf-path /share/temp/model_repos/internlm-chat-7b/ --tokenizer-path /share/temp/model_repos/internlm-chat-7b/ --tokenizer-kwargs padding_side='left' truncation='left' trust_remote_code=True --model-kwargs trust_remote_code=True device_map='auto' --max-seq-len 2048 --max-out-len 16 --batch-size 2 --num-gpus 1 --debug评测完成后，将会看到：
dataset version metric mode opencompass.models.huggingface.HuggingFace_model_repos_internlm-chat-7b
---------------------------------------------- --------- ------------- ------ -------------------------------------------------------------------------
ceval-computer_network db9ce2 accuracy gen 36.84
ceval-operating_system 1c2571 accuracy gen 36.84
ceval-computer_architecture a74dad accuracy gen 28.57
ceval-college_programming 4ca32a accuracy gen 32.43
ceval-college_physics 963fa8 accuracy gen 31.58
ceval-college_chemistry e78857 accuracy gen 16.67
ceval-advanced_mathematics ce03e2 accuracy gen 21.05
ceval-probability_and_statistics 65e812 accuracy gen 38.89
ceval-discrete_mathematics e894ae accuracy gen 18.75
ceval-electrical_engineer ae42b9 accuracy gen 35.14
ceval-metrology_engineer ee34ea accuracy gen 50
ceval-high_school_mathematics 1dc5bf accuracy gen 22.22
ceval-high_school_physics adf25f accuracy gen 31.58
ceval-high_school_chemistry 2ed27f accuracy gen 15.79
ceval-high_school_biology 8e2b9a accuracy gen 36.84
ceval-middle_school_mathematics bee8d5 accuracy gen 26.32
ceval-middle_school_biology 86817c accuracy gen 61.9
ceval-middle_school_physics 8accf6 accuracy gen 63.16
ceval-middle_school_chemistry 167a15 accuracy gen 65
ceval-veterinary_medicine b4e08d accuracy gen 47.83
ceval-college_economics f3f4e6 accuracy gen 38.18
ceval-business_administration c1614e accuracy gen 33.33
ceval-marxism cf874c accuracy gen 68.42
ceval-mao_zedong_thought 51c7a4 accuracy gen 70.83
ceval-education_science 591fee accuracy gen 58.62
ceval-teacher_qualification 4e4ced accuracy gen 68.18
ceval-high_school_politics 5c0de2 accuracy gen 26.32
ceval-high_school_geography 865461 accuracy gen 47.37
ceval-middle_school_politics 5be3e7 accuracy gen 52.38
ceval-middle_school_geography 8a63be accuracy gen 58.33
ceval-modern_chinese_history fc01af accuracy gen 73.91
ceval-ideological_and_moral_cultivation a2aa4a accuracy gen 63.16
ceval-logic f5b022 accuracy gen 31.82
ceval-law a110a1 accuracy gen 25
ceval-chinese_language_and_literature 0f8b68 accuracy gen 30.43
ceval-art_studies 2a1300 accuracy gen 60.61
ceval-professional_tour_guide 4e673e accuracy gen 62.07
ceval-legal_professional ce8787 accuracy gen 39.13
ceval-high_school_chinese 315705 accuracy gen 57.89
ceval-high_school_history 7eb30a accuracy gen 70
ceval-middle_school_history 48ab4a accuracy gen 59.09
ceval-civil_servant 87d061 accuracy gen 53.19
ceval-sports_science 70f27b accuracy gen 52.63
ceval-plant_protection 8941f9 accuracy gen 59.09
ceval-basic_medicine c409d6 accuracy gen 47.37
ceval-clinical_medicine 49e82d accuracy gen 40.91
ceval-urban_and_rural_planner 95b885 accuracy gen 45.65
ceval-accountant 002837 accuracy gen 26.53
ceval-fire_engineer bc23f5 accuracy gen 22.58
ceval-environmental_impact_assessment_engineer c64e2d accuracy gen 64.52
ceval-tax_accountant 3a5e3c accuracy gen 34.69
ceval-physician 6e277d accuracy gen 44.9
ceval-stem - naive_average gen 35.87
ceval-social-science - naive_average gen 52.2
ceval-humanities - naive_average gen 52.1
ceval-other - naive_average gen 44.73
ceval-hard - naive_average gen 24.57
ceval - naive_average gen 44.32进阶作业:
使用 OpenCompass 评测 InternLM2-Chat-7B 模型使用 LMDeploy 0.2.0 部署后在 C-Eval 数据集上的性能
因为LMDeploy不支持internlm2-chat-7b所有我们用internlm-chat-7b代替试一下
AssertionError: 'internlm2-chat-7b' is not supported. The supported models are: dict_keys(['base', 'llama', 'internlm', 'vicuna', 'wizardlM', 'internlm-chat-7b', 'internlm-chat', 'internlm-chat-7b-8k', 'internlm-chat-20b', 'internlm-20b', 'baichuan-7b', 'baichuan2-7b', 'puyu', 'llama2', 'qwen-7b', 'qwen-14b', 'codellama', 'solar', 'ultralm', 'ultracm', 'yi'])实验步骤
首先安装LMDeploy最新版本
pip install lmdeploy-0.2.1-cp310-cp310-manylinux2014_x86_64.whl然后转换internLM-Chat-7B为将模型转为 lmdeploy TurboMind 的格式
lmdeploy convert internlm-chat-7b /root/model/Shanghai_AI_Laboratory/internlm-chat-7b接着我们启动 TurboMind推理+API服务
lmdeploy serve api_server ./workspace \
--server_name 0.0.0.0 \
--server_port 23333 \
--instance_num 64 \
--tp 1
最后我们进行推理
python run.py configs/eval_internlm_chat_turbomind_api.py -w outputs/turbomind/internlm-chat-7b结果如下
dataset version metric mode internlm-chat-7b-turbomind
ceval-computer_network db9ce2 accuracy gen 63.16
ceval-operating_system 1c2571 accuracy gen 73.68
ceval-computer_architecture a74dad accuracy gen 57.14
ceval-college_programming 4ca32a accuracy gen 59.46
ceval-college_physics 963fa8 accuracy gen 52.63
ceval-college_chemistry e78857 accuracy gen 33.33
ceval-advanced_mathematics ce03e2 accuracy gen 26.32
ceval-probability_and_statistics 65e812 accuracy gen 38.89
ceval-discrete_mathematics e894ae accuracy gen 31.25
ceval-electrical_engineer ae42b9 accuracy gen 37.84
ceval-metrology_engineer ee34ea accuracy gen 66.67
ceval-high_school_mathematics 1dc5bf accuracy gen 38.89
ceval-high_school_physics adf25f accuracy gen 42.11
ceval-high_school_chemistry 2ed27f accuracy gen 47.37
ceval-high_school_biology 8e2b9a accuracy gen 36.84
ceval-middle_school_mathematics bee8d5 accuracy gen 47.37
ceval-middle_school_biology 86817c accuracy gen 80.95
ceval-middle_school_physics 8accf6 accuracy gen 68.42
ceval-middle_school_chemistry 167a15 accuracy gen 95.00
ceval-veterinary_medicine b4e08d accuracy gen 43.48
ceval-college_economics f3f4e6 accuracy gen 49.09
ceval-business_administration c1614e accuracy gen 60.61
ceval-marxism cf874c accuracy gen 84.21
ceval-mao_zedong_thought 51c7a4 accuracy gen 70.83
ceval-education_science 591fee accuracy gen 72.41
ceval-teacher_qualification 4e4ced accuracy gen 79.55
ceval-high_school_politics 5c0de2 accuracy gen 89.47
ceval-high_school_geography 865461 accuracy gen 63.16
ceval-middle_school_politics 5be3e7 accuracy gen 76.19
ceval-middle_school_geography 8a63be accuracy gen 75.00
ceval-modern_chinese_history fc01af accuracy gen 69.57
ceval-ideological_and_moral_cultivation a2aa4a accuracy gen 84.21
ceval-logic f5b022 accuracy gen 54.55
ceval-law a110a1 accuracy gen 50.00
ceval-chinese_language_and_literature 0f8b68 accuracy gen 56.52
ceval-art_studies 2a1300 accuracy gen 66.67
ceval-professional_tour_guide 4e673e accuracy gen 82.76
ceval-legal_professional ce8787 accuracy gen 52.17
ceval-high_school_chinese 315705 accuracy gen 73.68
ceval-high_school_history 7eb30a accuracy gen 75.00
ceval-middle_school_history 48ab4a accuracy gen 86.36
ceval-civil_servant 87d061 accuracy gen 63.83
ceval-sports_science 70f27b accuracy gen 78.95
ceval-plant_protection 8941f9 accuracy gen 77.27
ceval-basic_medicine c409d6 accuracy gen 63.16
ceval-clinical_medicine 49e82d accuracy gen 59.09
ceval-urban_and_rural_planner 95b885 accuracy gen 67.39
ceval-accountant 002837 accuracy gen 48.98
ceval-fire_engineer bc23f5 accuracy gen 54.84
ceval-environmental_impact_assessment_engineer c64e2d accuracy gen 58.06
ceval-tax_accountant 3a5e3c accuracy gen 44.90
ceval-physician 6e277d accuracy gen 57.14- 感谢你赐予我前进的力量
-
微信 
支付宝
赞赏者名单
因为你们的支持让我意识到写文章的价值🙏
本文是原创文章，采用 CC BY-NC-ND 4.0 协议，完整转载请注明来自 窝窝头
评论
匿名评论
隐私政策
你无需删除空行，直接评论以获取最佳展示效果
