[Dataset] Add R-Bench (ICML 2025)

dataset-index.yml

@@ -805,6 +805,12 @@
     paper: https://arxiv.org/pdf/1704.04683
     configpath: opencompass/configs/datasets/race/race_gen.py
     configpath_llmjudge: ''
+- rbench:
+    name: R-Bench
+    category: Reasoning
+    paper: https://arxiv.org/pdf/2505.02018
+    configpath: opencompass/configs/datasets/R-Bench/rbench_gen_37cbaf8.py
+    configpath_llmjudge: ''
 - realtoxicprompts:
     name: RealToxicPrompts
     category: Safety

opencompass/configs/datasets/R_Bench/R-Bench.md

@@ -0,0 +1,72 @@
+# R-Bench
+
+## Introduction
+
+The following introduction comes from the description on the [R-Bench official website](https://evalmodels.github.io/rbench/):
+
+```
+R-Bench is a graduate-level multi-disciplinary benchmark for evaluating the complex reasoning capabilities of Large Language Models (LLMs) and Multimodal Large Language Models (MLLMs). R stands for Reasoning.
+```
+
+According to statistics on R-Bench, the benchmark spans 19 departments, including mathematics, physics, biology, computer science, and chemistry, covering over 100 subjects such as Inorganic Chemistry, Chemical Reaction Kinetics, and Electromagnetism. It features 1,094 questions designed for testing language models and 665 questions specifically tailored for evaluating multimodal reasoning capabilities, available in both English and Chinese.
+
+These questions are meticulously curated to ensure rigorous difficulty calibration, subject balance, and cross-linguistic alignment, enabling the assessment to be an Olympiad-level multi-disciplinary benchmark.
+
+## Official Links
+
+### Paper
+
+[R-Bench: Graduate-level Multi-disciplinary Benchmarks for LLM & MLLM Complex Reasoning Evaluation](https://arxiv.org/abs/2505.02018)
+
+## Evaluation Results
+
+### Language Model Results
+
+```
+Model                     Source                                              Date       Average     RBench-T    RBench-T (zh)
+------------------------  --------------------------------------------------  ----------  ----------  ----------  ---------------
+OpenAI o1 🥇              https://openai.com/o1/                              2024-12-17  69.6        69.0        70.1
+Gemini2.0-Flash-Thinking 🥈 https://deepmind.google/technologies/gemini/flash-thinking/ 2025-01-21 68.0 68.4      67.5
+Doubao1.5Pro 🥉           https://www.volcengine.com/product/doubao           2025-01-21  62.7        62.0        63.4
+GPT-4o                    https://openai.com/index/hello-gpt-4o/              2024-11-20  52.6        53.6        51.6
+Claude3.5-sonnet          https://www.anthropic.com/news/claude-3-5-sonnet    2024-06-20  57.4        57.5        57.3
+Qwen2.5-72B               https://github.com/QwenLM/Qwen2.5                   2024-09-19  52.9        53.7        52.0
+Qwen2.5-32B               https://github.com/QwenLM/Qwen2.5                   2024-09-19  50.4        50.8        49.9
+Qwen2.5-7B                https://github.com/QwenLM/Qwen2.5                   2024-09-19  44.1        43.6        44.5
+```
+
+### Multimodal Model Results
+
+```
+Model                     Source                                              Date       Average     RBench-M    RBench-M (zh)
+------------------------  --------------------------------------------------  ----------  ----------  ----------  ---------------
+OpenAI o1 🥇              https://openai.com/o1/                              2024-12-17  53.1        53.2        53.0
+Doubao1.5Pro 🥈           https://www.volcengine.com/product/doubao           2025-01-21  40.2        37.9        42.4
+Claude-3-5-sonnet 🥉      https://www.anthropic.com/news/claude-3-5-sonnet    2025-04-10  39.0        39.7        38.3
+GPT-4o                    https://openai.com/index/hello-gpt-4o/              2024-11-20  33.3        33.4        33.2
+Qwen2.5-72B               https://github.com/QwenLM/Qwen2.5                   2024-09-19  25.4        25.1        25.7
+Qwen2.5-7B                https://github.com/QwenLM/Qwen2.5                   2024-09-19  21.0        19.6        22.3
+```
+
+Note:
+- RBench-T: Text-only questions for language models test
+- RBench-M: Multimodal questions for multimodal models test
+- The values in the table represent the Top-1 accuracy, in %
+- zh indicates the Chinese version
+
+## Reference
+
+```
+@inproceedings{
+  guo2025rbench,
+  title={RBench: Graduate-level Multi-disciplinary Benchmarks for
+    LLM & MLLM Complex Reasoning Evaluation},
+  author={Meng-Hao Guo, Jiajun Xu, Yi Zhang, Jiaxi Song, Haoyang Peng, Yi-Xuan Deng, 
+    Xinzhi Dong, Kiyohiro Nakayama, Zhengyang Geng, Chen Wang, Bolin Ni, Guo-Wei Yang, 
+    Yongming Rao, Houwen Peng, Han Hu, Gordon Wetzstein, Shi-min Hu},
+  year={2025},
+  eprint={2505.02018},
+  archivePrefix={arXiv},
+  primaryClass={cs.CV},
+  url={https://arxiv.org/abs/2505.02018}, 
+}

opencompass/configs/datasets/R_Bench/rbench_gen_544610.py

@@ -0,0 +1,76 @@
+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 AccwithDetailsEvaluator
+from opencompass.datasets import RBenchDataset
+from opencompass.utils.text_postprocessors import first_option_postprocess
+
+RBench_reader_cfg = dict(input_columns=[
+    'RBench_Question_Input', 'RBench_Option_A', 'RBench_Option_B',
+    'RBench_Option_C', 'RBench_Option_D', 'RBench_Option_E', 'RBench_Option_F'
+],
+                         output_column='target')
+
+RBench_datasets = []
+
+systemp_prompt_en = "Answer the following single choice question. The last line of your response should be of the following format: 'ANSWER: $LETTER' (without quotes) where LETTER is one of Options(e.g. one of ABCDEF). Think step by step before answering."
+
+systemp_prompt_zh = '回答以下单选题。答案的最后一行应采用以下格式：“答案是$LETTER”（不带引号），其中 LETTER 是选项之一（例如 ABCDEF 之一）。回答前请逐步思考。'
+
+RBench_infer_en_cfg = dict(
+    prompt_template=dict(
+        type=PromptTemplate,
+        template=dict(round=[
+            dict(
+                role='HUMAN',
+                prompt=
+                f'{systemp_prompt_en}\nQuestion: {{RBench_Question_Input}}\nA. {{RBench_Option_A}}\nB. {{RBench_Option_B}}\nC. {{RBench_Option_C}}\nD. {{RBench_Option_D}}\nE. {{RBench_Option_E}}\nF. {{RBench_Option_F}}\nAnswer: '
+            ),
+        ], ),
+    ),
+    retriever=dict(type=ZeroRetriever),
+    inferencer=dict(type=GenInferencer),
+)
+
+RBench_infer_zh_cfg = dict(
+    prompt_template=dict(
+        type=PromptTemplate,
+        template=dict(round=[
+            dict(
+                role='HUMAN',
+                prompt=
+                f'{systemp_prompt_zh}\n问题: {{RBench_Question_Input}}\nA. {{RBench_Option_A}}\nB. {{RBench_Option_B}}\nC. {{RBench_Option_C}}\nD. {{RBench_Option_D}}\nE. {{RBench_Option_E}}\nF. {{RBench_Option_F}}\n答案: '
+            ),
+        ], ),
+    ),
+    retriever=dict(type=ZeroRetriever),
+    inferencer=dict(type=GenInferencer),
+)
+
+RBench_eval_cfg = dict(evaluator=dict(type=AccwithDetailsEvaluator),
+                       pred_postprocessor=dict(type=first_option_postprocess,
+                                               options='ABCDEF'))
+
+RBench_datasets.append(
+    dict(
+        abbr=f'R-Bench_en',
+        type=RBenchDataset,
+        path='R-Bench/R-Bench',
+        name='R-Bench',
+        subset='en',
+        reader_cfg=RBench_reader_cfg,
+        infer_cfg=RBench_infer_en_cfg,
+        eval_cfg=RBench_eval_cfg,
+    ))
+
+RBench_datasets.append(
+    dict(
+        abbr=f'R-Bench_zh',
+        type=RBenchDataset,
+        path='R-Bench/R-Bench',
+        name='R-Bench',
+        subset='zh',
+        reader_cfg=RBench_reader_cfg,
+        infer_cfg=RBench_infer_zh_cfg,
+        eval_cfg=RBench_eval_cfg,
+    ))

opencompass/configs/datasets/R_Bench/rbench_llmjudge_gen_c89350.py

@@ -0,0 +1,107 @@
+from opencompass.openicl.icl_prompt_template import PromptTemplate
+from opencompass.openicl.icl_retriever import ZeroRetriever
+from opencompass.openicl.icl_inferencer import GenInferencer
+from opencompass.evaluator import GenericLLMEvaluator
+from opencompass.datasets import generic_llmjudge_postprocess
+from opencompass.datasets import RBenchDataset
+from opencompass.utils.text_postprocessors import first_option_postprocess
+
+RBench_reader_cfg = dict(
+    input_columns=[
+        'RBench_Question_Input',
+        'RBench_Option_A',
+        'RBench_Option_B',
+        'RBench_Option_C',
+        'RBench_Option_D',
+        'RBench_Option_E',
+        'RBench_Option_F',
+    ],
+    output_column='target',
+)
+
+RBench_datasets = []
+
+systemp_prompt = "Answer the following single choice question. The last line of your response should be of the following format: 'ANSWER: $LETTER' (without quotes) where LETTER is one of Options(e.g. one of ABCDEF). Think step by step before answering."
+
+GRADER_TEMPLATE = """
+    Please as a grading expert, judge whether the final answers given by the candidates below are consistent with the standard answers, that is, whether the candidates answered correctly. 
+
+    Here are some evaluation criteria:
+    1. Please refer to the given standard answer. You don't need to re-generate the answer to the question because the standard answer has been given. You only need to judge whether the candidate's answer is consistent with the standard answer according to the form of the question. Don't try to answer the original question. You can assume that the standard answer is definitely correct.
+    2. Because the candidate's answer may be different from the standard answer in the form of expression, before making a judgment, please understand the question and the standard answer first, and then judge whether the candidate's answer is correct, but be careful not to try to answer the original question.
+    3. Some answers may contain multiple items, such as multiple-choice questions, multiple-select questions, fill-in-the-blank questions, etc. As long as the answer is the same as the standard answer, it is enough. For multiple-select questions and multiple-blank fill-in-the-blank questions, the candidate needs to answer all the corresponding options or blanks correctly to be considered correct.
+    4. Some answers may be expressed in different ways, such as some answers may be a mathematical expression, some answers may be a textual description, as long as the meaning expressed is the same. And some formulas are expressed in different ways, but they are equivalent and correct.
+
+    Please judge whether the following answers are consistent with the standard answer based on the above criteria. Grade the predicted answer of this new question as one of:
+    A: CORRECT 
+    B: INCORRECT
+    Just return the letters "A" or "B", with no text around it.
+
+    Here is your task. Simply reply with either CORRECT, INCORRECT. Don't apologize or correct yourself if there was a mistake; we are just trying to grade the answer.
+
+    <Original Question Begin>: {systemp_prompt}\nQuestion: {{RBench_Question_Input}}\nA. {{RBench_Option_A}}\nB. {{RBench_Option_B}}\nC. {{RBench_Option_C}}\nD. {{RBench_Option_D}}\nE. {{RBench_Option_E}}\nF. {{RBench_Option_F}}\n<Original Question End>\n\n
+    <Gold Target Begin>: \n{target}\n<Gold Target End>\n\n
+    <Predicted Answer Begin>: \n{prediction}\n<Predicted End>\n\n
+    Judging the correctness of candidates' answers:
+""".strip()
+
+RBench_infer_cfg = dict(
+    prompt_template=dict(
+        type=PromptTemplate,
+        template=dict(
+            round=[
+                dict(
+                    role='HUMAN',
+                    prompt=f'{systemp_prompt}\nQuestion: {{RBench_Question_Input}}\nA. {{RBench_Option_A}}\nB. {{RBench_Option_B}}\nC. {{RBench_Option_C}}\nD. {{RBench_Option_D}}\nE. {{RBench_Option_E}}\nF. {{RBench_Option_F}}\nAnswer: ',
+                ),
+            ],
+        ),
+    ),
+    retriever=dict(type=ZeroRetriever),
+    inferencer=dict(type=GenInferencer),
+)
+RBench_datasets = []
+
+for subset in ['en', 'zh']:
+    RBench_eval_cfg = dict(
+        evaluator=dict(
+            type=GenericLLMEvaluator,
+            prompt_template=dict(
+                type=PromptTemplate,
+                template=dict(
+                    begin=[
+                        dict(
+                            role='SYSTEM',
+                            fallback_role='HUMAN',
+                            prompt="You are a helpful assistant who evaluates the correctness and quality of models' outputs.",
+                        )
+                    ],
+                    round=[
+                        dict(role='HUMAN', prompt=GRADER_TEMPLATE),
+                    ],
+                ),
+            ),
+            dataset_cfg=dict(
+                type=RBenchDataset,
+                path='R-Bench/R-Bench',
+                name='R-Bench',
+                subset=subset,
+                reader_cfg=RBench_reader_cfg,
+            ),
+            judge_cfg=dict(),
+            dict_postprocessor=dict(type=generic_llmjudge_postprocess),
+        ),
+    )
+
+    RBench_datasets.append(
+        dict(
+            abbr=f'R-Bench_{subset}',
+            type=RBenchDataset,
+            path='R-Bench/R-Bench',
+            name='R-Bench',
+            subset=subset,
+            reader_cfg=RBench_reader_cfg,
+            infer_cfg=RBench_infer_cfg,
+            eval_cfg=RBench_eval_cfg,
+        )
+    )

opencompass/datasets/__init__.py

@@ -127,6 +127,7 @@
 from .qaspercut import *  # noqa: F401, F403
 from .QuALITY import *  # noqa: F401, F403
 from .race import *  # noqa: F401, F403
+from .rbench import *  # noqa: F401, F403
 from .realtoxicprompts import *  # noqa: F401, F403
 from .reasonbench import ReasonBenchDataset  # noqa: F401, F403
 from .record import *  # noqa: F401, F403

opencompass/datasets/rbench.py

@@ -0,0 +1,37 @@
+from datasets import Dataset, load_dataset
+
+from opencompass.registry import LOAD_DATASET
+
+from .base import BaseDataset
+
+
+@LOAD_DATASET.register_module()
+class RBenchDataset(BaseDataset):
+
+    @staticmethod
+    def load_single(path, subset='en'):
+        raw_data = []
+        ds = load_dataset(path, f'rbench-t_{subset}')
+
+        for data in ds['test']:
+            raw_data.append({
+                'RBench_Question_Input': data['question'],
+                'RBench_Option_A': data['A'],
+                'RBench_Option_B': data['B'],
+                'RBench_Option_C': data['C'],
+                'RBench_Option_D': data['D'],
+                'RBench_Option_E': data['E'],
+                'RBench_Option_F': data['F'],
+                'target': data['answer'],
+            })
+        return Dataset.from_list(raw_data)
+
+    @staticmethod
+    def load(path, subset='en', **kwargs):
+        test_dataset = RBenchDataset.load_single(path=path, subset=subset)
+        return test_dataset
+
+
+if __name__ == '__main__':
+    dataset = RBenchDataset.load()
+    print(dataset)