add_mathbench

configs/datasets/MathBench/mathbench_gen.py

@@ -0,0 +1,4 @@
+from mmengine.config import read_base
+
+with read_base():
+    from .mathbench_gen_86de1c import mathbench_datasets  # noqa: F401, F403

configs/datasets/MathBench/mathbench_gen_86de1c.py

@@ -0,0 +1,115 @@
+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 CircularEvaluator, AccEvaluator
+from opencompass.datasets import MathBenchDataset, mathbench_postprocess
+from opencompass.utils.text_postprocessors import first_capital_postprocess
+
+
+
+single_choice_prompts = {
+    "single_choice_cn_with_reasoning": "以下是一道关于数学的单项选择题，请你一步一步推理并得到最终的答案选项。回答格式为如下：\n答案选项：A、B、C、D中你认为正确的一个选项\n计算过程：根据题目得到选项答案的一步步过程\n请严格按照上面的格式回答问题，下面是你要回答的题目：\n{question}\n答案选项：",
+    "single_choice_cn": "以下是一道关于数学的单项选择题，请你直接给出正确的答案选项。回答格式为如下：\n答案选项：A、B、C、D中你认为正确的选项\n下面是你要回答的题目：\n{question}\n答案选项：",
+    "single_choice_en_with_reasoning": "Here is a multiple-choice question about mathematics. Please provide the final answer option by step-by-step reasoning. Please answer in the following format:\nAnswer option: A, B, C, or D (the option you believe is correct)\nCalculation process: Step-by-step process to derive the answer option based on the question\nPlease strictly follow the above format to answer the question. Here is the question you need to answer:\n{question}\nAnswer option:",
+    "single_choice_en": "Here is a multiple-choice question about mathematics. Please provide the correct answer option directly. Please answer in the following format:\nAnswer option: A, B, C, or D (the option you believe is correct)\nHere is the question you need to answer:\n{question}\nAnswer option:",
+}
+
+cloze_prompts ={
+    "cloze_cn": [
+                dict(role='HUMAN', prompt='Q: 林中有15棵树。林务工人员今天将在林中种植树木。完成后，将有21棵树。林务工人员今天种植了多少棵树？'),
+                dict(role='BOT', prompt='A: 我们从15棵树开始。后来有21棵树。差值必定是他们种植的树木数量。所以，他们必须种植了21 - 15 = 6棵树。答案是 6\n'),
+                dict(role='HUMAN', prompt='Q: 如果停车场有3辆车，又有2辆车进来，停车场里有多少辆车？'),
+                dict(role='BOT', prompt='A: 停车场已经有3辆车。又进来了2辆车。现在有3 + 2 = 5辆车。答案是 5\n'),
+                dict(role='HUMAN', prompt='Q: Leah有32块巧克力，她的妹妹有42块。如果他们吃了35块，他们总共剩下多少块？'),
+                dict(role='BOT', prompt='A: Leah有32块巧克力，Leah的妹妹有42块。这意味着原本有32 + 42 = 74块巧克力。被吃掉了35块。所以他们总共还剩下74 - 35 = 39块巧克力。答案是 39\n'),
+                dict(role='HUMAN', prompt='Q: Jason有20个棒棒糖。他给Denny一些棒棒糖。现在Jason只剩下12个棒棒糖。Jason给Denny多少个棒棒糖？'),
+                dict(role='BOT', prompt='A: Jason有20个棒棒糖。因为他现在只剩下12个，所以他必须把剩下的都给了Denny。他给Denny的棒棒糖数量必定是20 - 12 = 8个。答案是 8\n'),
+                dict(role='HUMAN', prompt='Q: Shawn有五个玩具。在圣诞节，他从他的爸爸和妈妈那里各得到了两个玩具。现在他有多少个玩具？'),
+                dict(role='BOT', prompt='A: 他有5个玩具。他从妈妈那里得到了2个，所以之后他有5 + 2 = 7个玩具。然后他从爸爸那里得到了2个，所以总共他有7 + 2 = 9个玩具。答案是 9\n'),
+                dict(role='HUMAN', prompt='Q: 服务器房间里有九台电脑。从周一到周四每天增加五台电脑。现在服务器房里有多少台电脑？'),
+                dict(role='BOT', prompt='A: 从周一到周四有4天。每天增加5台电脑。这意味着总共增加了4 * 5 = 20台电脑。一开始有9台电脑，所以现在有9 + 20 = 29台电脑。答案是 29\n'),
+                dict(role='HUMAN', prompt='Q: Michael有58个高尔夫球。星期二，他丢失了23个高尔夫球。星期三，他又丢失了2个。星期三结束时他还剩下多少个高尔夫球？'),
+                dict(role='BOT', prompt='A: Michael一开始有58个球。星期二他丢失了23个，所以之后他还剩下58 - 23 = 35个球。星期三他又丢失了2个，所以现在他还剩下35 - 2 = 33个球。答案是 33\n'),
+                dict(role='HUMAN', prompt='Q: Olivia有23美元。她用每个3美元的价格买了五个百吉饼。她还剩下多少钱？'),
+                dict(role='BOT', prompt='A: 她以每个3美元的价格买了5个百吉饼。这意味着她在百吉饼上花费了5 * 3 = 15美元。她一开始有23美元，所以现在她还剩下23 - 15 = 8美元。答案是 8\n'),
+                dict(role='HUMAN', prompt='Q: {question}'),
+                dict(role='BOT', prompt='A: {answer}'),
+                ],
+    "cloze_en": [
+                dict(role='HUMAN', prompt='Q: There are 15 trees in the grove. Grove workers will plant trees in the grove today. After they are done, there will be 21 trees. How many trees did the grove workers plant today?'),
+                dict(role='BOT', prompt='A: We start with 15 trees. Later we have 21 trees. The difference must be the number of trees they planted. So, they must have planted 21 - 15 = 6 trees. The answer is 6.\n'),
+                dict(role='HUMAN', prompt='Q: If there are 3 cars in the parking lot and 2 more cars arrive, how many cars are in the parking lot?'),
+                dict(role='BOT', prompt='A: There are 3 cars in the parking lot already. 2 more arrive. Now there are 3 + 2 = 5 cars. The answer is 5.\n'),
+                dict(role='HUMAN', prompt='Q: Leah had 32 chocolates and her sister had 42. If they ate 35, how many pieces do they have left in total?'),
+                dict(role='BOT', prompt="A: Leah had 32 chocolates and Leah's sister had 42. That means there were originally 32 + 42 = 74 chocolates. 35 have been eaten. So in total they still have 74 - 35 = 39 chocolates. The answer is 39.\n"),
+                dict(role='HUMAN', prompt='Q: Jason had 20 lollipops. He gave Denny some lollipops. Now Jason has 12 lollipops. How many lollipops did Jason give to Denny?'),
+                dict(role='BOT', prompt='A: Jason had 20 lollipops. Since he only has 12 now, he must have given the rest to Denny. The number of lollipops he has given to Denny must have been 20 - 12 = 8 lollipops. The answer is 8.\n'),
+                dict(role='HUMAN', prompt='Q: Shawn has five toys. For Christmas, he got two toys each from his mom and dad. How many toys does he have now?'),
+                dict(role='BOT', prompt='A: He has 5 toys. He got 2 from mom, so after that he has 5 + 2 = 7 toys. Then he got 2 more from dad, so in total he has 7 + 2 = 9 toys. The answer is 9.\n'),
+                dict(role='HUMAN', prompt='Q: There were nine computers in the server room. Five more computers were installed each day, from monday to thursday. How many computers are now in the server room?'),
+                dict(role='BOT', prompt='A: There are 4 days from monday to thursday. 5 computers were added each day. That means in total 4 * 5 = 20 computers were added. There were 9 computers in the beginning, so now there are 9 + 20 = 29 computers. The answer is 29.\n'),
+                dict(role='HUMAN', prompt='Q: Michael had 58 golf balls. On tuesday, he lost 23 golf balls. On wednesday, he lost 2 more. How many golf balls did he have at the end of wednesday?'),
+                dict(role='BOT', prompt='A: Michael initially had 58 balls. He lost 23 on Tuesday, so after that he has 58 - 23 = 35 balls. On Wednesday he lost 2 more so now he has 35 - 2 = 33 balls. The answer is 33.\n'),
+                dict(role='HUMAN', prompt='Q: Olivia has $23. She bought five bagels for $3 each. How much money does she have left?'),
+                dict(role='BOT', prompt='A: She bought 5 bagels for $3 each. This means she spent 5 * $3 = $15 on the bagels. She had $23 in beginning, so now she has $23 - $15 = $8. The answer is 8.\n'),
+                dict(role='HUMAN', prompt='Q: {question}'),
+                dict(role='BOT', prompt='A: {answer}\n'),
+],
+}
+
+
+mathbench_sets = {
+    'college': ['single_choice_cn', 'cloze_en'],
+    'high': ['single_choice_cn', 'single_choice_en'],
+    'middle': ['single_choice_cn'],
+    'primary': ['single_choice_cn', 'cloze_cn'],
+}
+
+# Generate reasoning path if set True or just generate the final answer
+with_reasoning = True
+
+# Use circular evaluation or not
+with_circular_eval = True
+
+
+mathbench_datasets = []
+
+for _split in list(mathbench_sets.keys()):
+    for _name in mathbench_sets[_split]:
+        mathbench_infer_cfg = dict(
+            ice_template=dict(
+                type=PromptTemplate,
+                template=dict(
+                    begin="</E>",
+                    round=[
+                        dict(
+                            role="HUMAN",
+                            prompt= single_choice_prompts[_name + "_with_reasoning"] if with_reasoning else mathbench_sets[_name],
+                        ),
+                        dict(role="BOT", prompt="{answer}"),] if 'choice' in _name else cloze_prompts[_name],
+                    ),
+                ice_token="</E>",
+            ),
+            retriever=dict(type=ZeroRetriever),
+            inferencer=dict(type=GenInferencer),
+        )
+
+        mathbench_eval_cfg = dict(
+            evaluator=dict(type=CircularEvaluator if 'choice' in _name  else AccEvaluator),
+            pred_postprocessor=dict(type=first_capital_postprocess ) if 'single_choice' in _name else dict(type=mathbench_postprocess, name=_name),)
+
+        mathbench_datasets.append(
+            dict(
+                type=MathBenchDataset,
+                path=f"./data/mathbench/version_1/{_split}",
+                name=_name,
+                abbr="mathbench-" + _split + '-' + _name, 
+                reader_cfg=dict(
+                    input_columns=["question"],
+                    output_column="answer"
+                    ),
+                infer_cfg=mathbench_infer_cfg,
+                eval_cfg=mathbench_eval_cfg,
+            ))
+
+del _split, _name

opencompass/datasets/__init__.py

@@ -40,6 +40,7 @@
 from .leval import *  # noqa: F401, F403
 from .longbench import *  # noqa: F401, F403
 from .math import *  # noqa: F401, F403
+from .mathbench import *  # noqa: F401, F403
 from .mbpp import *  # noqa: F401, F403
 from .mmlu import *  # noqa: F401, F403
 from .multirc import *  # noqa: F401, F403

opencompass/datasets/mathbench.py

@@ -0,0 +1,78 @@
+import copy
+import json
+import os.path as osp
+import re
+
+from datasets import Dataset
+
+from opencompass.registry import LOAD_DATASET, TEXT_POSTPROCESSORS
+
+from .base import BaseDataset
+
+
+def get_number(options):
+
+    result_string = ''
+    for i, option in enumerate(options, start=65):
+        result_string += f'{chr(i)}. {option}\n'
+    return result_string
+
+
+@LOAD_DATASET.register_module()
+class MathBenchDataset(BaseDataset):
+
+    @staticmethod
+    def load(path: str, name: str):
+
+        circular_patterns = ['ABCD', 'BCDA', 'CDAB', 'DABC']
+
+        data = []
+        filename = osp.join(path, f'{name}.jsonl')
+        with open(filename, 'r') as infile:
+            for id, line in enumerate(infile):
+                entry = json.loads(line)
+                if 'cloze' in name:
+                    data.append({
+                        'question': entry['question'].strip(),
+                        'answer': entry['answer'].strip()
+                    })
+                else:
+                    for c in circular_patterns:
+                        line = copy.deepcopy(entry)
+                        options = []
+                        for i in range(4):
+                            options.append(line['options'][ord(c[i]) -
+                                                           ord('A')])
+                        line['options'] = options
+                        line['answer'] = {
+                            c[0]: 'A',
+                            c[1]: 'B',
+                            c[2]: 'C',
+                            c[3]: 'D'
+                        }[line['answer']]
+                        line['answer'] = str(
+                            id) + '--' + line['answer'] + '--' + c
+                        line['question'] = line['question'].strip(
+                        ) + '\n' + get_number(line['options'])
+                        data.append(line)
+
+        dataset = Dataset.from_list(data)
+        return dataset
+
+
+@TEXT_POSTPROCESSORS.register_module()
+def mathbench_postprocess(text: str, name: str) -> str:
+    ans = text
+    if '_cn' in name:
+        ans_line = ans.split('答案是')
+    else:
+        ans_line = ans.split('The answer is')
+    if len(ans_line) != 1:
+        ans = ans_line[1].strip()
+
+    output = re.sub(r'(\d),(\d)', r'\1\2', ans)
+    numbers = re.findall(r'-?\d*\.?\d+|\d+', output)
+    if numbers:
+        return numbers[-1]
+
+    return ans

opencompass/openicl/icl_evaluator/icl_circular_evaluator.py

@@ -0,0 +1,104 @@
+import collections
+
+from opencompass.registry import ICL_EVALUATORS
+
+from .icl_base_evaluator import BaseEvaluator
+
+
+@ICL_EVALUATORS.register_module()
+class CircularEvaluator(BaseEvaluator):
+    """Robust circular evaluator for multi-choice questions."""
+
+    def __init__(self) -> None:
+        super().__init__()
+        self.cp4 = ['ABCD', 'BCDA', 'CDAB', 'DABC']
+        self.cp1 = ['ABCD']
+
+    def score(self, predictions, references):
+        """Calculate the accuracy of predictions.
+
+        Args:
+            predictions (list): List of predictions.
+            references (list): List of references.
+
+        Returns:
+            dict: A dict of evaluation results.
+        """
+
+        self._metrics = {}
+        self._metrics.update({'acc_4': 0, 'acc_1': 0})
+        # Accuracy for patterns with no circular shift / 4 circular shifts
+        for pred, reference in zip(predictions, references):
+            index, ref, circular_pattern = reference.split('--')
+            if circular_pattern in self.cp4:
+                self._metrics['acc_4'] += 1 if pred == ref else 0
+            if circular_pattern in self.cp1:
+                self._metrics['acc_1'] += 1 if pred == ref else 0
+        for k in ['acc_4', 'acc_1']:
+            self._metrics[k] = self._metrics[k] / len(predictions) * 4 / int(
+                k.split('_')[-1]) * 100
+
+        # Accuracy for patterns with no circular shift / 4 circular shifts
+        details = {4: {}, 1: {}}
+        for pred, reference in zip(predictions, references):
+            index, ref, circular_pattern = reference.split('--')
+            if index not in details[4]:
+                details[4][index] = []
+                details[1][index] = []
+            if circular_pattern in self.cp4:
+                details[4][index].append(True if pred == ref else False)
+            if circular_pattern in self.cp1:
+                details[1][index].append(True if pred == ref else False)
+        # Calculate accuracy for having at least j correct out of i total
+        for i in [1, 4]:
+            for j in range(0, i + 1):
+                count, total = 0, 0
+                for index in details[i]:
+                    if sum(details[i][index]) >= j:
+                        count += 1
+                    total += 1
+                self._metrics[f'more_{i}_{j}'] = count / total * 100
+        # Consider fully correct as correct
+        for i in [1, 4]:
+            self._metrics[f'perf_{i}'] = self._metrics[f'more_{i}_{i}']
+
+        # Calculate voting accuracy
+        voting = {'vote_4': {}, 'vote_1': {}}
+        refs = {}
+        for pred, reference in zip(predictions, references):
+            index, ref, circular_pattern = reference.split('--')
+            c = circular_pattern
+            back_map = {'A': c[0], 'B': c[1], 'C': c[2], 'D': c[3]}
+            ref = back_map[ref]
+            if pred not in ['A', 'B', 'C', 'D']:
+                pred = '-'
+            else:
+                pred = back_map[pred]
+            if index not in voting['vote_4']:
+                voting['vote_4'][index] = collections.Counter()
+                voting['vote_1'][index] = collections.Counter()
+                refs[index] = ref
+
+            if c in self.cp4:
+                voting['vote_4'][index][pred] += 1
+            if c in self.cp1:
+                voting['vote_1'][index][pred] += 1
+        for k in ['vote_4', 'vote_1']:
+            voting_count = 0
+            for index in voting[k]:
+                if refs[index] == voting[k][index].most_common(1)[0][0]:
+                    voting_count += 1
+            self._metrics[k] = voting_count / len(voting[k]) * 100
+
+        # Calculate the frequency of ABCD in model predictions
+        prior_counts = {'A': 0, 'B': 0, 'C': 0, 'D': 0, '-': 0}
+        for pred, reference in zip(predictions, references):
+            if pred in ['A', 'B', 'C', 'D']:
+                prior_counts[pred] += 1
+            else:
+                prior_counts['-'] += 1
+        for k in ['A', 'B', 'C', 'D', '-']:
+            self._metrics[f'prior_{k}'] = prior_counts[k] / len(
+                predictions) * 100
+
+        return self._metrics