Merge branch 'open-compass:main' into hle_biomed

dataset-index.yml

@@ -122,12 +122,42 @@
     paper: https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=10778138
     configpath: opencompass/configs/datasets/MedBench/medbench_gen.py
     configpath_llmjudge: ''
+- MedCalc_Bench:
+    name: MedCalc_Bench
+    category: Knowledge / Medicine
+    paper: https://arxiv.org/abs/2406.12036
+    configpath: opencompass/configs/datasets/MedCalc_Bench/MedCalcBench_official_gen_a5155f.py
+    configpath_llmjudge: ''
+- MedXpertQA:
+    name: MedQA
+    category: Knowledge / Medicine
+    paper: https://arxiv.org/abs/2009.13081
+    configpath: opencompass/configs/datasets/MedQA/MedQA_gen.py
+    configpath_llmjudge: opencompass/configs/datasets/MedQA/MedQA_llmjudge_gen.py
 - MedXpertQA:
     name: MedXpertQA
     category: Knowledge / Medicine
     paper: https://arxiv.org/abs/2501.18362
     configpath: opencompass/configs/datasets/MedXpertQA/MedXpertQA_gen.py
     configpath_llmjudge: opencompass/configs/datasets/MedXpertQA/MedXpertQA_llmjudge_gen.py
+- ClinicBench:
+    name: ClinicBench
+    category: Knowledge / Medicine
+    paper: https://arxiv.org/abs/2405.00716
+    configpath: ''
+    configpath_llmjudge: opencompass/configs/datasets/ClinicBench/ClinicBench_llmjudge_gen.py
+- ScienceQA:
+    name: ScienceQA
+    category: Knowledge / Medicine
+    paper: https://arxiv.org/abs/2209.09513
+    configpath: ''
+    configpath_llmjudge: opencompass/configs/datasets/ScienceQA/ScienceQA_llmjudge_gen.py
+- PubMedQA:
+    name: PubMedQA
+    category: Knowledge / Medicine
+    paper: https://arxiv.org/abs/1909.06146
+    configpath: ''
+    configpath_llmjudge: opencompass/configs/datasets/PubMedQA/PubMedQA_llmjudge_gen.py
 - musr:
     name: MuSR
     category: Reasoning
@@ -343,6 +373,12 @@
     paper: https://arxiv.org/pdf/2004.05986
     configpath: opencompass/configs/datasets/CLUE_C3/CLUE_C3_gen.py
     configpath_llmjudge: ''
+- CARDBiomedBench:
+    name: CARDBiomedBench
+    category: Knowledge / Medicine
+    paper: https://www.biorxiv.org/content/10.1101/2025.01.15.633272v1
+    configpath: opencompass/configs/datasets/CARDBiomedBench
+    configpath_llmjudge: 'opencompass/configs/datasets/CARDBiomedBench/CARDBiomedBench_llmjudge_gen_99a231.py'
 - cb:
     name: SuperGLUE / CB
     category: Reasoning
@@ -739,6 +775,12 @@
     paper: https://arxiv.org/pdf/1911.11641v1
     configpath: opencompass/configs/datasets/piqa/piqa_gen.py
     configpath_llmjudge: ''
+- ProteinLMBench:
+    name: ProteinLMBench
+    category: Knowledge / Biology (Protein)
+    paper: https://arxiv.org/abs/2406.05540
+    configpath: opencompass/configs/datasets/ProteinLMBench/ProteinLMBench_gen.py
+    configpath_llmjudge: opencompass/configs/datasets/ProteinLMBench/ProteinLMBench_llmjudge_gen.py
 - py150:
     name: py150
     category: Code
@@ -1028,4 +1070,10 @@
     category: Reasoning/Code/Agent
     paper: ''
     configpath: opencompass/configs/datasets/internsandbox/internsandbox_gen_44b982.py
-    configpath_llmjudge: ''
+    configpath_llmjudge: ''
+- nejmaibench:
+    name: nejmaibench
+    category: Science /Medicine
+    paper: https://arxiv.org/pdf/2308.04709
+    configpath: opencompass/configs/datasets/nejm_ai_benchmark/nejmaibench_gen.py
+    configpath_llmjudge: opencompass/configs/datasets/nejm_ai_benchmark/nejmaibench_llmjudge_gen.py

opencompass/configs/datasets/CARDBiomedBench/CARDBiomedBench_llmjudge_gen_99a231.py

@@ -0,0 +1,101 @@
+from opencompass.datasets import CARDBiomedBenchDataset
+from opencompass.datasets import generic_llmjudge_postprocess
+from opencompass.openicl.icl_inferencer import GenInferencer
+from opencompass.openicl.icl_prompt_template import PromptTemplate
+from opencompass.openicl.icl_retriever import ZeroRetriever
+from opencompass.evaluator import GenericLLMEvaluator
+ZERO_SHOT_PROMPT = 'You are an expert in {expert}.\n{question}\n'
+
+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>: Q: You are an expert in {expert}.\n{question}\n<Original Question End>\n\n
+    <Gold Target Begin>: \n{answer}\n<Gold Target End>\n\n
+    <Predicted Answer Begin>: \n{prediction}\n<Predicted End>\n\n
+    Judging the correctness of candidates' answers:
+""".strip()
+
+
+# Reader configuration
+reader_cfg = dict(
+    input_columns=[
+        'question',
+        'answer',
+        'Bio_Category',
+        'SQL_Category',
+        'uuid',
+        'template uuid',
+        'expert',
+    ],
+    output_column='answer',
+)
+# Inference configuration
+infer_cfg = dict(
+    prompt_template=dict(
+        type=PromptTemplate,
+        template=dict(
+            round=[
+                dict(
+                    
+                    role='HUMAN',
+                    prompt=ZERO_SHOT_PROMPT, # prompt mode: zero-shot
+                ),
+            ],
+        ),
+    ),
+    retriever=dict(type=ZeroRetriever),
+    inferencer=dict(type=GenInferencer),
+)
+
+# Evaluation configuration
+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=CARDBiomedBenchDataset,
+            path='NIH-CARD/CARDBiomedBench',
+            prompt_mode='zero-shot',
+            reader_cfg=reader_cfg,
+        ),
+        judge_cfg=dict(),
+        dict_postprocessor=dict(type=generic_llmjudge_postprocess),
+    ),
+)
+cardbiomedbench_dataset = dict(
+    type=CARDBiomedBenchDataset,
+    abbr='cardbiomedbench',
+    path='NIH-CARD/CARDBiomedBench',
+    prompt_mode='zero-shot',
+    reader_cfg=reader_cfg,
+    infer_cfg=infer_cfg,
+    eval_cfg=eval_cfg,
+)
+cardbiomedbench_datasets = [cardbiomedbench_dataset]

opencompass/configs/datasets/ClinicBench/ClinicBench_llmjudge_gen.py

@@ -0,0 +1,4 @@
+from mmengine.config import read_base
+
+with read_base():
+    from .ClinicBench_llmjudge_gen_d09668 import ClinicBench_datasets

opencompass/configs/datasets/ClinicBench/ClinicBench_llmjudge_gen_d09668.py

@@ -0,0 +1,100 @@
+from mmengine.config import read_base
+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.ClinicBench import ClinicBenchDataset
+
+
+QUERY_TEMPLATE = """
+Answer the following multiple 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 ABCDEFGHIJKLMNOP). Think step by step before answering.
+
+Question:\n
+{question}
+
+Options:\n
+{choices}
+
+""".strip()
+
+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>: {question}\n {choices} \n<Original Question End>\n\n
+    <Gold Target Begin>: \n{label}\n<Gold Target End>\n\n
+    <Predicted Answer Begin>: \n{prediction}\n<Predicted End>\n\n
+    Judging the correctness of candidates' answers:
+""".strip()
+
+ClinicBench_datasets = []
+
+ClinicBench_reader_cfg = dict(
+    input_columns=['question', 'choices'],
+    output_column='label',
+)
+
+ClinicBench_infer_cfg = dict(
+    prompt_template=dict(
+        type=PromptTemplate,
+        template=dict(
+            round=[
+                dict(role='HUMAN', prompt=QUERY_TEMPLATE),
+            ],
+        ),
+    ),
+    retriever=dict(type=ZeroRetriever),
+    inferencer=dict(type=GenInferencer),
+)
+
+ClinicBench_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=ClinicBenchDataset,
+            path='xuxuxuxuxu/Pharmacology-QA',
+            reader_cfg=ClinicBench_reader_cfg,
+        ),
+        judge_cfg=dict(),
+        dict_postprocessor=dict(type=generic_llmjudge_postprocess),
+    ),
+)
+
+ClinicBench_datasets.append(
+    dict(
+        abbr=f'ClinicBench',
+        type=ClinicBenchDataset,
+        path='xuxuxuxuxu/Pharmacology-QA',
+        reader_cfg=ClinicBench_reader_cfg,
+        infer_cfg=ClinicBench_infer_cfg,
+        eval_cfg=ClinicBench_eval_cfg,
+    )
+)

opencompass/configs/datasets/MedCalc_Bench/MedCalcBench_official_gen_a5155f.py

@@ -0,0 +1,57 @@
+from opencompass.datasets import MedCalc_BenchDataset, MedCalcOfficial_Evaluator
+from opencompass.openicl.icl_inferencer import GenInferencer
+from opencompass.openicl.icl_prompt_template import PromptTemplate
+from opencompass.openicl.icl_retriever import ZeroRetriever
+
+ZERO_SHOT_PROMPT = 'You are a helpful assistant for calculating a score for a given patient note. Please think step-by-step to solve the question and then generate the required score. Your output should only contain a JSON dict formatted as {"step_by_step_thinking": str(your_step_by_step_thinking_procress_to_solve_the_question), "answer": str(short_and_direct_answer_of_the_question)}. \n Here is the patient note:\n{patient_note}\n\nHere is the task:\n{question}\n\nPlease directly output the JSON dict formatted as {"step_by_step_thinking": str(your_step_by_step_thinking_procress_to_solve_the_question), "answer": str(short_and_direct_answer_of_the_question)}:'
+# Reader configuration
+reader_cfg = dict(
+        input_columns=[
+        'row_number',
+        'calculator_id',
+        'calculator_name',
+        'category',
+        'note_id',
+        'output_type',
+        'note_type',
+        'patient_note',
+        'question',
+        'relevant_entities',
+        'ground_truth_answer',
+        'lower_limit',
+        'upper_limit',
+        'ground_truth_explanation'
+    ],
+    output_column='ground_truth_answer',
+)
+
+
+# Inference configuration
+infer_cfg = dict(
+    prompt_template=dict(
+        type=PromptTemplate,
+        template=dict(
+            round=[
+            dict(role='HUMAN',prompt=ZERO_SHOT_PROMPT),
+        ])
+    ),
+    retriever=dict(type=ZeroRetriever),
+    inferencer=dict(type=GenInferencer),
+)
+
+# Evaluation configuration
+eval_cfg = dict(
+    evaluator=dict(type=MedCalcOfficial_Evaluator),
+    pred_role='BOT',
+)
+medcal_bench_dataset = dict(
+    type=MedCalc_BenchDataset,
+    abbr='medcal_bench_official_zero_shot_eval',
+    path='ncbi/MedCalc-Bench-v1.0',
+    prompt_mode='zero-shot',
+    reader_cfg=reader_cfg,
+    infer_cfg=infer_cfg,
+    eval_cfg=eval_cfg,
+)
+
+medcal_bench_datasets = [medcal_bench_dataset]

opencompass/configs/datasets/MedQA/MedQA_gen_3bf756.py

@@ -0,0 +1,63 @@
+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 AccEvaluator
+from opencompass.utils.text_postprocessors import first_option_postprocess
+from opencompass.datasets.MedQA import MedQADataset
+
+
+QUERY_TEMPLATE = """
+Answer the following multiple 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 ABCDEFGHIJKLMNOP). Think step by step before answering.
+
+Question:\n
+{question}
+
+Options:\n
+{choices}
+
+""".strip()
+
+
+MedQA_datasets = []
+
+MedQA_reader_cfg = dict(
+    input_columns=['question', 'choices'],
+    output_column='label',
+)
+
+MedQA_infer_cfg = dict(
+    prompt_template=dict(
+        type=PromptTemplate,
+        template=dict(
+            round=[
+                dict(role='HUMAN', prompt=QUERY_TEMPLATE),
+            ],
+        ),
+    ),
+    retriever=dict(type=ZeroRetriever),
+    inferencer=dict(type=GenInferencer),
+)
+
+MedQA_subsets = {
+    'US': 'xuxuxuxuxu/MedQA_US_test',
+    'Mainland': 'xuxuxuxuxu/MedQA_Mainland_test',
+    'Taiwan': 'xuxuxuxuxu/MedQA_Taiwan_test',
+}
+
+for split in list(MedQA_subsets.keys()):
+
+    MedQA_eval_cfg = dict(
+        evaluator=dict(type=AccEvaluator),
+        pred_postprocessor=dict(type=first_option_postprocess, options='ABCD')
+    )
+
+    MedQA_datasets.append(
+        dict(
+            abbr=f'MedQA_{split}',
+            type=MedQADataset,
+            path=MedQA_subsets[split],
+            reader_cfg=MedQA_reader_cfg,
+            infer_cfg=MedQA_infer_cfg,
+            eval_cfg=MedQA_eval_cfg,
+        )
+    )

opencompass/configs/datasets/MedQA/MedQA_llmjudge_gen_3bf756.py

@@ -0,0 +1,108 @@
+from mmengine.config import read_base
+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.MedQA import MedQADataset
+
+
+QUERY_TEMPLATE = """
+Answer the following multiple 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 ABCDEFGHIJKLMNOP). Think step by step before answering.
+
+Question:\n
+{question}
+
+Options:\n
+{choices}
+
+""".strip()
+
+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>: {question}\n {choices} \n<Original Question End>\n\n
+    <Gold Target Begin>: \n{label}\n<Gold Target End>\n\n
+    <Predicted Answer Begin>: \n{prediction}\n<Predicted End>\n\n
+    Judging the correctness of candidates' answers:
+""".strip()
+
+MedQA_datasets = []
+
+MedQA_reader_cfg = dict(
+    input_columns=['question', 'choices'],
+    output_column='label',
+)
+
+MedQA_infer_cfg = dict(
+    prompt_template=dict(
+        type=PromptTemplate,
+        template=dict(
+            round=[
+                dict(role='HUMAN', prompt=QUERY_TEMPLATE),
+            ],
+        ),
+    ),
+    retriever=dict(type=ZeroRetriever),
+    inferencer=dict(type=GenInferencer),
+)
+
+MedQA_subsets = {
+    'US': 'xuxuxuxuxu/MedQA_US_test',
+    'Mainland': 'xuxuxuxuxu/MedQA_Mainland_test',
+    'Taiwan': 'xuxuxuxuxu/MedQA_Taiwan_test',
+}
+
+for split in list(MedQA_subsets.keys()):
+
+    MedQA_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=MedQADataset,
+                path=MedQA_subsets[split],
+                reader_cfg=MedQA_reader_cfg,
+            ),
+            judge_cfg=dict(),
+            dict_postprocessor=dict(type=generic_llmjudge_postprocess),
+        ),
+    )
+
+    MedQA_datasets.append(
+        dict(
+            abbr=f'MedQA_{split}',
+            type=MedQADataset,
+            path=MedQA_subsets[split],
+            reader_cfg=MedQA_reader_cfg,
+            infer_cfg=MedQA_infer_cfg,
+            eval_cfg=MedQA_eval_cfg,
+        )
+    )

opencompass/configs/datasets/ProteinLMBench/ProteinLMBench_gen_a67965.py

@@ -0,0 +1,46 @@
+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 AccEvaluator
+from opencompass.datasets.ProteinLMBench import ProteinLMBenchDataset, ProteinLMBenchEvaluator
+
+QUERY_TEMPLATE = "Answer the following multiple choice question. There is only one correct answer. The last line of your response should be in the format 'Answer: $LETTER' (without quotes), where LETTER is the letter among {start} through {end}.\n{question}"
+
+
+# Reader configuration
+reader_cfg = dict(
+    input_columns=['question', 'start', 'end', 'options'],
+    output_column='label',
+)
+
+# Inference configuration
+infer_cfg = dict(
+    prompt_template=dict(
+        type=PromptTemplate,
+        template=dict(
+            round=[
+                dict(
+                    role='HUMAN',
+                    prompt=QUERY_TEMPLATE
+                )
+            ], ),
+    ),
+    retriever=dict(type=ZeroRetriever),
+    inferencer=dict(type=GenInferencer),
+)
+
+# Evaluation configuration
+eval_cfg = dict(
+    evaluator=dict(type=ProteinLMBenchEvaluator),
+)
+
+proteinlmbench_dataset = dict(
+    abbr='ProteinLMBench',
+    type=ProteinLMBenchDataset,
+    path='tsynbio/ProteinLMBench',
+    reader_cfg=reader_cfg,
+    infer_cfg=infer_cfg,
+    eval_cfg=eval_cfg
+)
+
+proteinlmbench_datasets = [proteinlmbench_dataset]

opencompass/configs/datasets/ProteinLMBench/ProteinLMBench_llmjudge_gen_a67965.py

@@ -0,0 +1,89 @@
+from mmengine.config import read_base
+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.ProteinLMBench import ProteinLMBenchDataset
+
+QUERY_TEMPLATE = "Answer the following multiple choice question. There is only one correct answer. The last line of your response should be in the format 'Answer: $LETTER' (without quotes), where LETTER is the letter among {start} through {end}.\n{question}"
+
+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>: {question}\n<Original Question End>\n\n
+    <Gold Target Begin>: \n{label}\n<Gold Target End>\n\n
+    <Predicted Answer Begin>: \n{prediction}\n<Predicted End>\n\n
+    Judging the correctness of candidates' answers:
+""".strip()
+
+
+reader_cfg = dict(
+    input_columns=['question', 'start', 'end', 'options'],
+    output_column='label',
+)
+
+infer_cfg = dict(
+    prompt_template=dict(
+        type=PromptTemplate,
+        template=dict(
+            round=[
+                dict(role='HUMAN', prompt=QUERY_TEMPLATE),
+            ],
+        ),
+    ),
+    retriever=dict(type=ZeroRetriever),
+    inferencer=dict(type=GenInferencer),
+)
+
+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=ProteinLMBenchDataset,
+            path='tsynbio/ProteinLMBench',
+            reader_cfg=reader_cfg,
+        ),
+        judge_cfg=dict(),
+        dict_postprocessor=dict(type=generic_llmjudge_postprocess),
+    ),
+)
+
+proteinlmbench_dataset = dict(
+    abbr='ProteinLMBench',
+    type=ProteinLMBenchDataset,
+    path='tsynbio/ProteinLMBench',
+    reader_cfg=reader_cfg,
+    infer_cfg=infer_cfg,
+    eval_cfg=eval_cfg
+)
+
+proteinlmbench_datasets = [proteinlmbench_dataset]

opencompass/configs/datasets/PubMedQA/PubMedQA_llmjudge_gen.py

@@ -0,0 +1,4 @@
+from mmengine.config import read_base
+
+with read_base():
+    from .PubMedQA_llmjudge_gen_f00302 import PubMedQA_datasets

opencompass/configs/datasets/PubMedQA/PubMedQA_llmjudge_gen_f00302.py

@@ -0,0 +1,94 @@
+from mmengine.config import read_base
+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.PubMedQA import PubMedQADataset
+
+
+QUERY_TEMPLATE = """
+Answer the following multiple 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 ABCDEFGHIJKLMNOP). Think step by step before answering.
+Question:\n
+{question}
+Options:\n
+{choices}
+""".strip()
+
+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>: {question}\n {choices} \n<Original Question End>\n\n
+    <Gold Target Begin>: \n{label}\n<Gold Target End>\n\n
+    <Predicted Answer Begin>: \n{prediction}\n<Predicted End>\n\n
+    Judging the correctness of candidates' answers:
+""".strip()
+
+PubMedQA_datasets = []
+
+PubMedQA_reader_cfg = dict(
+    input_columns=['question', 'choices'],
+    output_column='label',
+)
+
+PubMedQA_infer_cfg = dict(
+    prompt_template=dict(
+        type=PromptTemplate,
+        template=dict(
+            round=[
+                dict(role='HUMAN', prompt=QUERY_TEMPLATE),
+            ],
+        ),
+    ),
+    retriever=dict(type=ZeroRetriever),
+    inferencer=dict(type=GenInferencer),
+)
+
+PubMedQA_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=PubMedQADataset,
+            path='qiaojin/PubMedQA',
+            reader_cfg=PubMedQA_reader_cfg,
+        ),
+        judge_cfg=dict(),
+        dict_postprocessor=dict(type=generic_llmjudge_postprocess),
+    ),
+)
+
+PubMedQA_datasets.append(
+    dict(
+        abbr=f'PubMedQA',
+        type=PubMedQADataset,
+        path='qiaojin/PubMedQA',
+        reader_cfg=PubMedQA_reader_cfg,
+        infer_cfg=PubMedQA_infer_cfg,
+        eval_cfg=PubMedQA_eval_cfg,
+    )
+)

opencompass/configs/datasets/ScienceQA/ScienceQA_llmjudge_gen.py

@@ -0,0 +1,4 @@
+from mmengine.config import read_base
+
+with read_base():
+    from .ScienceQA_llmjudge_gen_f00302 import ScienceQA_datasets

opencompass/configs/datasets/ScienceQA/ScienceQA_llmjudge_gen_f00302.py

@@ -0,0 +1,94 @@
+from mmengine.config import read_base
+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.ScienceQA import ScienceQADataset
+
+
+QUERY_TEMPLATE = """
+Answer the following multiple 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 ABCDEFGHIJKLMNOP). Think step by step before answering.
+Question:\n
+{question}
+Options:\n
+{choices}
+""".strip()
+
+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>: {question}\n {choices} \n<Original Question End>\n\n
+    <Gold Target Begin>: \n{label}\n<Gold Target End>\n\n
+    <Predicted Answer Begin>: \n{prediction}\n<Predicted End>\n\n
+    Judging the correctness of candidates' answers:
+""".strip()
+
+ScienceQA_datasets = []
+
+ScienceQA_reader_cfg = dict(
+    input_columns=['question', 'choices'],
+    output_column='label',
+)
+
+ScienceQA_infer_cfg = dict(
+    prompt_template=dict(
+        type=PromptTemplate,
+        template=dict(
+            round=[
+                dict(role='HUMAN', prompt=QUERY_TEMPLATE),
+            ],
+        ),
+    ),
+    retriever=dict(type=ZeroRetriever),
+    inferencer=dict(type=GenInferencer),
+)
+
+ScienceQA_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=ScienceQADataset,
+            path='derek-thomas/ScienceQA',
+            reader_cfg=ScienceQA_reader_cfg,
+        ),
+        judge_cfg=dict(),
+        dict_postprocessor=dict(type=generic_llmjudge_postprocess),
+    ),
+)
+
+ScienceQA_datasets.append(
+    dict(
+        abbr=f'ScienceQA',
+        type=ScienceQADataset,
+        path='derek-thomas/ScienceQA',
+        reader_cfg=ScienceQA_reader_cfg,
+        infer_cfg=ScienceQA_infer_cfg,
+        eval_cfg=ScienceQA_eval_cfg,
+    )
+)

opencompass/configs/datasets/mmlu_pro/mmlu_pro_biomed_0shot_cot_gen_057927.py

@@ -0,0 +1,60 @@
+from mmengine.config import read_base
+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 AccEvaluator
+from opencompass.datasets import MMLUProDataset
+from opencompass.utils.text_postprocessors import match_answer_pattern
+
+categories = [
+    'health',
+]
+
+QUERY_TEMPLATE = """
+Answer the following multiple 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 ABCDEFGHIJKLMNOP). Think step by step before answering.
+Question:\n
+{question}
+Options:\n
+{options_str}
+""".strip()
+
+mmlu_pro_datasets = []
+
+for category in categories:
+    mmlu_pro_reader_cfg = dict(
+        input_columns=['question', 'cot_content', 'options_str'],
+        output_column='answer',
+        train_split='validation',
+        test_split='test',
+    )
+    mmlu_pro_infer_cfg = dict(
+        prompt_template=dict(
+            type=PromptTemplate,
+            template=dict(
+                round=[
+                    dict(role='HUMAN',
+                         prompt=QUERY_TEMPLATE),
+                ],
+            ),
+        ),
+        retriever=dict(type=ZeroRetriever),
+        inferencer=dict(type=GenInferencer),
+    )
+
+    mmlu_pro_eval_cfg = dict(
+        evaluator=dict(type=AccEvaluator),
+        pred_postprocessor=dict(
+            type=match_answer_pattern,
+            answer_pattern=r'(?i)ANSWER\s*:\s*([A-P])')
+    )
+
+    mmlu_pro_datasets.append(
+        dict(
+            abbr=f'mmlu_pro_{category.replace(" ", "_")}',
+            type=MMLUProDataset,
+            path='opencompass/mmlu_pro',
+            category=category,
+            reader_cfg=mmlu_pro_reader_cfg,
+            infer_cfg=mmlu_pro_infer_cfg,
+            eval_cfg=mmlu_pro_eval_cfg,
+        ))

opencompass/configs/datasets/mmlu_pro/mmlu_pro_biomed_0shot_nocot_genericllmeval_gen_057927.py

@@ -0,0 +1,101 @@
+from mmengine.config import read_base
+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 MMLUProDataset, generic_llmjudge_postprocess
+
+categories = [
+    'health',
+]
+
+
+QUERY_TEMPLATE = """
+Answer the following multiple 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 ABCDEFGHIJKLMNOP). Think step by step before answering.
+Question:\n
+{question}
+Options:\n
+{options_str}
+""".strip()
+
+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>: {question}\n {options_str} \n<Original Question End>\n\n
+    <Gold Target Begin>: \n{answer}\n<Gold Target End>\n\n
+    <Predicted Answer Begin>: \n{prediction}\n<Predicted End>\n\n
+    Judging the correctness of candidates' answers:
+""".strip()
+
+mmlu_pro_datasets = []
+
+for category in categories:
+    mmlu_pro_reader_cfg = dict(
+        input_columns=['question', 'cot_content', 'options_str'],
+        output_column='answer',
+        train_split='validation',
+        test_split='test',
+    )
+    mmlu_pro_infer_cfg = dict(
+        prompt_template=dict(
+            type=PromptTemplate,
+            template=dict(
+                round=[
+                    dict(role='HUMAN', prompt=QUERY_TEMPLATE),
+                ],
+            ),
+        ),
+        retriever=dict(type=ZeroRetriever),
+        inferencer=dict(type=GenInferencer),
+    )
+
+    mmlu_pro_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=MMLUProDataset,
+                path='opencompass/mmlu_pro',
+                category=category,
+                reader_cfg=mmlu_pro_reader_cfg,
+            ),
+            judge_cfg=dict(),
+            dict_postprocessor=dict(type=generic_llmjudge_postprocess),
+        ),
+    )
+
+    mmlu_pro_datasets.append(
+        dict(
+            abbr=f'mmlu_pro_{category.replace(" ", "_")}',
+            type=MMLUProDataset,
+            path='opencompass/mmlu_pro',
+            category=category,
+            reader_cfg=mmlu_pro_reader_cfg,
+            infer_cfg=mmlu_pro_infer_cfg,
+            eval_cfg=mmlu_pro_eval_cfg,
+        )
+    )

opencompass/configs/datasets/nejm_ai_benchmark/nejmaibench_gen.py

@@ -0,0 +1,4 @@
+from mmengine.config import read_base
+
+with read_base():
+    from .nejmaibench_gen_60c8f5 import nejmaibench_datasets  # noqa: F401, F403

opencompass/configs/datasets/nejm_ai_benchmark/nejmaibench_gen_60c8f5.py

@@ -0,0 +1,59 @@
+from opencompass.datasets import NejmaibenchDataset, NejmaibenchEvaluator
+from opencompass.openicl.icl_inferencer import GenInferencer
+from opencompass.openicl.icl_prompt_template import PromptTemplate
+from opencompass.openicl.icl_retriever import ZeroRetriever
+
+import os
+
+SYSTEM_PROMPT = 'You are a helpful medical assistant.\n\n' # Where to put this?
+ZERO_SHOT_PROMPT = 'Q: {question}\n Please select the correct answer from the options above and output only the corresponding letter (A, B, C, D, or E) without any explanation or additional text.\n'
+
+# Reader configuration
+reader_cfg = dict(
+    input_columns=[
+        'question',
+        'options',
+        'Subject',
+        'prompt_mode',
+        
+    ],
+    output_column='label',
+)
+
+# Inference configuration
+infer_cfg = dict(
+    prompt_template=dict(
+        type=PromptTemplate,
+        template=dict(
+            begin=[
+                dict(role='SYSTEM', fallback_role='HUMAN', prompt=SYSTEM_PROMPT),
+            ],
+            round=[
+                dict(
+                    role='HUMAN',
+                    prompt=ZERO_SHOT_PROMPT, # prompt mode: zero-shot
+                ),
+            ],
+        ),
+    ),
+    retriever=dict(type=ZeroRetriever),
+    inferencer=dict(type=GenInferencer),
+)
+
+# Evaluation configuration
+eval_cfg = dict(
+    evaluator=dict(type=NejmaibenchEvaluator),
+    pred_role='BOT',
+)
+nejmaibench_dataset = dict(
+    type=NejmaibenchDataset,
+    abbr='nejmaibench',
+    path='opencompass/nejmaibench',
+    prompt_mode='zero-shot',
+    reader_cfg=reader_cfg,
+    infer_cfg=infer_cfg,
+    eval_cfg=eval_cfg,
+    
+)
+
+nejmaibench_datasets = [nejmaibench_dataset]

opencompass/configs/datasets/nejm_ai_benchmark/nejmaibench_llmjudge_gen.py

@@ -0,0 +1,4 @@
+from mmengine.config import read_base
+
+with read_base():
+    from .nejmaibench_llmjudge_gen_60c8f5 import nejmaibench_datasets  # noqa: F401, F403

opencompass/configs/datasets/nejm_ai_benchmark/nejmaibench_llmjudge_gen_60c8f5.py

@@ -0,0 +1,108 @@
+from opencompass.datasets import NejmaibenchDataset
+from opencompass.openicl.icl_inferencer import GenInferencer
+from opencompass.openicl.icl_prompt_template import PromptTemplate
+from opencompass.openicl.icl_retriever import ZeroRetriever
+from opencompass.datasets import generic_llmjudge_postprocess
+from opencompass.evaluator import GenericLLMEvaluator
+import os
+
+SYSTEM_PROMPT = 'You are a helpful medical assistant.\n\n' # Where to put this?
+ZERO_SHOT_PROMPT = 'Q: {question}\n Please select the correct answer from the options above and output only the corresponding letter (A, B, C, D, or E) without any explanation or additional text.\n'
+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>: Q: {question}\nPlease select the correct answer from the options above and output only the corresponding letter (A, B, C, D, or E) without any explanation or additional text.\n\n<Original Question End>\n\n
+    <Gold Target Begin>: \n{label}\n<Gold Target End>\n\n
+    <Predicted Answer Begin>: \n{prediction}\n<Predicted End>\n\n
+    Judging the correctness of candidates' answers:
+""".strip()
+
+# Reader configuration
+reader_cfg = dict(
+    input_columns=[
+        'question',
+        'options',
+        'Subject',
+        'prompt_mode',
+        
+    ],
+    output_column='label',
+)
+
+
+# Inference configuration
+infer_cfg = dict(
+    prompt_template=dict(
+        type=PromptTemplate,
+        template=dict(
+            begin=[
+                dict(role='SYSTEM', fallback_role='HUMAN', prompt=SYSTEM_PROMPT),
+            ],
+            round=[
+                dict(
+                    role='HUMAN',
+                    prompt=ZERO_SHOT_PROMPT, # prompt mode: zero-shot
+                ),
+            ],
+        ),
+    ),
+    retriever=dict(type=ZeroRetriever),
+    inferencer=dict(type=GenInferencer),
+)
+
+# Evaluation configuration
+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=NejmaibenchDataset,
+            path='opencompass/nejmaibench',
+            prompt_mode='zero-shot',
+            reader_cfg=reader_cfg,
+        ),
+        judge_cfg=dict(),
+        dict_postprocessor=dict(type=generic_llmjudge_postprocess),
+    ),
+)
+
+
+nejmaibench_dataset = dict(
+    type=NejmaibenchDataset,
+    abbr='nejmaibench',
+    path='opencompass/nejmaibench',
+    prompt_mode='zero-shot',
+    reader_cfg=reader_cfg,
+    infer_cfg=infer_cfg,
+    eval_cfg=eval_cfg,
+    
+)
+
+nejmaibench_datasets = [nejmaibench_dataset]

opencompass/configs/models/baichuan/hf_baichuan_m1_14b_base.py

@@ -0,0 +1,14 @@
+import torch
+from opencompass.models import HuggingFaceBaseModel
+
+models = [
+    dict(
+        type=HuggingFaceBaseModel,
+        abbr='baichuan-m1-14b-base-hf',
+        path='baichuan-inc/Baichuan-M1-14B-Base',
+        max_out_len=1024,
+        batch_size=8,
+        model_kwargs=dict(device_map='auto', trust_remote_code=True, torch_dtype=torch.bfloat16),
+        run_cfg=dict(num_gpus=1),
+    )
+]

opencompass/configs/models/baichuan/hf_baichuan_m1_14b_instruct.py

@@ -0,0 +1,14 @@
+import torch
+from opencompass.models import HuggingFacewithChatTemplate
+
+models = [
+    dict(
+        type=HuggingFacewithChatTemplate,
+        abbr='baichuan-m1-14b-instruct-hf',
+        path='baichuan-inc/Baichuan-M1-14B-Instruct',
+        max_out_len=2048,
+        batch_size=8,
+        model_kwargs=dict(device_map='auto', trust_remote_code=True, torch_dtype=torch.bfloat16),
+        run_cfg=dict(num_gpus=1),
+    )
+]

opencompass/configs/models/gemma/vllm_gemma_3_12b_it.py

@@ -0,0 +1,16 @@
+from opencompass.models import VLLMwithChatTemplate
+
+models = [
+    dict(
+        type=VLLMwithChatTemplate,
+        abbr='gemma-3-12b-it-vllm',
+        path='google/gemma-3-12b-it',
+        model_kwargs=dict(tensor_parallel_size=4,
+                          # for long context
+                          rope_scaling={'factor': 8.0, 'rope_type': 'linear'}),
+        max_out_len=4096,
+        batch_size=1,
+        generation_kwargs=dict(temperature=0),
+        run_cfg=dict(num_gpus=4),
+    )
+]

opencompass/configs/models/gemma/vllm_gemma_3_27b_it.py

@@ -0,0 +1,16 @@
+from opencompass.models import VLLMwithChatTemplate
+
+models = [
+    dict(
+        type=VLLMwithChatTemplate,
+        abbr='gemma-3-27b-it-vllm',
+        path='google/gemma-3-27b-it',
+        model_kwargs=dict(tensor_parallel_size=4,
+                          # for long context
+                          rope_scaling={'factor': 8.0, 'rope_type': 'linear'}),
+        max_out_len=4096,
+        batch_size=1,
+        generation_kwargs=dict(temperature=0),
+        run_cfg=dict(num_gpus=4),
+    )   
+]

opencompass/configs/models/gemma/vllm_gemma_3_4b_it.py

@@ -0,0 +1,17 @@
+from opencompass.models import VLLMwithChatTemplate
+
+models = [
+    dict(
+        type=VLLMwithChatTemplate,
+        abbr='gemma-3-4b-it-vllm',
+        path='google/gemma-3-4b-it',
+        model_kwargs=dict(tensor_parallel_size=2, 
+                          # for long context
+                          rope_scaling={'factor': 8.0, 'rope_type': 'linear'}),
+        max_seq_len=140000,
+        max_out_len=4096,
+        batch_size=1,
+        generation_kwargs=dict(temperature=0),
+        run_cfg=dict(num_gpus=2),
+    )
+]

opencompass/configs/models/hf_internlm/lmdeploy_internlm3_8b_instruct_128k.py

@@ -0,0 +1,19 @@
+from opencompass.models import TurboMindModelwithChatTemplate
+
+models = [
+    dict(
+        type=TurboMindModelwithChatTemplate,
+        abbr='internlm3-8b-instruct-turbomind',
+        path='internlm/internlm3-8b-instruct',
+        engine_config=dict(session_len=142000, max_batch_size=1, tp=2,
+                           # for long context
+                           rope_scaling_factor=6.0),
+        gen_config=dict(
+            top_k=1, temperature=1e-6, top_p=0.9, max_new_tokens=8192
+        ),
+        max_seq_len=142000,
+        max_out_len=8192,
+        batch_size=1,
+        run_cfg=dict(num_gpus=2),
+    )
+]

opencompass/configs/models/hf_internlm/lmdeploy_oreal_32b.py

@@ -0,0 +1,20 @@
+from opencompass.models import TurboMindModelwithChatTemplate
+from opencompass.utils.text_postprocessors import extract_non_reasoning_content
+
+models = [
+    dict(
+        type=TurboMindModelwithChatTemplate,
+        abbr='OREAL-32B',
+        path='internlm/OREAL-32B',
+        engine_config=dict(session_len=32768, max_batch_size=16, tp=4),
+        gen_config=dict(top_k=1,
+                        temperature=1e-6,
+                        top_p=0.9,
+                        max_new_tokens=32768),
+        max_seq_len=32768,
+        max_out_len=32768,
+        batch_size=16,
+        run_cfg=dict(num_gpus=4),
+        pred_postprocessor=dict(type=extract_non_reasoning_content)
+    )
+]

opencompass/configs/models/huatuogpt/hf_huatuogpt2_13b.py

@@ -0,0 +1,17 @@
+from opencompass.models import HuggingFacewithChatTemplate
+
+models = [
+    dict(
+        type=HuggingFacewithChatTemplate,
+        abbr='huatuogpt2-13b-hf',
+        path='FreedomIntelligence/HuatuoGPT2-13B',
+        tokenizer_kwargs=dict(padding_side='left',
+                              truncation_side='left',
+                              trust_remote_code=True,
+                              use_fast=True,),
+        max_out_len=1024,
+        batch_size=8,
+        model_kwargs=dict(device_map='auto', trust_remote_code=True),
+        run_cfg=dict(num_gpus=4),
+    )
+]

opencompass/configs/models/huatuogpt/hf_huatuogpt2_7b.py

@@ -0,0 +1,13 @@
+from opencompass.models import HuggingFacewithChatTemplate
+
+models = [
+    dict(
+        type=HuggingFacewithChatTemplate,
+        abbr='huatuogpt2-7b-hf',
+        path='FreedomIntelligence/HuatuoGPT2-7B',
+        max_out_len=1024,
+        batch_size=8,
+        model_kwargs=dict(device_map='auto', trust_remote_code=True),
+        run_cfg=dict(num_gpus=1),
+    )
+]

opencompass/configs/models/huatuogpt/hf_huatuogpt_o1_7b.py

@@ -0,0 +1,15 @@
+from opencompass.models import HuggingFacewithChatTemplate
+from opencompass.utils.text_postprocessors import extract_non_reasoning_content
+
+models = [
+    dict(
+        type=HuggingFacewithChatTemplate,
+        abbr='huatuogpt-o1-7b-hf',
+        path='FreedomIntelligence/HuatuoGPT-o1-7B',
+        max_out_len=2048,
+        batch_size=8,
+        model_kwargs=dict(device_map='auto', trust_remote_code=True),
+        run_cfg=dict(num_gpus=1),
+        pred_postprocessor=dict(type=extract_non_reasoning_content, think_start_token='## Thinking', think_end_token='## Final Response'),
+    )
+]

opencompass/configs/models/huatuogpt/hf_huatuogpt_o1_8b.py

@@ -0,0 +1,15 @@
+from opencompass.models import HuggingFacewithChatTemplate
+from opencompass.utils.text_postprocessors import extract_non_reasoning_content
+
+models = [
+    dict(
+        type=HuggingFacewithChatTemplate,
+        abbr='huatuogpt-o1-8b-hf',
+        path='FreedomIntelligence/HuatuoGPT-o1-8B',
+        max_out_len=2048,
+        batch_size=8,
+        model_kwargs=dict(device_map='auto', trust_remote_code=True),
+        run_cfg=dict(num_gpus=1),
+        pred_postprocessor=dict(type=extract_non_reasoning_content, think_start_token='## Thinking', think_end_token='## Final Response'),
+    )
+]

opencompass/configs/models/qwen2_5/vllm_qwen2_5_14b_instruct_128k.py

@@ -0,0 +1,21 @@
+from opencompass.models import VLLMwithChatTemplate
+
+models = [
+    dict(
+        type=VLLMwithChatTemplate,
+        abbr='qwen2.5-14b-instruct-vllm',
+        path='Qwen/Qwen2.5-14B-Instruct',
+        model_kwargs=dict(
+            tensor_parallel_size=4,
+            rope_scaling={
+                'factor': 4.0,
+                'original_max_position_embeddings': 32768,
+                'rope_type': 'yarn'
+            },
+        ),
+        max_out_len=4096,
+        batch_size=1,
+        generation_kwargs=dict(temperature=0),
+        run_cfg=dict(num_gpus=4),
+    )
+]

opencompass/configs/models/qwen2_5/vllm_qwen2_5_32b_instruct_128k.py

@@ -0,0 +1,21 @@
+from opencompass.models import VLLMwithChatTemplate
+
+models = [
+    dict(
+        type=VLLMwithChatTemplate,
+        abbr='qwen2.5-32b-instruct-vllm',
+        path='Qwen/Qwen2.5-32B-Instruct',
+        model_kwargs=dict(
+            tensor_parallel_size=8,
+            rope_scaling={
+                'factor': 4.0,
+                'original_max_position_embeddings': 32768,
+                'rope_type': 'yarn'
+            },
+        ),
+        max_out_len=4096,
+        batch_size=1,
+        generation_kwargs=dict(temperature=0),
+        run_cfg=dict(num_gpus=8),
+    )
+]

opencompass/configs/models/qwen2_5/vllm_qwen2_5_72b_instruct_128k.py

@@ -0,0 +1,21 @@
+from opencompass.models import VLLMwithChatTemplate
+
+models = [
+    dict(
+        type=VLLMwithChatTemplate,
+        abbr='qwen2_5-72b-instruct-vllm',
+        path='Qwen/Qwen2.5-72B-Instruct',
+        model_kwargs=dict(
+            tensor_parallel_size=8,
+            rope_scaling={
+                'factor': 4.0,
+                'original_max_position_embeddings': 32768,
+                'rope_type': 'yarn'
+            },
+        ),
+        max_out_len=4096,
+        batch_size=1,
+        generation_kwargs=dict(temperature=0),
+        run_cfg=dict(num_gpus=8),
+    )
+]

opencompass/configs/models/qwen2_5/vllm_qwen2_5_7b_instruct_128k.py

@@ -0,0 +1,21 @@
+from opencompass.models import VLLMwithChatTemplate
+
+models = [
+    dict(
+        type=VLLMwithChatTemplate,
+        abbr='qwen2.5-7b-instruct-vllm',
+        path='Qwen/Qwen2.5-7B-Instruct',
+        model_kwargs=dict(
+            tensor_parallel_size=4,
+            rope_scaling={
+                'factor': 4.0,
+                'original_max_position_embeddings': 32768,
+                'rope_type': 'yarn'
+            },
+        ),
+        max_out_len=4096,
+        batch_size=1,
+        generation_kwargs=dict(temperature=0),
+        run_cfg=dict(num_gpus=4),
+    )
+]

opencompass/datasets/CARDBiomedBench.py

@@ -0,0 +1,30 @@
+from datasets import load_dataset
+
+from opencompass.registry import LOAD_DATASET
+
+from .base import BaseDataset
+
+
+def _parse(item, prompt_mode):
+    item['expert'] = item['Bio_Category']
+    item['start'] = chr(65)
+    item['end'] = chr(65 + len(item.get('choices', {'label': []})['label']) -
+                      1)
+    item['prompt_mode'] = prompt_mode
+    return item
+
+
+@LOAD_DATASET.register_module()
+class CARDBiomedBenchDataset(BaseDataset):
+
+    @staticmethod
+    def load(path: str, prompt_mode: str, **kwargs):
+        data_files = {'test': 'data/CARDBiomedBench.csv'}
+        dataset = load_dataset(path, data_files=data_files, split='test')
+        # dataset = dataset.select(range(200))
+        if prompt_mode == 'zero-shot':
+            dataset = dataset.map(lambda item: _parse(item, prompt_mode),
+                                  load_from_cache_file=False)
+        elif prompt_mode == 'few-shot':
+            pass  # TODO: Implement few-shot prompt
+        return dataset

opencompass/datasets/ClinicBench.py

@@ -0,0 +1,19 @@
+from datasets import load_dataset
+
+from opencompass.registry import LOAD_DATASET
+
+from .base import BaseDataset
+
+
+@LOAD_DATASET.register_module()
+class ClinicBenchDataset(BaseDataset):
+
+    @staticmethod
+    def load_single(path):
+        dataset = load_dataset(path)['train']
+        return dataset
+
+    @staticmethod
+    def load(path):
+        dataset = ClinicBenchDataset.load_single(path)
+        return dataset

opencompass/datasets/MedCalc_Bench.py

@@ -0,0 +1,323 @@
+import math
+import re
+from datetime import datetime
+
+import numpy as np
+from datasets import load_dataset
+
+from opencompass.openicl import BaseEvaluator
+from opencompass.registry import LOAD_DATASET
+
+from .base import BaseDataset
+
+
+def check_correctness(answer: str, ground_truth, calid, upper_limit,
+                      lower_limit):
+    """"""
+    calid = int(calid)
+
+    if calid in [13, 68]:
+        # Output Type: date
+
+        if datetime.strptime(
+                answer,
+                '%m/%d/%Y').strftime('%-m/%-d/%Y') == datetime.strptime(
+                    ground_truth, '%m/%d/%Y').strftime('%-m/%-d/%Y'):
+            correctness = 1
+        else:
+            correctness = 0
+    elif calid in [69]:
+        # Output Type: integer (A, B)
+        match = re.search(
+            r"\(?[\"\']?(\d+)\s*(weeks?)?[\"\']?,?"
+            r"\s*[\"\']?(\d+)\s*(days?)?[\"\']?\s*\)?", ground_truth)
+        ground_truth = f'({match.group(1)}, {match.group(3)})'
+        match = re.search(
+            r"\(?[\"\']?(\d+)\s*(weeks?)?[\"\']?,?"
+            r"\s*[\"\']?(\d+)\s*(days?)?[\"\']?\s*\)?", answer)
+        if match:
+            weeks = match.group(1)
+            days = match.group(3)
+            answer = f'({weeks}, {days})'
+            if eval(answer) == eval(ground_truth):
+                correctness = 1
+            else:
+                correctness = 0
+        else:
+            correctness = 0
+    elif calid in [
+            4, 15, 16, 17, 18, 20, 21, 25, 27, 28, 29, 32, 33, 36, 43, 45, 48,
+            51, 69
+    ]:
+        # Output Type: integer A
+        answer = round(eval(answer))
+        if answer == eval(ground_truth):
+            correctness = 1
+        else:
+            correctness = 0
+    elif calid in [
+            2, 3, 5, 6, 7, 8, 9, 10, 11, 19, 22, 23, 24, 26, 30, 31, 38, 39,
+            40, 44, 46, 49, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67
+    ]:
+        # Output Type: decimal
+        answer = eval(answer)
+        if answer >= eval(lower_limit) and answer <= eval(upper_limit):
+            correctness = 1
+        else:
+            correctness = 0
+    else:
+        raise ValueError(f'Unknown calculator ID: {calid}')
+    return correctness
+
+
+def extract_answer(answer, calid):
+
+    calid = int(calid)
+    extracted_answer = re.findall(r'[Aa]nswer":\s*(.*?)\}', answer)
+    matches = re.findall(
+        r'"step_by_step_thinking":\s*"'
+        r'([^"]+)"\s*,\s*"[Aa]nswer"', answer)
+
+    if matches:
+        # Select the last match
+        last_match = matches[-1]
+        explanation = last_match
+    else:
+        explanation = 'No Explanation'
+
+    if len(extracted_answer) == 0:
+        extracted_answer = 'Not Found'
+    else:
+        extracted_answer = extracted_answer[-1].strip().strip('"')
+        if extracted_answer == 'str(short_and_direct\
+                _answer_of_the_question)':
+            extracted_answer = 'Not Found'
+        if extracted_answer == 'str(value which is\
+                the answer to the question)':
+            extracted_answer = 'Not Found'
+        if extracted_answer == 'X.XX':
+            extracted_answer = 'Not Found'
+
+    if calid in [13, 68]:
+        # Output Type: date
+        match = re.search(
+            r'^(0?[1-9]|1[0-2])\/(0?[1-9]'
+            r'|[12][0-9]|3[01])\/(\d{4})', extracted_answer)
+        if match:
+            month = int(match.group(1))
+            day = int(match.group(2))
+            year = match.group(3)
+            answer = f'{month:02}/{day:02}/{year}'
+        else:
+            answer = 'N/A'
+
+    elif calid in [69]:
+        # Output Type: integer (A, B)
+        match = re.search(
+            r"\(?[\"\']?(\d+)\s*(weeks?)?[\"\']?,"
+            r"\?\s*[\"\']?(\d+)\s*(days?)?[\"\']?\s*\)?", extracted_answer)
+        extracted_answer = extracted_answer.replace('[', '(').replace(
+            ']', ')').replace("'", '').replace('"', '')
+        match = re.search(
+            r"\(?[\"\']?(\d+)\s*(weeks?)?[\"\']?,"
+            r"?\s*[\"\']?(\d+)\s*(days?)?[\"\']?\s*\)?", extracted_answer)
+        if match:
+            weeks = match.group(1)
+            days = match.group(3)
+            answer = f'({weeks}, {days})'
+        else:
+            answer = 'N/A'
+    elif calid in [
+            4, 15, 16, 17, 18, 20, 21, 25, 27, 28, 29, 32, 33, 36, 43, 45, 48,
+            51, 69
+    ]:
+        # Output Type: integer A
+        match = re.search(r'(\d+) out of', extracted_answer)
+        if match:  # cases like "3 out of 5"
+            answer = match.group(1)
+        else:
+            match = re.search(r'-?\d+(, ?-?\d+)+', extracted_answer)
+            if match:  # cases like "3, 4, 5"
+                answer = str(len(match.group(0).split(',')))
+            else:
+                # match = re.findall(r"(?<!-)\d+", extracted_answer)
+                match = re.findall(r'(-?\d+(\.\d+)?)', extracted_answer)
+                # match = re.findall(r"-?\d+", extracted_answer)
+                if len(match) > 0:  # find the last integer
+                    answer = match[-1][0]
+                    # answer = match[-1].lstrip("0")
+                else:
+                    answer = 'N/A'
+    elif calid in [
+            2, 3, 5, 6, 7, 8, 9, 10, 11, 19, 22, 23, 24, 26, 30, 31, 38, 39,
+            40, 44, 46, 49, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67
+    ]:
+        # Output Type: decimal
+        match = re.search(r'str\((.*)\)', extracted_answer)
+        if match:
+            expression = match.group(1).replace('^', '**').replace(
+                'is odd', '% 2 == 1').replace('is even', '% 2 == 0').replace(
+                    'sqrt', 'math.sqrt').replace('.math', '').replace(
+                        'weight',
+                        '').replace('height', '').replace('mg/dl', '').replace(
+                            'g/dl', '').replace('mmol/L', '').replace(
+                                'kg', '').replace('g',
+                                                  '').replace('mEq/L', '')
+            expression = expression.split('#')[0]
+            if expression.count('(') > expression.count(')'):  # add missing ')
+                expression += ')' * (expression.count('(') -
+                                     expression.count(')'))
+            elif expression.count(')') > expression.count(
+                    '('):  # add missing (
+                expression = '(' * (expression.count(')') -
+                                    expression.count('(')) + expression
+            try:
+                answer = eval(expression, {'__builtins__': None}, {
+                    'min': min,
+                    'pow': pow,
+                    'round': round,
+                    'abs': abs,
+                    'int': int,
+                    'float': float,
+                    'math': math,
+                    'np': np,
+                    'numpy': np
+                })
+            except Exception:
+                print(f'Error in evaluating expression: {expression}')
+                answer = 'N/A'
+        else:
+            match = re.search(r'(-?\d+(\.\d+)?)\s*mL/min/1.73',
+                              extracted_answer)
+            if match:  # cases like "8.1 mL/min/1.73 m\u00b2"
+                answer = eval(match.group(1))
+            else:
+                match = re.findall(r'(-?\d+(\.\d+)?)\%', extracted_answer)
+                if len(match) > 0:  # cases like "53.1%"
+                    answer = eval(match[-1][0]) / 100
+                else:
+                    match = re.findall(r'(-?\d+(\.\d+)?)', extracted_answer)
+                    if len(
+                            match
+                    ) > 0:  # cases like "8.1 mL/min/1.73 m\u00b2" or "11.1"
+                        answer = eval(match[-1][0])
+                    else:
+                        answer = 'N/A'
+        if answer != 'N/A':
+            answer = str(answer)
+
+    return answer, explanation
+
+
+def _parse(item, prompt_mode):
+    item['row_number'] = item['Row Number']
+    item['calculator_id'] = item['Calculator ID']
+    item['calculator_name'] = item['Calculator Name']
+    item['category'] = item['Category']
+    item['output_type'] = item['Output Type']
+    item['note_id'] = item['Note ID']
+    item['note_type'] = item['Note Type']
+    item['patient_note'] = item['Patient Note']
+    item['question'] = item['Question']
+    item['relevant_entities'] = item['Relevant Entities']
+    item['ground_truth_answer'] = item['Ground Truth Answer']
+    item['lower_limit'] = item['Lower Limit']
+    item['upper_limit'] = item['Upper Limit']
+    item['ground_truth_explanation'] = item['Ground Truth Explanation']
+    return item
+
+
+@LOAD_DATASET.register_module()
+class MedCalc_BenchDataset(BaseDataset):
+
+    @staticmethod
+    def load(path: str, prompt_mode: str, **kwargs):
+        data_files = {
+            'test': 'data/test-00000-of-00001.parquet',
+            'train': 'data/train-00000-of-00001.parquet'
+        }
+        dataset = load_dataset(path, data_files=data_files, split='test')
+        # dataset = dataset.select(range(2))
+        if prompt_mode == 'zero-shot':
+            dataset = dataset.map(lambda item: _parse(item, prompt_mode),
+                                  load_from_cache_file=False)
+        elif prompt_mode == 'few-shot':
+            pass  # TODO: Implement few-shot prompt
+        return dataset
+
+
+class MedCalcOfficial_Evaluator(BaseEvaluator):
+
+    def score(self, predictions, references, test_set):
+
+        if len(predictions) != len(references):
+            return {'error': 'preds and refrs have different length'}
+
+        correct = 0
+        count = 0
+        details = []
+        for idx, (i, j) in enumerate(zip(predictions, references)):
+            calculator_id = test_set['calculator_id'][idx]
+            lower_limit = test_set['lower_limit'][idx]
+            upper_limit = test_set['upper_limit'][idx]
+            row_number = test_set['row_number'][idx]
+            note_id = test_set['note_id'][idx]
+            category = test_set['category'][idx]
+            question = test_set['question'][idx]
+            calculator_name = test_set['calculator_name'][idx]
+            patient_note = test_set['patient_note'][idx]
+            ground_truth_explanation = test_set['ground_truth_explanation'][
+                idx]
+            ground_truth_answer = test_set['ground_truth_answer'][idx]
+            try:
+                answer_value, explanation = extract_answer(
+                    i, int(calculator_id))
+
+                print(answer_value)
+                print(explanation)
+
+                correctness = check_correctness(answer_value,
+                                                ground_truth_answer,
+                                                calculator_id, upper_limit,
+                                                lower_limit)
+
+                status = 'Correct' if correctness else 'Incorrect'
+
+                outputs = {
+                    'Row Number': int(row_number),
+                    'Calculator Name': calculator_name,
+                    'Calculator ID': calculator_id,
+                    'Category': category,
+                    'Note ID': note_id,
+                    'Patient Note': patient_note,
+                    'Question': question,
+                    'LLM Answer': answer_value,
+                    'LLM Explanation': explanation,
+                    'Ground Truth Answer': ground_truth_answer,
+                    'Ground Truth Explanation': ground_truth_explanation,
+                    'Result': status
+                }
+
+            except Exception as e:
+                outputs = {
+                    'Row Number': int(row_number),
+                    'Calculator Name': calculator_name,
+                    'Calculator ID': calculator_id,
+                    'Category': category,
+                    'Note ID': note_id,
+                    'Patient Note': patient_note,
+                    'Question': question,
+                    'LLM Answer': str(e),
+                    'LLM Explanation': str(e),
+                    'Ground Truth Answer': ground_truth_answer,
+                    'Ground Truth Explanation': ground_truth_explanation,
+                    'Result': 'Incorrect'
+                }
+                status = 'Incorrect'
+            count += 1
+            if status == 'Correct':
+                correct += 1
+            details.append(outputs)
+
+        result = {'accuracy': 100 * correct / count, 'details': details}
+        return result

opencompass/datasets/MedQA.py

@@ -0,0 +1,29 @@
+from datasets import Dataset, load_dataset
+
+from opencompass.registry import LOAD_DATASET
+
+from .base import BaseDataset
+
+
+@LOAD_DATASET.register_module()
+class MedQADataset(BaseDataset):
+
+    @staticmethod
+    def load_single(path):
+        dataset = []
+        ds = load_dataset(path)
+        for data in ds['train']:
+            data['label'] = data['answer_idx']
+            choices = ''
+            for option in data['options']:
+                choices += option + '. ' + data['options'][option] + '\n'
+            data['choices'] = choices
+
+            dataset.append(data)
+
+        return Dataset.from_list(dataset)
+
+    @staticmethod
+    def load(path):
+        dataset = MedQADataset.load_single(path)
+        return dataset

opencompass/datasets/ProteinLMBench.py

@@ -0,0 +1,58 @@
+from datasets import load_dataset
+
+from opencompass.openicl import BaseEvaluator
+from opencompass.registry import LOAD_DATASET
+from opencompass.utils.text_postprocessors import first_option_postprocess
+
+from .base import BaseDataset
+
+
+def _parse(item):
+    item['start'] = chr(65)
+    item['end'] = chr(65 + len(item.get('options', [])) - 1)
+    new_options = []
+    choices = ''
+    for i in range(len(item['options'])):
+        new_options.append(item['options'][i].split(': ')[-1])
+        choices += chr(65 +
+                       i) + '. ' + item['options'][i].split(': ')[-1] + '\n'
+    item['question'] = (f'\nQuestion: {item["question"]}\n'
+                        f'Answer Choices: \n{choices}')
+    item['options'] = new_options
+    item['label'] = chr(65 + int(item['answer'].split(' ')[-1]) -
+                        1)  # Index from 1 in answer
+    return item
+
+
+@LOAD_DATASET.register_module()
+class ProteinLMBenchDataset(BaseDataset):
+
+    @staticmethod
+    def load(path: str, **kwargs):
+        dataset = load_dataset(path, 'evaluation', split='train')
+        dataset = dataset.map(lambda item: _parse(item))
+
+        return dataset
+
+
+class ProteinLMBenchEvaluator(BaseEvaluator):
+
+    def score(self, predictions, references, test_set):
+        if len(predictions) != len(references):
+            return {'error': 'preds and refrs have different length'}
+        correct = 0
+        count = 0
+        details = []
+        for idx, (prediction,
+                  reference) in enumerate(zip(predictions, references)):
+            options = ''.join(
+                [chr(65 + i) for i in range(len(test_set['options'][idx]))])
+            predict = first_option_postprocess(prediction, options)
+            detail = {'pred': predict, 'answer': reference, 'correct': False}
+            count += 1
+            if predict == reference:
+                correct += 1
+                detail['correct'] = True
+            details.append(detail)
+        result = {'accuracy': 100 * correct / count, 'details': details}
+        return result

opencompass/datasets/PubMedQA.py

@@ -0,0 +1,34 @@
+from datasets import Dataset, load_dataset
+
+from opencompass.registry import LOAD_DATASET
+
+from .base import BaseDataset
+
+
+@LOAD_DATASET.register_module()
+class PubMedQADataset(BaseDataset):
+
+    @staticmethod
+    def load_single(path):
+        dataset = []
+        ds = load_dataset(path, 'pqa_labeled')
+        for data in ds['train']:
+            data['question'] = (f"CONTEXTS: {data['context']}\n"
+                                f"QUESTION: {data['question']}")
+            choices = 'A. yes\nB. no\nC. maybe'
+            data['choices'] = choices
+            if data['final_decision'] == 'yes':
+                data['label'] = 'A. yes'
+            elif data['final_decision'] == 'no':
+                data['label'] = 'B. no'
+            else:
+                data['label'] = 'C. maybe'
+
+            dataset.append(data)
+
+        return Dataset.from_list(dataset)
+
+    @staticmethod
+    def load(path):
+        dataset = PubMedQADataset.load_single(path)
+        return dataset

opencompass/datasets/ScienceQA.py

@@ -0,0 +1,32 @@
+from datasets import Dataset, load_dataset
+
+from opencompass.registry import LOAD_DATASET
+
+from .base import BaseDataset
+
+
+@LOAD_DATASET.register_module()
+class ScienceQADataset(BaseDataset):
+
+    @staticmethod
+    def load_single(path):
+        dataset = []
+        ds = load_dataset(path)
+        for data in ds['test']:
+            if data['image'] is None:
+                data['label'] = chr(65 + data['answer']
+                                    ) + '. ' + data['choices'][data['answer']]
+                choices = ''
+                for i in range(len(data['choices'])):
+                    choices += chr(65 + i) + '. ' + data['choices'][i] + '\n'
+                data['choices'] = choices
+                # print(data)
+
+                dataset.append(data)
+
+        return Dataset.from_list(dataset)
+
+    @staticmethod
+    def load(path):
+        dataset = ScienceQADataset.load_single(path)
+        return dataset

opencompass/datasets/__init__.py

@@ -16,6 +16,7 @@ from .boolq import *  # noqa: F401, F403
 from .bustum import *  # noqa: F401, F403
 from .c3 import *  # noqa: F401, F403
 from .calm import *  # noqa: F401, F403
+from .CARDBiomedBench import CARDBiomedBenchDataset  # noqa: F401
 from .cb import *  # noqa: F401, F403
 from .ceval import *  # noqa: F401, F403
 from .charm import *  # noqa: F401, F403
@@ -96,6 +97,9 @@ from .math_intern import *  # noqa: F401, F403
 from .mathbench import *  # noqa: F401, F403
 from .mbpp import *  # noqa: F401, F403
 from .medbench import *  # noqa: F401, F403
+from .MedCalc_Bench import MedCalc_BenchDataset  # noqa: F401
+from .MedCalc_Bench import MedCalcOfficial_Evaluator  # noqa: F401
+from .MedQA import *  # noqa: F401, F403
 from .MedXpertQA import *  # noqa: F401, F403
 from .mgsm import *  # noqa: F401, F403
 from .mmlu import *  # noqa: F401, F403
@@ -109,6 +113,7 @@ from .musr import *  # noqa: F401, F403
 from .narrativeqa import *  # noqa: F401, F403
 from .natural_question import *  # noqa: F401, F403
 from .natural_question_cn import *  # noqa: F401, F403
+from .nejmaibench import *  # noqa: F401, F403
 from .NPHardEval import *  # noqa: F401, F403
 from .obqa import *  # noqa: F401, F403
 from .olymmath import *  # noqa: F401, F403
@@ -116,6 +121,7 @@ from .OlympiadBench import *  # noqa: F401, F403
 from .OpenFinData import *  # noqa: F401, F403
 from .physics import *  # noqa: F401, F403
 from .piqa import *  # noqa: F401, F403
+from .ProteinLMBench import *  # noqa: F401, F403
 from .py150 import *  # noqa: F401, F403
 from .qasper import *  # noqa: F401, F403
 from .qaspercut import *  # noqa: F401, F403

opencompass/datasets/base.py

@@ -23,7 +23,8 @@ class BaseDataset:
                 'idx': idx
             },
                                   with_indices=True,
-                                  writer_batch_size=16)
+                                  writer_batch_size=16,
+                                  load_from_cache_file=False)
             dataset = concatenate_datasets([dataset] * n)
             self.dataset = dataset
         else:
@@ -34,7 +35,8 @@ class BaseDataset:
                     'idx': idx
                 },
                                                 with_indices=True,
-                                                writer_batch_size=16)
+                                                writer_batch_size=16,
+                                                load_from_cache_file=False)
                 dataset[key] = concatenate_datasets([dataset[key]] * n)
                 self.dataset[key] = dataset[key]
         self._init_reader(**reader_cfg)

opencompass/datasets/nejmaibench.py

@@ -0,0 +1,139 @@
+import re
+
+import pandas as pd
+from datasets import Dataset
+
+from opencompass.openicl import BaseEvaluator
+from opencompass.registry import LOAD_DATASET, TEXT_POSTPROCESSORS
+from opencompass.utils import get_data_path
+
+from .base import BaseDataset
+
+
+def _parse(item, prompt_mode):
+    # 1. 从 Choices 字符串里按行拆分出每个选项
+    raw_choices = item.get('Choices', '')
+    # 去掉首尾空白并按行分割，过滤掉空行
+    lines = [
+        line.strip() for line in raw_choices.strip().splitlines()
+        if line.strip()
+    ]
+
+    # 2. 用正则去掉行首的 "A. "/"B. " 等前缀，只保留选项内容
+    options_list = [re.sub(r'^[A-Z]\.\s*', '', line) for line in lines]
+
+    # 3. 写回 item
+    item['options'] = options_list
+
+    # 4. 重建带标号的选项字符串
+    options_str = '\n'.join(f'{chr(65 + i)}. {opt}'
+                            for i, opt in enumerate(options_list))
+
+    # 5. 构造 question、label、prompt_mode、start、end
+    item['question'] = f"{item['Question']}\n{options_str}"
+    item['label'] = item['Answer']
+    item['prompt_mode'] = prompt_mode
+    item['start'] = chr(65)
+    item['end'] = chr(65 + len(options_list) - 1)
+    return item
+
+
+@LOAD_DATASET.register_module()
+class NejmaibenchDataset(BaseDataset):
+
+    @staticmethod
+    def load(path: str, prompt_mode: str = 'zero-shot', **kwargs):
+        # 读取 CSV 文件为 DataFrame，并将 NaN 转为空字符串
+        path = get_data_path(path)
+        df = pd.read_csv(path, encoding='utf-8')
+        df = df.fillna('')
+
+        # 转换为字典列表
+        data_list = df.to_dict(orient='records')
+
+        # 将数据列表包装为 Dataset
+        dataset = Dataset.from_list(data_list)
+
+        # 根据提示模式进行解析
+        if prompt_mode == 'zero-shot':
+            dataset = dataset.map(lambda item: _parse(item, prompt_mode))
+        elif prompt_mode == 'few-shot':
+            pass  # TODO: Implement few-shot prompt handling
+        return dataset
+
+
+class NejmaibenchEvaluator(BaseEvaluator):
+
+    def score(self, predictions, references, test_set):
+        method = test_set['prompt_mode'][0]
+
+        if len(predictions) != len(references):
+            return {'error': 'preds and refrs have different length'}
+        correct = 0
+        count = 0
+        details = []
+        for idx, (i, j) in enumerate(zip(predictions, references)):
+            i = answer_cleansing(method, i, test_set['options'][idx],
+                                 test_set['label'][idx])
+            detail = {
+                'pred': i,
+                'answer': j,
+                'correct': False,
+                'Subject': test_set['Subject'][idx],
+            }
+            count += 1
+            if i == j:
+                correct += 1
+                detail['correct'] = True
+            details.append(detail)
+        result = {'accuracy': 100 * correct / count, 'details': details}
+        return result
+
+
+@TEXT_POSTPROCESSORS.register_module()
+def answer_cleansing(
+    method: str,
+    prediction: str,
+    options: list,
+    label: str,
+) -> str:
+
+    # Clean up unwanted phrases in the prediction
+    for unwanted_phrase in [
+            'I understand',
+            'A through J',
+            'A through E',
+            'A through D',
+    ]:
+        prediction = prediction.replace(unwanted_phrase, '')
+
+    options_num = len(options)
+    options = [chr(65 + i) for i in range(options_num)]
+    options_str = r'\b(' + '|'.join(options) + r')\b'
+    prediction = re.findall(options_str, prediction)
+
+    if len(prediction) == 0:
+        prediction = []
+        return prediction
+    else:
+        # If there is a "label" and its length is 1,
+        # process prediction accordingly
+        if len(label) == 1:
+            if method == 'few-shot':
+                answer_flag = True if len(prediction) > 1 else False
+                # choose the first or last element based on the answer_flag
+                if answer_flag:
+                    prediction = [prediction[0]]
+                else:
+                    prediction = [prediction[-1]]
+            elif method == 'zero-shot':
+                # choose the first element in list
+                prediction = [prediction[0]]
+            else:
+                raise ValueError('Method is not properly defined ...')
+
+            # Remove trailing period if it exists
+            if prediction[0] and prediction[0].endswith('.'):
+                prediction[0] = prediction[0][:-1]
+
+        return prediction[0]

opencompass/utils/datasets_info.py

@@ -446,6 +446,11 @@ DATASETS_MAPPING = {
         "hf_id": "",
         "local": "./data/ChemBench4K",
     },
+    "opencompass/nejmaibench": {
+        "ms_id": "",
+        "hf_id": "",
+        "local": "./data/nejmaibench/NEJM_All_Questions_And_Answers.csv",
+    },
 
 }
 
@@ -798,6 +803,11 @@ DATASETS_URL = {
         "url":
         "http://opencompass.oss-cn-shanghai.aliyuncs.com/datasets/data/ChemBench4K.zip",
         "md5": "fc23fd21b2566a5dbbebfa4601d7779c"
+    },
+    "nejmaibench": {
+        "url":
+        "http://opencompass.oss-cn-shanghai.aliyuncs.com/datasets/data/nejmaibench.zip",
+        "md5": "e6082cae3596b3ebea73e23ba445b99e"
     }
     
 }