一、环境配置
我们来配置OpenCompass实验环境:
conda create -n opencompass python=3.10
conda activate opencompass
conda install pytorch==2.1.2 torchvision==0.16.2 torchaudio==2.1.2 pytorch-cuda=12.1 -c pytorch -c nvidia -y
cd /root
git clone -b 0.2.4 https://github.com/open-compass/opencompass
cd opencompass
pip install -e .
apt-get update
apt-get install cmake
pip install -r requirements.txt
pip install protobuf
二、数据准备
我们需要准备ceval
数据集,开发机中已经包含了数据压缩包,因此我们将数据解压到opencompass
目录:
cp /share/temp/datasets/OpenCompassData-core-20231110.zip /root/opencompass/
unzip OpenCompassData-core-20231110.zip
三、启动评测
打开 opencompass
文件夹下configs/models/hf_internlm/
的hf_internlm2_chat_1_8b.py
文件,修改文件如下:
from opencompass.models import HuggingFaceCausalLM
models = [
dict(
type=HuggingFaceCausalLM,
abbr='internlm2-1.8b-hf',
path="/share/new_models/Shanghai_AI_Laboratory/internlm2-chat-1_8b",
tokenizer_path='/share/new_models/Shanghai_AI_Laboratory/internlm2-chat-1_8b',
model_kwargs=dict(
trust_remote_code=True,
device_map='auto',
),
tokenizer_kwargs=dict(
padding_side='left',
truncation_side='left',
use_fast=False,
trust_remote_code=True,
),
max_out_len=100,
min_out_len=1,
max_seq_len=2048,
batch_size=8,
run_cfg=dict(num_gpus=1, num_procs=1),
)
]
通过以下命令评测 InternLM2-Chat-1.8B
模型在 C-Eval
数据集上的性能:
#环境变量配置
export MKL_SERVICE_FORCE_INTEL=1
python run.py --datasets ceval_gen --models hf_internlm2_chat_1_8b --debug
评测过程时间较长,我在10%A100算力的开发机中评测时间约4.5-5小时左右。评测结果如下: