superglue.py 14.3 KB
Newer Older
Jason Phang's avatar
Jason Phang committed
1
"""
2
3
4
5
6
7
8
9
10
SuperGLUE: A Stickier Benchmark for General-Purpose Language Understanding Systems
https://w4ngatang.github.io/static/papers/superglue.pdf

SuperGLUE is a benchmark styled after GLUE with a new set of more difficult language
understanding tasks.

Homepage: https://super.gluebenchmark.com/

TODO: WSC requires free-form generation.
Jason Phang's avatar
Jason Phang committed
11
"""
Jason Phang's avatar
Jason Phang committed
12
import numpy as np
13
14
import sklearn
import transformers.data.metrics.squad_metrics as squad_metrics
Jonathan Tow's avatar
Jonathan Tow committed
15
16
17
from lm_eval.base import rf, Task
from lm_eval.metrics import mean, acc_all, metric_max_over_ground_truths, yesno
from lm_eval.utils import general_detokenize
Jason Phang's avatar
Jason Phang committed
18

Jason Phang's avatar
Jason Phang committed
19

20
21
22
23
24
25
26
27
28
29
30
31
32
_CITATION = """
@inproceedings{NEURIPS2019_4496bf24,
    author = {Wang, Alex and Pruksachatkun, Yada and Nangia, Nikita and Singh, Amanpreet and Michael, Julian and Hill, Felix and Levy, Omer and Bowman, Samuel},
    booktitle = {Advances in Neural Information Processing Systems},
    editor = {H. Wallach and H. Larochelle and A. Beygelzimer and F. d\textquotesingle Alch\'{e}-Buc and E. Fox and R. Garnett},
    pages = {},
    publisher = {Curran Associates, Inc.},
    title = {SuperGLUE: A Stickier Benchmark for General-Purpose Language Understanding Systems},
    url = {https://proceedings.neurips.cc/paper/2019/file/4496bf24afe7fab6f046bf4923da8de6-Paper.pdf},
    volume = {32},
    year = {2019}
}
"""
Jason Phang's avatar
Jason Phang committed
33

Jason Phang's avatar
Jason Phang committed
34

Jonathan Tow's avatar
Jonathan Tow committed
35
class BoolQ(Task):
36
    VERSION = 1
Leo Gao's avatar
Leo Gao committed
37
38
    DATASET_PATH = "super_glue"
    DATASET_NAME = "boolq"
Jason Phang's avatar
Jason Phang committed
39
40
41
42
43
44
45
46

    def has_training_docs(self):
        return True

    def has_validation_docs(self):
        return True

    def has_test_docs(self):
47
        return False
Jason Phang's avatar
Jason Phang committed
48

Jonathan Tow's avatar
Jonathan Tow committed
49
50
51
52
53
54
55
56
    def training_docs(self):
        if self._training_docs is None:
            self._training_docs = list(self.dataset["train"])
        return self._training_docs

    def validation_docs(self):
        return self.dataset["validation"]

Leo Gao's avatar
Update  
Leo Gao committed
57
    def doc_to_text(self, doc):
58
        return f"{doc['passage']}\nQuestion: {doc['question']}?\nAnswer:"
59
60
61
62
63
64

    def should_decontaminate(self):
        return True

    def doc_to_decontamination_query(self, doc):
        return doc['passage']
Leo Gao's avatar
Update  
Leo Gao committed
65
66
    
    def doc_to_target(self, doc):
67
        return " " + yesno(doc['label']) 
Jason Phang's avatar
Jason Phang committed
68

69
    def construct_requests(self, doc, ctx):
Leo Gao's avatar
Update  
Leo Gao committed
70

71
        ll_yes, _ = rf.loglikelihood(ctx, ' yes')
Jason Phang's avatar
Jason Phang committed
72
        ll_no, _ = rf.loglikelihood(ctx, ' no')
Leo Gao's avatar
Update  
Leo Gao committed
73
74
75
76
77
78
79
80
81

        return ll_yes, ll_no

    def process_results(self, doc, results):
        ll_yes, ll_no = results
        gold = doc["label"]

        acc = 1. if (ll_yes > ll_no) == gold else 0.

82
83
84
85
86
87
88
89
90
91
92
93
94
        return {
            "acc": acc
        }
    
    def higher_is_better(self):
        return {
            "acc": True
        }
    
    def aggregation(self):
        return {
            "acc": mean
        }
Jason Phang's avatar
Jason Phang committed
95

Jason Phang's avatar
Jason Phang committed
96

Jonathan Tow's avatar
Jonathan Tow committed
97
class CommitmentBank(Task):
thomasw21's avatar
thomasw21 committed
98
    VERSION = 1
Leo Gao's avatar
Leo Gao committed
99
100
    DATASET_PATH = "super_glue"
    DATASET_NAME = "cb"
Jason Phang's avatar
Jason Phang committed
101
102
103
104
105
106
107
108

    def has_training_docs(self):
        return True

    def has_validation_docs(self):
        return True

    def has_test_docs(self):
109
        return False
Jason Phang's avatar
Jason Phang committed
110

Jonathan Tow's avatar
Jonathan Tow committed
111
112
113
114
115
116
117
118
    def training_docs(self):
        if self._training_docs is None:
            self._training_docs = list(self.dataset["train"])
        return self._training_docs

    def validation_docs(self):
        return self.dataset["validation"]

119
    def doc_to_text(self, doc):
Leo Gao's avatar
Leo Gao committed
120
        return "{}\nQuestion: {}. True, False or Neither?\nAnswer:".format(
Jason Phang's avatar
Jason Phang committed
121
122
123
            doc["premise"],
            doc["hypothesis"],
        )
124

thefazzer's avatar
thefazzer committed
125
    def doc_to_target(self, doc):
126
127
128
        # True = entailment
        # False = contradiction
        # Neither = neutral
thomasw21's avatar
Fix CB  
thomasw21 committed
129
        return " {}".format({0: "True", 1: "False", 2: "Neither"}[doc["label"]])
Jason Phang's avatar
Jason Phang committed
130

thefazzer's avatar
thefazzer committed
131
    def construct_requests(self, doc, ctx):
Leo Gao's avatar
Leo Gao committed
132
133
        ll_true, _ = rf.loglikelihood(ctx, ' True')
        ll_false, _ = rf.loglikelihood(ctx, ' False')
thomasw21's avatar
Fix CB  
thomasw21 committed
134
        ll_neither, _ = rf.loglikelihood(ctx, ' Neither')
135

thomasw21's avatar
Fix CB  
thomasw21 committed
136
        return ll_true, ll_false, ll_neither
thefazzer's avatar
thefazzer committed
137
138
139

    def process_results(self, doc, results):
        gold = doc["label"]
thefazzer's avatar
thefazzer committed
140
141
        pred = np.argmax(results)
        acc = 1. if pred == gold else 0.
Jason Phang's avatar
Jason Phang committed
142

thefazzer's avatar
thefazzer committed
143
        return {
thefazzer's avatar
thefazzer committed
144
145
            "acc": acc,
            "f1": (pred, gold)
thefazzer's avatar
thefazzer committed
146
147
148
149
        }
    
    def higher_is_better(self):
        return {
150
151
            "acc": True,
            "f1": True
thefazzer's avatar
thefazzer committed
152
        }
Jason Phang's avatar
Jason Phang committed
153
154
155
156
157
158
159
160
161
162
163

    @classmethod
    def cb_multi_fi(cls, items):
        preds, golds = zip(*items)
        preds = np.array(preds)
        golds = np.array(golds)
        f11 = sklearn.metrics.f1_score(y_true=golds == 0, y_pred=preds == 0)
        f12 = sklearn.metrics.f1_score(y_true=golds == 1, y_pred=preds == 1)
        f13 = sklearn.metrics.f1_score(y_true=golds == 2, y_pred=preds == 2)
        avg_f1 = mean([f11, f12, f13])
        return avg_f1
thefazzer's avatar
thefazzer committed
164
165
166
    
    def aggregation(self):
        return {
thefazzer's avatar
thefazzer committed
167
            "acc": mean,
Jason Phang's avatar
Jason Phang committed
168
            "f1": self.cb_multi_fi,
thefazzer's avatar
thefazzer committed
169
        }
Jason Phang's avatar
Jason Phang committed
170

Jason Phang's avatar
Jason Phang committed
171

Jonathan Tow's avatar
Jonathan Tow committed
172
class Copa(Task):
Leo Gao's avatar
Leo Gao committed
173
    VERSION = 0
Leo Gao's avatar
Leo Gao committed
174
175
    DATASET_PATH = "super_glue"
    DATASET_NAME = "copa"
Jason Phang's avatar
Jason Phang committed
176
177
178
179
180
181
182
183

    def has_training_docs(self):
        return True

    def has_validation_docs(self):
        return True

    def has_test_docs(self):
184
        return False
Jason Phang's avatar
Jason Phang committed
185

Jonathan Tow's avatar
Jonathan Tow committed
186
187
188
189
190
191
192
193
    def training_docs(self):
        if self._training_docs is None:
            self._training_docs = list(self.dataset["train"])
        return self._training_docs

    def validation_docs(self):
        return self.dataset["validation"]

194
    def doc_to_text(self, doc):
Jason Phang's avatar
Jason Phang committed
195
        # Drop the period
Jason Phang's avatar
Jason Phang committed
196
197
198
199
        connector = {
            "cause": "because",
            "effect": "therefore",
        }[doc["question"]]
200
        return doc["premise"].strip()[:-1] + f" {connector}"
Jason Phang's avatar
Jason Phang committed
201

thefazzer's avatar
thefazzer committed
202
    def doc_to_target(self, doc):
203
204
        correct_choice = doc["choice1"] if doc["label"] == 0 else doc["choice2"]
        # Connect the sentences
205
        return " " + self.convert_choice(correct_choice)
thefazzer's avatar
thefazzer committed
206
207

    def construct_requests(self, doc, ctx):
thefazzer's avatar
thefazzer committed
208
209
        choice1 = " " + self.convert_choice(doc["choice1"])
        choice2 = " " + self.convert_choice(doc["choice2"])
thefazzer's avatar
thefazzer committed
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
        
        ll_choice1, _ = rf.loglikelihood(ctx, choice1)
        ll_choice2, _ = rf.loglikelihood(ctx, choice2)

        return ll_choice1, ll_choice2

    def process_results(self, doc, results):
        gold = doc["label"]
        pred = np.argmax(results)
        acc = 1. if pred == gold else 0.

        return {
            "acc": acc
        }
    
    def higher_is_better(self):
        return {
            "acc": True
        }
    
    def aggregation(self):
        return {
            "acc": mean
        }
Jason Phang's avatar
Jason Phang committed
234
235
236
237
238
239

    @staticmethod
    def convert_choice(choice):
        return choice[0].lower() + choice[1:]


Jonathan Tow's avatar
Jonathan Tow committed
240
class MultiRC(Task):
241
    VERSION = 1
Leo Gao's avatar
Leo Gao committed
242
243
    DATASET_PATH = "super_glue"
    DATASET_NAME = "multirc"
Jason Phang's avatar
multirc  
Jason Phang committed
244
245
246
247
248
249
250
251

    def has_training_docs(self):
        return True

    def has_validation_docs(self):
        return True

    def has_test_docs(self):
252
        return False
Jason Phang's avatar
multirc  
Jason Phang committed
253

Jonathan Tow's avatar
Jonathan Tow committed
254
255
256
257
258
259
260
261
    def training_docs(self):
        if self._training_docs is None:
            self._training_docs = list(self.dataset["train"])
        return self._training_docs

    def validation_docs(self):
        return self.dataset["validation"]

262
    def doc_to_text(self, doc):
Leo Gao's avatar
Leo Gao committed
263
        return f"{doc['paragraph']}\nQuestion: {doc['question']}\nAnswer:"
264
265

    def doc_to_target(self, doc):
Leo Gao's avatar
Leo Gao committed
266
        return " " + self.format_answer(answer=doc["answer"], label=doc["label"])
Jason Phang's avatar
multirc  
Jason Phang committed
267
268
269

    @staticmethod
    def format_answer(answer, label):
Leo Gao's avatar
Fix  
Leo Gao committed
270
        label_str = "yes" if label else "no"
thomasw21's avatar
thomasw21 committed
271
        return f"{answer}\nIs the answer correct? {label_str}"
Jason Phang's avatar
multirc  
Jason Phang committed
272

thefazzer's avatar
thefazzer committed
273
274
275
276
277
278
279
280
281
282
    def construct_requests(self, doc, ctx):
        true_choice = self.format_answer(answer=doc["answer"], label=True)
        false_choice = self.format_answer(answer=doc["answer"], label=False)
        
        ll_true_choice, _ = rf.loglikelihood(ctx, f' {true_choice}')
        ll_false_choice, _ = rf.loglikelihood(ctx, f' {false_choice}')

        return ll_true_choice, ll_false_choice

    def process_results(self, doc, results):
thomasw21's avatar
thomasw21 committed
283
284
        ll_true_choice, ll_false_choice = results
        pred = ll_true_choice > ll_false_choice
Jason Phang's avatar
multirc  
Jason Phang committed
285
        return {
thefazzer's avatar
thefazzer committed
286
287
288
289
290
291
292
293
294
295
296
            "acc": (pred, doc)
        }
    
    def higher_is_better(self):
        return {
            "acc": True
        }
    
    def aggregation(self):
        return {
            "acc": acc_all
Jason Phang's avatar
multirc  
Jason Phang committed
297
298
        }

Jason Phang's avatar
Jason Phang committed
299

Jonathan Tow's avatar
Jonathan Tow committed
300
class ReCoRD(Task):
Leo Gao's avatar
Leo Gao committed
301
    VERSION = 0
Jason Phang's avatar
Jason Phang committed
302
303
304
305
306
307
308
309
310
311
    DATASET_PATH = "super_glue"
    DATASET_NAME = "record"

    def has_training_docs(self):
        return True

    def has_validation_docs(self):
        return True

    def has_test_docs(self):
Leo Gao's avatar
Leo Gao committed
312
        return False
Jason Phang's avatar
Jason Phang committed
313
314
315
316

    def training_docs(self):
        # In ReCoRD, each doc manifests multiple "examples" in the context of few shot example packing.
        # Each doc consists of multiple answer candidates, each of which is scored yes/no.
317
318
        if self._training_docs is None:
            self._training_docs = []
Jonathan Tow's avatar
Jonathan Tow committed
319
            for doc in self.dataset["train"]:
Jason Phang's avatar
Jason Phang committed
320
                self._training_docs.append(self._process_doc(doc))
321
322
323
        return self._training_docs

    def validation_docs(self):
Jason Phang's avatar
Jason Phang committed
324
        # See: training_docs
Jonathan Tow's avatar
Jonathan Tow committed
325
        for doc in self.dataset["validation"]:
Jason Phang's avatar
Jason Phang committed
326
327
328
329
330
331
332
333
334
335
            yield self._process_doc(doc)

    @classmethod
    def _process_doc(cls, doc):
        return {
            "passage": doc["passage"],
            "query": doc["query"],
            "entities": sorted(list(set(doc["entities"]))),
            "answers": sorted(list(set(doc["answers"]))),
        }
Jason Phang's avatar
Jason Phang committed
336
337
338
339
340
341
342
343
344
345
346
347
348

    def doc_to_text(self, doc):
        initial_text, *highlights = doc["passage"].strip().split("\n@highlight\n")
        text = initial_text + "\n\n"
        for highlight in highlights:
            text += f"  - {highlight}.\n"
        return text

    @classmethod
    def format_answer(cls, query, entity):
        return f'  - {query}'.replace("@placeholder", entity)

    def doc_to_target(self, doc):
Jason Phang's avatar
Jason Phang committed
349
350
        # We only output the first correct entity in a doc
        return self.format_answer(query=doc["query"], entity=doc["answers"][0])
Jason Phang's avatar
Jason Phang committed
351
352
353
354

    def construct_requests(self, doc, ctx):
        requests = [
            rf.loglikelihood(ctx, self.format_answer(query=doc["query"], entity=entity))
Jason Phang's avatar
Jason Phang committed
355
            for entity in doc["entities"]
Jason Phang's avatar
Jason Phang committed
356
357
358
359
360
361
362
363
        ]
        return requests

    def process_results(self, doc, results):
        # ReCoRD's evaluation is actually deceptively simple:
        # - Pick the maximum likelihood prediction entity
        # - Evaluate the accuracy and token F1 PER EXAMPLE
        # - Average over all examples
Jason Phang's avatar
Jason Phang committed
364
        max_idx = np.argmax(np.array([result[0] for result in results]))
Leo Gao's avatar
Leo Gao committed
365

Jason Phang's avatar
Jason Phang committed
366
        prediction = doc["entities"][max_idx]
Jason Phang's avatar
Jason Phang committed
367
        gold_label_set = doc["answers"]
Jason Phang's avatar
Jason Phang committed
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
        f1 = metric_max_over_ground_truths(squad_metrics.compute_f1, prediction, gold_label_set)
        em = metric_max_over_ground_truths(squad_metrics.compute_exact, prediction, gold_label_set)

        return {
            "f1": f1,
            "em": em,
        }

    def higher_is_better(self):
        return {
            "f1": True,
            "em": True,
        }

    def aggregation(self):
        return {
            "f1": mean,
            "em": mean,
        }


Jonathan Tow's avatar
Jonathan Tow committed
389
class WordsInContext(Task):
Leo Gao's avatar
Leo Gao committed
390
    VERSION = 0
Leo Gao's avatar
Leo Gao committed
391
392
    DATASET_PATH = "super_glue"
    DATASET_NAME = "wic"
Jason Phang's avatar
Jason Phang committed
393
394
395
396
397
398
399
400

    def has_training_docs(self):
        return True

    def has_validation_docs(self):
        return True

    def has_test_docs(self):
401
        return False
Jason Phang's avatar
Jason Phang committed
402

Jonathan Tow's avatar
Jonathan Tow committed
403
404
405
406
407
408
409
410
    def training_docs(self):
        if self._training_docs is None:
            self._training_docs = list(self.dataset["train"])
        return self._training_docs

    def validation_docs(self):
        return self.dataset["validation"]

411
    def doc_to_text(self, doc):
Leo Gao's avatar
Leo Gao committed
412
413
        return "Sentence 1: {}\nSentence 2: {}\nQuestion: Is the word '{}' used in the same way in the" \
               " two sentences above?\nAnswer:".format(
Jason Phang's avatar
Jason Phang committed
414
415
416
417
                    doc["sentence1"],
                    doc["sentence2"],
                    doc["sentence1"][doc["start1"]:doc["end1"]],
                )
418
419
420

    def doc_to_target(self, doc):
        return " {}".format({0: "no", 1: "yes"}[doc["label"]])
Jason Phang's avatar
Jason Phang committed
421

Jason Phang's avatar
Jason Phang committed
422
423
424
425
426
    def construct_requests(self, doc, ctx):
        ll_yes, _ = rf.loglikelihood(ctx, ' yes')
        ll_no, _ = rf.loglikelihood(ctx, ' no')

        return ll_yes, ll_no
427

Jason Phang's avatar
Jason Phang committed
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
    def process_results(self, doc, results):
        ll_yes, ll_no = results
        gold = doc["label"]

        acc = 1. if (ll_yes > ll_no) == gold else 0.

        return {
            "acc": acc
        }

    def higher_is_better(self):
        return {
            "acc": True
        }

    def aggregation(self):
        return {
            "acc": mean
        }
Jason Phang's avatar
Jason Phang committed
447
448


Jonathan Tow's avatar
Jonathan Tow committed
449
class SGWinogradSchemaChallenge(Task):
Leo Gao's avatar
Leo Gao committed
450
    VERSION = 0
Jason Phang's avatar
wsc  
Jason Phang committed
451
452
    # Note: This implementation differs from Fig G.32 because this is the SuperGLUE,
    #       binary version of the task.
Leo Gao's avatar
Leo Gao committed
453
454
    DATASET_PATH = "super_glue"
    DATASET_NAME = "wsc"
Jason Phang's avatar
Jason Phang committed
455
456
457
458
459
460
461
462

    def has_training_docs(self):
        return True

    def has_validation_docs(self):
        return True

    def has_test_docs(self):
463
        return False
Jason Phang's avatar
Jason Phang committed
464
465
466
467

    def training_docs(self):
        if self.has_training_docs():
            if self._training_docs is None:
Jason Phang's avatar
Jason Phang committed
468
                # GPT-3 Paper's format only uses positive examples for fewshot "training"
Jason Phang's avatar
Jason Phang committed
469
470
                self._training_docs = [
                    doc for doc in
Jonathan Tow's avatar
Jonathan Tow committed
471
                    self.dataset["train"]
Jason Phang's avatar
Jason Phang committed
472
473
474
475
                    if doc["label"]
                ]
            return self._training_docs

Jonathan Tow's avatar
Jonathan Tow committed
476
477
478
    def validation_docs(self):
        return self.dataset["validation"]

479
    def doc_to_text(self, doc):
Jason Phang's avatar
Jason Phang committed
480
        raw_passage = doc["text"]
Jonathan Tow's avatar
Jonathan Tow committed
481
482
483
        # NOTE: HuggingFace span indices are word-based not character-based.
        pre = " ".join(raw_passage.split()[:doc["span2_index"]])
        post = raw_passage[len(pre) + len(doc["span2_text"]) + 1:]
Leo Gao's avatar
Leo Gao committed
484
        passage = general_detokenize(pre + " *{}*".format(doc['span2_text']) + post)
Jason Phang's avatar
wsc  
Jason Phang committed
485
        noun = doc["span1_text"]
Jason Phang's avatar
Jason Phang committed
486
487
488
        pronoun = doc["span2_text"]
        text = (
            f"Passage: {passage}\n"
Jason Phang's avatar
wsc  
Jason Phang committed
489
            + f"Question: In the passage above, does the pronoun \"*{pronoun}*\" refer to \"*{noun}*\"?\n"
Jason Phang's avatar
Jason Phang committed
490
491
492
493
            + "Answer:"
        )
        return text

494
    def doc_to_target(self, doc):
Leo Gao's avatar
Leo Gao committed
495
        return " " + yesno(doc['label'])
496

Leo Gao's avatar
Leo Gao committed
497
    def construct_requests(self, doc, ctx):
Jason Phang's avatar
wsc  
Jason Phang committed
498
499
500
501
502

        ll_yes, _ = rf.loglikelihood(ctx, ' yes')
        ll_no, _ = rf.loglikelihood(ctx, ' no')

        return ll_yes, ll_no
503

Jason Phang's avatar
Jason Phang committed
504
    def process_results(self, doc, results):
Jason Phang's avatar
wsc  
Jason Phang committed
505
506
507
508
509
510
511
512
        ll_yes, ll_no = results
        gold = doc["label"]

        acc = 1. if (ll_yes > ll_no) == gold else 0.

        return {
            "acc": acc
        }
Anish Thite's avatar
Anish Thite committed
513

Leo Gao's avatar
Leo Gao committed
514
    def higher_is_better(self):
Jason Phang's avatar
Jason Phang committed
515
516
517
518
519
520
521
522
        return {
            "acc": True
        }

    def aggregation(self):
        return {
            "acc": mean
        }