[Feature] Add CompassArena (#828)

* add compass arena

* add compass_arena

* add compass arena

* Update opencompass/summarizers/subjective/compass_arena.py

Co-authored-by: Songyang Zhang <tonysy@users.noreply.github.com>

* Update opencompass/summarizers/subjective/__init__.py

Co-authored-by: Songyang Zhang <tonysy@users.noreply.github.com>

* Update opencompass/datasets/subjective/compass_arena.py

Co-authored-by: Songyang Zhang <tonysy@users.noreply.github.com>

* Update opencompass/datasets/subjective/__init__.py

Co-authored-by: Songyang Zhang <tonysy@users.noreply.github.com>

* Update configs/eval_subjective_compassarena.py

Co-authored-by: Songyang Zhang <tonysy@users.noreply.github.com>

* Update configs/datasets/subjective/compassarena/compassarena_compare.py

Co-authored-by: Songyang Zhang <tonysy@users.noreply.github.com>

* Update configs/eval_subjective_compassarena.py

Co-authored-by: Songyang Zhang <tonysy@users.noreply.github.com>

* Update configs/datasets/subjective/compassarena/compassarena_compare.py

Co-authored-by: Songyang Zhang <tonysy@users.noreply.github.com>

* fix check position bias

---------

Co-authored-by: Songyang Zhang <tonysy@users.noreply.github.com>

configs/datasets/subjective/compassarena/compassarena_compare.py

@@ -0,0 +1,160 @@
+from opencompass.openicl.icl_prompt_template import PromptTemplate
+from opencompass.openicl.icl_retriever import ZeroRetriever
+from opencompass.openicl.icl_inferencer import GenInferencer
+from opencompass.openicl.icl_evaluator import LMEvaluator
+from opencompass.datasets import CompassArenaDataset
+
+subjective_reader_cfg = dict(
+    input_columns=['question', 'ref'],
+    output_column='judge',
+    )
+
+data_path ="data/subjective/"
+
+subjective_datasets = []
+
+base_prompt = """
+
+[回答1开始]
+{prediction}
+[回答1结束]
+
+[回答2开始]
+{prediction2}
+[回答2结束]
+
+根据评分要求，在以下 3 个选项中做出选择:
+A. 回答1更好
+B. 回答2更好
+C. 回答1、2平局
+并提供你的解释原因。
+
+如果你认为回答1更好，你的输出应形如：
+选择：A
+原因：blahblah blahblah\n
+
+如果你认为回答2更好，你的输出应形如：
+选择：B
+原因：blahblah blahblah\n
+
+如果你认为回答1、2打成平手，你的输出应形如：
+选择：C
+原因：blahblah blahblah\n
+"""
+
+knowledge_prompt = """
+请根据提供的 评分要求，用户问题，参考答案 以及 相应的两个回答（回答1，回答2），判断两个回答中哪一个更好。
+评分要求（重要性依次递减）:
+1. 更好的回答能与参考答案吻合或表明参考答案的意思。
+2. 在都准确答对问题的前提下，更好的回答能对知识点进行额外补充，且补充的知识准确无误。
+3. 更好的回答更加符合与人类对话的习惯，包括语气、情调等。
+
+[用户问题]
+{question}
+
+[参考答案]
+{ref}
+""" + base_prompt
+
+
+language_prompt = """
+请根据提供的 评分要求，用户问题 以及 相应的两个回答（回答1，回答2），判断两个回答中哪一个更好。
+评分要求（重要性依次递减）:
+1. 在有明确的参考答案的情况下，越贴近参考答案或表明了参考答案的意思的回答越好。
+2. 更好的回答在语言表达上更流畅，更加符合与人类对话的习惯，包括语气、情调等
+3. 在都准确答对问题的前提下，更好的回答能进行额外补充，且补充的内容准确无误。
+
+[用户问题]
+{question}
+
+[参考答案]
+{ref}
+""" + base_prompt
+
+
+math_prompt = """
+请根据提供的 评分要求，用户问题，参考答案 以及 相应的两个回答（回答1，回答2），判断两个回答中哪一个更好。
+评分要求（重要性依次递减）:
+1. 更好的回答的答案能和参考答案一致。
+2. 若两个回答的答案都与参考答案不一致，则更好的回答的推理过程应更加合理。
+3. 更好的回答更加符合与人类对话的习惯，包括语气、情调等。
+
+[用户问题]
+{question}
+
+[参考答案]
+{ref}
+""" + base_prompt
+
+reason_prompt = math_prompt
+
+qa_prompt = """
+请根据提供的 评分要求，用户问题 以及 相应的两个回答（回答1，回答2），判断两个回答中哪一个更好。
+评分要求（重要性依次递减）:
+1. 好的回答必须首先具有事实正确性，即除了想象的内容外，所引用或阐述的各种信息都是真实正确的
+2. 好的回答必须具有逻辑连贯性，围绕一个中心进行回答，且前后连贯，逻辑没有问题
+3. 在都准确答对问题的前提下，更好的回答能进行额外补充，且补充的内容准确无误
+
+[用户问题]
+{question}
+""" + base_prompt
+
+
+
+creation_prompt = """
+请根据提供的 评分要求，用户问题 以及 相应的两个回答（回答1，回答2），判断两个回答中哪一个更好。
+评分要求（重要性依次递减）:
+1. 好的回答必须首先符合用户问题里的各种需求，不能跑题 
+2. 好的回答必须具有逻辑连贯性，围绕一个中心进行回答
+3. 好的回答必须具有创造性的词语和表达丰富度
+
+[用户问题]
+{question}
+""" + base_prompt
+
+
+subjective_all_sets = ["knowledge", "language", "math", "reason", "qa", "creationv2_zh"]
+prompt_all_sets = [knowledge_prompt, language_prompt, math_prompt, reason_prompt, qa_prompt, creation_prompt]
+
+for _name,_prompt in zip(subjective_all_sets, prompt_all_sets):
+    subjective_infer_cfg = dict(
+            prompt_template=dict(
+                type=PromptTemplate,
+                template=dict(round=[
+                    dict(
+                        role='HUMAN',
+                        prompt="{question}"
+                    ),
+                ]),
+            ),
+            retriever=dict(type=ZeroRetriever),
+            inferencer=dict(type=GenInferencer, max_seq_len=4096, max_out_len=2048),
+        )
+
+    subjective_eval_cfg = dict(
+        evaluator=dict(
+            type=LMEvaluator,
+            infer_order='double',
+            prompt_template=dict(
+                type=PromptTemplate,
+                template=dict(round=[
+                    dict(
+                        role='HUMAN',
+                        prompt = _prompt
+                    ),
+                ]),
+            ),
+        ),
+        pred_role="BOT",
+    )
+
+    subjective_datasets.append(
+        dict(
+            abbr=f"{_name}",
+            type=CompassArenaDataset,
+            path=data_path,
+            name=_name,
+            reader_cfg=subjective_reader_cfg,
+            infer_cfg=subjective_infer_cfg,
+            eval_cfg=subjective_eval_cfg
+        ))

configs/eval_subjective_compassarena.py

@@ -0,0 +1,95 @@
+from os import getenv as gv
+from opencompass.models import HuggingFaceCausalLM
+from mmengine.config import read_base
+with read_base():
+    from .models.chatglm.hf_chatglm3_6b_32k import models as chatglm3_6b_32k_model
+    from .models.yi.hf_yi_6b_chat import models as yi_6b_chat_model
+    from .datasets.subjective.compassarena.compassarena_compare import subjective_datasets
+
+from opencompass.models import HuggingFaceCausalLM, HuggingFace, HuggingFaceChatGLM3, OpenAI
+from opencompass.models.openai_api import OpenAIAllesAPIN
+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 CompassArenaSummarizer
+
+infer = dict(
+    #partitioner=dict(type=NaivePartitioner),
+    partitioner=dict(type=SizePartitioner, max_task_size=10000),
+    runner=dict(
+        type=SlurmSequentialRunner,
+        partition='llm_dev2',
+        quotatype='auto',
+        max_num_workers=256,
+        task=dict(type=OpenICLInferTask)),
+)
+
+api_meta_template = dict(
+    round=[
+        dict(role='HUMAN', api_role='HUMAN'),
+        dict(role='BOT', api_role='BOT', generate=True),
+    ]
+)
+
+gpt4 = dict(
+        abbr='gpt4-turbo',
+        type=OpenAI, path='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=1,
+        max_out_len=2048,
+        max_seq_len=4096,
+        batch_size=4,
+        retry=20,
+        temperature = 1
+)
+models = [*chatglm3_6b_32k_model, *yi_6b_chat_model]
+datasets = [*subjective_datasets]
+
+
+
+work_dir = 'outputs/compass_arena/'
+
+# -------------Inferen Stage ----------------------------------------
+
+judge_model = dict(
+        abbr='GPT4-Turbo',
+        type=OpenAI, path='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=1,
+        max_out_len=1024,
+        max_seq_len=4096,
+        batch_size=2,
+        retry=20,
+        temperature = 0
+)
+## ------------- Evaluation Configuration
+eval = dict(
+    partitioner=dict(
+        type=SubjectiveSizePartitioner,
+        strategy='split',
+        max_task_size=10000,
+        mode='m2n',
+        base_models = [gpt4],
+        compare_models = [*chatglm3_6b_32k_model, *yi_6b_chat_model, ]
+    ),
+    runner=dict(
+        type=SlurmSequentialRunner,
+        partition='llm_dev2',
+        quotatype='auto',
+        max_num_workers=32,
+        task=dict(
+            type=SubjectiveEvalTask,
+            judge_cfg=judge_model
+        )),
+)
+
+
+summarizer = dict(
+    type=CompassArenaSummarizer
+)

opencompass/datasets/subjective/__init__.py

@@ -1,4 +1,5 @@
 from .alignbench import AlignmentBenchDataset  # noqa: F401, F403
+from .compass_arena import CompassArenaDataset  # noqa: F401, F403
 from .corev2 import Corev2Dataset  # noqa: F401, F403
 from .creationbench import CreationBenchDataset  # noqa: F401, F403
 from .information_retrival import IRDataset  # noqa: F401, F403

opencompass/datasets/subjective/compass_arena.py

@@ -0,0 +1,28 @@
+from datasets import Dataset
+
+from opencompass.registry import LOAD_DATASET
+
+from .subjective_cmp import SubjectiveCmpDataset
+
+
+@LOAD_DATASET.register_module()
+class CompassArenaDataset(SubjectiveCmpDataset):
+
+    def load(
+        self,
+        path: str,
+        name: str,
+    ):
+        dataset = list(super().load(path, name))
+        creation_dataset = []
+        for data in dataset:
+            if 'reference' in data['others']:
+                if data['others']['reference'] is not None:
+                    data['ref'] = data['others']['reference']
+                else:
+                    data['ref'] = '满足用户需求，言之有理即可'
+            else:
+                data['ref'] = '满足用户需求，言之有理即可'
+            creation_dataset.append(data)
+        dataset = Dataset.from_list(creation_dataset)
+        return dataset

opencompass/datasets/subjective/subjective_cmp.py

@@ -26,7 +26,8 @@ class SubjectiveCmpDataset(BaseDataset):
                     'capability': capability,
                     'others': others,
                     'judge': {
-                        'capability': capability
+                        'capability': capability,
+                        'question': question
                     }
                 })
         dataset = Dataset.from_list(raw_data)

opencompass/summarizers/subjective/__init__.py

@@ -1,5 +1,6 @@
 # flake8: noqa: F401, E501
 from .alignmentbench import AlignmentBenchSummarizer
+from .compass_arena import CompassArenaSummarizer
 from .corev2 import Corev2Summarizer
 from .creationbench import CreationBenchSummarizer
 from .information_retrival import IRSummarizer

opencompass/summarizers/subjective/compass_arena.py

@@ -0,0 +1,174 @@
+# flake8: noqa: E501
+import ast
+import csv
+import os
+import os.path as osp
+import re
+from collections import defaultdict
+from datetime import datetime
+from itertools import product
+
+import mmengine
+from mmengine import ConfigDict
+from prettytable import from_csv
+
+from opencompass.partitioners.sub_naive import remove_duplicate_pairs
+from opencompass.utils import dataset_abbr_from_cfg, model_abbr_from_cfg
+
+from .utils import get_judgeanswer_and_reference, get_outdir
+
+
+def post_process_compass_arena(s):
+    if result := re.findall('(?:选择：|Choice: )([ABC])', s):
+        return result[0]
+    else:
+        return None
+
+
+def check_position_bias(judged_answers, references, banned_choice=['C']):
+    """Check position bias for judgellm's judgement.
+
+    Args:
+        judged_answers: The successfully extracted judgement.
+        references: The references contains original question, which is used to located the same question for different position judgement.
+    """
+    position_bias_flag = 0
+    position_bias_dict = {}
+    for judge, ref in zip(judged_answers, references):
+        question = ref['others']['question']
+        question_hash = hash(question)
+        if question_hash not in position_bias_dict:
+            position_bias_dict[question_hash] = {
+                'question': question,
+                'judge': judge
+            }
+        else:
+            first_judge = position_bias_dict[question_hash]['judge']
+            if judge == first_judge and first_judge not in banned_choice and judge not in banned_choice:
+                # If second choice is same with first choice, there has position bias.
+                position_bias_flag += 1
+    return position_bias_flag
+
+
+class CompassArenaSummarizer:
+    """Do the subjectivity analyze based on evaluation results.
+
+    Args:
+        config (ConfigDict): The configuration object of the evaluation task.
+            It's expected to be filled out at runtime.
+    """
+
+    def __init__(self, config: ConfigDict, judge_type='general') -> None:
+        self.tasks = []
+        self.cfg = config
+        self.base_models = self.cfg['eval']['partitioner']['base_models']
+        self.compare_models = self.cfg['eval']['partitioner']['compare_models']
+        self.judge_abbr = model_abbr_from_cfg(self.cfg['judge_model'])
+        self.judge_type = judge_type
+        assert self.judge_type in ['general']
+        self.judge_map = {
+            'general': post_process_compass_arena,
+        }
+        self.judge_function = self.judge_map[self.judge_type]
+
+    def summarize(self,
+                  time_str: str = datetime.now().strftime('%Y%m%d_%H%M%S'),
+                  check_pos_bias=True):
+        """Summarize the subjectivity analysis based on evaluation results.
+
+        Args:
+            time_str (str): Timestamp for file naming.
+
+        Returns:
+            pd.DataFrame: The summary results.
+        """
+        dataset_cfgs = self.cfg['datasets']
+        output_dir, results_folder = get_outdir(self.cfg, time_str)
+        model_combinations = list(
+            product(self.base_models, self.compare_models))
+        unique_combinations = remove_duplicate_pairs(
+            [combo for combo in model_combinations if combo[0] != combo[1]])
+        fout_list = []
+        for model_pair in unique_combinations:
+            model1, model2, judge_model = model_pair[0]['abbr'], model_pair[1][
+                'abbr'], self.judge_abbr
+            subdir = model1 + '_' + model2 + '_judged-by--' + self.judge_abbr
+            subdir_path = os.path.join(results_folder, subdir)
+            if os.path.isdir(subdir_path):
+                for dataset in dataset_cfgs:
+                    dataset_abbr = dataset_abbr_from_cfg(dataset)
+                    fout = osp.join(
+                        output_dir, 'judged-by--' + judge_model + '-' +
+                        dataset_abbr + '-report.csv')
+                    fout_list.append(fout)
+                    judged_answers, references = get_judgeanswer_and_reference(
+                        dataset,
+                        subdir_path,
+                        self.judge_function,
+                    )
+                    if check_pos_bias:
+                        bias_num = check_position_bias(judged_answers,
+                                                       references)
+                    else:
+                        bias_num = 0
+                    win_model1, win_model2, categories = defaultdict(
+                        float), defaultdict(float), defaultdict(float)
+                    model1, model2 = references[0]['answer1'], references[0][
+                        'answer2']
+                    for prediction, reference in zip(judged_answers,
+                                                     references):
+                        if dataset_abbr == 'zhihu_hot_0113':
+                            reference['capability'] = 'QA'
+                        categories['total'] += 1
+                        categories[reference['capability']] += 1
+                        if prediction == 'A':
+                            if reference['answer1'] == model1:
+                                win_model1[reference['capability']] += 1
+                                win_model1['total'] += 1
+                            else:
+                                win_model2[reference['capability']] += 1
+                                win_model2['total'] += 1
+                        elif prediction == 'B':
+                            if reference['answer1'] == model1:
+                                win_model2[reference['capability']] += 1
+                                win_model2['total'] += 1
+                            else:
+                                win_model1[reference['capability']] += 1
+                                win_model1['total'] += 1
+                    for capability in categories:
+                        if capability not in win_model1:
+                            win_model1[capability] = 0.0
+                        else:
+                            win_model1[capability] = round(
+                                (win_model1[capability] /
+                                 categories[capability]) * 100, 2)
+                        if capability not in win_model2:
+                            win_model2[capability] = 0.0
+                        else:
+                            win_model2[capability] = round(
+                                (win_model2[capability] /
+                                 categories[capability]) * 100, 2)
+                    win_model1['position_bias'] = bias_num
+                    win_model2['position_bias'] = bias_num
+                    scores = {
+                        'win_' + model1: win_model1,
+                        'win_' + model2: win_model2
+                    }
+                    rows = list(scores.keys())
+                    columns = list(scores[rows[0]].keys())
+                    columns.insert(0, columns.pop(columns.index('total')))
+                    columns.insert(1,
+                                   columns.pop(columns.index('position_bias')))
+                    with open(fout, 'a+', newline='') as csvfile:
+                        writer = csv.writer(csvfile)
+                        writer.writerow([model1 + '_vs_' + model2] + columns)
+                        for row in rows:
+                            writer.writerow(
+                                [row] +
+                                [scores[row][column] for column in columns])
+            else:
+                print(subdir_path + ' is not exist! please check!')
+        for fout in fout_list:
+            with open(fout, 'r') as f:
+                x = from_csv(f)
+            print(x)