evalscope 0.11.0__py3-none-any.whl → 0.12.0__py3-none-any.whl

evalscope/third_party/thinkbench/eval.py

@@ -0,0 +1,264 @@
+import json
+import os
+import pandas as pd
+import plotly.graph_objects as go
+import re
+from collections import defaultdict
+from functools import lru_cache
+from modelscope import AutoTokenizer
+from plotly.subplots import make_subplots
+from tqdm.contrib.concurrent import thread_map
+from typing import List
+
+from evalscope.third_party.thinkbench.tools.llm import request_url
+from evalscope.third_party.thinkbench.tools.utils import extract_answer
+from evalscope.utils.io_utils import dump_jsonl_data
+
+cur_path = os.path.dirname(os.path.abspath(__file__))
+
+class EvalThink:
+    def __init__(self, report_path, tokenizer_path, model_name, dataset_name, subsets, split_strategies='llm', judge_config=None):
+        self.report_path = report_path
+        self.reformat_template = open(os.path.join(cur_path, 'resources/reformat_template.txt'), 'r').read()
+        self.critique_template = open(os.path.join(cur_path, 'resources/critique_template.txt'), 'r').read()
+        self.switch_tokens = ['alternatively', 'but wait', 'let me reconsider', 'another way', 'another approach', 'another method', 'another angle']
+        self.subset_dict = defaultdict(lambda: defaultdict(list))
+        self.think_end_token = '</think>'
+        self.tokenizer = AutoTokenizer.from_pretrained(tokenizer_path)
+        self.model_name = model_name
+        self.dataset_name = dataset_name
+        self.subsets = subsets
+        self.metrics = ['token_efficiency', 'completion_len', 'thought_num', 'accuracy']
+        self.split_strategies = split_strategies  # split by llm, keywords, separator
+        self.judge_config = judge_config
+
+    @lru_cache(maxsize=None)
+    def get_think_part(self, text):
+        last_think_end = text.rfind(self.think_end_token)
+        return text[:last_think_end].lower()
+
+    @lru_cache(maxsize=None)
+    def cal_tokens(self, text: str):
+        return len(self.tokenizer.encode(text, add_special_tokens=False))
+
+    def process_choice(self, choice, problem):
+        think_part = self.get_think_part(choice['message']['content'])
+        answer = choice['review']['gold']
+        tokens = self.cal_tokens(think_part)
+        switch_count = sum(think_part.count(token) for token in self.switch_tokens)
+        useful_tokens = self.cal_tokens(self.get_first_correct(think_part, problem, answer))
+        score = choice['review']['result']
+        return tokens, switch_count, useful_tokens, score
+
+    def process_item(self, item):
+        problem = item['raw_input'].get('question') or item['raw_input'].get('problem') or ''
+        results = []
+        for choice in item['choices']:
+            results.append(self.process_choice(choice, problem))
+            break  # only process the first choice
+
+        tokens, switch_counts, useful_tokens, scores = zip(*results)
+
+        avg_tokens = sum(tokens) / len(tokens)
+        avg_thought_num = sum(switch_counts) / len(switch_counts)
+        avg_token_efficiency = sum(useful_tokens) / sum(tokens)
+        avg_accuracy = sum(scores) / len(scores)
+
+        return avg_tokens, avg_thought_num, avg_token_efficiency, avg_accuracy
+
+    def split_by_llm(self, response, problem) -> List[str]:
+        response = response.replace('\n', ' ') # remove newline characters
+        prompt = self.reformat_template.format(problem=problem, response=response)
+        llm_response = request_url(self.judge_config, prompt)
+        return llm_response.split('\n\n')
+
+    def split_by_keywords(self, text) -> List[str]:
+        pattern = r'(?=\b(?:{})\b)'.format('|'.join(map(re.escape, self.switch_tokens)))
+        segments = re.split(pattern, text)
+        # remove empty segments
+        segments = [segment.strip() for segment in segments if segment.strip()]
+
+        return segments if segments else [text]
+
+    def split_by_separator(self, text) -> List[str]:
+        return text.split('\n\n')
+
+    def get_answer_index(self, response: List[str], problem: str, answer: str) -> int:
+        tagged_response = ''
+        for sdx, step in enumerate(response):
+            tagged_response += f'<paragraph_{sdx}>\n{step}\n</paragraph_{sdx}>\n\n'
+        tagged_response = tagged_response.strip()
+
+        prompt = self.critique_template.format(problem=problem, answer=answer, tagged_response=tagged_response)
+        llm_response = request_url(self.judge_config, prompt)
+        answer_index = extract_answer(llm_response)
+
+        dump_jsonl_data({'prompt': prompt, 'response': llm_response, 'answer_index': answer_index},
+                        os.path.join(self.report_path, 'answer_index.jsonl'),
+                        dump_mode='append')
+        try:
+            answer_index = int(answer_index)
+        except Exception:
+            answer_index = -1
+        return answer_index
+
+    def get_first_correct(self, response: str, problem: str, answer: str) -> str:
+        if self.split_strategies == 'llm':
+            text_list = self.split_by_llm(response, problem)
+        elif self.split_strategies == 'keywords':
+            text_list = self.split_by_keywords(response)
+        else:
+            text_list = self.split_by_separator(response)
+
+        answer_index = self.get_answer_index(text_list, problem, answer)
+
+        if answer_index == -1:  # no correct answer found
+            first_correct = ''
+        else:
+            first_correct = '\n\n'.join(text_list[: answer_index])
+        return first_correct
+
+    def plot_metrics(self, results, output_dir):
+        fig = make_subplots(rows=1, cols=len(self.metrics),
+                            subplot_titles=('Token Efficiency', 'Completion Length', 'Thought Num', 'Accuracy'),
+                            shared_xaxes=True, x_title='Subsets')
+
+
+        for i, metric in enumerate(self.metrics, start=1):
+            y_values = [results[metric][subset] for subset in self.subsets]
+            fig.add_trace(
+                go.Scatter(x=list(range(len(self.subsets))), y=y_values,
+                           mode='lines+markers',
+                           name=metric.replace('_', ' ').title()),
+                row=1, col=i
+            )
+            # Add annotations for each data point
+            for j, y in enumerate(y_values):
+                fig.add_annotation(
+                    x=j,
+                    y=y,
+                    text=f'{y:.2f}',
+                    showarrow=False,
+                    yshift=10,
+                    row=1,
+                    col=i
+                )
+
+        fig.update_layout(
+            height=500,
+            width=1500,
+            title_text=f'Evaluation Metrics for {self.model_name} on {self.dataset_name}',
+            legend=dict(orientation='h', yanchor='bottom', y=1.02, xanchor='right', x=1)
+        )
+
+        for i in range(1, len(self.metrics) + 1):
+            fig.update_xaxes(
+                ticktext=self.subsets,
+                tickvals=list(range(len(self.subsets))),
+                row=1, col=i
+            )
+            fig.update_yaxes(title_text=self.metrics[i-1].replace('_', ' ').title(), row=1, col=i)
+        # Update y-axis ranges
+        fig.update_yaxes(range=[0, 1], row=1, col=1)  # Token Efficiency
+        fig.update_yaxes(range=[0, 13], row=1, col=3)  # Switch Frequency
+        fig.update_yaxes(range=[0, 1], row=1, col=4)  # Accuracy
+
+        os.makedirs(output_dir, exist_ok=True)
+        output_path = os.path.join(output_dir, f'{self.model_name}_{self.dataset_name}_metrics.png')
+        fig.write_image(output_path)
+        print(f'save figure to: {output_path}')
+
+
+
+    def filter_df(self, df, response_len: int = 8000, count: int=10):
+        def is_valid_row(row):
+            return all(self.cal_tokens(choice['message']['content']) <= response_len for choice in row['choices'])
+
+        bools = df.apply(is_valid_row, axis=1)
+
+        return df[bools].head(count)
+
+
+    def evaluate(self, output_dir, max_tokens=8000, count=50):
+        for subset in self.subsets:
+            review_path = os.path.join(self.report_path, 'reviews', self.model_name, f'{self.dataset_name}_{subset}.jsonl')
+            review_df = pd.read_json(review_path, lines=True)
+
+            review_df = self.filter_df(review_df, response_len=max_tokens, count=count)
+
+            results = thread_map(
+                self.process_item,
+                (item for _, item in review_df.iterrows()),
+                desc=f'Evaluating {subset}',
+                total=len(review_df),
+                max_workers=16
+            )
+
+            avg_tokens, avg_thought_num, avg_token_efficiency, avg_accuracy = zip(*results)
+
+            self.subset_dict[subset]['completion_len'] = sum(avg_tokens) / len(avg_tokens)
+            self.subset_dict[subset]['thought_num'] = sum(avg_thought_num) / len(avg_thought_num)
+            self.subset_dict[subset]['token_efficiency'] = sum(avg_token_efficiency) / len(avg_token_efficiency)
+            self.subset_dict[subset]['accuracy'] = sum(avg_accuracy) / len(avg_accuracy)
+
+
+        results = {metric: {subset: self.subset_dict[subset][metric] for subset in self.subsets}
+                   for metric in self.metrics}
+
+        self.plot_metrics(results, output_dir)
+
+        return results
+
+def run_task(config, output_dir='outputs', max_tokens=8000, count=50):
+    evaluator = EvalThink(**config,)
+    results = evaluator.evaluate(output_dir, max_tokens, count)
+    print(results)
+
+judge_config = dict(
+    api_key='EMPTY',
+    base_url='http://0.0.0.0:8801/v1',
+    model_name='Qwen2.5-72B-Instruct',
+)
+
+distill_qwen_config = dict(
+    report_path = '/mnt/data/data/user/maoyunlin.myl/eval-scope/outputs/20250218_180219',
+    model_name = 'DeepSeek-R1-Distill-Qwen-7B',
+    tokenizer_path = 'deepseek-ai/DeepSeek-R1-Distill-Qwen-7B',
+    dataset_name = 'math_500',
+    subsets = ['Level 1', 'Level 2', 'Level 3', 'Level 4', 'Level 5'],
+    split_strategies='separator',
+    judge_config=judge_config
+)
+
+math_qwen_config = dict(
+    report_path = '/mnt/data/data/user/maoyunlin.myl/eval-scope/outputs/20250219_202358',
+    model_name = 'Qwen2.5-Math-7B-Instruct',
+    tokenizer_path = 'Qwen/Qwen2.5-Math-7B-Instruct',
+    dataset_name = 'math_500',
+    subsets = ['Level 1', 'Level 2', 'Level 3', 'Level 4', 'Level 5'],
+    split_strategies='separator'
+)
+
+r1_config = dict(
+    report_path = '/mnt/data/data/user/maoyunlin.myl/eval-scope/outputs/20250221_104202',
+    model_name = 'deepseek-r1',
+    tokenizer_path = 'deepseek-ai/DeepSeek-R1',
+    dataset_name = 'math_500',
+    subsets = ['Level 1', 'Level 2', 'Level 3', 'Level 4', 'Level 5'],
+    split_strategies='separator'
+)
+
+qwq_config = dict(
+    report_path = '/mnt/data/data/user/maoyunlin.myl/eval-scope/outputs/20250221_105911',
+    model_name = 'qwq-32b-preview',
+    tokenizer_path = 'Qwen/QwQ-32B-Preview',
+    dataset_name = 'math_500',
+    subsets = ['Level 1', 'Level 2', 'Level 3', 'Level 4', 'Level 5'],
+    split_strategies='separator'
+)
+
+if __name__ == '__main__':
+    run_task(distill_qwen_config)
+    # run_task(math_qwen_config)
+    # run_task(r1_config)
+    # run_task(qwq_config)

evalscope/third_party/thinkbench/infer.py

@@ -0,0 +1,100 @@
+import os
+
+from evalscope import TaskConfig, run_task
+
+
+def eval_distill_qwen():
+    model_name = 'DeepSeek-R1-Distill-Qwen-7B'
+    dataset_name = 'math_500'
+    subsets = ['Level 1', 'Level 2', 'Level 3', 'Level 4', 'Level 5']
+
+    task_config = TaskConfig(
+        api_url='http://0.0.0.0:8801/v1/chat/completions',
+        model=model_name,
+        eval_type='service',
+        datasets=[dataset_name],
+        dataset_args={dataset_name: {'few_shot_num': 0, 'subset_list': subsets}},
+        eval_batch_size=32,
+        generation_config={
+            'max_tokens': 20000,  # avoid exceed max length
+            'temperature': 0.6,
+            'top_p': 0.95,
+            'n': 1,
+        },
+    )
+    run_task(task_config)
+
+
+def eval_math_qwen():
+    model_name = 'Qwen2.5-Math-7B-Instruct'
+    dataset_name = 'math_500'
+    subsets = ['Level 1', 'Level 2', 'Level 3', 'Level 4', 'Level 5']
+
+    task_config = TaskConfig(
+        api_url='http://0.0.0.0:8801/v1/chat/completions',
+        model=model_name,
+        eval_type='service',
+        datasets=[dataset_name],
+        dataset_args={dataset_name: {'few_shot_num': 0, 'subset_list': subsets}},
+        eval_batch_size=32,
+        generation_config={
+            'max_tokens': 3000,  # avoid exceed max length
+            'temperature': 0.6,
+            'top_p': 0.95,
+            'n': 3,
+        },
+    )
+    run_task(task_config)
+
+def eval_r1():
+    model_name = 'deepseek-r1'
+    dataset_name = 'math_500'
+    subsets = ['Level 1', 'Level 2', 'Level 3', 'Level 4', 'Level 5']
+
+    task_config = TaskConfig(
+        api_url='https://dashscope.aliyuncs.com/compatible-mode/v1/chat/completions',
+        api_key=os.environ['DASHSCOPE_API_KEY'],
+        model=model_name,
+        eval_type='service',
+        datasets=[dataset_name],
+        dataset_args={dataset_name: {'few_shot_num': 0, 'subset_list': subsets}},
+        eval_batch_size=3,
+        generation_config={
+            'max_tokens': 12000,  # avoid exceed max length
+            'temperature': 0.6,
+            'top_p': 0.95,
+            'n': 1,
+        },
+        limit=50,
+        use_cache='/mnt/data/data/user/maoyunlin.myl/eval-scope/outputs/20250221_104202'
+    )
+    run_task(task_config)
+
+def eval_qwq():
+    model_name = 'qwq-32b-preview'
+    dataset_name = 'math_500'
+    subsets = ['Level 1', 'Level 2', 'Level 3', 'Level 4', 'Level 5']
+
+    task_config = TaskConfig(
+        api_url='https://dashscope.aliyuncs.com/compatible-mode/v1/chat/completions',
+        api_key=os.environ['DASHSCOPE_API_KEY'],
+        model=model_name,
+        eval_type='service',
+        datasets=[dataset_name],
+        dataset_args={dataset_name: {'few_shot_num': 0, 'subset_list': subsets}},
+        eval_batch_size=32,
+        generation_config={
+            'max_tokens': 8000,  # avoid exceed max length
+            'temperature': 0.6,
+            'top_p': 0.95,
+            'n': 1,
+        },
+        use_cache='/mnt/data/data/user/maoyunlin.myl/eval-scope/outputs/20250221_105911'
+    )
+    run_task(task_config)
+
+if __name__ == '__main__':
+    # eval_distill_qwen()
+    # eval_math_qwen()
+    # eval_r1()
+    eval_qwq()

evalscope/third_party/thinkbench/resources/critique_template.txt

@@ -0,0 +1,17 @@
+The following is a math problem and a solution (split into paragraphs, enclosed with tags and indexed from 0):
+
+[Math Problem]
+
+{problem}
+
+[Correct Answer]
+
+{answer}
+
+[Solution]
+
+{tagged_response}
+
+Your task is to review and critique the solution paragraph by paragraph. Once you identify an correct answer in a paragraph, return the index of the paragraph where the earliest correct answer occurs. Otherwise, return the index of -1 (which typically denotes "not found").
+
+Please put your final answer (i.e., the index) in \boxed{{}}.

evalscope/third_party/thinkbench/resources/reformat_template.txt

@@ -0,0 +1,31 @@
+I will present you with a solution to a math problem. Unfortunately, the solution lacks proper paragraphing, making it hard to read. Your task is to improve readability by reformatting the solution into well-structured paragraphs. Follow these specific guidelines:
+
+* Insert \n\n for paragraph breaks within the original solution. Do **NOT** alter any content of the original solution (the only exception is for itemized lists; see below).
+
+  - Each paragraph should represent a distinct, concise reasoning step that logically advances the solution.
+
+  - Reasoning steps can include case discussions, formula simplifications, or formula derivations. Each of these should be treated as an individual reasoning step and paragraphed accordingly.
+
+  - If an introductory analysis exists in the original solution, treat it as an initial reasoning step and place it as the first paragraph.
+
+  - Do **NOT** place any mathematical formulas in their own separate paragraphs; instead, include them within the same paragraph as the preceding text to form a cohesive reasoning step.
+
+* For any itemized lists (ordered or unordered), convert them into a written format, such as "First/Second/Third." This is the **ONLY** content modification allowed.
+
+* Avoid making paragraphs too lengthy, as long paragraphs might contain multiple reasoning steps that should be paragraphed separately.
+
+* Disregard the accuracy of the solution content. Do **NOT** alter any of the original solution's content; focus solely on structuring it into logical, readable paragraphs.
+
+* Reply with the reformatted solution directly.
+
+--------------------------------------------------
+
+Here is the math problem, and the solution that needs to be reformatted:
+
+[Math Problem]
+
+{problem}
+
+[Solution]
+
+{response}

evalscope/third_party/thinkbench/tools/llm.py

@@ -0,0 +1,47 @@
+import os
+from openai import OpenAI
+
+
+def request_url(llm_config, content):
+    try:
+        client = OpenAI(
+            api_key=llm_config['api_key'],
+            base_url=llm_config['base_url'],
+        )
+        completion = client.chat.completions.create(
+            model=llm_config['model_name'],
+            messages=[{'role': 'user', 'content': content}]
+        )
+        return completion.choices[0].message.content
+    except Exception as e:
+        print(e)
+
+def request_qwen(content):
+    try:
+        client = OpenAI(
+            api_key=os.getenv('DASHSCOPE_API_KEY'),
+            base_url='https://dashscope.aliyuncs.com/compatible-mode/v1',
+        )
+
+        completion = client.chat.completions.create(
+            model='qwen-max',
+            messages=[{'role': 'user', 'content': content}]
+        )
+        return completion.choices[0].message.content
+    except Exception as e:
+        print(e)
+
+
+def request_local(content):
+    try:
+        client = OpenAI(
+            api_key='EMPTY',
+            base_url='http://0.0.0.0:8801/v1',
+        )
+        completion = client.chat.completions.create(
+            model='Qwen2.5-72B-Instruct',
+            messages=[{'role': 'user', 'content': content}]
+        )
+        return completion.choices[0].message.content
+    except Exception as e:
+        print(e)

evalscope/third_party/thinkbench/tools/utils.py

@@ -0,0 +1,13 @@
+import re
+
+
+def extract_answer(solution_text: str):
+    boxed_pattern = r'\\boxed\{([^}]*)\}'
+    matches = re.findall(boxed_pattern, solution_text)
+    if matches:
+        last_boxed_content = matches[-1]
+        number_pattern = r'-?\d+'
+        number_matches = re.findall(number_pattern, last_boxed_content)
+        if number_matches:
+            return number_matches[-1].strip()
+    return None

evalscope/utils/model_utils.py

@@ -1,5 +1,6 @@
+import os
 from enum import Enum
-from typing import TYPE_CHECKING
+from typing import TYPE_CHECKING, Optional, Tuple, Union
 
 if TYPE_CHECKING:
     from transformers import GenerationConfig
@@ -22,3 +23,18 @@ def fix_do_sample_warning(generation_config: 'GenerationConfig') -> None:
         generation_config.temperature = 1.
         generation_config.top_p = 1.
         generation_config.top_k = 50
+
+
+def get_device() -> str:
+    from transformers.utils import is_torch_cuda_available, is_torch_mps_available, is_torch_npu_available
+
+    if is_torch_npu_available():
+        device = 'npu'
+    elif is_torch_mps_available():
+        device = 'mps'
+    elif is_torch_cuda_available():
+        device = 'cuda'
+    else:
+        device = 'cpu'
+
+    return device

evalscope/utils/utils.py

@@ -101,50 +101,50 @@ class ResponseParser:
         options_concat = '|'.join([str(i) for i in options])
 
         patterns = [
-            f'答案是?\s?([{options_concat}])',
-            f'答案是?\s?：([{options_concat}])',
-            f'答案是?\s?:([{options_concat}])',
-            f'答案应该?是\s?([{options_concat}])',
-            f'答案应该?选\s?([{options_concat}])',
-            f'答案为\s?([{options_concat}])',
-            f'答案选\s?([{options_concat}])',
-            f'选择?\s?([{options_concat}])',
-            f'故选?\s?([{options_concat}])'
-            f'只有选?项?\s?([{options_concat}])\s?是?对',
-            f'只有选?项?\s?([{options_concat}])\s?是?错',
-            f'只有选?项?\s?([{options_concat}])\s?不?正确',
-            f'只有选?项?\s?([{options_concat}])\s?错误',
-            f'说法不?对选?项?的?是\s?([{options_concat}])',
-            f'说法不?正确选?项?的?是\s?([{options_concat}])',
-            f'说法错误选?项?的?是\s?([{options_concat}])',
-            f'([{options_concat}])\s?是正确的',
-            f'([{options_concat}])\s?是正确答案',
-            f'选项\s?([{options_concat}])\s?正确',
-            f'所以答\s?([{options_concat}])',
-            f'所以\s?([{options_concat}][.。$]?$)',
-            f'所有\s?([{options_concat}][.。$]?$)',
-            f'[\s，：:,]([{options_concat}])[。，,\.]?$',
-            f'[\s，,：:][故即]([{options_concat}])[。\.]?$',
-            f'[\s，,：:]因此([{options_concat}])[。\.]?$',
-            f'[是为。]\s?([{options_concat}])[。\.]?$',
-            f'因此\s?([{options_concat}])[。\.]?$',
-            f'显然\s?([{options_concat}])[。\.]?$',
-            f'答案是\s?(\S+)(?:。|$)',
-            f'答案应该是\s?(\S+)(?:。|$)',
-            f'答案为\s?(\S+)(?:。|$)',
-            f'答案是(.*?)[{options_concat}]',
-            f'答案为(.*?)[{options_concat}]',
-            f'固选(.*?)[{options_concat}]',
-            f'答案应该是(.*?)[{options_concat}]',
-            f'[Tt]he answer is \(?[{options_concat}]\)?',
-            f'[Tt]he correct answer is [{options_concat}]',
-            f'[Tt]he correct answer is:\n[{options_concat}]',
-            f'(\s|^)[{options_concat}][\s。，,\.$]',  # noqa
-            f'^选项\s?([{options_concat}])',
-            f'^([{options_concat}])\s?选?项',
-            f'(\s|^)[{options_concat}][\s。，,：:\.$]',
-            f'(\s|^)[{options_concat}](\s|$)',
-            f'[{options_concat}]',
+            rf'答案是?\s?([{options_concat}])',
+            rf'答案是?\s?：([{options_concat}])',
+            rf'答案是?\s?:([{options_concat}])',
+            rf'答案应该?是\s?([{options_concat}])',
+            rf'答案应该?选\s?([{options_concat}])',
+            rf'答案为\s?([{options_concat}])',
+            rf'答案选\s?([{options_concat}])',
+            rf'选择?\s?([{options_concat}])',
+            rf'故选?\s?([{options_concat}])'
+            rf'只有选?项?\s?([{options_concat}])\s?是?对',
+            rf'只有选?项?\s?([{options_concat}])\s?是?错',
+            rf'只有选?项?\s?([{options_concat}])\s?不?正确',
+            rf'只有选?项?\s?([{options_concat}])\s?错误',
+            rf'说法不?对选?项?的?是\s?([{options_concat}])',
+            rf'说法不?正确选?项?的?是\s?([{options_concat}])',
+            rf'说法错误选?项?的?是\s?([{options_concat}])',
+            rf'([{options_concat}])\s?是正确的',
+            rf'([{options_concat}])\s?是正确答案',
+            rf'选项\s?([{options_concat}])\s?正确',
+            rf'所以答\s?([{options_concat}])',
+            rf'所以\s?([{options_concat}][.。$]?$)',
+            rf'所有\s?([{options_concat}][.。$]?$)',
+            rf'[\s，：:,]([{options_concat}])[。，,\.]?$',
+            rf'[\s，,：:][故即]([{options_concat}])[。\.]?$',
+            rf'[\s，,：:]因此([{options_concat}])[。\.]?$',
+            rf'[是为。]\s?([{options_concat}])[。\.]?$',
+            rf'因此\s?([{options_concat}])[。\.]?$',
+            rf'显然\s?([{options_concat}])[。\.]?$',
+            rf'答案是\s?(\S+)(?:。|$)',
+            rf'答案应该是\s?(\S+)(?:。|$)',
+            rf'答案为\s?(\S+)(?:。|$)',
+            rf'答案是(.*?)[{options_concat}]',
+            rf'答案为(.*?)[{options_concat}]',
+            rf'固选(.*?)[{options_concat}]',
+            rf'答案应该是(.*?)[{options_concat}]',
+            rf'[Tt]he answer is \(?[{options_concat}]\)?',
+            rf'[Tt]he correct answer is [{options_concat}]',
+            rf'[Tt]he correct answer is:\n[{options_concat}]',
+            rf'(\s|^)[{options_concat}][\s。，,\.$]',  # noqa
+            rf'^选项\s?([{options_concat}])',
+            rf'^([{options_concat}])\s?选?项',
+            rf'(\s|^)[{options_concat}][\s。，,：:\.$]',
+            rf'(\s|^)[{options_concat}](\s|$)',
+            rf'[{options_concat}]',
         ]
 
         regexes = [re.compile(pattern) for pattern in patterns]
@@ -166,8 +166,8 @@ class ResponseParser:
             text: The text to parse.
         """
         patterns = [
-            r'[Aa]nswer:\s*(\w+)',
             r'answer is \(?(\w+)\)?',
+            r'[Aa]nswer:\s*(\w+)',
             r'[Tt]he correct answer is:\s*(\w+)',
             r'[Tt]he correct answer is:\n\s*(\w+)',
             r'[Tt]he correct answer is:\n\n-\s*(\w+)',

evalscope/version.py

@@ -1,4 +1,4 @@
 # Copyright (c) Alibaba, Inc. and its affiliates.
 
-__version__ = '0.11.0'
-__release_datetime__ = '2025-02-13 12:00:00'
+__version__ = '0.12.0'
+__release_datetime__ = '2025-02-27 21:00:00'