superglue.py 11.6 KB
Newer Older
Leo Gao's avatar
Leo Gao committed
1
2
# REMINDER: this code needs to be rewritten for the new framework. Remove this comment when the code is fully converted.

Jason Phang's avatar
Jason Phang committed
3
4
import numpy as np
from tqdm import auto as tqdm_lib
5
from . common import HFTask, simple_accuracy_metric, yesno
Leo Gao's avatar
Update  
Leo Gao committed
6
from lm_eval.base import rf, mean
Jason Phang's avatar
Jason Phang committed
7

8
class BoolQ(HFTask):
Leo Gao's avatar
Leo Gao committed
9
10
    DATASET_PATH = "super_glue"
    DATASET_NAME = "boolq"
Jason Phang's avatar
Jason Phang committed
11
12
13
14
15
16
17
18
19
20
21

    def has_training_docs(self):
        return True

    def has_validation_docs(self):
        return True

    def has_test_docs(self):
        return True

    def fewshot_description(self):
22
        # TODO: figure out actual description
Jason Phang's avatar
Jason Phang committed
23
24
        return "Read the following passages and answer each question with a yes or a no."

Leo Gao's avatar
Update  
Leo Gao committed
25
26
27
28
29
    def doc_to_text(self, doc):
        return f"{doc['passage']}\nquestion: {doc['question']}\nanswer: "
    
    def doc_to_target(self, doc):
        return yesno(doc['label']) 
Jason Phang's avatar
Jason Phang committed
30

31
    def construct_requests(self, doc, ctx):
Leo Gao's avatar
Update  
Leo Gao committed
32

33
34
        ll_yes, _ = rf.loglikelihood(ctx, ' yes')
        ll_no , _ = rf.loglikelihood(ctx, ' no')
Leo Gao's avatar
Update  
Leo Gao committed
35
36
37
38
39
40
41
42
43

        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.

44
45
46
47
48
49
50
51
52
53
54
55
56
        return {
            "acc": acc
        }
    
    def higher_is_better(self):
        return {
            "acc": True
        }
    
    def aggregation(self):
        return {
            "acc": mean
        }
Jason Phang's avatar
Jason Phang committed
57
58


59
class CommitmentBank(HFTask):
Leo Gao's avatar
Leo Gao committed
60
61
    DATASET_PATH = "super_glue"
    DATASET_NAME = "cb"
Jason Phang's avatar
Jason Phang committed
62
63
64
65
66
67
68
69
70
71

    def has_training_docs(self):
        return True

    def has_validation_docs(self):
        return True

    def has_test_docs(self):
        return True

72
73
    def doc_to_text(self, doc):
        return "{}\nquestion:\t{}\ttrue, false or neither?\nanswer:".format(
Jason Phang's avatar
Jason Phang committed
74
75
76
            doc["premise"],
            doc["hypothesis"],
        )
77
78
79
80
81
82

    def doc_to_target(self, doc):
        # True = entailment
        # False = contradiction
        # Neither = neutral
        return " {}".format({0: "true", 1: "neither", 2: "false"}[doc["label"]])
Jason Phang's avatar
Jason Phang committed
83
84

    def evaluate(self, docs, lm, provide_description, num_fewshot):
85
86
87
88
89
        # TODO: Implement evaluation code using new framework

        # ***IMPORTANT***: this evaluation function needs to be rewritten for the new framework. 
        # For more info, check out the interface in base.py and the example BoolQ implementation in superglue.py. 
        # Remove this comment when the evaluation code is implemented.
Jason Phang's avatar
Jason Phang committed
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
        golds = [doc["label"] for doc in docs]
        preds = []
        for doc in tqdm_lib.tqdm(docs):
            ctx = self.fewshot_context(
                doc=doc,
                provide_description=provide_description,
                num_fewshot=num_fewshot,
            )
            probs = np.array([
                lm.loglikelihood(ctx, ' true'),
                lm.loglikelihood(ctx, ' neither'),
                lm.loglikelihood(ctx, ' false'),
            ])
            preds.append(np.argmax(probs))
        return simple_accuracy_metric(preds=preds, golds=golds)


107
class Copa(HFTask):
Leo Gao's avatar
Leo Gao committed
108
109
    DATASET_PATH = "super_glue"
    DATASET_NAME = "copa"
Jason Phang's avatar
Jason Phang committed
110
111
112
113
114
115
116
117
118
119

    def has_training_docs(self):
        return True

    def has_validation_docs(self):
        return True

    def has_test_docs(self):
        return True

120
    def doc_to_text(self, doc):
Jason Phang's avatar
Jason Phang committed
121
        # Drop the period
Jason Phang's avatar
Jason Phang committed
122
123
124
125
        connector = {
            "cause": "because",
            "effect": "therefore",
        }[doc["question"]]
126
127
128
129
130
131
        return doc["premise"].strip()[:-1] + f" {connector} "

    def doc_to_target(self, doc):
        correct_choice = doc["choice1"] if doc["label"] == 0 else doc["choice2"]
        # Connect the sentences
        return self.convert_choice(correct_choice)
Jason Phang's avatar
Jason Phang committed
132
133

    def evaluate(self, docs, lm, provide_description, num_fewshot):
134
135
136
137
138
        # TODO: Implement evaluation code using new framework

        # ***IMPORTANT***: this evaluation function needs to be rewritten for the new framework. 
        # For more info, check out the interface in base.py and the example BoolQ implementation in superglue.py. 
        # Remove this comment when the evaluation code is implemented.
Jason Phang's avatar
Jason Phang committed
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
        golds = [doc["label"] for doc in docs]
        preds = []
        for doc in tqdm_lib.tqdm(docs):
            ctx = self.fewshot_context(
                doc=doc,
                provide_description=provide_description,
                num_fewshot=num_fewshot,
            )
            choice1 = " " + self.convert_choice(doc["choice1"])
            choice2 = " " + self.convert_choice(doc["choice2"])
            preds.append(lm.loglikelihood(ctx, choice2) > lm.loglikelihood(ctx, choice1))
        return simple_accuracy_metric(preds=preds, golds=golds)

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


157
class MultiRC(HFTask):
Leo Gao's avatar
Leo Gao committed
158
159
    DATASET_PATH = "super_glue"
    DATASET_NAME = "multirc"
Jason Phang's avatar
multirc  
Jason Phang committed
160
161
162
163
164
165
166
167
168
169
170
171
172

    def has_training_docs(self):
        return True

    def has_validation_docs(self):
        return True

    def has_test_docs(self):
        return True

    def fewshot_description(self):
        return "READING COMPREHENSION ANSWER KEY"

173
174
175
176
177
    def doc_to_text(self, doc):
        return f"{doc['paragraph']}\n\n{doc['question']}\n"

    def doc_to_target(self, doc):
        return self.format_answer(answer=doc["answer"], label=doc["label"])
Jason Phang's avatar
multirc  
Jason Phang committed
178
179
180
181
182
183
184

    @staticmethod
    def format_answer(answer, label):
        label_str = "True" if label else "False"
        return f"[{label_str}] {answer}"

    def evaluate(self, docs, lm, provide_description, num_fewshot):
185
186
187
188
189
        # TODO: Implement evaluation code using new framework

        # ***IMPORTANT***: this evaluation function needs to be rewritten for the new framework. 
        # For more info, check out the interface in base.py and the example BoolQ implementation in superglue.py. 
        # Remove this comment when the evaluation code is implemented.
Jason Phang's avatar
multirc  
Jason Phang committed
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
        preds = []
        for doc in docs:
            ctx = self.fewshot_context(
                doc=doc,
                provide_description=provide_description,
                num_fewshot=num_fewshot,
            )
            true_choice = self.format_answer(answer=doc["answer"], label=True)
            false_choice = self.format_answer(answer=doc["answer"], label=False)
            preds.append(
                lm.loglikelihood(ctx, f' {true_choice}')
                > lm.loglikelihood(ctx, f' {false_choice}')
            )

        # Only count as correct if all answers are labeled correctly for each question
        question_scoring_dict = {}
        for doc, pred in zip(docs, preds):
            question_id = doc["idx"]["question"]
            if question_id not in question_scoring_dict:
                question_scoring_dict[question_id] = []
            gold_label = doc["label"] == 1
            question_scoring_dict[question_id].append(gold_label == pred)
        acc = np.mean([int(all(x)) for x in question_scoring_dict.values()])
        return {
            "major": acc,
            "minor": {"acc": acc},
            "higher_is_better": True,
        }


220
class WordsInContext(HFTask):
Leo Gao's avatar
Leo Gao committed
221
222
    DATASET_PATH = "super_glue"
    DATASET_NAME = "wic"
Jason Phang's avatar
Jason Phang committed
223
224
225
226
227
228
229
230
231
232

    def has_training_docs(self):
        return True

    def has_validation_docs(self):
        return True

    def has_test_docs(self):
        return True

233
234
    def doc_to_text(self, doc):
        return "{}\n{}\nquestion\tIs the word '{}' used in the same way in the" \
Jason Phang's avatar
Jason Phang committed
235
236
237
238
239
               " two sentences above?\nanswer:".format(
                    doc["sentence1"],
                    doc["sentence2"],
                    doc["sentence1"][doc["start1"]:doc["end1"]],
                )
240
241
242

    def doc_to_target(self, doc):
        return " {}".format({0: "no", 1: "yes"}[doc["label"]])
Jason Phang's avatar
Jason Phang committed
243
244

    def evaluate(self, docs, lm, provide_description, num_fewshot):
245
246
247
248
249
        # TODO: Implement evaluation code using new framework

        # ***IMPORTANT***: this evaluation function needs to be rewritten for the new framework. 
        # For more info, check out the interface in base.py and the example BoolQ implementation in superglue.py. 
        # Remove this comment when the evaluation code is implemented.
Jason Phang's avatar
Jason Phang committed
250
251
252
253
254
255
256
257
258
259
260
261
        golds = [doc["label"] for doc in docs]
        preds = []
        for doc in tqdm_lib.tqdm(docs):
            ctx = self.fewshot_context(
                doc=doc,
                provide_description=provide_description,
                num_fewshot=num_fewshot,
            )
            preds.append(lm.loglikelihood(ctx, ' yes') > lm.loglikelihood(ctx, ' no'))
        return simple_accuracy_metric(preds=preds, golds=golds)


262
class SGWinogradSchemaChallenge(HFTask):
Leo Gao's avatar
Leo Gao committed
263
264
    DATASET_PATH = "super_glue"
    DATASET_NAME = "wsc"
Jason Phang's avatar
Jason Phang committed
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

    def has_training_docs(self):
        return True

    def has_validation_docs(self):
        return True

    def has_test_docs(self):
        return True

    def training_docs(self):
        if self.has_training_docs():
            if self._training_docs is None:
                # GPT-3 Paper's format only uses positive examples
                self._training_docs = [
                    doc for doc in
                    self._load_nlp_dataset()["train"]
                    if doc["label"]
                ]
            return self._training_docs

    def fewshot_description(self):
        return "Final Exam with Answer Key\n" \
           "Instructions: Please carefully read the following passages. " \
           "For each passage, you must identify which noun the pronoun marked in *bold*" \
           " refers to.\n====="

292
    def doc_to_text(self, doc):
Jason Phang's avatar
Jason Phang committed
293
294
295
296
297
298
299
300
301
302
303
304
305
306
        raw_passage = doc["text"]
        passage = (
            raw_passage[:doc["span2_index"]]
            + "*{}*".format(doc["span2_text"])
            + raw_passage[doc["span2_index"] + len(doc["span2_text"]):]
        )
        pronoun = doc["span2_text"]
        text = (
            f"Passage: {passage}\n"
            + f"Question: In the passage above, what does the pronoun \"*{pronoun}*\" refer to?\n"
            + "Answer:"
        )
        return text

307
308
309
    def doc_to_target(self, doc):
        return " {}".format(doc["span1_text"])

Jason Phang's avatar
Jason Phang committed
310
    def evaluate(self, docs, lm, provide_description, num_fewshot):
311
312
313
314
315
        # TODO: Implement evaluation code using new framework

        # ***IMPORTANT***: this evaluation function needs to be rewritten for the new framework. 
        # For more info, check out the interface in base.py and the example BoolQ implementation in superglue.py. 
        # Remove this comment when the evaluation code is implemented.
Jason Phang's avatar
Jason Phang committed
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
        golds = [doc["label"] for doc in docs]
        preds = []
        for doc in tqdm_lib.tqdm(docs):
            ctx = self.fewshot_context(
                doc=doc,
                provide_description=provide_description,
                num_fewshot=num_fewshot,
            )
            to_predict = " " + doc["span1_text"]
            num_tokens = len(lm.tokenizer.tokenize(to_predict))
            generated = lm.generate(
                context=ctx,
                max_gen_length=num_tokens,
            )
            preds.append(1 if generated == to_predict else 0)
        return simple_accuracy_metric(preds=preds, golds=golds)
Anish Thite's avatar
Anish Thite committed
332
333
334
335
336
337
338
339
340

class RTE(HFTask):
    DATASET_PATH = "super_glue"
    DATASET_NAME = "rte"

    def fewshot_description(self):
        #TODO: implement
        pass

341
342
343
344
345
346
    def doc_to_text(self, doc):
        return ''.join([doc['premise'], '\nquestion: ',doc['hypothesis'], ' True or False?\nanswer: '])

    def doc_to_target(self, doc):
        return 'True' if doc['label'] == 0 else 'False'

347
348
349
350
351
    # TODO: Implement evaluation code

    # ***IMPORTANT***: this evaluation function needs to be written for the new framework. 
    # For more info, check out the interface in base.py and the example BoolQ implementation in superglue.py. 
    # Remove this comment when the evaluation code is implemented.
Anish Thite's avatar
Anish Thite committed
352