jiangzhs/OpenCompass
1
0
Fork 0
mirror of https://github.com/open-compass/opencompass.git synced 2025-05-30 16:03:24 +08:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

fix

Browse Source
This commit is contained in:
Myhs-phz 2025-03-17 09:40:57 +00:00
parent 51f5792f7c
commit 53c6725d19
9 changed files with 202 additions and 338 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
opencompass/configs/datasets/aime2024/aime2024_gen.py
View File

@ -30,10 +30,10 @@ aime2024_eval_cfg = dict(
aime2024_datasets = [
dict(
abbr='aime2024',
type=Aime2024Dataset,
path='opencompass/aime2024',
type=CustomDataset,
path='opencompass/aime2025',
reader_cfg=aime2024_reader_cfg,
infer_cfg=aime2024_infer_cfg,
eval_cfg=aime2024_eval_cfg
eval_cfg=aime2024_eval_cfg,
)
]

66
opencompass/configs/datasets/aime2024/aime2024_llm_judge_gen.py
View File

@ -1,14 +1,26 @@
from opencompass.openicl.icl_prompt_template import PromptTemplate
from opencompass.openicl.icl_retriever import ZeroRetriever
from opencompass.openicl.icl_inferencer import GenInferencer
from opencompass.datasets import Aime2024Dataset, MATHEvaluator, math_postprocess_v2
from opencompass.datasets import CustomDataset
from opencompass.evaluator import GenericLLMEvaluator
from opencompass.datasets import generic_llmjudge_postprocess
aime2024_reader_cfg = dict(input_columns=['question'], output_column='answer')
aime2024_reader_cfg = dict(
input_columns=['question'],
output_column='answer'
aime2024_infer_cfg = dict(
prompt_template=dict(
type=PromptTemplate,
template=dict(
round=[
dict(
role='HUMAN',
prompt='{question}\nRemember to put your final answer within \\boxed{}.',
),
],
),
),
retriever=dict(type=ZeroRetriever),
inferencer=dict(type=GenInferencer),
)
GRADER_TEMPLATE = """
@ -19,35 +31,19 @@ GRADER_TEMPLATE = """
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.
5. If the prediction is given with \\boxed{}, please ignore the \\boxed{} and only judge whether the candidate's answer is consistent with the standard answer.
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 A) {A}\n B) {B}\n C) {C}\n D) {D}\n<Original Question End>\n\n
<Original Question Begin>: \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()
aime2024_infer_cfg = dict(
prompt_template=dict(
type=PromptTemplate,
template=dict(
round=[
dict(role='HUMAN', prompt='{question}\nPlease reason step by step, and put your final answer within \\boxed{}.'),
],
)
),
retriever=dict(type=ZeroRetriever),
inferencer=dict(type=GenInferencer, max_out_len=2048)
)
aime2024_eval_cfg = dict(
evaluator=dict(
type=GenericLLMEvaluator,
@ -58,33 +54,31 @@ aime2024_eval_cfg = dict(
dict(
role='SYSTEM',
fallback_role='HUMAN',
prompt="You are a helpful assistant who evaluates the correctness and quality of models' outputs.")
prompt="You are a helpful assistant who evaluates the correctness and quality of models' outputs.",
)
],
round=[
dict(
role='HUMAN',
prompt=GRADER_TEMPLATE
),
]),
dict(role='HUMAN', prompt=GRADER_TEMPLATE),
],
),
),
dataset_cfg=dict(
type=Aime2024Dataset,
path='opencompass/aime2024',
type=CustomDataset,
path='opencompass/aime2025',
reader_cfg=aime2024_reader_cfg,
),
judge_cfg=dict(),
dict_postprocessor=dict(type=generic_llmjudge_postprocess),
),
pred_role='BOT',
)
)
aime2024_datasets = [
dict(
abbr='aime2024',
type=Aime2024Dataset,
path='opencompass/aime2024',
type=CustomDataset,
path='opencompass/aime2025',
reader_cfg=aime2024_reader_cfg,
infer_cfg=aime2024_infer_cfg,
eval_cfg=aime2024_eval_cfg
eval_cfg=aime2024_eval_cfg,
)
]

27
opencompass/configs/datasets/cmmlu/cmmlu_llm_judge_gen.py
View File

@ -4,7 +4,6 @@ from opencompass.openicl.icl_inferencer import GenInferencer
from opencompass.openicl.icl_evaluator import AccEvaluator
from opencompass.datasets import CMMLUDataset
from opencompass.utils.text_postprocessors import match_answer_pattern
from opencompass.openicl.icl_evaluator import LMEvaluator
from opencompass.evaluator import GenericLLMEvaluator
from opencompass.datasets import generic_llmjudge_postprocess
@ -75,14 +74,12 @@ cmmlu_subject_mapping = {
'traditional_chinese_medicine': '中医中药',
'virology': '病毒学',
'world_history': '世界历史',
'world_religions': '世界宗教'
'world_religions': '世界宗教',
}
QUERY_TEMPLATE = """
你回答的最后一行**必须**是以下格式 '答案: $选项' (不带引号), 其中选项是ABCD之一.
{question}
A) {A}
B) {B}
C) {C}
@ -97,14 +94,11 @@ GRADER_TEMPLATE = """
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>: \n {question}\n A) {A}\n B) {B}\n C) {C}\n D) {D}\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
@ -140,14 +134,13 @@ for _name in cmmlu_all_sets:
dict(
role='SYSTEM',
fallback_role='HUMAN',
prompt="You are a helpful assistant who evaluates the correctness and quality of models' outputs.")
prompt="You are a helpful assistant who evaluates the correctness and quality of models' outputs.",
)
],
round=[
dict(
role='HUMAN',
prompt=GRADER_TEMPLATE
),
]),
dict(role='HUMAN', prompt=GRADER_TEMPLATE),
],
),
),
dataset_cfg=dict(
type=CMMLUDataset,
@ -157,7 +150,7 @@ for _name in cmmlu_all_sets:
input_columns=['question', 'A', 'B', 'C', 'D'],
output_column='answer',
train_split='dev',
test_split='test'
test_split='test',
),
),
judge_cfg=dict(),
@ -175,10 +168,12 @@ for _name in cmmlu_all_sets:
input_columns=['question', 'A', 'B', 'C', 'D'],
output_column='answer',
train_split='dev',
test_split='test'),
test_split='test',
),
infer_cfg=cmmlu_infer_cfg,
eval_cfg=cmmlu_eval_cfg,
mode='singlescore',
))
)
)
del _name, _ch_name

53
opencompass/configs/datasets/drop/drop_llm_judge_gen.py
View File

@ -2,8 +2,9 @@ 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.datasets import DropOpenAIDataset, DropOpenAIEvaluator, generic_llmjudge_postprocess
from opencompass.datasets import DropOpenAIDataset
from opencompass.evaluator import GenericLLMEvaluator
from opencompass.datasets import generic_llmjudge_postprocess
with read_base():
from .drop_examples import drop_examples # noqa: F401, F403
@ -17,35 +18,33 @@ drop_reader_cfg = dict(
template = f'You will be asked to read a passage and answer a question. Some examples of passages and Q&A are provided below.\n\n{drop_examples}\n\n# Your Task\n\n---\n{{prompt}}\n\nThink step by step, then write a line of the form "Answer: $ANSWER" at the end of your response.'
drop_infer_cfg = dict(
prompt_template=dict(type=PromptTemplate, template=dict(round=[dict(role='HUMAN', prompt=template)])),
retriever=dict(type=ZeroRetriever),
inferencer=dict(type=GenInferencer))
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.
5. If the prediction is given with \\boxed{}, please ignore the \\boxed{} and only judge whether the candidate's answer is consistent with the standard answer.
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
<Original Question Begin>: {prompt}\n \n<Original Question End>\n\n
<Gold Target Begin>: \n{answers}\n<Gold Target End>\n\n
<Predicted Answer Begin>: \n{prediction}\n<Predicted End>\n\n
Judging the correctness of candidates' answers:
""".strip()
drop_infer_cfg = dict(
prompt_template=dict(
type=PromptTemplate,
template=dict(round=[dict(role='HUMAN', prompt=template)]),
),
retriever=dict(type=ZeroRetriever),
inferencer=dict(type=GenInferencer),
)
drop_eval_cfg = dict(
evaluator=dict(
@ -53,18 +52,17 @@ drop_eval_cfg = dict(
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.")
],
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
),
]),
dict(role='HUMAN', prompt=GRADER_TEMPLATE),
],
),
),
dataset_cfg=dict(
type=DropOpenAIDataset,
@ -76,8 +74,6 @@ drop_eval_cfg = dict(
),
pred_role='BOT',
)
drop_datasets = [
dict(
abbr='drop',
@ -85,5 +81,6 @@ drop_datasets = [
path='data/drop_simple_eval/dev.jsonl',
reader_cfg=drop_reader_cfg,
infer_cfg=drop_infer_cfg,
eval_cfg=drop_eval_cfg)
eval_cfg=drop_eval_cfg,
)
]

52
opencompass/configs/datasets/hellaswag/hellaswag_llm_judge_gen.py
View File

@ -1,10 +1,11 @@
from opencompass.openicl.icl_prompt_template import PromptTemplate
from opencompass.openicl.icl_retriever import FixKRetriever
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 HellaswagDatasetwithICE, generic_llmjudge_postprocess
from opencompass.datasets import HellaswagDatasetwithICE
from opencompass.utils.text_postprocessors import first_option_postprocess
from opencompass.evaluator import GenericLLMEvaluator
from opencompass.datasets import generic_llmjudge_postprocess
hellaswag_reader_cfg = dict(
input_columns=['ctx', 'A', 'B', 'C', 'D'],
@ -13,6 +14,14 @@ hellaswag_reader_cfg = dict(
test_split='val',
)
align_prompt = """Continue the following text without adding any additional information or formatting:
{ctx}
A) {A}
B) {B}
C) {C}
D) {D}
What is the right option?'"""
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.
@ -21,45 +30,27 @@ GRADER_TEMPLATE = """
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
<Original Question Begin>: {ctx}\n A) {A}\n B) {B}\n C) {C}\n D) {D}\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()
hellaswag_infer_cfg = dict(
ice_template=dict(
type=PromptTemplate,
template=dict(
round=[
dict(role='HUMAN', prompt=f'{{ctx}}\nA) {{A}}\nB) {{B}}\nC) {{C}}\nD) {{D}}\nWhat is the right option?'),
dict(role='BOT', prompt='{label}\n'),
]
),
),
prompt_template=dict(
type=PromptTemplate,
template=dict(
begin=[
dict(role='HUMAN', prompt='Continue the following text without adding any additional information or formatting:\n'),
'</E>',
],
round=[
dict(role='HUMAN', prompt=f'{{ctx}}\nA) {{A}}\nB) {{B}}\nC) {{C}}\nD) {{D}}\nWhat is the right option?'),
dict(role='BOT', prompt='{label}\n'),
dict(role='HUMAN', prompt=align_prompt),
],
),
ice_token='</E>',
),
retriever=dict(type=FixKRetriever, fix_id_list=list(range(10))),
retriever=dict(type=ZeroRetriever),
inferencer=dict(type=GenInferencer),
)
@ -73,14 +64,13 @@ hellaswag_eval_cfg = dict(
dict(
role='SYSTEM',
fallback_role='HUMAN',
prompt="You are a helpful assistant who evaluates the correctness and quality of models' outputs.")
prompt="You are a helpful assistant who evaluates the correctness and quality of models' outputs.",
)
],
round=[
dict(
role='HUMAN',
prompt=GRADER_TEMPLATE
),
]),
dict(role='HUMAN', prompt=GRADER_TEMPLATE),
],
),
),
dataset_cfg=dict(
type=HellaswagDatasetwithICE,
@ -90,8 +80,6 @@ hellaswag_eval_cfg = dict(
judge_cfg=dict(),
dict_postprocessor=dict(type=generic_llmjudge_postprocess),
),
pred_role='BOT',
)
hellaswag_datasets = [

30
opencompass/configs/datasets/mmlu/mmlu_llm_judge_gen.py
View File

@ -15,10 +15,8 @@ with read_base():
# Please download the dataset from https://people.eecs.berkeley.edu/~hendrycks/data.tar
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 ABCD. Think step by step before answering.
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 ABCD.
{input}
A) {A}
B) {B}
C) {C}
@ -33,16 +31,13 @@ GRADER_TEMPLATE = """
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 A) {A}\n B) {B}\n C) {C}\n D) {D}\n<Original Question End>\n\n
<Gold Target Begin>: \n{answer}\n<Gold Target End>\n\n
<Original Question Begin>: {input}\n A) {A}\n B) {B}\n C) {C}\n D) {D}\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()
@ -50,7 +45,8 @@ GRADER_TEMPLATE = """
mmlu_reader_cfg = dict(
input_columns=['input', 'A', 'B', 'C', 'D'],
output_column='target',
train_split='dev')
train_split='dev',
)
mmlu_datasets = []
for name in mmlu_all_sets:
@ -77,14 +73,13 @@ for name in mmlu_all_sets:
dict(
role='SYSTEM',
fallback_role='HUMAN',
prompt="You are a helpful assistant who evaluates the correctness and quality of models' outputs.")
prompt="You are a helpful assistant who evaluates the correctness and quality of models' outputs.",
)
],
round=[
dict(
role='HUMAN',
prompt=GRADER_TEMPLATE
),
]),
dict(role='HUMAN', prompt=GRADER_TEMPLATE),
],
),
),
dataset_cfg=dict(
type=MMLUDataset,
@ -97,7 +92,6 @@ for name in mmlu_all_sets:
),
pred_role='BOT',
)
mmlu_datasets.append(
dict(
abbr=f'lukaemon_mmlu_{name}',
@ -107,4 +101,6 @@ for name in mmlu_all_sets:
reader_cfg=mmlu_reader_cfg,
infer_cfg=mmlu_infer_cfg,
eval_cfg=mmlu_eval_cfg,
))
mode='singlescore',
)
)

6
opencompass/configs/datasets/mmlu_pro/mmlu_pro_llm_judge_gen.py
View File

@ -10,13 +10,10 @@ with read_base():
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 = """
@ -27,14 +24,11 @@ GRADER_TEMPLATE = """
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

268
opencompass/configs/datasets/musr/musr_llm_judge_gen.py
View File

@ -1,8 +1,6 @@
from opencompass.datasets import MusrDataset, MusrEvaluator
from opencompass.openicl import PromptTemplate, ZeroRetriever, GenInferencer
from opencompass.datasets import generic_llmjudge_postprocess
from opencompass.datasets import MusrDataset, generic_llmjudge_postprocess
from opencompass.evaluator import GenericLLMEvaluator
from opencompass.openicl import PromptTemplate, ZeroRetriever, GenInferencer
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.
@ -12,217 +10,80 @@ GRADER_TEMPLATE = """
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.
5. If the prediction is given with \\boxed{}, please ignore the \\boxed{} and only judge whether the candidate's answer is consistent with the standard answer.
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
<Original Question Begin>: {system_prompt}\n{prompt}\n<Original Question End>\n\n
<Gold Target Begin>: \n{gold_answer}\n<Gold Target End>\n\n
<Predicted Answer Begin>: \n{prediction}\n<Predicted End>\n\n
Judging the correctness of candidates' answers:
""".strip()
# Common configuration components
reader_cfg = dict(
input_columns=[
'context',
'question_text',
'question',
'answer',
'choices',
'choices_str',
'intermediate_trees',
'intermediate_data',
'prompt',
'system_prompt',
'gold_answer',
'scidx',
'self_consistency_n',
'ablation_name',
],
output_column='gold_answer',
)
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='{prompt}'),
],
),
),
retriever=dict(type=ZeroRetriever),
inferencer=dict(type=GenInferencer),
)
# Dataset configurations
DATASET_CONFIGS = {
'murder_mysteries': {
'abbr': 'musr_murder_mysteries',
'name': 'murder_mysteries',
'path': 'opencompass/musr',
'reader_cfg': dict(
input_columns=['context', 'question_text', 'question', 'answer', 'choices', 'choices_str', 'intermediate_trees', 'intermediate_data', 'prompt', 'system_prompt', 'gold_answer', 'scidx', 'self_consistency_n', 'ablation_name'],
output_column='gold_answer',
),
'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='{prompt}'
),
]
),
),
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=MusrDataset,
path='opencompass/musr',
name='murder_mysteries',
reader_cfg=dict(
input_columns=['context', 'question_text', 'question', 'answer', 'choices', 'choices_str', 'intermediate_trees', 'intermediate_data', 'prompt', 'system_prompt', 'gold_answer', 'scidx', 'self_consistency_n', 'ablation_name'],
output_column='gold_answer',
),
),
judge_cfg=dict(),
dict_postprocessor=dict(type=generic_llmjudge_postprocess),
),
pred_role='BOT',
)
},
'object_placements': {
'abbr': 'musr_object_placements',
'name': 'object_placements',
'path': 'opencompass/musr',
'reader_cfg': dict(
input_columns=['context', 'question_text', 'question', 'answer', 'choices', 'choices_str', 'intermediate_trees', 'intermediate_data', 'prompt', 'system_prompt', 'gold_answer', 'scidx', 'self_consistency_n', 'ablation_name'],
output_column='gold_answer',
),
'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='{prompt}'
),
]
),
),
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=MusrDataset,
path='opencompass/musr',
name='object_placements',
reader_cfg=dict(
input_columns=['context', 'question_text', 'question', 'answer', 'choices', 'choices_str', 'intermediate_trees', 'intermediate_data', 'prompt', 'system_prompt', 'gold_answer', 'scidx', 'self_consistency_n', 'ablation_name'],
output_column='gold_answer',
),
),
judge_cfg=dict(),
dict_postprocessor=dict(type=generic_llmjudge_postprocess),
),
pred_role='BOT',
)
},
'team_allocation': {
'abbr': 'musr_team_allocation',
'name': 'team_allocation',
'path': 'opencompass/musr',
'reader_cfg': dict(
input_columns=['context', 'question_text', 'question', 'answer', 'choices', 'choices_str', 'intermediate_trees', 'intermediate_data', 'prompt', 'system_prompt', 'gold_answer', 'scidx', 'self_consistency_n', 'ablation_name'],
output_column='gold_answer',
),
'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='{prompt}'
),
]
),
),
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=MusrDataset,
path='opencompass/musr',
name='team_allocation',
reader_cfg=dict(
input_columns=['context', 'question_text', 'question', 'answer', 'choices', 'choices_str', 'intermediate_trees', 'intermediate_data', 'prompt', 'system_prompt', 'gold_answer', 'scidx', 'self_consistency_n', 'ablation_name'],
output_column='gold_answer',
),
),
judge_cfg=dict(),
dict_postprocessor=dict(type=generic_llmjudge_postprocess),
),
pred_role='BOT',
)
},
}
# Create dataset configurations
musr_datasets = []
for config in DATASET_CONFIGS.values():
@ -231,8 +92,35 @@ for config in DATASET_CONFIGS.values():
type=MusrDataset,
path=config['path'],
name=config['name'],
reader_cfg=config['reader_cfg'],
infer_cfg=config['infer_cfg'],
eval_cfg=config['eval_cfg'],
reader_cfg=reader_cfg,
infer_cfg=infer_cfg,
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=MusrDataset,
path=config['path'],
name=config['name'],
reader_cfg=reader_cfg,
),
judge_cfg=dict(),
dict_postprocessor=dict(type=generic_llmjudge_postprocess),
),
),
)
musr_datasets.append(dataset)

32
opencompass/summarizers/default.py
View File

@ -186,17 +186,29 @@ class DefaultSummarizer:
eval_modes.append(dataset_eval_mode.get(dataset_abbr, 'unknown'))
else:
group_metrics = list(functools.reduce(lambda a, b: a & b, [set(dataset_metrics[dataset_abbr]) for dataset_abbr in sg['subsets']]))
if need_smart_metric and len(group_metrics) > 1:
for metric in group_metrics:
for dataset_abbr in sg['subsets']:
scores.setdefault(metric, {})[dataset_abbr + '@' + metric] = parsed_results[model_abbr][dataset_abbr][metric]
eval_modes.append(dataset_eval_mode.get(sg['subsets'][0], 'unknown'))
else:
group_metrics = [default_metric]
group_metrics.append(default_metric)
for metric in group_metrics:
for dataset_abbr in sg['subsets']:
metric = dataset_metrics[dataset_abbr][0]
scores.setdefault(default_metric, {})[dataset_abbr + '@' + metric] = parsed_results[model_abbr][dataset_abbr][metric]
eval_modes.append(dataset_eval_mode.get(dataset_abbr, 'unknown'))
if metric == default_metric:
metric_default = dataset_metrics[dataset_abbr][0]
scores.setdefault(default_metric, {})[dataset_abbr + '@' + metric_default] = \
parsed_results[model_abbr][dataset_abbr][metric_default]
eval_modes.append(dataset_eval_mode.get(dataset_abbr, 'unknown'))
else:
scores.setdefault(metric, {})[dataset_abbr + '@' + metric] = \
parsed_results[model_abbr][dataset_abbr][metric]
eval_modes.append(dataset_eval_mode.get(sg['subsets'][0], 'unknown'))
# if need_smart_metric and len(group_metrics) > 1:
# for metric in group_metrics:
# for dataset_abbr in sg['subsets']:
# scores.setdefault(metric, {})[dataset_abbr + '@' + metric] = parsed_results[model_abbr][dataset_abbr][metric]
# eval_modes.append(dataset_eval_mode.get(sg['subsets'][0], 'unknown'))
# else:
# group_metrics = [default_metric]
# for dataset_abbr in sg['subsets']:
# metric = dataset_metrics[dataset_abbr][0]
# scores.setdefault(default_metric, {})[dataset_abbr + '@' + metric] = parsed_results[model_abbr][dataset_abbr][metric]
# eval_modes.append(dataset_eval_mode.get(dataset_abbr, 'unknown'))
result = {}
for metric in scores: