跑MT-bench-101数据集时报错 No module named 'rouge'

[chaserRen]

执行指令 python run.py ./configs/eval_subjective_mtbench101.py
我在opencompass目录下放置了mtbench101数据集，路径：opencompass/data/subjective/mtbench101.jsonl
报错如下

(opencompass) (base) [qa-caif-cicd@SH-IDCA1404-10-140-54-48 opencompass]$ python run.py ./configs/eval_subjective_mtbench101.py 
/nvme/envs/opencompass/lib/python3.10/site-packages/torch/cuda/__init__.py:129: UserWarning: CUDA initialization: The NVIDIA driver on your system is too old (found version 11070). Please update your GPU driver by downloading and installing a new version from the URL: http://www.nvidia.com/Download/index.aspx Alternatively, go to: https://pytorch.org to install a PyTorch version that has been compiled with your version of the CUDA driver. (Triggered internally at /opt/conda/conda-bld/pytorch_1729647382455/work/c10/cuda/CUDAFunctions.cpp:108.)
  return torch._C._cuda_getDeviceCount() > 0
Traceback (most recent call last):
  File "/home/qa-caif-cicd/renzhifei/opencompass/run.py", line 1, in <module>
    from opencompass.cli.main import main
  File "/home/qa-caif-cicd/renzhifei/opencompass/opencompass/cli/main.py", line 15, in <module>
    from opencompass.utils.run import (fill_eval_cfg, fill_infer_cfg,
  File "/home/qa-caif-cicd/renzhifei/opencompass/opencompass/utils/run.py", line 9, in <module>
    from opencompass.datasets.custom import make_custom_dataset_config
  File "/home/qa-caif-cicd/renzhifei/opencompass/opencompass/datasets/__init__.py", line 65, in <module>
    from .longbench import *  # noqa: F401, F403
  File "/home/qa-caif-cicd/renzhifei/opencompass/opencompass/datasets/longbench/__init__.py", line 1, in <module>
    from .evaluators import LongBenchClassificationEvaluator  # noqa: F401, F403
  File "/home/qa-caif-cicd/renzhifei/opencompass/opencompass/datasets/longbench/evaluators.py", line 9, in <module>
    from rouge import Rouge
ModuleNotFoundError: No module named 'rouge'

配置文件

from mmengine.config import read_base

with read_base():
    from .datasets.subjective.multiround.mtbench101_judge import subjective_datasets

from opencompass.models import HuggingFaceCausalLM, HuggingFace, HuggingFaceChatGLM3, OpenAI
from opencompass.partitioners import NaivePartitioner, SizePartitioner
from opencompass.partitioners.sub_naive import SubjectiveNaivePartitioner
from opencompass.partitioners.sub_size import SubjectiveSizePartitioner
from opencompass.runners import LocalRunner
from opencompass.runners import SlurmSequentialRunner
from opencompass.tasks import OpenICLInferTask
from opencompass.tasks.subjective_eval import SubjectiveEvalTask
from opencompass.summarizers import MTBench101Summarizer

# ---------------------------------------------------------------------------------------------------------

api_meta_template = dict(
    round=[
        dict(role='SYSTEM', api_role='SYSTEM'),
        dict(role='HUMAN', api_role='HUMAN'),
        dict(role='BOT', api_role='BOT', generate=True),
    ]
)

# -------------Inference Stage ----------------------------------------
# For subjective evaluation, we often set do sample for models
models = [
    dict(
        type=HuggingFaceChatGLM3,
        abbr='chatglm3-6b-hf',
        path='THUDM/chatglm3-6b',
        tokenizer_path='THUDM/chatglm3-6b',
        model_kwargs=dict(
            device_map='auto',
            trust_remote_code=True,
        ),
        tokenizer_kwargs=dict(
            padding_side='left',
            truncation_side='left',
            trust_remote_code=True,
        ),
        generation_kwargs=dict(
            do_sample=True,
        ),
        meta_template=api_meta_template,
        max_out_len=4096,
        max_seq_len=4096,
        batch_size=1,
        run_cfg=dict(num_gpus=2, num_procs=1),
    )
]

datasets = [*subjective_datasets]

infer = dict(
    partitioner=dict(type=SizePartitioner, max_task_size=10000),
    runner=dict(
        type=SlurmSequentialRunner,
        partition='llm_dev2',
        quotatype='auto',
        max_num_workers=32,
        task=dict(type=OpenICLInferTask),
    ),
)

# -------------Evalation Stage ----------------------------------------

## ------------- JudgeLLM Configuration
judge_models = [dict(
    abbr='GPT4-Turbo',
    type=OpenAI,
    path='gpt-4-1106-preview', # To compare with the official leaderboard, please use gpt-4-1106-preview
    key='',  # The key will be obtained from $OPENAI_API_KEY, but you can write down your key here as well
    meta_template=api_meta_template,
    query_per_second=16,
    max_out_len=4096,
    max_seq_len=4096,
    batch_size=8,
    temperature=0.8,
)]

## ------------- Evaluation Configuration



eval = dict(
    partitioner=dict(type=SubjectiveSizePartitioner, max_task_size=100000, mode='singlescore', models=models, judge_models=judge_models),
    runner=dict(type=LocalRunner, max_num_workers=32, task=dict(type=SubjectiveEvalTask)),
)

summarizer = dict(type=MTBench101Summarizer, judge_type='single')

work_dir = 'outputs/mtbench101/'

[bittersweet1999]

这个应该是没装好opencompass，再pip install -e 一下试试呢

[sir3mat]

mtbench101 summarizer doesn't work with this setup. What i am missing?

from opencompass.models import OpenAISDK
from mmengine.config import read_base
from opencompass.partitioners import SizePartitioner
from opencompass.runners import LocalRunner
from opencompass.tasks import OpenICLInferTask
from opencompass.tasks.subjective_eval import SubjectiveEvalTask
from opencompass.tasks.subjective_eval import SubjectiveEvalTask
from opencompass.summarizers.subjective.mtbench101 import MTBench101Summarizer
from opencompass.partitioners.sub_size import SubjectiveSizePartitioner

with read_base():

    from opencompass.configs.datasets.subjective.multiround.mtbench101_judge_new import (
        mtbench101_datasets,
    )

work_dir = "./outputs/mtbench"

datasets = [*mtbench101_datasets]

api_meta_template = dict(
    round=[
        dict(role="SYSTEM", api_role="SYSTEM"),
        dict(role="HUMAN", api_role="HUMAN"),
        dict(role="BOT", api_role="BOT", generate=True),
    ]
)

models = [
    dict(
        abbr="casperhansen/llama-3.3-70b-instruct-awq",
        type=OpenAISDK,
        key="EMPTY",  # API key
        openai_api_base="http://172.18.21.142:8000/v1",  # Service address
        path="casperhansen/llama-3.3-70b-instruct-awq",  # Model name for service request
        tokenizer_path="casperhansen/llama-3.3-70b-instruct-awq",  # The tokenizer name or path, if set to `None`, uses the default `gpt-4` tokenizer
        rpm_verbose=True,  # Whether to print request rate
        meta_template=api_meta_template,  # Service request template
        # Service request parameters
        query_per_second=1,  # Service request rate
        max_seq_len=120000,  # Context length
        max_out_len=64000,  # Maximum output length
        temperature=0.1,  # Generation temperature
        batch_size=64,  # Batch size
        retry=3,  # Number of retries
    )
]


infer = dict(
    partitioner=dict(type=SizePartitioner, max_task_size=5000, gen_task_coef=20),
    runner=dict(
        type=LocalRunner, max_num_workers=32, task=dict(type=OpenICLInferTask), retry=5
    ),
)

## ------------- JudgeLLM Configuration
judge_models = [
    dict(
        abbr="casperhansen/llama-3.3-70b-instruct-awq",
        type=OpenAISDK,
        key="EMPTY",  # API key
        openai_api_base="http://<ipaddress>/v1",  # Service address
        path="casperhansen/llama-3.3-70b-instruct-awq",  # Model name for service request
        tokenizer_path="casperhansen/llama-3.3-70b-instruct-awq",  # The tokenizer name or path, if set to `None`, uses the default `gpt-4` tokenizer
        rpm_verbose=True,  # Whether to print request rate
        meta_template=api_meta_template,  # Service request template
        # Service request parameters
        query_per_second=1,  # Service request rate
        max_seq_len=120000,  # Context length
        max_out_len=64000,  # Maximum output length
        temperature=0.1,  # Generation temperature
        batch_size=64,  # Batch size
        retry=3,  # Number of retries
    )
]


eval = dict(
    partitioner=dict(
        type=SubjectiveSizePartitioner,
        max_task_size=100000,
        models=models,
        judge_models=judge_models,
    ),
    runner=dict(
        type=LocalRunner, max_num_workers=32, task=dict(type=SubjectiveEvalTask)
    ),
)


summarizer = dict(type=MTBench101Summarizer, judge_type='single')

i have the results here: outputs/mtbench/20241226_065941/results/casperhansen/llama-3.3-70b-instruct-awq_judged-by--casperhansen/llama-3.3-70b-instruct-awq/mtbench101.json

"avg": 8.628,
    "GR": 7.676056338028169,
    "IC": 8.606666666666667,
    "AR": 8.431372549019608,
    "FR": 9.175675675675675,
    "MR": 7.351851851851852,
    "CC": 9.585034013605442,
    "TS": 9.0,
    "CR": 8.926470588235293,
    "SA": 9.301369863013699,
    "SI": 8.651006711409396,
    "CM": 7.825,
    "PI": 7.942528735632184,
    "SC": 9.688311688311689,
    "details": {
        "0": {
            "origin_prompt": [
                {
                    "role": "SYSTEM",
                    "prompt": "Please act as an impartial judge follow this instructions: ..... \n"
                },
                {
                    "role": "HUMAN",
                    "prompt": "The dialogue need to be judged is: \n *** \n \n\n Human: .... \n"
                }
            ],
            "prediction": "The AI assistant's response accurately addresses the problem .... Rating: [[10]]",
            "gold": {
                "multi_id": 1,
                "task": "GR",
                "turn_id": "1"
            }
        },
        .....

How to setup the summarizer correctly with mtbench101_new

[bittersweet1999]

plz just use from opencompass.configs.datasets.subjective.multiround.mtbench101_judge when you use from opencompass.summarizers.subjective.mtbench101 import MTBench101Summarizer as summarizer.

If you want to use 'from opencompass.configs.datasets.subjective.multiround.mtbench101_judge'_new, plz use from opencompass.summarizers import DefaultSubjectiveSummarizer and summarizer = dict(type=DefaultSubjectiveSummarizer)

[sir3mat]

thanks :))) !!!
what are the differences between _new and not new?

[bittersweet1999]

we are planning to merge llm-based evaluation and rule-based evaluation on summarizer. So the new config is adapt to a new summarizer(DefaultSubjectiveSummarizer), which is same with DefaultSummarizer

[sir3mat]

okok love to hear this news!