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

Merge branch 'main' of github.com:open-compass/opencompass into tmp_olmpbench

Browse Source
This commit is contained in:
liushz 2025-03-25 02:48:09 +00:00
commit 21a92d4c14
29 changed files with 1618 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
.github/workflows/pr-stage-check.yml vendored
View File

@ -20,7 +20,7 @@ jobs:
matrix:
python-version: ['3.10']
include:
- torch: 2.0.0
- torch: 2.5.1
steps:
- uses: actions/checkout@v3
- name: Set up Python ${{ matrix.python-version }}
@ -30,7 +30,7 @@ jobs:
- name: Upgrade pip
run: python -m pip install --upgrade pip
- name: Install PyTorch
run: pip install torch==${{matrix.torch}}+cpu -f https://download.pytorch.org/whl/cpu/torch_stable.html
run: pip install torch==${{matrix.torch}} -f https://download.pytorch.org/whl/cpu/torch_stable.html
- name: Install system dependencies
run: |
sudo sed -i '$ a deb http://th.archive.ubuntu.com/ubuntu jammy main' /etc/apt/sources.list
@ -106,7 +106,7 @@ jobs:
- name: Upgrade pip
run: python -m pip install pip --upgrade
- name: Install PyTorch
run: pip install torch==2.0.0+${{matrix.platform}} -f https://download.pytorch.org/whl/${{matrix.platform}}/torch_stable.html
run: pip install torch==2.5.1 -f https://download.pytorch.org/whl/cpu/torch_stable.html
- name: Install opencompass dependencies
run: |
pip install -r requirements.txt

5
dataset-index.yml
View File

@ -529,6 +529,11 @@
category: Understanding
paper: https://proceedings.neurips.cc/paper_files/paper/2019/file/4496bf24afe7fab6f046bf4923da8de6-Paper.pdf
configpath: opencompass/configs/datasets/SuperGLUE_MultiRC
- multipl_e:
name: MultiPL-E
category: Code
paper: https://arxiv.org/pdf/2210.14868
configpath: opencompass/configs/datasets/multipl_e
- narrativeqa:
name: NarrativeQA
category: Understanding

56
opencompass/configs/datasets/ARC_Prize_Public_Evaluation/arc_prize_public_evaluation_gen_fedd04.py Normal file
View File

@ -0,0 +1,56 @@
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.arc_prize_public_evaluation import ARCPrizeDataset, ARCPrizeEvaluator
# The system_prompt defines the initial instructions for the model,
# setting the context for solving ARC tasks.
system_prompt = '''You are a puzzle solving wizard. You are given a puzzle from the abstraction and reasoning corpus developed by Francois Chollet.'''
# User message template is a template for creating user prompts. It includes placeholders for training data and test input data,
# guiding the model to learn the rule and apply it to solve the given puzzle.
user_message_template = '''Here are the example input and output pairs from which you should learn the underlying rule to later predict the output for the given test input:
----------------------------------------
{training_data}
----------------------------------------
Now, solve the following puzzle based on its input grid by applying the rules you have learned from the training data.:
----------------------------------------
[{{'input': {input_test_data}, 'output': [[]]}}]
----------------------------------------
What is the output grid? Only provide the output grid in the form as in the example input and output pairs. Do not provide any additional information:'''
arc_prize_public_evaluation_reader_cfg = dict(
input_columns=['training_data', 'input_test_data'],
output_column='output_test_data'
)
arc_prize_public_evaluation_infer_cfg = dict(
prompt_template=dict(
type=PromptTemplate,
template=dict(
round=[
dict(role='SYSTEM',fallback_role='HUMAN', prompt=system_prompt),
dict(role='HUMAN', prompt=user_message_template),
],
)
),
retriever=dict(type=ZeroRetriever),
inferencer=dict(type=GenInferencer)
)
arc_prize_public_evaluation_eval_cfg = dict(
evaluator=dict(type=ARCPrizeEvaluator)
)
arc_prize_public_evaluation_datasets = [
dict(
abbr='ARC_Prize_Public_Evaluation',
type=ARCPrizeDataset,
path='opencompass/arc_prize_public_evaluation',
reader_cfg=arc_prize_public_evaluation_reader_cfg,
infer_cfg=arc_prize_public_evaluation_infer_cfg,
eval_cfg=arc_prize_public_evaluation_eval_cfg
)
]

45
opencompass/configs/datasets/GaokaoBench/GaokaoBench_no_subjective_gen_d16acb.py Normal file
View File

@ -0,0 +1,45 @@
import os
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 GaokaoBenchDataset
from mmengine.config import read_base
with read_base():
from .GaokaoBench_prompts import MCQ_prompts, FBQ_prompts
GaokaoBench_datasets = []
for folder, prompts in [
('Multiple-choice_Questions', MCQ_prompts),
('Fill-in-the-blank_Questions', FBQ_prompts),
]:
for p in prompts:
reader_cfg = {
'input_columns': ['question'],
'output_column': 'answer',
}
infer_cfg = {
'ice_template': {
'type': PromptTemplate,
'template': {'round': [{'role': 'HUMAN', 'prompt': p['prefix_prompt'] + '{question}'}]},
'ice_token': '</E>',
},
'retriever': {'type': ZeroRetriever},
'inferencer': {'type': GenInferencer},
}
eval_cfg = {
'evaluator': {'type': 'GaokaoBenchEvaluator' + '_' + p['type']},
'pred_role': 'BOT',
}
_base_path = 'opencompass/GAOKAO-BENCH'
dataset = {
'type': GaokaoBenchDataset,
'abbr': 'GaokaoBench_' + p['keyword'],
'path': _base_path,
'filename': '/' + folder + '/' + p['keyword'] + '.json',
'name': p['keyword'],
'reader_cfg': reader_cfg,
'infer_cfg': infer_cfg,
'eval_cfg': eval_cfg,
}
GaokaoBench_datasets.append(dataset)

81
opencompass/configs/datasets/MathBench/mathbench_2024_gen_4b8f28.py Normal file
View File

@ -0,0 +1,81 @@
from mmengine.config import read_base
from copy import deepcopy
from opencompass.openicl.icl_prompt_template import PromptTemplate
from opencompass.openicl.icl_retriever import ZeroRetriever
from opencompass.openicl.icl_inferencer import GenInferencer, PPLInferencer
from opencompass.openicl.icl_evaluator import CircularEvaluator, AccEvaluator
from opencompass.datasets import MathBenchDataset, math_postprocess_v2
from opencompass.utils.text_postprocessors import first_option_postprocess
with read_base():
from .mathbench_prompt import zero_shot_prompts, few_shot_prompts, mathbench_sets
# Max for this dataset is 4
num_shot = 0
# Generate reasoning path or not, only for single choice
with_reasoning = True
# Use circular evaluation or not
with_circular_eval = True
# Use PPL mode in single choice test or not
use_ppl_single_choice = False
assert 0 <= num_shot <= 4
if num_shot == 0:
prompts = zero_shot_prompts
else:
prompts = {name: p[- 2 * num_shot - 2:] for name, p in few_shot_prompts.items()}
mathbench_datasets = []
for _split in mathbench_sets:
for _name in mathbench_sets[_split]:
if 'single_choice' in _name:
if with_reasoning:
template_round = prompts[_name + '_with_reasoning']
else:
template_round = prompts[_name]
else:
template_round = prompts[_name]
if 'single_choice' in _name:
pred_postprocessor = dict(type=first_option_postprocess, options='ABCD')
else:
pred_postprocessor = dict(type=math_postprocess_v2)
if 'single_choice' in _name and with_circular_eval:
evaluator = dict(type=CircularEvaluator)
else:
evaluator = dict(type=AccEvaluator)
# assemble the final config
mathbench_reader_cfg = dict(input_columns=['question'], output_column='answer')
if use_ppl_single_choice and 'single_choice' in _name and not with_reasoning:
template = {}
for answer in ['A', 'B', 'C', 'D']:
one_template_round = deepcopy(template_round)
one_template_round['round'][-1]['prompt'] = one_template_round['round'][-1]['prompt'].format(answer=answer)
template[answer] = dict(round=one_template_round)
mathbench_infer_cfg = dict(
prompt_template=dict(type=PromptTemplate, template=template),
retriever=dict(type=ZeroRetriever),
inferencer=dict(type=PPLInferencer),
)
else:
mathbench_infer_cfg = dict(
prompt_template=dict(type=PromptTemplate, template=dict(round=template_round)),
retriever=dict(type=ZeroRetriever),
inferencer=dict(type=GenInferencer),
)
mathbench_eval_cfg = dict(evaluator=evaluator, pred_postprocessor=pred_postprocessor)
mathbench_datasets.append(
dict(
abbr='mathbench-' + _split + '-' + _name,
type=MathBenchDataset,
path=f'data/mathbench_v1/{_split}',
name=_name,
with_circular=with_circular_eval,
reader_cfg=mathbench_reader_cfg,
infer_cfg=mathbench_infer_cfg,
eval_cfg=mathbench_eval_cfg,
)
)

189
opencompass/configs/datasets/bbh/bbh_llmjudge_gen_b5bdf1.py Normal file
View File

@ -0,0 +1,189 @@
# flake8: noqa
import os
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 BBHDataset
from opencompass.datasets.generic import generic_llmjudge_academic_postprocess
bbh_reader_cfg = dict(input_columns=['input'], output_column='target')
bbh_multiple_choice_sets = [
'temporal_sequences',
'disambiguation_qa',
'date_understanding',
'tracking_shuffled_objects_three_objects',
'penguins_in_a_table',
'geometric_shapes',
'snarks',
'ruin_names',
'tracking_shuffled_objects_seven_objects',
'tracking_shuffled_objects_five_objects',
'logical_deduction_three_objects',
'hyperbaton',
'logical_deduction_five_objects',
'logical_deduction_seven_objects',
'movie_recommendation',
'salient_translation_error_detection',
'reasoning_about_colored_objects',
]
bbh_free_form_sets = [
'multistep_arithmetic_two',
'navigate',
'dyck_languages',
'word_sorting',
'sports_understanding',
'boolean_expressions',
'object_counting',
'formal_fallacies',
'causal_judgement',
'web_of_lies',
]
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>: \n{input}\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()
bbh_sets = bbh_multiple_choice_sets + bbh_free_form_sets
# For zero shot inference in bbh
bbh_datasets = []
for _name in bbh_sets:
bbh_infer_cfg = dict(
prompt_template=dict(
type=PromptTemplate,
template=dict(round=[
dict(
role='HUMAN',
prompt=f"Question: {{input}}\n You must give your final answer by starting with 'So the answer is' "
)
])),
retriever=dict(type=ZeroRetriever),
inferencer=dict(type=GenInferencer))
bbh_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=BBHDataset,
name=_name,
path='opencompass/bbh',
reader_cfg=bbh_reader_cfg,
),
judge_cfg=dict(),
dict_postprocessor=dict(type=generic_llmjudge_academic_postprocess, metric_name='score'),
),
pred_role='BOT',
)
bbh_datasets.append(
dict(
type=BBHDataset,
path='opencompass/bbh',
name=_name,
abbr='bbh-' + _name,
reader_cfg=bbh_reader_cfg,
infer_cfg=bbh_infer_cfg.copy(),
eval_cfg=bbh_eval_cfg.copy())
)
# For original 3 shot inference in bbh
bbh_3_shot_datasets = []
for _name in bbh_sets:
with open(os.path.join(os.path.dirname(__file__), 'lib_prompt', f'{_name}.txt'), 'r') as f:
_hint = f.read()
bbh_infer_cfg = dict(
prompt_template=dict(
type=PromptTemplate,
template=dict(round=[
dict(
role='HUMAN',
prompt=f"Follow the given examples and answer the question.\n{_hint}\n\nQ: {{input}}\nA: Let's think step by step."
)
])),
retriever=dict(type=ZeroRetriever),
inferencer=dict(type=GenInferencer))
bbh_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=BBHDataset,
name=_name,
path='opencompass/bbh',
reader_cfg=bbh_reader_cfg,
),
judge_cfg=dict(),
dict_postprocessor=dict(type=generic_llmjudge_academic_postprocess, metric_name='score'),
),
pred_role='BOT',
)
bbh_3_shot_datasets.append(
dict(
type=BBHDataset,
path='opencompass/bbh',
name=_name,
abbr='bbh-' + _name,
reader_cfg=bbh_reader_cfg,
infer_cfg=bbh_infer_cfg.copy(),
eval_cfg=bbh_eval_cfg.copy()))

45
opencompass/configs/datasets/bigcodebench/bigcodebench_hard_complete_gen_2888d3.py Normal file
View File

@ -0,0 +1,45 @@
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 (BigCodeBenchDataset, BigCodeBenchEvaluator)
bigcodebench_hard_reader_cfg = dict(
input_columns=['complete_prompt'],
output_column='test',
)
bigcodebench_hard_infer_cfg = dict(prompt_template=dict(
type=PromptTemplate,
template=dict(
begin=[dict(role='system', fallback_role='HUMAN', prompt='')],
round=[
dict(role='HUMAN', prompt='{complete_prompt}'),
])),
retriever=dict(type=ZeroRetriever),
inferencer=dict(type=GenInferencer))
bigcodebench_hard_eval_cfg = dict(
evaluator=dict(
type=BigCodeBenchEvaluator,
release_version='v0.1.2',
eval_type='complete',
# remote_execute_api='https://bigcode-bigcodebench-evaluator.hf.space/',
remote_execute_api=
'https://opencompass-opencompass-bigcodebench-evaluator.hf.space', # noqa: E501
dataset_version='hard',
),
pred_role='BOT',
)
bigcodebench_hard_complete_datasets = [
dict(
abbr='bigcodebench_hard_complete',
type=BigCodeBenchDataset,
path='opencompass/bigcodebench',
reader_cfg=bigcodebench_hard_reader_cfg,
infer_cfg=bigcodebench_hard_infer_cfg,
eval_cfg=bigcodebench_hard_eval_cfg,
release_version='v0.1.2',
dataset_version='hard',
)
]

45
opencompass/configs/datasets/bigcodebench/bigcodebench_hard_instruct_gen_c3d5ad.py Normal file
View File

@ -0,0 +1,45 @@
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 (BigCodeBenchDataset, BigCodeBenchEvaluator)
bigcodebench_hard_reader_cfg = dict(
input_columns=['instruct_prompt'],
output_column='test',
)
bigcodebench_hard_infer_cfg = dict(prompt_template=dict(
type=PromptTemplate,
template=dict(
begin=[dict(role='system', fallback_role='HUMAN', prompt='')],
round=[
dict(role='HUMAN', prompt='{instruct_prompt}'),
])),
retriever=dict(type=ZeroRetriever),
inferencer=dict(type=GenInferencer))
bigcodebench_hard_eval_cfg = dict(
evaluator=dict(
type=BigCodeBenchEvaluator,
release_version='v0.1.2',
eval_type='instruct',
# remote_execute_api='https://bigcode-bigcodebench-evaluator.hf.space/',
remote_execute_api=
'https://opencompass-opencompass-bigcodebench-evaluator.hf.space', # noqa: E501
dataset_version='hard',
),
pred_role='BOT',
)
bigcodebench_hard_instruct_datasets = [
dict(
abbr='bigcodebench_hard_instruct',
type=BigCodeBenchDataset,
path='opencompass/bigcodebench',
reader_cfg=bigcodebench_hard_reader_cfg,
infer_cfg=bigcodebench_hard_infer_cfg,
eval_cfg=bigcodebench_hard_eval_cfg,
release_version='v0.1.2',
dataset_version='hard',
)
]

39
opencompass/configs/datasets/cmo_fib/cmo_fib_gen_2783e5.py Normal file
View File

@ -0,0 +1,39 @@
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 CMOFibDataset, MATHEvaluator, math_postprocess_v2
cmo_fib_reader_cfg = dict(
input_columns=['question'],
output_column='answer'
)
cmo_fib_infer_cfg = dict(
prompt_template=dict(
type=PromptTemplate,
template=dict(
round=[
dict(role='HUMAN', prompt='{question}\n请一步一步地推理，并将最终答案写入\\boxed{}.'),
],
)
),
retriever=dict(type=ZeroRetriever),
inferencer=dict(type=GenInferencer)
)
cmo_fib_eval_cfg = dict(
evaluator=dict(type=MATHEvaluator, version='v2'), pred_postprocessor=dict(type=math_postprocess_v2)
)
cmo_fib_datasets = [
dict(
abbr='cmo_fib',
type=CMOFibDataset,
path='opencompass/cmo_fib',
reader_cfg=cmo_fib_reader_cfg,
infer_cfg=cmo_fib_infer_cfg,
eval_cfg=cmo_fib_eval_cfg
)
]

37
opencompass/configs/datasets/gsm8k/gsm8k_0shot_v2_gen_17d799.py Normal file
View File

@ -0,0 +1,37 @@
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 GSM8KDataset, gsm8k_postprocess, gsm8k_dataset_postprocess, Gsm8kEvaluator
from opencompass.datasets import MATHEvaluator, math_postprocess_v2
gsm8k_reader_cfg = dict(input_columns=['question'], output_column='answer')
gsm8k_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),
)
gsm8k_eval_cfg = dict(
evaluator=dict(type=MATHEvaluator, version='v2'),
pred_postprocessor=dict(type=math_postprocess_v2),
dataset_postprocessor=dict(type=gsm8k_dataset_postprocess),
)
gsm8k_datasets = [
dict(
abbr='gsm8k',
type=GSM8KDataset,
path='opencompass/gsm8k',
reader_cfg=gsm8k_reader_cfg,
infer_cfg=gsm8k_infer_cfg,
eval_cfg=gsm8k_eval_cfg,
)
]

117
opencompass/configs/datasets/korbench/korbench_llmjudge_gen_17854d.py Normal file
View File

@ -0,0 +1,117 @@
from opencompass.datasets.korbench.korbench import korbenchDataset, korbenchEvaluator
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
from opencompass.datasets import generic_llmjudge_postprocess
categories = ['cipher', 'counterfactual', 'logic', 'operation', 'puzzle']
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>: \n{prompt}\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()
korbench_0shot_single_datasets = []
for category in categories:
# Prompt template
prompt_template = dict(
type=PromptTemplate,
template=dict(
begin=[
dict(
role='HUMAN',
prompt=''
)
],
round=[
dict(
role='HUMAN',
prompt='{prompt}' # f-string
)
]
)
)
# Reader configuration
reader_cfg = dict(
input_columns=['prompt'],
output_column='answer',
)
# Inference configuration
infer_cfg = dict(
prompt_template=prompt_template,
retriever=dict(type=ZeroRetriever),
inferencer=dict(type=GenInferencer, max_out_len=1024),
)
# 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=korbenchDataset,
path='opencompass/korbench',
prompt_mode='0_shot',
category=category,
reader_cfg=reader_cfg,
),
judge_cfg=dict(),
dict_postprocessor=dict(type=generic_llmjudge_postprocess),
),
pred_role='BOT',
)
# Dataset
korbench_dataset = dict(
type=korbenchDataset,
abbr=f'korbench_{category}',
path='opencompass/korbench',
prompt_mode='0_shot',
category=category,
reader_cfg=reader_cfg,
infer_cfg=infer_cfg,
eval_cfg=eval_cfg,
mode='singlescore',
)
korbench_0shot_single_datasets.append(korbench_dataset)

117
opencompass/configs/datasets/korbench/korbench_llmjudge_gen_56cf43.py Normal file
View File

@ -0,0 +1,117 @@
from opencompass.datasets.korbench.korbench import korbenchDataset, korbenchEvaluator
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
from opencompass.datasets import generic_llmjudge_postprocess
categories = ['cipher', 'counterfactual', 'logic', 'operation', 'puzzle']
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>: \n{prompt}\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()
korbench_0shot_single_datasets = []
for category in categories:
# Prompt template
prompt_template = dict(
type=PromptTemplate,
template=dict(
begin=[
dict(
role='HUMAN',
prompt=''
)
],
round=[
dict(
role='HUMAN',
prompt='{prompt}' # f-string
)
]
)
)
# Reader configuration
reader_cfg = dict(
input_columns=['prompt'],
output_column='answer',
)
# Inference configuration
infer_cfg = dict(
prompt_template=prompt_template,
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=korbenchDataset,
path='opencompass/korbench',
prompt_mode='0_shot',
category=category,
reader_cfg=reader_cfg,
),
judge_cfg=dict(),
dict_postprocessor=dict(type=generic_llmjudge_postprocess),
),
pred_role='BOT',
)
# Dataset
korbench_dataset = dict(
type=korbenchDataset,
abbr=f'korbench_{category}',
path='opencompass/korbench',
prompt_mode='0_shot',
category=category,
reader_cfg=reader_cfg,
infer_cfg=infer_cfg,
eval_cfg=eval_cfg,
mode='singlescore',
)
korbench_0shot_single_datasets.append(korbench_dataset)

96
opencompass/configs/datasets/math/math_500_llmjudge_gen_6ff468.py Normal file
View File

@ -0,0 +1,96 @@
from opencompass.openicl.icl_prompt_template import PromptTemplate
from opencompass.openicl.icl_retriever import ZeroRetriever
from opencompass.openicl.icl_inferencer import GenInferencer
from opencompass.evaluator import GenericLLMEvaluator
from opencompass.datasets import generic_llmjudge_postprocess
from opencompass.datasets import MATHDataset
# ----------------------------- Detailed Config -----------------------------
math_reader_cfg = dict(input_columns=['problem'], output_column='solution')
math_infer_cfg = dict(
prompt_template=dict(
type=PromptTemplate,
template=dict(
round=[
dict(role='HUMAN', prompt='{problem}\nRemember to put your final answer within \\boxed{}.'),
]
),
),
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>: \n{problem}\n<Original Question End>\n\n
<Gold Target Begin>: \n{solution}\n<Gold Target End>\n\n
<Predicted Answer Begin>: \n{prediction}\n<Predicted End>\n\n
Judging the correctness of candidates' answers:
""".strip()
# Evaluation configuration
math_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=MATHDataset,
path='opencompass/math',
file_name = 'test_prm800k_500.json',
reader_cfg=math_reader_cfg,
),
judge_cfg=dict(),
dict_postprocessor=dict(type=generic_llmjudge_postprocess),
),
pred_role='BOT',
)
math_datasets = [
dict(
type=MATHDataset,
abbr='math_prm800k_500-llmjudge',
path='opencompass/math',
file_name = 'test_prm800k_500.json',
reader_cfg=math_reader_cfg,
infer_cfg=math_infer_cfg,
eval_cfg=math_eval_cfg,
mode='singlescore',
)
]

56
opencompass/configs/datasets/multipl_e/multiple_top_ten_gen.py Normal file
View File

@ -0,0 +1,56 @@
# Select the 10 most popular programming languages from MultiPL-E to compose the test set.
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 MultiplEDataset, MultiplEEvaluator
_TOP_TEN_LANGUAGE_ = ['cpp', 'cs', 'go', 'java', 'rb', 'js', 'php', 'r', 'rs', 'sh']
multiple_reader_cfg = dict(input_columns=['language', 'prompt'], output_column='tests')
multiple_infer_cfg = dict(
prompt_template=dict(type=PromptTemplate, template='Based on the provided {language} code snippet, complete the subsequent content. The initial part of the completed code must match the provided code snippet exactly:\n{prompt}'),
retriever=dict(type=ZeroRetriever),
inferencer=dict(type=GenInferencer),
)
multiple_eval_cfg = {
lang: dict(
evaluator=dict(
type=MultiplEEvaluator,
language=lang,
ip_address='https://opencompass-multiple-evaluator.hf.space',
),
pred_role='BOT',
) for lang in _TOP_TEN_LANGUAGE_
}
multiple_datasets = [
dict(
type=MultiplEDataset,
abbr=f'humaneval-multiple-{lang}',
language=lang,
num_repeats=1,
path='opencompass/multipl_e',
tag='humaneval',
reader_cfg=multiple_reader_cfg,
infer_cfg=multiple_infer_cfg,
eval_cfg=multiple_eval_cfg[lang],
) for lang in _TOP_TEN_LANGUAGE_
]
multiple_datasets += [
dict(
type=MultiplEDataset,
abbr=f'mbpp-multiple-{lang}',
language=lang,
num_repeats=1,
path='opencompass/multipl_e',
tag='mbpp',
reader_cfg=multiple_reader_cfg,
infer_cfg=multiple_infer_cfg,
eval_cfg=multiple_eval_cfg[lang],
) for lang in _TOP_TEN_LANGUAGE_
]

2
opencompass/configs/datasets/nq/nq_open_1shot_gen_2e45e5.py
View File

@ -20,7 +20,7 @@ for k in [1]:
)
),
retriever=dict(type=ZeroRetriever),
inferencer=dict(type=GenInferencer, max_out_len=50)
inferencer=dict(type=GenInferencer)
)
else:
nq_infer_cfg = dict(

29
opencompass/configs/datasets/scicode/scicode_gen_62c139.py Normal file
View File

@ -0,0 +1,29 @@
from opencompass.openicl.icl_prompt_template import PromptTemplate
from opencompass.openicl.icl_retriever import ZeroRetriever
from opencompass.openicl.icl_inferencer import ChatInferencer
from opencompass.datasets import SciCodeDataset, SciCodeEvaluator
SciCode_reader_cfg = dict(input_columns=['prompt'], output_column=None)
SciCode_infer_cfg = dict(
ice_template=dict(
type=PromptTemplate,
template='',
),
retriever=dict(type=ZeroRetriever),
inferencer=dict(type=ChatInferencer, infer_mode='every'))
SciCode_eval_cfg = dict(evaluator=dict(type=SciCodeEvaluator, dataset_path='./data/scicode', with_bg=False))
SciCode_datasets = [
dict(
abbr='SciCode',
type=SciCodeDataset,
path='./data/scicode',
with_bg=False,
reader_cfg=SciCode_reader_cfg,
infer_cfg=SciCode_infer_cfg,
eval_cfg=SciCode_eval_cfg)
]

4
opencompass/configs/datasets/supergpqa/supergpqa_llmjudge_gen_12b8bc.py
View File

@ -1,5 +1,5 @@
from opencompass.datasets.supergpqa.supergpqa import (
SuperGPQADataset,
SuperGPQADataset, supergpqa_llmjudge_postprocess
)
from opencompass.openicl.icl_inferencer import GenInferencer
from opencompass.openicl.icl_prompt_template import PromptTemplate
@ -87,7 +87,7 @@ eval_cfg = dict(
reader_cfg=reader_cfg,
),
judge_cfg=dict(),
dict_postprocessor=dict(type=generic_llmjudge_postprocess),
dict_postprocessor=dict(type=supergpqa_llmjudge_postprocess),
),
)
supergpqa_dataset = dict(

62
opencompass/configs/datasets/triviaqa/triviaqa_wiki_1shot_gen_c87d61.py Normal file
View File

@ -0,0 +1,62 @@
from opencompass.openicl.icl_prompt_template import PromptTemplate
from opencompass.openicl.icl_retriever import ZeroRetriever, FixKRetriever
from opencompass.openicl.icl_inferencer import GenInferencer
from opencompass.datasets import TriviaQADatasetV2, TriviaQAEvaluator
triviaqa_datasets = []
for k in [1]:
triviaqa_reader_cfg = dict(
input_columns=['question'], output_column='answer', train_split='train', test_split='validation')
if k == 0:
triviaqa_infer_cfg = dict(
prompt_template=dict(
type=PromptTemplate,
template=dict(
round=[
dict(role='HUMAN', prompt='Q: {question}'),
dict(role='BOT', prompt='A:'),
]
)
),
retriever=dict(type=ZeroRetriever),
inferencer=dict(type=GenInferencer)
)
else:
triviaqa_infer_cfg = dict(
ice_template=dict(
type=PromptTemplate,
template=dict(
round=[
dict(role='HUMAN', prompt='Q: {question}'),
dict(role='BOT', prompt='A: {answer}.\n'),
]
),
),
prompt_template=dict(
type=PromptTemplate,
template=dict(
begin='</E>',
round=[
dict(role='HUMAN', prompt='Q: {question}'),
dict(role='BOT', prompt='A:'),
]
),
ice_token='</E>',
),
retriever=dict(type=FixKRetriever, fix_id_list=list(range(k))),
inferencer=dict(type=GenInferencer),
)
triviaqa_eval_cfg = dict(evaluator=dict(type=TriviaQAEvaluator), pred_role='BOT')
triviaqa_datasets.append(
dict(
type=TriviaQADatasetV2,
abbr=f'triviaqa_wiki_{k}shot',
path='opencompass/trivia_qa',
reader_cfg=triviaqa_reader_cfg,
infer_cfg=triviaqa_infer_cfg,
eval_cfg=triviaqa_eval_cfg)
)

12
opencompass/configs/models/phi/hf_phi_4.py Normal file
View File

@ -0,0 +1,12 @@
from opencompass.models import HuggingFacewithChatTemplate
models = [
dict(
type=HuggingFacewithChatTemplate,
abbr='phi-4',
path='microsoft/phi-4',
max_out_len=1024,
batch_size=8,
run_cfg=dict(num_gpus=2),
)
]

17
opencompass/configs/models/qwq/lmdeploy_qwq_32b.py Normal file
View File

@ -0,0 +1,17 @@
from opencompass.models import TurboMindModelwithChatTemplate
from opencompass.utils.text_postprocessors import extract_non_reasoning_content
models = [
dict(
type=TurboMindModelwithChatTemplate,
abbr='QwQ-32B',
path='Qwen/QwQ-32B',
engine_config=dict(session_len=32768, max_batch_size=16, tp=2),
gen_config=dict(top_k=1, temperature=1e-6, top_p=0.9, max_new_tokens=8192),
max_seq_len=32768,
max_out_len=8192,
batch_size=16,
run_cfg=dict(num_gpus=2),
pred_postprocessor=dict(type=extract_non_reasoning_content)
)
]

1
opencompass/datasets/__init__.py
View File

@ -98,6 +98,7 @@ from .mmlu_cf import * # noqa: F401, F403
from .mmlu_pro import * # noqa: F401, F403
from .MMLUArabic import * # noqa: F401, F403
from .mmmlu import * # noqa: F401, F403
from .multipl_e import * # noqa: F401, F403
from .multirc import * # noqa: F401, F403
from .musr import * # noqa: F401, F403
from .narrativeqa import * # noqa: F401, F403

27
opencompass/datasets/custom.py
View File

@ -183,6 +183,33 @@ class CustomDataset(BaseDataset):
return Dataset.from_list(data)
@LOAD_DATASET.register_module()
class CodeCustomDataset(BaseDataset):
@staticmethod
def load(path, file_name=None, local_mode=False, num_repeats=1, **kwargs):
path = get_data_path(path, local_mode=local_mode)
if file_name is not None:
path = os.path.join(path, file_name)
data = []
if path.endswith('.jsonl'):
with open(path, 'r', encoding='utf-8') as f:
for line in f:
data.extend(
[json.loads(line.strip()) for _ in range(num_repeats)])
elif path.endswith('.csv'):
with open(path, 'r', encoding='utf-8-sig') as f:
reader = csv.reader(f)
header = next(reader)
for row in reader:
data.extend(
[dict(zip(header, row)) for _ in range(num_repeats)])
else:
raise ValueError(f'Unsupported file format: {path}')
return Dataset.from_list(data)
class CircularCustomDataset(CustomDataset, metaclass=CircularDatasetMeta):
dataset_class = CustomDataset

103
opencompass/datasets/multipl_e.py Normal file
View File

@ -0,0 +1,103 @@
import json
import os.path as osp
from datasets import Dataset
from opencompass.openicl.icl_evaluator.code_evaluator import CodeEvaluator
from opencompass.registry import LOAD_DATASET
from opencompass.utils import get_data_path
from .base import BaseDataset
# currently supporting languages
_HUMANEVAL_LANGUAGE_ = [
'adb', 'clj', 'cpp', 'cs', 'd', 'dart', 'elixir', 'go', 'hs', 'java', 'jl',
'js', 'lua', 'ml', 'php', 'pl', 'py', 'r', 'rb', 'rkt', 'rs', 'scala',
'sh', 'swift', 'ts'
]
_MBPP_LANGUAGE_ = [
'adb', 'clj', 'cpp', 'cs', 'd', 'elixir', 'go', 'hs', 'java', 'jl', 'js',
'lua', 'ml', 'php', 'pl', 'py', 'r', 'rb', 'rkt', 'rs', 'scala', 'sh',
'swift', 'ts'
]
@LOAD_DATASET.register_module()
class MultiplEDataset(BaseDataset):
@staticmethod
def load(path: str,
language: str,
num_repeats: int = 1,
tag: str = 'humaneval',
local_mode: bool = False):
"""Load dataset for pass k mode.
Args:
path(str): The path to the dataset.
language(str): The language of the dataset.
num_repeats(int): Number of repetition for this dataset to get.
tag(str): The tag of the dataset.
local_mode(bool): Whether to load the dataset in local mode.
Returns:
Dataset: A PyTorch dataset.
"""
path = get_data_path(path, local_mode=local_mode)
assert tag in ['humaneval',
'mbpp'], 'tag must be in ["humaneval", "mbpp"]'
if tag == 'humaneval':
assert language in _HUMANEVAL_LANGUAGE_, (
f'language must be in {_HUMANEVAL_LANGUAGE_}')
else:
assert language in _MBPP_LANGUAGE_, (
f'language must be in {_MBPP_LANGUAGE_}')
file_path = osp.join(path, f'{tag}-{language}.jsonl')
dataset = []
with open(file_path, 'r', encoding='utf-8') as f:
for line in f:
dataset.extend(
[json.loads(line.strip()) for _ in range(num_repeats)])
return Dataset.from_list(dataset)
class MultiplEEvaluator(CodeEvaluator):
def _stop_at_stop_token(self, decoded_string, stop_tokens):
"""Produces the prefix of decoded_string that ends at the first
occurrence of a stop_token.
WARNING: the decoded_string *must not* include the prompt,
which may have stop tokens itself.
Args:
decoded_string: A string generated by the model.
stop_tokens: A list of strings, where each string is a stop token.
Returns:
The decoded_string, truncated at the first occurrence of a stop
token.
"""
min_stop_index = len(decoded_string)
for stop_token in stop_tokens:
stop_index = decoded_string.find(stop_token)
if stop_index != -1 and stop_index < min_stop_index:
min_stop_index = stop_index
return decoded_string[:min_stop_index]
def _process_completions(self, test_case, completions):
"""Process completions with a test case.
Args:
test_case: A test case.
completions: A list of completions.
Returns:
A list of processed completions.
"""
processed_completions = []
for comp in completions:
comp = self._extract_code(comp)
post_comp = self._remove_prefix(test_case['prompt'], comp)
post_comp = self._stop_at_stop_token(post_comp,
test_case['stop_tokens'])
processed_completions.append(post_comp)
return processed_completions

132
opencompass/datasets/supergpqa/supergpqa.py
View File

@ -1,4 +1,5 @@
import os
import re
from datasets import Dataset, load_dataset
@ -7,6 +8,7 @@ from opencompass.datasets.supergpqa.supergpqa_eval import (
from opencompass.datasets.supergpqa.supergpqa_utils import load_yaml
from opencompass.openicl.icl_evaluator import BaseEvaluator
from opencompass.registry import ICL_EVALUATORS, LOAD_DATASET
from opencompass.utils import get_logger
from ..base import BaseDataset
@ -180,3 +182,133 @@ class SuperGPQAEvaluator(BaseEvaluator):
'details':
details,
}
def _generic_llmjudge_postprocess(judgement: str):
match = re.search(r'(A|B)', judgement)
grade_letter = (match.group(0) if match else 'B'
) # Default to "INCORRECT" if no match
return grade_letter
def supergpqa_llmjudge_postprocess(
output: dict,
output_path: str,
dataset: Dataset,
) -> dict:
# Get the original dataset
original_dataset = dataset.reader.dataset['test']
judged_answers = []
original_responses = []
references = []
details = []
# Initialize statistics dictionaries
stats = {'discipline': {}, 'field': {}, 'subfield': {}}
total_correct = 0
total_count = 0
# Process each sample
for k, v in output.items():
idx = int(k) # Convert key to integer for indexing
original_responses.append(v['prediction'])
processed_judge = _generic_llmjudge_postprocess(v['prediction'])
# Get category information from the dataset
sample = original_dataset[idx]
discipline = sample.get('discipline', 'unknown')
field = sample.get('field', 'unknown')
subfield = sample.get('subfield', 'unknown')
# Initialize category stats if not exists
for level, key in [
('discipline', discipline),
('field', f'{discipline}/{field}'),
('subfield', f'{discipline}/{field}/{subfield}'),
]:
if key not in stats[level]:
stats[level][key] = {'correct': 0, 'total': 0}
# Record the judgment
if processed_judge is not None:
judged_answers.append(processed_judge)
try:
gold = v['gold']
references.append(gold)
except KeyError:
get_logger().warning(
f'No gold answer for {k}, use empty string as reference!')
gold = ''
references.append('')
# Check if the answer is correct (A means correct)
is_correct = processed_judge == 'A'
total_count += 1
if is_correct:
total_correct += 1
# Update category stats
for level, key in [
('discipline', discipline),
('field', f'{discipline}/{field}'),
('subfield', f'{discipline}/{field}/{subfield}'),
]:
stats[level][key]['correct'] += 1
# Update category totals
for level, key in [
('discipline', discipline),
('field', f'{discipline}/{field}'),
('subfield', f'{discipline}/{field}/{subfield}'),
]:
stats[level][key]['total'] += 1
# Add to details
details.append({
'id': k,
'question': sample['question'],
'options': sample['options'],
'origin_prompt': v['origin_prompt'],
'llm_judge': processed_judge,
'gold': gold,
'is_correct': is_correct,
'discipline': discipline,
'field': field,
'subfield': subfield,
})
# Calculate overall accuracy with two decimal places
overall_accuracy = (round(
(total_correct / total_count * 100), 2) if total_count > 0 else 0.00)
# Initialize results dictionary
results = {
'accuracy': overall_accuracy,
'total_correct': total_correct,
'total_count': total_count,
'details': details,
}
# Calculate accuracy for each category and flatten into results
for level in stats:
for key, value in stats[level].items():
if value['total'] > 0:
# Calculate accuracy with two decimal places
accuracy = round((value['correct'] / value['total'] * 100), 2)
# Create a flattened key for the category
flat_key = f'SuperGPQA-{level}'
if level == 'discipline':
flat_key = f'SuperGPQA-{key}'
elif level == 'field':
discipline, field = key.split('/')
flat_key = f'SuperGPQA-{discipline}-{field}'
elif level == 'subfield':
discipline, field, subfield = key.split('/')
flat_key = f'SuperGPQA-{discipline}-{field}-{subfield}'
# Add to results
results[flat_key] = accuracy
return results

17
opencompass/evaluator/generic_llm_evaluator.py
View File

@ -84,6 +84,8 @@ class GenericLLMEvaluator(BaseEvaluator):
references: Optional[List] = None,
) -> Dict:
"""Apply to single-model scoring."""
assert len(predictions) == len(
references), 'predictions and references must have the same length'
# -------------- Build Inferencer ----------------
self.build_inferencer()
@ -127,7 +129,7 @@ class GenericLLMEvaluator(BaseEvaluator):
prompt_template=self.prompt_template)
output = mmengine.load(self.output_path)
return self.output_postprocess(output)
return self.output_postprocess(output, dataset)
def pred_postprocess(self, predictions: List) -> Dict:
if self.pred_postprocessor is None:
@ -137,15 +139,24 @@ class GenericLLMEvaluator(BaseEvaluator):
proc = TEXT_POSTPROCESSORS.get(kwargs.pop('type'))
return [proc(pred, **kwargs) for pred in predictions]
def output_postprocess(self, output: Dict) -> Dict:
def output_postprocess(self, output: Dict, dataset=None) -> Dict:
"""Postprocess output by adding necessary statistics or data into
it."""
import inspect
if self.dict_postprocessor is None:
return output
else:
kwargs = self.dict_postprocessor
proc = DICT_POSTPROCESSORS.get(kwargs.pop('type'))
return proc(output, self.output_path, **kwargs)
sig = inspect.signature(proc)
if 'dataset' in sig.parameters:
return proc(output,
self.output_path,
dataset=dataset,
**kwargs)
else:
return proc(output, self.output_path, **kwargs)
@property
def default_judge_cfg(self):

267
opencompass/openicl/icl_evaluator/code_evaluator.py Normal file
View File

@ -0,0 +1,267 @@
# flake8: noqa: E501
import difflib
import os
import re
import tempfile
import time
from typing import Any, Dict, List, Optional, Tuple, Union
from datasets import Dataset
from gradio_client import Client
from opencompass.openicl.icl_evaluator import BaseEvaluator
from opencompass.registry import ICL_EVALUATORS
@ICL_EVALUATORS.register_module()
class CodeEvaluator(BaseEvaluator):
"""Evaluator for code generation tasks.
This evaluator sends code to a remote evaluation service to test its
functionality against provided test cases. It handles code extraction,
processing, and result analysis.
"""
def __init__(self,
language: str,
ip_address: str = 'localhost',
retry: int = 3) -> None:
"""Initialize the CodeEvaluator.
Args:
language (str): Programming language of the code to evaluate.
ip_address (str, optional): IP address of the evaluation service. Defaults to 'localhost'.
retry (int, optional): Number of retry attempts for failed connections. Defaults to 3.
"""
self.language = language
self.retry = retry
self.client = Client(ip_address)
super().__init__()
def _extract_code(self, text: str) -> str:
"""Extract code from markdown-formatted text.
Args:
text (str): Text that may contain code blocks in markdown format.
Returns:
str: Extracted code from the last code block, or the original text if no code blocks found.
"""
blocks = re.findall(r'```\w*\n(.*?)```', text, re.DOTALL)
if len(blocks) >= 1:
text = blocks[0]
return text
def _code_eval_service(
self, input_data: Union[Dict, List,
str]) -> Tuple[bool, Union[Dict, List, Any]]:
"""Send code to the remote evaluation service using gradio_client and
get the results.
Args:
input_data: Can be one of:
- dict: Dictionary containing code information for a single test case
- list: List of dictionaries for batch evaluation
- str: File path to code file
Returns:
tuple: (succeed, output)
- succeed (bool): Whether the request was successful
- output (dict/list/str): Evaluation results or error message
"""
try:
temp_file_path = None
# Handle file path input
if isinstance(input_data, str):
with tempfile.NamedTemporaryFile(suffix=f'.{self.language}',
delete=False) as temp_file:
temp_file_path = temp_file.name
with open(input_data, 'r') as src_file:
content = src_file.read()
temp_file.write(content.encode())
input_data = temp_file_path
# Send to evaluation service
result = self.client.predict(input_data, api_name='/evaluate')
# Process the result
if isinstance(result, (dict, list)):
return True, result
else:
# Try to parse the result as JSON if it's a string
try:
import json
parsed_result = json.loads(result)
return True, parsed_result
except: # noqa: E722
return True, {'status': 'unknown', 'raw_result': result}
except Exception as e:
return False, str(e)
finally:
# Clean up temporary file if it was created
if temp_file_path and os.path.exists(temp_file_path):
try:
os.unlink(temp_file_path)
except: # noqa: E722
pass
def _remove_prefix(self,
prompt: str,
completion: str,
threshold: float = 0.95) -> str:
"""Determine the truncation point in the completion based on the last
line of the prompt, remove all content before that line in the
completion, and return the completion string after removing the prefix.
This is done to convert chatbot-style inference mode to completion
mode.
Args:
prompt (str): The prompt text.
completion (str): The completion text.
threshold (float): Line similarity threshold.
Returns:
str: The completion string after removing the prefix.
"""
prompt_lines = prompt.splitlines()
completion_lines = completion.splitlines()
if not prompt_lines:
return completion
last_prompt_line = prompt_lines[-1]
cut_index = -1
for i, completion_line in enumerate(completion_lines):
similarity = difflib.SequenceMatcher(None, last_prompt_line,
completion_line).ratio()
if similarity >= threshold:
cut_index = i
break
if cut_index != -1:
return '\n'.join(completion_lines[cut_index + 1:])
else:
return completion
def _process_completions(self, test_case: dict, completions: list) -> list:
"""Process code completion list, which typically involves extracting
code, removing repetitive prefixes caused by chatbot mode, and other
steps to ensure the model-generated code can be compiled successfully.
Args:
test_case (dict): Dictionary containing test case information including:
completions (list): List of code completions generated by the model.
Returns:
list: Processed code completion list.
"""
processed_completions = []
for comp in completions:
comp = self._extract_code(comp)
post_comp = self._remove_prefix(test_case['prompt'], comp)
processed_completions.append(post_comp)
return processed_completions
def _evaluate(
self, input_data: Union[Dict, List]
) -> Tuple[bool, Optional[Union[Dict, List]], Optional[str]]:
"""Evaluate code with retry mechanism.
Args:
input_data: Can be either:
- dict: Dictionary containing code and test information for a single test case
- list: List of dictionaries for batch evaluation
Returns:
tuple: (success, output, error_message)
- success (bool): Whether the evaluation was successful
- output (dict or list): Evaluation output (if successful)
- error_message (str): Error message (if failed)
"""
num_retry = 0
while num_retry < self.retry:
succeed, output = self._code_eval_service(input_data)
if not succeed:
num_retry += 1
time.sleep(10)
else:
break
if not succeed:
return False, None, f'code eval service connection failed: {output}'
return True, output, None
def score(self, predictions: List, references: List,
test_set: Dataset) -> Dict:
"""Score code generation predictions against references.
Args:
predictions (list): List of model-generated code completions.
references (list): List of reference solutions (not directly used in evaluation).
test_set (Dataset): Dataset containing test cases and other metadata.
Returns:
dict: Evaluation results including:
- accuracy: Percentage of correctly solved problems
- details: Detailed results for each test case
- error: Error message if evaluation failed
"""
if len(predictions) != len(references):
return {
'error':
'predictions and references have different '
f'length. len(predictions): {len(predictions)}, '
f'len(references): {len(references)}'
}
test_set = test_set.to_pandas()
# Use the first column as the unique identifier
test_set_origin = test_set.drop_duplicates(subset=test_set.columns[0])
num_repeats = int(len(test_set) / len(test_set_origin))
# 1. Prepare data for all test cases
all_test_cases = []
for i in range(len(test_set_origin)):
test_case = test_set_origin.iloc[i]
completions = predictions[i * num_repeats:(i + 1) * num_repeats]
# Process code completions
processed_completions = self._process_completions(
test_case, completions)
result_dict = {
'name': test_case['name'],
'language': test_case['language'],
'prompt': test_case['prompt'],
'tests': test_case['tests'],
'processed_completions': processed_completions,
'completions': completions
}
all_test_cases.append(result_dict)
# 2. Send all test cases to the evaluation service
success, outputs, error_message = self._evaluate(all_test_cases)
if not success:
return {'error': error_message}
# 3. Process the returned results
details = []
correct = 0
for output in outputs:
if output.get('status') == 'OK':
output['correct'] = True
correct += 1
else:
output['correct'] = False
details.append(output)
return {
f'pass@{num_repeats}': 100 * correct / len(test_set_origin),
'details': details
}

8
opencompass/openicl/icl_evaluator/icl_base_evaluator.py
View File

@ -89,6 +89,14 @@ class BaseEvaluator:
original_dataset: Dataset,
**score_kwargs,
):
# Check if predictions and references have the
# same length if both are provided
if 'predictions' in score_kwargs and 'references' in score_kwargs:
if len(score_kwargs['predictions']) != len(
score_kwargs['references']):
raise ValueError(
'Predictions and references must have the same length')
real_size = len(original_dataset) // n
all_details = []
all_results = []

2
opencompass/runners/volc.py
View File

@ -256,7 +256,7 @@ class VOLCRunner(BaseRunner):
with open(config_path) as fp:
volc_cfg = yaml.safe_load(fp)
if num_gpus <= 0:
flavor = 'ml.c3i.2xlarge'
flavor = 'ml.r3i.2xlarge'
elif num_gpus == 1:
flavor = 'ml.pni2l.3xlarge'
elif num_gpus == 2:

11
opencompass/utils/datasets_info.py
View File

@ -193,6 +193,12 @@ DATASETS_MAPPING = {
"hf_id": "",
"local": "./data/mmlu_pro",
},
# MultiPL-E
"opencompass/multipl_e": {
"ms_id": "",
"hf_id": "",
"local": "./data/multipl_e",
},
# NQ
"opencompass/natural_question": {
"ms_id": "opencompass/natural_question",
@ -627,6 +633,11 @@ DATASETS_URL = {
"http://opencompass.oss-cn-shanghai.aliyuncs.com/datasets/data/mmlu_pro.zip",
"md5": "e3200c7380f4cea5f13c768f2815fabb",
},
"multipl_e": {
"url":
"http://opencompass.oss-cn-shanghai.aliyuncs.com/datasets/data/multipl_e.zip",
"md5": "24462aac7a38a4a62f5c5e89eb614e20",
},
"/Longbench": {
"url":
"http://opencompass.oss-cn-shanghai.aliyuncs.com/datasets/data/Longbench.zip",