File size: 17,294 Bytes
256a159 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 |
import copy
import csv
import json
import os
from typing import List
from datasets import Dataset
from opencompass.datasets.circular import (CircularDatasetMeta,
CircularEvaluator)
from opencompass.openicl.icl_evaluator import AccEvaluator, BaseEvaluator
from opencompass.openicl.icl_inferencer import GenInferencer, PPLInferencer
from opencompass.openicl.icl_prompt_template import PromptTemplate
from opencompass.openicl.icl_retriever import ZeroRetriever
from opencompass.registry import LOAD_DATASET
from .base import BaseDataset
class OptionSimAccEvaluator(BaseEvaluator):
def __init__(self, options) -> None:
super().__init__()
if not all((isinstance(i, str) and i.isupper() and len(i) == 1)
for i in options):
raise ValueError(
f'Each options should be single upper letter, got {options}')
self.options = options
def match_any_label(self, pred, test_item):
from rapidfuzz.distance import Levenshtein as L
from opencompass.utils.text_postprocessors import \
first_option_postprocess
pred = pred.strip()
if any([pred == i for i in self.options]):
parsed = pred
else:
parsed = ''
if parsed == '':
parsed = first_option_postprocess(pred,
''.join(self.options),
cushion=False)
if parsed == '':
possible_options = []
for opt in self.options:
opt_str = test_item[opt]
if opt_str is not None and opt_str.lower() in pred.lower():
possible_options.append(opt)
if len(possible_options) == 1:
parsed = possible_options[0]
if parsed == '':
dists = []
for opt in self.options:
opt_str = test_item[opt]
if opt_str is None:
continue
cands = [opt, opt_str, opt + '. ' + opt_str]
d = min(L.distance(pred, cand) for cand in cands)
dists.append((d, opt))
if len(dists) > 0:
parsed = min(dists)[1]
return parsed
def score(self, predictions: List, references: List, test_set) -> dict:
assert len(predictions) == len(references)
num_correct, num_total = 0, 0
details = {}
for index in range(len(predictions)):
pred = predictions[index]
refr = references[index]
parsed = self.match_any_label(pred, test_set[index])
num_correct += 1 if parsed == refr else 0
num_total += 1
details[str(index)] = {}
details[str(index)]['pred'] = pred
details[str(index)]['parsed'] = parsed
details[str(index)]['refr'] = refr
details[str(index)]['correct'] = parsed == refr
return {'accuracy': num_correct / num_total * 100, 'details': details}
# TODO: DO NOT COPY YOURSELF!!!
class CircularOptionSimAccEvaluator(OptionSimAccEvaluator):
def __init__(self, options, circular_pattern='circular'):
super().__init__(options)
self.circular_pattern = circular_pattern
def score(self, predictions, references, test_set):
from opencompass.datasets.circular import (get_all_possible_patterns,
get_circular_patterns,
get_origin_patterns)
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):
parsed = self.match_any_label(pred, origin_item)
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 parsed == 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']
parsed = self.match_any_label(pred, origin_item)
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 parsed == 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
# make details
details = {}
for index in range(len(predictions)):
parsed = self.match_any_label(predictions[index], test_set[index])
details[str(index)] = {}
if 'question' in test_set[index]:
details[str(index)]['question'] = test_set[index]['question']
details[str(index)]['pred'] = predictions[index]
details[str(index)]['parsed'] = parsed
details[str(index)]['refr'] = references[index]
details[str(index)]['correct'] = parsed == references[index]
metrics['details'] = details
return metrics
@LOAD_DATASET.register_module()
class CustomDataset(BaseDataset):
@staticmethod
def load(path):
if path.endswith('.jsonl'):
with open(path, 'r', encoding='utf-8-sig') as f:
data = [json.loads(line) for line in f]
elif path.endswith('.csv'):
with open(path, 'r', encoding='utf-8-sig') as f:
reader = csv.reader(f)
header = next(reader)
data = [dict(zip(header, row)) for row in reader]
else:
raise ValueError(f'Unsupported file format: {path}')
return Dataset.from_list(data)
class CircularCustomDataset(CustomDataset, metaclass=CircularDatasetMeta):
dataset_class = CustomDataset
def stringfy_types(obj):
for k, v in obj.items():
if k == 'type':
obj[k] = f'{v.__module__}.{v.__name__}'
elif isinstance(v, dict):
stringfy_types(v)
return obj
def make_mcq_gen_config(meta):
if meta.get('template', None) is None:
_human_prompt = 'Question: {question}' + ''.join(
[f'\n{item}. {{{item}}}' for item in meta['options']])
human_prompt = meta.get('human_prompt', _human_prompt)
_bot_prompt = f'Answer: {{{meta["output_column"]}}}'
bot_prompt = meta.get('bot_prompt', _bot_prompt)
template = dict(round=[
dict(role='HUMAN', prompt=human_prompt),
dict(role='BOT', prompt=bot_prompt),
])
else:
template = meta['template']
reader_cfg = dict(
input_columns=meta['input_columns'],
output_column=meta['output_column'],
)
infer_cfg = dict(
prompt_template=dict(
type=PromptTemplate,
template=template,
),
retriever=dict(type=ZeroRetriever),
inferencer=dict(type=GenInferencer),
)
eval_cfg = dict(
evaluator=dict(type=meta.get('evaluator', OptionSimAccEvaluator),
**meta.get('evaluator_kwargs',
{'options': meta['options']})),
pred_role='BOT',
)
dataset = dict(
abbr=meta['abbr'],
type=CustomDataset,
path=meta['path'],
reader_cfg=reader_cfg,
infer_cfg=infer_cfg,
eval_cfg=eval_cfg,
)
return dataset
def make_circular_mcq_gen_config(meta):
if meta.get('template', None) is None:
_human_prompt = 'Question: {question}' + ''.join(
[f'\n{item}. {{{item}}}' for item in meta['options']])
human_prompt = meta.get('human_prompt', _human_prompt)
_bot_prompt = f'Answer: {{{meta["output_column"]}}}'
bot_prompt = meta.get('bot_prompt', _bot_prompt)
template = dict(round=[
dict(role='HUMAN', prompt=human_prompt),
dict(role='BOT', prompt=bot_prompt),
])
else:
template = meta['template']
reader_cfg = dict(
input_columns=meta['input_columns'],
output_column=meta['output_column'],
)
infer_cfg = dict(
prompt_template=dict(
type=PromptTemplate,
template=template,
),
retriever=dict(type=ZeroRetriever),
inferencer=dict(type=GenInferencer),
)
eval_cfg = dict(
evaluator=dict(type=meta.get('evaluator',
CircularOptionSimAccEvaluator),
**meta.get('evaluator_kwargs',
{'options': meta['options']})),
pred_role='BOT',
)
dataset = dict(
abbr=meta['abbr'],
type=CircularCustomDataset,
option_keys=meta['options'],
answer_key=meta['output_column'],
path=meta['path'],
reader_cfg=reader_cfg,
infer_cfg=infer_cfg,
eval_cfg=eval_cfg,
)
return dataset
def make_qa_gen_config(meta):
if meta.get('template', None) is None:
human_prompt = meta.get('human_prompt', '{question}')
if meta['output_column'] is None:
template = dict(round=[
dict(role='HUMAN', prompt=human_prompt),
])
else:
bot_prompt = meta.get('bot_prompt', f'{{{meta["output_column"]}}}')
template = dict(round=[
dict(role='HUMAN', prompt=human_prompt),
dict(role='BOT', prompt=bot_prompt),
])
else:
template = meta['template']
reader_cfg = dict(
input_columns=meta['input_columns'],
output_column=meta['output_column'],
)
infer_cfg = dict(
prompt_template=dict(
type=PromptTemplate,
template=template,
),
retriever=dict(type=ZeroRetriever),
inferencer=dict(type=GenInferencer),
)
eval_cfg = dict(
evaluator=dict(type=meta.get('evaluator', AccEvaluator),
**meta.get('evaluator_kwargs', {})),
pred_role='BOT',
)
dataset = dict(
abbr=meta['abbr'],
type=CustomDataset,
path=meta['path'],
reader_cfg=reader_cfg,
infer_cfg=infer_cfg,
eval_cfg=eval_cfg,
)
return dataset
def make_mcq_ppl_config(meta):
if meta.get('template', None) is None:
_human_prompt = 'Question: {question}' + ''.join(
[f'\n{item}. {{{item}}}' for item in meta['options']])
human_prompt = meta.get('human_prompt', _human_prompt)
_bot_prompt = f'Answer: {{{meta["output_column"]}}}'
bot_prompt = meta.get('bot_prompt', _bot_prompt)
template = {
answer: dict(round=[
dict(role='HUMAN', prompt=human_prompt),
dict(role='BOT',
prompt=bot_prompt.format(
**{meta['output_column']: answer})),
], )
for answer in meta['options']
}
else:
template = meta['template']
reader_cfg = dict(
input_columns=meta['input_columns'],
output_column=meta['output_column'],
)
infer_cfg = dict(
prompt_template=dict(
type=PromptTemplate,
template=template,
),
retriever=dict(type=ZeroRetriever),
inferencer=dict(type=PPLInferencer),
)
eval_cfg = dict(evaluator=dict(type=meta.get('evaluator', AccEvaluator),
**meta.get('evaluator_kwargs', {})))
dataset = dict(
abbr=meta['abbr'],
type=CustomDataset,
path=meta['path'],
reader_cfg=reader_cfg,
infer_cfg=infer_cfg,
eval_cfg=eval_cfg,
)
return dataset
def make_circular_mcq_ppl_config(meta):
if meta.get('template', None) is None:
_human_prompt = 'Question: {question}' + ''.join(
[f'\n{item}. {{{item}}}' for item in meta['options']])
human_prompt = meta.get('human_prompt', _human_prompt)
_bot_prompt = f'Answer: {{{meta["output_column"]}}}'
bot_prompt = meta.get('bot_prompt', _bot_prompt)
template = {
answer: dict(round=[
dict(role='HUMAN', prompt=human_prompt),
dict(role='BOT',
prompt=bot_prompt.format(
**{meta['output_column']: answer})),
], )
for answer in meta['options']
}
else:
template = meta['template']
reader_cfg = dict(
input_columns=meta['input_columns'],
output_column=meta['output_column'],
)
infer_cfg = dict(
prompt_template=dict(
type=PromptTemplate,
template=template,
),
retriever=dict(type=ZeroRetriever),
inferencer=dict(type=PPLInferencer),
)
eval_cfg = dict(
evaluator=dict(type=meta.get('evaluator', CircularEvaluator),
**meta.get('evaluator_kwargs', {})))
dataset = dict(
abbr=meta['abbr'],
type=CircularCustomDataset,
option_keys=meta['options'],
answer_key=meta['output_column'],
path=meta['path'],
reader_cfg=reader_cfg,
infer_cfg=infer_cfg,
eval_cfg=eval_cfg,
)
return dataset
def parse_example_dataset(config):
# config -> .meta.jsonl -> parsed_results
path = config['path']
# load sample and get parsed_meta
parsed_meta = {}
if path.endswith('.jsonl'):
with open(path, 'r', encoding='utf-8') as f:
data_item = json.loads(f.readline())
elif path.endswith('.csv'):
with open(path, 'r', encoding='utf-8') as f:
reader = csv.reader(f)
header = next(reader)
row = next(reader)
data_item = dict(zip(header, row))
else:
raise ValueError(f'Unsupported ext: {path}, .jsonl or .csv required')
parsed_meta['path'] = path
input_columns = [i for i in data_item.keys() if i != 'answer']
parsed_meta['input_columns'] = input_columns
output_column = 'answer' if 'answer' in data_item else None
parsed_meta['output_column'] = output_column
options = []
for i in range(26):
i = chr(ord('A') + i)
if i in data_item:
options.append(i)
else:
break
parsed_meta['options'] = options
abbr = os.path.basename(path).split('.')[0]
parsed_meta['abbr'] = abbr
parsed_meta['data_type'] = 'mcq' if len(options) > 1 else 'qa'
parsed_meta['infer_method'] = 'gen'
# try to read meta json
meta_path = config.get('meta_path', path + '.meta.json')
if os.path.exists(meta_path):
with open(meta_path, 'r', encoding='utf-8') as f:
read_from_file_meta = json.load(f)
else:
read_from_file_meta = {}
# get config meta
config_meta = copy.deepcopy(config)
# merge meta
meta = {}
meta.update(parsed_meta)
meta.update(read_from_file_meta)
meta.update(config_meta)
return meta
def make_custom_dataset_config(config):
# considered as a custom dataset
meta = parse_example_dataset(config)
make_config_func = {
('mcq', 'gen'): make_mcq_gen_config,
('mcq', 'ppl'): make_mcq_ppl_config,
('qa', 'gen'): make_qa_gen_config,
('circular-mcq', 'gen'): make_circular_mcq_gen_config,
('circular-mcq', 'ppl'): make_circular_mcq_ppl_config,
}.get((meta['data_type'], meta['infer_method']), None)
if make_config_func is None:
raise ValueError(f'Unsupported dataset data_type: {meta["data_type"]}'
f' and infer_method: {meta["infer_method"]}')
dataset = make_config_func(meta)
dataset = stringfy_types(dataset)
return dataset
|