TwT-6's picture
Upload 2667 files
256a159 verified
raw
history blame
15 kB
import copy
import itertools
from typing import Callable, List, Optional, Union
from datasets import Dataset, DatasetDict
from opencompass.openicl.icl_evaluator import BaseEvaluator
from .arc import ARCDataset
from .ceval import CEvalDataset
from .cmmlu import CMMLUDataset
from .commonsenseqa import commonsenseqaDataset
from .hellaswag import hellaswagDataset_V2
from .mmlu import MMLUDataset
from .obqa import OBQADataset
from .piqa import piqaDataset_V2
from .race import RaceDataset
from .siqa import siqaDataset_V3
from .xiezhi import XiezhiDataset
def get_origin_patterns(option_keys):
return [tuple(option_keys)]
def get_circular_patterns(option_keys):
double_option_keys = option_keys + option_keys
circular_patterns = [
tuple(double_option_keys[i:i + len(option_keys)])
for i in range(len(option_keys))
]
return circular_patterns
def get_all_possible_patterns(option_keys):
circular_patterns = list(itertools.permutations(option_keys))
return circular_patterns
class CircularDatasetMeta(type):
"""This Meta Class is designed to transform a class that reads datasets
into one that supports reading datasets required for CircularEval. It
overloads an existing load method for the original class.
The Meta Class should possess the following attributes:
- `dataset_class` (class): The class for reading datasets, such as
`CEvalDataset`.
- `default_circular_splits` (list, optional): The default splits of the
dataset that need to undergo CircularEval, like ['val', 'test']. If a
`Dataset` is loaded originally, this field will be ignored.
- `default_option_keys` (list): The keys for options in the dataset, such
as ['A', 'B', 'C', 'D'].
- `default_answer_key` (str, optional): The key for answers in the dataset,
like 'answer'. This is an alternative to
`default_answer_key_switch_method`.
- `default_answer_key_switch_method` (function, optional): The method to
transform the key for answers in the dataset. This is an alternative to
`default_answer_key`.
"""
@staticmethod
def make_circular_items(
origin_item,
circular_patterns,
option_keys,
answer_key,
answer_key_switch_method,
qid,
):
items = []
for circular_pattern in circular_patterns:
item = copy.deepcopy(origin_item)
for i in range(len(option_keys)):
item[circular_pattern[i]] = origin_item[option_keys[i]]
if answer_key_switch_method is None:
if origin_item[answer_key] in option_keys:
item[answer_key] = circular_pattern[option_keys.index(
origin_item[answer_key])]
else:
pass
else:
item = answer_key_switch_method(item, circular_pattern)
item['qid'] = qid
item['circular_pattern'] = tuple(circular_pattern)
items.append(item)
return items
@staticmethod
def make_circular_dataset(dataset, circular_patterns, option_keys,
answer_key, answer_key_switch_method):
circulated_items = []
for i, item in enumerate(dataset):
item = CircularDatasetMeta.make_circular_items(
item,
circular_patterns,
option_keys,
answer_key,
answer_key_switch_method,
i,
)
circulated_items.extend(item)
return Dataset.from_list(circulated_items)
def make_circular(
dataset: Union[Dataset, DatasetDict],
circular_splits: Optional[List[str]] = ['test'],
circular_patterns: str = 'circular',
option_keys: List[str] = ['A', 'B', 'C', 'D'],
answer_key: Optional[str] = 'answer',
answer_key_switch_method: Optional[Callable] = None,
):
"""Transform the dataset into one that is compatible with CircularEval.
In CircularEval, the original multiple-choice questions with options
ABCD are augmented by shuffling the order of options, such as BCDA,
CDAB, DABC, etc. A model is considered correct only if it answers all
augmented questions correctly. This method effectively prevents models
from memorizing answers.
Args:
datasets: The dataset to be augmented.
circular_splits: List of splits to make circular. This is only
effective when the dataset is a DatasetDict.
circular_patterns: Method for circular processing, can be 'circular'
for single cycle or 'all_possible' for all permutations, default
is 'circular'.
option_keys: List of keys for options, default to ['A', 'B', 'C', 'D'].
answer_key: Key for the answer, default to 'answer'. When specified,
ensure that the content of answer_key is among the option_keys.
It is an alternative to specifying answer_key_switch_method.
answer_key_switch_method: Function to modify the answer_key. It is an
alternative to specifying answer_key.
"""
if isinstance(circular_patterns, str):
if circular_patterns == 'circular':
circular_patterns = get_circular_patterns(option_keys)
elif circular_patterns == 'all_possible':
circular_patterns = get_all_possible_patterns(option_keys)
else:
raise ValueError(
f'Unknown circular_patterns: {circular_patterns}')
else:
assert isinstance(circular_patterns, list)
assert all([isinstance(i, list) for i in circular_patterns])
# TODO: other necessary sanity checks
raise NotImplementedError(
'circular_patterns int list of list has not been tested yet')
if answer_key is None and answer_key_switch_method is None:
raise ValueError(
'answer_key and answer_key_switch_method cannot be both None')
if answer_key is not None and answer_key_switch_method is not None:
raise ValueError(
'either answer_key or answer_key_switch_method should be None')
if isinstance(dataset, Dataset):
dataset = CircularDatasetMeta.make_circular_dataset(
dataset,
circular_patterns,
option_keys,
answer_key,
answer_key_switch_method,
)
else:
assert isinstance(dataset, DatasetDict)
dataset_dict = {}
for split in dataset:
if circular_splits is not None and split in circular_splits:
dataset_dict[
split] = CircularDatasetMeta.make_circular_dataset(
dataset[split],
circular_patterns,
option_keys,
answer_key,
answer_key_switch_method,
)
else:
dataset_dict[split] = dataset[split]
dataset = DatasetDict(dataset_dict)
return dataset
def __new__(cls, name, bases, dct):
new_cls = super().__new__(cls, name, bases, dct)
def load(cls, circular_patterns='circular', *args, **kwargs):
circular_splits = getattr(cls, 'default_circular_splits', None)
option_keys = getattr(cls, 'default_option_keys', None)
if 'option_keys' in kwargs:
option_keys = kwargs.pop('option_keys')
assert option_keys is not None, 'option_keys cannot be None'
answer_key = getattr(cls, 'default_answer_key', None)
if 'answer_key' in kwargs:
answer_key = kwargs.pop('answer_key')
answer_key_switch_method = getattr(
cls, 'default_answer_key_switch_method', None)
dataset = cls.dataset_class.load(*args, **kwargs)
return CircularDatasetMeta.make_circular(
dataset,
circular_splits,
circular_patterns,
option_keys,
answer_key,
answer_key_switch_method,
)
setattr(new_cls, 'load', classmethod(load))
return new_cls
class CircularCEvalDataset(CEvalDataset, metaclass=CircularDatasetMeta):
dataset_class = CEvalDataset
default_circular_splits = ['val', 'test']
default_option_keys = ['A', 'B', 'C', 'D']
default_answer_key = 'answer'
class CircularMMLUDataset(MMLUDataset, metaclass=CircularDatasetMeta):
dataset_class = MMLUDataset
default_circular_splits = ['test']
default_option_keys = ['A', 'B', 'C', 'D']
default_answer_key = 'target'
class CircularCMMLUDataset(CMMLUDataset, metaclass=CircularDatasetMeta):
dataset_class = CMMLUDataset
default_circular_splits = ['test']
default_option_keys = ['A', 'B', 'C', 'D']
default_answer_key = 'answer'
class CircularCSQADataset(commonsenseqaDataset, metaclass=CircularDatasetMeta):
dataset_class = commonsenseqaDataset
default_circular_splits = ['validation']
default_option_keys = ['A', 'B', 'C', 'D', 'E']
default_answer_key = 'answerKey'
class CircularARCDataset(ARCDataset, metaclass=CircularDatasetMeta):
dataset_class = ARCDataset
default_circular_splits = None
default_option_keys = ['textA', 'textB', 'textC', 'textD']
def default_answer_key_switch_method(item, circular_pattern):
circular_pattern = tuple(i[-1] for i in circular_pattern)
item['answerKey'] = circular_pattern['ABCD'.index(item['answerKey'])]
return item
class CircularHSWAGDataset(hellaswagDataset_V2, metaclass=CircularDatasetMeta):
dataset_class = hellaswagDataset_V2
default_circular_splits = None
default_option_keys = ['A', 'B', 'C', 'D']
default_answer_key = 'label'
class CircularOBQADataset(OBQADataset, metaclass=CircularDatasetMeta):
dataset_class = OBQADataset
default_circular_splits = None
default_option_keys = ['A', 'B', 'C', 'D']
default_answer_key = 'answerKey'
class CircularRaceDataset(RaceDataset, metaclass=CircularDatasetMeta):
dataset_class = RaceDataset
default_circular_splits = ['test']
default_option_keys = ['A', 'B', 'C', 'D']
default_answer_key = 'answer'
class CircularXiezhiDataset(XiezhiDataset, metaclass=CircularDatasetMeta):
dataset_class = XiezhiDataset
default_circular_splits = None
default_option_keys = ['A', 'B', 'C', 'D']
default_answer_key = 'answer'
class CircularsiqaDataset(siqaDataset_V3, metaclass=CircularDatasetMeta):
dataset_class = siqaDataset_V3
default_circular_splits = ['validation']
default_option_keys = ['A', 'B', 'C']
default_answer_key = 'answer'
class CircularpiqaDataset(piqaDataset_V2, metaclass=CircularDatasetMeta):
dataset_class = piqaDataset_V2
default_circular_splits = ['validation']
default_option_keys = ['sol1', 'sol2']
def default_answer_key_switch_method(item, circular_pattern):
circular_pattern = tuple(int(i[-1]) - 1 for i in circular_pattern)
item['answer'] = 'AB'[circular_pattern['AB'.index(item['answer'])]]
return item
class CircularEvaluator(BaseEvaluator):
"""This Evaluator assesses datasets post-Circular processing, generating
the following evaluation metrics:
- `acc_{origin|circular|all_possible}`: Treats each question with shuffled
answer options as separate, calculating accuracy.
- `perf_{origin|circular|all_possible}`: According Circular logic, a
question is considered correct only if all its variations with shuffled
options are answered correctly, calculating accuracy. perf is short for
perfect.
- `more_{num}_{origin|circular|all_possible}`: According to Circular logic,
a question is considered correct only if the number of its variations
answered correctly is greater than or equal to `num`, calculating
accuracy.
Note that when the `all_possible` method is used to shuffle option order,
it naturally includes the Circular method, and its metrics will also be
output.
Args:
circular_pattern: The method of shuffling options, either 'circular' or
'all_possible', defaulting to 'circular'.
"""
def __init__(self, circular_pattern='circular'):
super().__init__()
self.circular_pattern = circular_pattern
def score(self, predictions, references, test_set):
circular_patterns = {}
circular_patterns['origin'] = get_origin_patterns(
test_set[0]['circular_pattern'])
circular_patterns['circular'] = get_circular_patterns(
test_set[0]['circular_pattern'])
if self.circular_pattern == 'all_possible':
circular_patterns['all_possible'] = get_all_possible_patterns(
test_set[0]['circular_pattern'])
metrics = {}
tmp_metrics = {}
tmp_metrics.update({f'correct_{k}': 0 for k in circular_patterns})
tmp_metrics.update({f'count_{k}': 0 for k in circular_patterns})
# calculate the original accuracy
for pred, refr, origin_item in zip(predictions, references, test_set):
circular_pattern = origin_item['circular_pattern']
for k in circular_patterns:
if tuple(circular_pattern) in circular_patterns[k]:
tmp_metrics[f'correct_{k}'] += 1 if pred == refr else 0
tmp_metrics[f'count_{k}'] += 1
for k in circular_patterns:
metrics[f'acc_{k}'] = (tmp_metrics[f'correct_{k}'] /
tmp_metrics[f'count_{k}'] * 100)
# calculate the circular accuracy
_details = {k: {} for k in circular_patterns}
for pred, refr, origin_item in zip(predictions, references, test_set):
index = origin_item['qid']
circular_pattern = origin_item['circular_pattern']
for k in circular_patterns:
if tuple(circular_pattern) in circular_patterns[k]:
_details[k].setdefault(
index, []).append(True if pred == refr else False)
for k in _details:
_details[k] = {
index: sum(_details[k][index])
for index in _details[k]
}
for k in _details:
for j in range(1, len(circular_patterns[k]) + 1):
count = sum([_details[k][index] >= j for index in _details[k]])
total = len(_details[k])
if j != len(circular_patterns[k]):
metrics[f'more_{j}_{k}'] = count / total * 100
else:
metrics[f'perf_{k}'] = count / total * 100
return metrics