base.py 33.9 KB
Newer Older
Leo Gao's avatar
Leo Gao committed
1
import abc
2
from typing import Iterable
thefazzer's avatar
thefazzer committed
3
import numpy as np
4
import random
Leo Gao's avatar
Leo Gao committed
5
import re
6
7
8
import os
import json
import hashlib
Jonathan Tow's avatar
Jonathan Tow committed
9
import datasets
10
from sqlitedict import SqliteDict
11
from tqdm import tqdm
12
import torch
Leo Gao's avatar
Leo Gao committed
13
import torch.nn.functional as F
14
from accelerate import find_executable_batch_size
&'s avatar
& committed
15

16
from lm_eval.metrics import mean, weighted_perplexity, weighted_mean, bits_per_byte
17
from lm_eval import utils
18
from abc import abstractmethod
Jason Phang's avatar
gpt3  
Jason Phang committed
19

Jason Phang's avatar
Jason Phang committed
20

Leo Gao's avatar
Leo Gao committed
21
class LM(abc.ABC):
Leo Gao's avatar
Leo Gao committed
22
23
24
    def __init__(self):
        self.cache_hook = CacheHook(None)

25
    @abstractmethod
Leo Gao's avatar
Leo Gao committed
26
    def loglikelihood(self, requests):
Leo Gao's avatar
Leo Gao committed
27
        """Compute log-likelihood of generating a continuation from a context.
Fabrizio Milo's avatar
Fabrizio Milo committed
28
        Downstream tasks should attempt to use loglikelihood instead of other
Leo Gao's avatar
Leo Gao committed
29
        LM calls whenever possible.
Jason Phang's avatar
gpt3  
Jason Phang committed
30

Leo Gao's avatar
Leo Gao committed
31
32
33
        :param requests: list
            A list of pairs (context, continuation)
            context: str
Fabrizio Milo's avatar
Fabrizio Milo committed
34
                Context string. Implementations of LM must be able to handle an
Leo Gao's avatar
Leo Gao committed
35
                empty context string.
Leo Gao's avatar
Leo Gao committed
36
            continuation: str
Fabrizio Milo's avatar
Fabrizio Milo committed
37
38
                The continuation over which log likelihood will be calculated. If
                there is a word boundary, the space should be in the continuation.
Leo Gao's avatar
Leo Gao committed
39
40
41
42
                For example, context="hello" continuation=" world" is correct.
        :return: list
            A list of pairs (logprob, isgreedy)
            logprob: float
Jason Phang's avatar
Jason Phang committed
43
                The log probability of `continuation`
Leo Gao's avatar
Leo Gao committed
44
            isgreedy:
Jason Phang's avatar
Jason Phang committed
45
46
47
48
                Whether `continuation` would be generated by greedy sampling from `context`
        """
        pass

49
    @abstractmethod
Leo Gao's avatar
Leo Gao committed
50
    def loglikelihood_rolling(self, requests):
Jason Phang's avatar
Jason Phang committed
51
52
53
54
        """Compute full log-likelihood of a string, with no truncation, for perplexity computation
        - We will use the full max context length of the model.
        - For inputs that exceed the max context length, we divide the tokenized string into chunks of up to
        the max context length.
Fabrizio Milo's avatar
Fabrizio Milo committed
55
        - IMPORTANT: Each document's loglikelihood/perplexity is computed *separately*, unlike other implementations
Jason Phang's avatar
Jason Phang committed
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
          which may simply concatenate multiple documents together.
        - IMPORTANT: We maximize the amount of context for each prediction. Specifically, for inputs that we break into
          multiple chunks, the last input will still a full-sized context.
          Example:
            Input tokens: [ 0 1 2 3 4 5 6 7 8 9 ]
            Prefix: EOT
            Max context length: 4
            Resulting input/prediction pairs:

                INPUT:  EOT   0   1   2
                PRED:     0   1   2   3

                INPUT:    3   4   5   6
                PRED:     4   5   6   7

                INPUT:    5   6   7   8
                PRED:             8   9

          Observe that:
            1. Each token is predicted exactly once
            2. For the last pair, we provide the full context, but only score the last two tokens

        :param requests: list
            A list of strings
            string: str
                String for which we are computing per-toke  loglikelihood
        :return: list
            A list of pairs (logprob, isgreedy)
            logprob: float
                The log probability of `continuation`
            isgreedy:
                Whether `continuation` would be generated by greedy sampling from `context`
Leo Gao's avatar
Leo Gao committed
88
89
90
        """
        pass

&'s avatar
& committed
91
    # TODO: Add an optional max length
92
    @abstractmethod
Leo Gao's avatar
Update  
Leo Gao committed
93
    def greedy_until(self, requests):
Leo Gao's avatar
Leo Gao committed
94
95
96
97
98
99
        """Generate greedily until a stopping sequence

        :param requests: list
            A list of pairs (context, until)
            context: str
                Context string
Leo Gao's avatar
Leo Gao committed
100
            until: [str]
Fabrizio Milo's avatar
Fabrizio Milo committed
101
                The string sequences to generate until. These string sequences
Leo Gao's avatar
Leo Gao committed
102
                may each span across multiple tokens, or may be part of one token.
Leo Gao's avatar
Leo Gao committed
103
104
105
106
        :return: list
            A list of strings continuation
            continuation: str
                The generated continuation.
Jason Phang's avatar
gpt3  
Jason Phang committed
107
        """
Leo Gao's avatar
Leo Gao committed
108
109
        pass

Jason Phang's avatar
gpt3  
Jason Phang committed
110
    @classmethod
111
112
    def create_from_arg_string(cls, arg_string, additional_config=None):
        additional_config = {} if additional_config is None else additional_config
113
114
115
        args = utils.simple_parse_args_string(arg_string)
        args2 = {k: v for k, v in additional_config.items() if v is not None}
        return cls(**args, **args2)
Jason Phang's avatar
gpt3  
Jason Phang committed
116

Leo Gao's avatar
Leo Gao committed
117
118
119
    def set_cache_hook(self, cache_hook):
        self.cache_hook = cache_hook

Leo Gao's avatar
Leo Gao committed
120

121
class BaseLM(LM):
122
123
124
125
126
127
    def __init__(self):
        super().__init__()
        self.batch_schedule = 1
        self.batch_sizes = {}
        self.max_batch_size = 512

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
    @property
    @abstractmethod
    def eot_token_id(self):
        pass

    @property
    @abstractmethod
    def max_length(self):
        pass

    @property
    @abstractmethod
    def max_gen_toks(self):
        pass

    @property
    @abstractmethod
    def batch_size(self):
        pass

    @property
    @abstractmethod
    def device(self):
        pass

153
    @abstractmethod
Fabrizio Milo's avatar
Fabrizio Milo committed
154
155
156
    def tok_encode(self, string: str):
        pass

157
    @abstractmethod
Fabrizio Milo's avatar
Fabrizio Milo committed
158
159
    def tok_decode(self, tokens: Iterable[int]):
        pass
Jason Phang's avatar
gpt3  
Jason Phang committed
160

161
    @abstractmethod
Fabrizio Milo's avatar
Fabrizio Milo committed
162
163
    def _model_generate(self, context, max_length, eos_token_id):
        pass
Jason Phang's avatar
gpt3  
Jason Phang committed
164

165
166
    @abstractmethod
    def _model_call(self, inps):
Jason Phang's avatar
gpt3  
Jason Phang committed
167
        """
168
169
        inps: a torch tensor of shape [batch, sequence]
        the size of sequence may vary from call to call
Jason Phang's avatar
gpt3  
Jason Phang committed
170

171
        returns: a torch tensor of shape [batch, sequence, vocab] with the
172
        logits returned from the model
173
174
        """
        pass
175

176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
    def _detect_batch_size(self, requests=None, pos=0):
        if requests:
            _, context_enc, continuation_enc = requests[pos]
            max_length = len((context_enc + continuation_enc)[-(self.max_length + 1) :][:-1])
        else:
            max_length = self.max_length

        # if OOM, then halves batch_size and tries again
        @find_executable_batch_size(starting_batch_size=self.max_batch_size)
        def forward_batch(batch_size):
            test_batch = torch.ones((batch_size, max_length), device=self.device).long()
            for _ in range(5):
                _ = F.log_softmax(self._model_call(test_batch), dim=-1).cpu()
            return batch_size

        batch_size = forward_batch()
        utils.clear_torch_cache()

        return batch_size

Leo Gao's avatar
Leo Gao committed
196
    # subclass must implement properties vocab_size, eot_token_id, max_gen_toks, batch_size, device, max_length.
197
198
    # TODO: enforce this somehow

gakada's avatar
gakada committed
199
    def _encode_pair(self, context, continuation):
200
201
202
203
        n_spaces = len(context) - len(context.rstrip())
        if n_spaces > 0:
            continuation = context[-n_spaces:] + continuation
            context = context[:-n_spaces]
gakada's avatar
gakada committed
204
205
206
        whole_enc = self.tok_encode(context + continuation)
        context_enc = self.tok_encode(context)
        context_enc_len = len(context_enc)
207
        continuation_enc = whole_enc[context_enc_len:]
gakada's avatar
gakada committed
208
209
        return context_enc, continuation_enc

210
211
212
213
214
    def loglikelihood(self, requests):
        new_reqs = []
        for context, continuation in requests:
            if context == "":
                # end of text as context
gakada's avatar
gakada committed
215
                context_enc, continuation_enc = [self.eot_token_id], self.tok_encode(continuation)
216
            else:
gakada's avatar
gakada committed
217
                context_enc, continuation_enc = self._encode_pair(context, continuation)
218
219
220
221
222
223
224

            new_reqs.append(((context, continuation), context_enc, continuation_enc))

        return self._loglikelihood_tokens(new_reqs)

    def loglikelihood_rolling(self, requests):
        # TODO: Implement caching once we've confirmed the perplexity implementation
225

226
227
        # automatic batch size detection for vectorization
        adaptive_batch_size = None
228
        if self.batch_size == "auto":
229
            # using rolling window with maximum context
230
            print("Passed argument batch_size = auto. Detecting largest batch size")
231
            batch_size = self._detect_batch_size()
232
233
            print(f"Determined Largest batch size: {batch_size}")
            adaptive_batch_size = batch_size
234
235

        loglikelihoods = []
Fabrizio Milo's avatar
Fabrizio Milo committed
236
237
238
239
240
241
242
243
244
245
246
247
        for (string,) in tqdm(requests):
            rolling_token_windows = list(
                map(
                    utils.make_disjoint_window,
                    utils.get_rolling_token_windows(
                        token_list=self.tok_encode(string),
                        prefix_token=self.eot_token_id,
                        max_seq_len=self.max_length,
                        context_len=1,
                    ),
                )
            )
248
249
250

            rolling_token_windows = [(None,) + x for x in rolling_token_windows]

251
252
            # TODO: extract out this call so it only gets called once and also somehow figure out partial caching for
            # that
Fabrizio Milo's avatar
Fabrizio Milo committed
253
            string_nll = self._loglikelihood_tokens(
254
255
256
                rolling_token_windows,
                disable_tqdm=True,
                override_bs=adaptive_batch_size,
Fabrizio Milo's avatar
Fabrizio Milo committed
257
258
            )

259
260
            # discard is_greedy
            string_nll = [x[0] for x in string_nll]
Fabrizio Milo's avatar
Fabrizio Milo committed
261

262
263
264
265
266
            string_nll = sum(string_nll)
            loglikelihoods.append(string_nll)

        return loglikelihoods

267
    def _loglikelihood_tokens(self, requests, disable_tqdm=False, override_bs=None):
268
269
270
271
272
273
        # TODO: implement some kind of efficient-request-middleware that lumps together requests with the same context
        res = []

        def _collate(x):
            # the negative sign on len(toks) sorts descending - this has a few advantages:
            # - time estimates will always be over not underestimates, which is more useful for planning
274
275
276
            # - to know the size of a batch when going through the list, you know the first one is always the batch
            #   padded context length. this is useful to simplify the batching logic and more importantly to make
            #   automatic adaptive batches much much easier to implement
277
278
279
            # - any OOMs will happen right away rather than near the end

            toks = x[1] + x[2]
280
            return -len(toks), tuple(toks)
Fabrizio Milo's avatar
Fabrizio Milo committed
281

Fabrizio Milo's avatar
Fabrizio Milo committed
282
        re_ord = utils.Reorderer(requests, _collate)
283

284
285
286
        reordered_requests = re_ord.get_reordered()
        n_reordered_requests = len(reordered_requests)

287
288
        # automatic (variable) batch size detection for vectorization
        # pull longest context sample from request
289
290
291
292
293
294
295
296
        def _batch_scheduler(pos):
            sched = pos // int(n_reordered_requests / self.batch_schedule)
            if sched in self.batch_sizes:
                return self.batch_sizes[sched]
            print(f"Passed argument batch_size = auto:{self.batch_schedule}. Detecting largest batch size")
            self.batch_sizes[sched] = self._detect_batch_size(reordered_requests, pos)
            print(f"Determined largest batch size: {self.batch_sizes[sched]}")
            return self.batch_sizes[sched]
297

Fabrizio Milo's avatar
Fabrizio Milo committed
298
        for chunk in utils.chunks(
299
300
301
            tqdm(reordered_requests, disable=disable_tqdm),
            n=self.batch_size if self.batch_size != "auto" else override_bs if override_bs is not None else 0,
            fn=_batch_scheduler if self.batch_size == "auto" and n_reordered_requests > 0 else None,
Fabrizio Milo's avatar
Fabrizio Milo committed
302
        ):
303
            inps = []
304
            cont_toks_list = []
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
            inplens = []

            padding_length = None

            # because vectorizing is annoying, we first convert each (context, continuation) pair to padded
            # tensors, then we pack them together into a batch, call the model, and then pick it all apart
            # again because vectorizing is annoying

            for _, context_enc, continuation_enc in chunk:
                # sanity check
                assert len(context_enc) > 0
                assert len(continuation_enc) > 0
                assert len(continuation_enc) <= self.max_length

                # how this all works:
                #          CTX      CONT
321
                # inp    0 1 2 3|4 5 6 7 8 9   <- last token is deleted by inp[:, :-1]
322
                # gpt2    \               \
323
324
                # logits   1 2 3|4 5 6 7 8 9   <- the ctx half gets tossed out by the
                # cont_toks      4 5 6 7 8 9      [:, -len(continuation_enc):, :self.vocab_size] slice
325
326
327

                # when too long to fit in context, truncate from the left
                inp = torch.tensor(
Fabrizio Milo's avatar
Fabrizio Milo committed
328
329
                    (context_enc + continuation_enc)[-(self.max_length + 1) :][:-1],
                    dtype=torch.long,
330
                ).to(self.device)
Fabrizio Milo's avatar
Fabrizio Milo committed
331
                (inplen,) = inp.shape
332
333
334
335

                cont = continuation_enc

                # since in _collate we make sure length is descending, the longest is always the first one.
Fabrizio Milo's avatar
Fabrizio Milo committed
336
337
338
                padding_length = (
                    padding_length if padding_length is not None else inplen
                )
339

340
                # pad length from seq to padding_length
Fabrizio Milo's avatar
Fabrizio Milo committed
341
342
343
344
345
346
347
348
349
                inp = torch.cat(
                    [
                        inp,  # [seq]
                        torch.zeros(padding_length - inplen, dtype=torch.long).to(
                            inp.device
                        ),  # [padding_length - seq]
                    ],
                    dim=0,
                )
350

351
352
                inps.append(inp.unsqueeze(0))  # [1, padding_length]
                cont_toks_list.append(cont)
353
354
                inplens.append(inplen)

355
            batched_inps = torch.cat(inps, dim=0)  # [batch, padding_length
Fabrizio Milo's avatar
Fabrizio Milo committed
356
357
358
            multi_logits = F.log_softmax(
                self._model_call(batched_inps), dim=-1
            ).cpu()  # [batch, padding_length, vocab]
359

Fabrizio Milo's avatar
Fabrizio Milo committed
360
361
362
            for (cache_key, _, _), logits, inp, inplen, cont_toks in zip(
                chunk, multi_logits, inps, inplens, cont_toks_list
            ):
363

364
365
                # Slice to original seq length
                contlen = len(cont_toks)
Fabrizio Milo's avatar
Fabrizio Milo committed
366
367
368
                logits = logits[inplen - contlen : inplen].unsqueeze(
                    0
                )  # [1, seq, vocab]
369

370
                # Check if per-token argmax is exactly equal to continuation
371
                greedy_tokens = logits.argmax(dim=-1)
Fabrizio Milo's avatar
Fabrizio Milo committed
372
373
374
                cont_toks = torch.tensor(cont_toks, dtype=torch.long).unsqueeze(
                    0
                )  # [1, seq]
375
376
                max_equal = (greedy_tokens == cont_toks).all()

377
378
                # Obtain log-probs at the corresponding continuation token indices
                # last_token_slice = logits[:, -1, :].squeeze(0).tolist()
Fabrizio Milo's avatar
Fabrizio Milo committed
379
380
381
                logits = torch.gather(logits, 2, cont_toks.unsqueeze(-1)).squeeze(
                    -1
                )  # [1, seq]
382

383
                # Answer: (log prob, is-exact-match)
384
385
386
387
388
389
390
391
                answer = (float(logits.sum()), bool(max_equal))

                # partial caching
                if cache_key is not None:
                    self.cache_hook.add_partial("loglikelihood", cache_key, answer)

                res.append(answer)

Fabrizio Milo's avatar
Fabrizio Milo committed
392
        return re_ord.get_original(res)
Fabrizio Milo's avatar
Fabrizio Milo committed
393

394
    def greedy_until(self, requests):
Fabrizio Milo's avatar
Fabrizio Milo committed
395
        # TODO: implement fully general `until` that handles until that are
396
        #       multiple tokens or that span multiple tokens correctly
397
398
399
400
401

        # TODO: extract to TokenizedLM?
        res = []

        def _collate(x):
402
403
404
405
406
407
408
            # the negative sign on len(toks) sorts descending - this has a few advantages:
            # - time estimates will always be over not underestimates, which is more useful for planning
            # - to know the size of a batch when going through the list, you know the first one is always the batch
            #   padded context length. this is useful to simplify the batching logic and more importantly to make
            #   automatic adaptive batches much much easier to implement
            # - any OOMs will happen right away rather than near the end

409
            toks = self.tok_encode(x[0])
410
            return -len(toks), x[0]
Fabrizio Milo's avatar
Fabrizio Milo committed
411

Fabrizio Milo's avatar
Fabrizio Milo committed
412
        re_ord = utils.Reorderer(requests, _collate)
413

414
        for context, request_args in tqdm(re_ord.get_reordered()):
415
            until = request_args["until"]
416
417
            if isinstance(until, str):
                until = [until]
418

419
420
421
422
            if until:
                (primary_until,) = self.tok_encode(until[0])
            else:
                primary_until = None
423

Fabrizio Milo's avatar
Fabrizio Milo committed
424
425
426
            context_enc = torch.tensor(
                [self.tok_encode(context)[self.max_gen_toks - self.max_length :]]
            ).to(self.device)
427

428
            max_gen_tokens = min(
429
                self.max_gen_toks, request_args.get("max_length", self.max_gen_toks)
430
            )
Fabrizio Milo's avatar
Fabrizio Milo committed
431
            cont = self._model_generate(
432
                context_enc, context_enc.shape[1] + max_gen_tokens, primary_until
Fabrizio Milo's avatar
Fabrizio Milo committed
433
434
435
            )

            s = self.tok_decode(cont[0].tolist()[context_enc.shape[1] :])
436
437
438

            for term in until:
                s = s.split(term)[0]
Fabrizio Milo's avatar
Fabrizio Milo committed
439

440
441
            # partial caching
            self.cache_hook.add_partial("greedy_until", (context, until), s)
Fabrizio Milo's avatar
Fabrizio Milo committed
442

443
            res.append(s)
Fabrizio Milo's avatar
Fabrizio Milo committed
444

Fabrizio Milo's avatar
Fabrizio Milo committed
445
        return re_ord.get_original(res)
Leo Gao's avatar
Leo Gao committed
446

Leo Gao's avatar
Leo Gao committed
447

448
class Task(abc.ABC):
&'s avatar
&amp; committed
449
450
451
452
453
454
455
456
    """A task represents an entire benchmark including its dataset, problems,
    answers, and evaluation methods. See BoolQ for a simple example implementation

    A `doc` can be any python object which represents one instance of evaluation.
    This is usually a dictionary e.g.
        {"question": ..., "answer": ...} or
        {"question": ..., question, answer)
    """
Jonathan Tow's avatar
Jonathan Tow committed
457

Jon Tow's avatar
Jon Tow committed
458
459
    # The name of the `Task` benchmark as denoted in the HuggingFace datasets Hub
    # or a path to a custom `datasets` loading script.
Jonathan Tow's avatar
Jonathan Tow committed
460
461
462
463
464
    DATASET_PATH: str = None

    # The name of a subset within `DATASET_PATH`.
    DATASET_NAME: str = None

Jon Tow's avatar
Jon Tow committed
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
    def __init__(self, data_dir=None, cache_dir=None, download_mode=None):
        """
        :param data_dir: str
            Stores the path to a local folder containing the `Task`'s data files.
            Use this to specify the path to manually downloaded data (usually when
            the dataset is not publicly accessible).
        :param cache_dir: str
            The directory to read/write the `Task` dataset. This follows the
            HuggingFace `datasets` API with the default cache directory located at:
                `~/.cache/huggingface/datasets`
            NOTE: You can change the cache location globally for a given process
            by setting the shell environment variable, `HF_DATASETS_CACHE`,
            to another directory:
                `export HF_DATASETS_CACHE="/path/to/another/directory"`
        :param download_mode: datasets.DownloadMode
            How to treat pre-existing `Task` downloads and data.
            - `datasets.DownloadMode.REUSE_DATASET_IF_EXISTS`
                Reuse download and reuse dataset.
            - `datasets.DownloadMode.REUSE_CACHE_IF_EXISTS`
                Reuse download with fresh dataset.
            - `datasets.DownloadMode.FORCE_REDOWNLOAD`
                Fresh download and fresh dataset.
        """
        self.download(data_dir, cache_dir, download_mode)
489
        self._training_docs = None
490
        self._fewshot_docs = None
sdtblck's avatar
sdtblck committed
491

Jon Tow's avatar
Jon Tow committed
492
    def download(self, data_dir=None, cache_dir=None, download_mode=None):
Fabrizio Milo's avatar
Fabrizio Milo committed
493
        """Downloads and returns the task dataset.
Jonathan Tow's avatar
Jonathan Tow committed
494
495
        Override this method to download the dataset from a custom API.

Jon Tow's avatar
Jon Tow committed
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
        :param data_dir: str
            Stores the path to a local folder containing the `Task`'s data files.
            Use this to specify the path to manually downloaded data (usually when
            the dataset is not publicly accessible).
        :param cache_dir: str
            The directory to read/write the `Task` dataset. This follows the
            HuggingFace `datasets` API with the default cache directory located at:
                `~/.cache/huggingface/datasets`
            NOTE: You can change the cache location globally for a given process
            by setting the shell environment variable, `HF_DATASETS_CACHE`,
            to another directory:
                `export HF_DATASETS_CACHE="/path/to/another/directory"`
        :param download_mode: datasets.DownloadMode
            How to treat pre-existing `Task` downloads and data.
            - `datasets.DownloadMode.REUSE_DATASET_IF_EXISTS`
                Reuse download and reuse dataset.
            - `datasets.DownloadMode.REUSE_CACHE_IF_EXISTS`
                Reuse download with fresh dataset.
            - `datasets.DownloadMode.FORCE_REDOWNLOAD`
                Fresh download and fresh dataset.
Jonathan Tow's avatar
Jonathan Tow committed
516
517
518
519
        """
        self.dataset = datasets.load_dataset(
            path=self.DATASET_PATH,
            name=self.DATASET_NAME,
Jon Tow's avatar
Jon Tow committed
520
521
            data_dir=data_dir,
            cache_dir=cache_dir,
Fabrizio Milo's avatar
Fabrizio Milo committed
522
            download_mode=download_mode,
Jonathan Tow's avatar
Jonathan Tow committed
523
        )
sdtblck's avatar
sdtblck committed
524

525
526
527
528
    def should_decontaminate(self):
        """Whether this task supports decontamination against model training set."""
        return False

529
    @abstractmethod
530
    def has_training_docs(self):
Jason Phang's avatar
checkin  
Jason Phang committed
531
        """Whether the task has a training set"""
532
        pass
533

534
    @abstractmethod
535
    def has_validation_docs(self):
Jason Phang's avatar
checkin  
Jason Phang committed
536
537
538
        """Whether the task has a validation set"""
        pass

539
    @abstractmethod
Jason Phang's avatar
checkin  
Jason Phang committed
540
541
    def has_test_docs(self):
        """Whether the task has a test set"""
542
543
        pass

Leo Gao's avatar
Leo Gao committed
544
    def training_docs(self):
Jason Phang's avatar
checkin  
Jason Phang committed
545
546
547
548
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
Leo Gao's avatar
Leo Gao committed
549
        return []
550

Leo Gao's avatar
Leo Gao committed
551
    def validation_docs(self):
552
553
554
555
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
Leo Gao's avatar
Leo Gao committed
556
        return []
557

Leo Gao's avatar
Leo Gao committed
558
    def test_docs(self):
559
560
561
562
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
Leo Gao's avatar
Leo Gao committed
563
        return []
Leo Gao's avatar
Leo Gao committed
564

Jon Tow's avatar
Jon Tow committed
565
566
567
568
    def _process_doc(self, doc):
        """
        Override this to process (detokenize, strip, replace, etc.) individual
        documents. This can be used in a map over documents of a data split.
Jon Tow's avatar
Jon Tow committed
569
        E.g. `map(self._process_doc, self.dataset["validation"])`
Jon Tow's avatar
Jon Tow committed
570
571
572
573
574
575

        :return: dict
            The processed version of the specified `doc`.
        """
        return doc

576
    def fewshot_examples(self, k, rnd):
577
578
        if self._training_docs is None:
            self._training_docs = list(self.training_docs())
579

Leo Gao's avatar
Leo Gao committed
580
        return rnd.sample(self._training_docs, k)
Leo Gao's avatar
Leo Gao committed
581

582
    def doc_to_decontamination_query(self, doc):
Fabrizio Milo's avatar
Fabrizio Milo committed
583
584
585
586
        print(
            "Override doc_to_decontamination_query with document specific decontamination query."
        )
        assert False
587

588
    @abstractmethod
Leo Gao's avatar
Update  
Leo Gao committed
589
590
591
    def doc_to_text(self, doc):
        pass

592
    @abstractmethod
Leo Gao's avatar
Update  
Leo Gao committed
593
    def doc_to_target(self, doc):
Leo Gao's avatar
Leo Gao committed
594
        pass
Leo Gao's avatar
Leo Gao committed
595

596
    @abstractmethod
597
    def construct_requests(self, doc, ctx):
Fabrizio Milo's avatar
Fabrizio Milo committed
598
        """Uses RequestFactory to construct Requests and returns an iterable of
Leo Gao's avatar
Leo Gao committed
599
600
        Requests which will be sent to the LM.

601
602
        :param doc:
            The document as returned from training_docs, validation_docs, or test_docs.
Leo Gao's avatar
Leo Gao committed
603
        :param ctx: str
Fabrizio Milo's avatar
Fabrizio Milo committed
604
            The context string, generated by fewshot_context. This includes the natural
605
            language description, as well as the few shot examples, and the question
Fabrizio Milo's avatar
Fabrizio Milo committed
606
            part of the document for `doc`.
Leo Gao's avatar
Leo Gao committed
607
        """
Leo Gao's avatar
Leo Gao committed
608
        pass
609

610
    @abstractmethod
Leo Gao's avatar
Leo Gao committed
611
    def process_results(self, doc, results):
Fabrizio Milo's avatar
Fabrizio Milo committed
612
613
        """Take a single document and the LM results and evaluates, returning a
        dict where keys are the names of submetrics and values are the values of
614
        the metric for that one document
Leo Gao's avatar
Leo Gao committed
615
616
617
618
619

        :param doc:
            The document as returned from training_docs, validation_docs, or test_docs.
        :param results:
            The results of the requests created in construct_requests.
Jason Phang's avatar
checkin  
Jason Phang committed
620
        """
Leo Gao's avatar
Leo Gao committed
621
        pass
Jason Phang's avatar
gpt3  
Jason Phang committed
622

623
    @abstractmethod
624
625
    def aggregation(self):
        """
&'s avatar
&amp; committed
626
        :returns: {str: [metric_score] -> float}
Fabrizio Milo's avatar
Fabrizio Milo committed
627
            A dictionary where keys are the names of submetrics and values are
&'s avatar
&amp; committed
628
            functions that aggregate a list of metric scores
629
630
631
        """
        pass

632
    @abstractmethod
633
634
635
    def higher_is_better(self):
        """
        :returns: {str: bool}
Fabrizio Milo's avatar
Fabrizio Milo committed
636
            A dictionary where keys are the names of submetrics and values are
637
638
639
640
            whether a higher value of the submetric is better
        """
        pass

Jason Phang's avatar
Jason Phang committed
641
    def fewshot_description(self):
642
        import warnings
Fabrizio Milo's avatar
Fabrizio Milo committed
643

644
        warnings.warn(
Jonathan Tow's avatar
Jonathan Tow committed
645
            "`fewshot_description` will be removed in futures versions. Pass "
646
            "any custom descriptions to the `evaluate` function instead.",
Fabrizio Milo's avatar
Fabrizio Milo committed
647
648
            DeprecationWarning,
        )
Jason Phang's avatar
checkin  
Jason Phang committed
649
650
        return ""

651
    @utils.positional_deprecated
Fabrizio Milo's avatar
Fabrizio Milo committed
652
653
654
655
    def fewshot_context(
        self, doc, num_fewshot, provide_description=None, rnd=None, description=None
    ):
        """Returns a fewshot context string that is made up of a prepended description
656
657
658
659
660
661
662
663
664
665
        (if provided), the `num_fewshot` number of examples, and an appended prompt example.

        :param doc: str
            The document as returned from training_docs, validation_docs, or test_docs.
        :param num_fewshot: int
            The number of fewshot examples to provide in the returned context string.
        :param provide_description: bool
            Not implemented, and this option is deprecated and will be removed in a future version in favor of a different description providing method
        :param rnd: random.Random
            The pseudo-random number generator used to randomly sample examples.
666
            WARNING: This is currently a required arg although it's optionalized with a default `None`.
667
668
669
670
671
        :param description: str
            The task's description that will be prepended to the fewshot examples.
        :returns: str
            The fewshot context.
        """
Fabrizio Milo's avatar
Fabrizio Milo committed
672
673
674
        assert (
            rnd is not None
        ), "A `random.Random` generator argument must be provided to `rnd`"
Jonathan Tow's avatar
Merge  
Jonathan Tow committed
675
        assert not provide_description, (
Jonathan Tow's avatar
Jonathan Tow committed
676
            "The `provide_description` arg will be removed in future versions. To prepend "
677
            "a custom description to the context, supply the corresponding string via the "
Jonathan Tow's avatar
Jonathan Tow committed
678
            "`description` arg."
Jonathan Tow's avatar
Merge  
Jonathan Tow committed
679
        )
680
681
        if provide_description is not None:
            # nudge people to not specify it at all
Fabrizio Milo's avatar
Fabrizio Milo committed
682
683
684
            print(
                "WARNING: provide_description is deprecated and will be removed in a future version in favor of description_dict"
            )
685

686
        description = description + "\n\n" if description else ""
687

688
689
        if num_fewshot == 0:
            labeled_examples = ""
690
        else:
691
692
693
694
695
            # for sets with no training docs, draw from other set *but ensure no overlap with current doc*
            if self.has_training_docs():
                fewshotex = self.fewshot_examples(k=num_fewshot, rnd=rnd)
            else:
                if self._fewshot_docs is None:
696
                    self._fewshot_docs = list(
Fabrizio Milo's avatar
Fabrizio Milo committed
697
698
699
                        self.validation_docs()
                        if self.has_validation_docs()
                        else self.test_docs()
700
                    )
701

702
                fewshotex = rnd.sample(self._fewshot_docs, num_fewshot + 1)
703

704
705
                # get rid of the doc that's the one we're evaluating, if it's in the fewshot
                fewshotex = [x for x in fewshotex if x != doc][:num_fewshot]
706

Fabrizio Milo's avatar
Fabrizio Milo committed
707
708
709
710
711
712
713
714
715
            labeled_examples = (
                "\n\n".join(
                    [
                        self.doc_to_text(doc) + self.doc_to_target(doc)
                        for doc in fewshotex
                    ]
                )
                + "\n\n"
            )
Leo Gao's avatar
Update  
Leo Gao committed
716

717
        example = self.doc_to_text(doc)
Leo Gao's avatar
Leo Gao committed
718
719
720
        return description + labeled_examples + example


Jon Tow's avatar
Jon Tow committed
721
class MultipleChoiceTask(Task):
Leo Gao's avatar
Leo Gao committed
722
    def doc_to_target(self, doc):
Fabrizio Milo's avatar
Fabrizio Milo committed
723
        return " " + doc["choices"][doc["gold"]]
Leo Gao's avatar
Leo Gao committed
724

Leo Gao's avatar
Leo Gao committed
725
726
    def construct_requests(self, doc, ctx):
        lls = [
Fabrizio Milo's avatar
Fabrizio Milo committed
727
            rf.loglikelihood(ctx, " {}".format(choice))[0] for choice in doc["choices"]
Leo Gao's avatar
Leo Gao committed
728
729
730
731
732
733
734
        ]

        return lls

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

Fabrizio Milo's avatar
Fabrizio Milo committed
735
        acc = 1.0 if np.argmax(results) == gold else 0.0
736
        completion_len = np.array([float(len(i)) for i in doc["choices"]])
Fabrizio Milo's avatar
Fabrizio Milo committed
737
        acc_norm = 1.0 if np.argmax(results / completion_len) == gold else 0.0
Leo Gao's avatar
Leo Gao committed
738
739

        return {
Leo Gao's avatar
Leo Gao committed
740
741
            "acc": acc,
            "acc_norm": acc_norm,
Leo Gao's avatar
Leo Gao committed
742
        }
Fabrizio Milo's avatar
Fabrizio Milo committed
743

Leo Gao's avatar
Leo Gao committed
744
745
    def higher_is_better(self):
        return {
Leo Gao's avatar
Leo Gao committed
746
747
            "acc": True,
            "acc_norm": True,
Leo Gao's avatar
Leo Gao committed
748
        }
Fabrizio Milo's avatar
Fabrizio Milo committed
749

Leo Gao's avatar
Leo Gao committed
750
751
    def aggregation(self):
        return {
Leo Gao's avatar
Leo Gao committed
752
753
            "acc": mean,
            "acc_norm": mean,
Leo Gao's avatar
Leo Gao committed
754
755
756
        }


Jason Phang's avatar
Jason Phang committed
757
class PerplexityTask(Task, abc.ABC):
758
759
760
761
    def should_decontaminate(self):
        """Whether this task supports decontamination against model training set."""
        return True

Jason Phang's avatar
Jason Phang committed
762
763
764
765
766
767
768
    def has_training_docs(self):
        return False

    def fewshot_examples(self, k, rnd):
        assert k == 0
        return []

Fabrizio Milo's avatar
Fabrizio Milo committed
769
770
771
772
773
774
775
776
777
    def fewshot_context(
        self, doc, num_fewshot, provide_description=None, rnd=None, description=None
    ):
        assert (
            num_fewshot == 0
        ), "The number of fewshot examples must be 0 for perplexity tasks."
        assert (
            rnd is not None
        ), "A `random.Random` generator argument must be provided to `rnd`."
Jonathan Tow's avatar
Merge  
Jonathan Tow committed
778
        assert not provide_description, (
Jonathan Tow's avatar
Jonathan Tow committed
779
            "The `provide_description` arg will be removed in future versions. To prepend "
Jonathan Tow's avatar
Jonathan Tow committed
780
            "a custom description to the context, supply the corresponding string via the "
Jonathan Tow's avatar
Jonathan Tow committed
781
            "`description` arg."
Jonathan Tow's avatar
Merge  
Jonathan Tow committed
782
        )
783
784
        if provide_description is not None:
            # nudge people to not specify it at all
Fabrizio Milo's avatar
Fabrizio Milo committed
785
786
787
            print(
                "WARNING: provide_description is deprecated and will be removed in a future version in favor of description_dict"
            )
788

Jason Phang's avatar
Jason Phang committed
789
790
791
        return ""

    def higher_is_better(self):
Leo Gao's avatar
Leo Gao committed
792
793
794
795
796
        return {
            "word_perplexity": False,
            "byte_perplexity": False,
            "bits_per_byte": False,
        }
Jason Phang's avatar
Jason Phang committed
797

798
799
800
    def doc_to_decontamination_query(self, doc):
        return doc

Jason Phang's avatar
Jason Phang committed
801
    def doc_to_text(self, doc):
802
        return ""
Jason Phang's avatar
Jason Phang committed
803
804

    def doc_to_target(self, doc):
805
        return doc
Jason Phang's avatar
Jason Phang committed
806
807
808

    def construct_requests(self, doc, ctx):
        assert not ctx
Leo Gao's avatar
Leo Gao committed
809
        req = rf.loglikelihood_rolling(self.doc_to_target(doc))
Jason Phang's avatar
Jason Phang committed
810
811
812
        return req

    def process_results(self, doc, results):
Fabrizio Milo's avatar
Fabrizio Milo committed
813
        (loglikelihood,) = results
Leo Gao's avatar
Leo Gao committed
814
        words = self.count_words(doc)
815
        bytes_ = self.count_bytes(doc)
Jason Phang's avatar
Jason Phang committed
816
        return {
Leo Gao's avatar
Leo Gao committed
817
            "word_perplexity": (loglikelihood, words),
818
            "byte_perplexity": (loglikelihood, bytes_),
819
            "bits_per_byte": (loglikelihood, bytes_),
Jason Phang's avatar
Jason Phang committed
820
821
822
823
        }

    def aggregation(self):
        return {
Leo Gao's avatar
Leo Gao committed
824
825
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
826
            "bits_per_byte": bits_per_byte,
Jason Phang's avatar
Jason Phang committed
827
828
        }

829
830
    @classmethod
    def count_bytes(cls, doc):
Leo Gao's avatar
Leo Gao committed
831
        return len(doc.encode("utf-8"))
832
833
834

    @classmethod
    def count_words(cls, doc):
Fabrizio Milo's avatar
Fabrizio Milo committed
835
        """Downstream tasks with custom word boundaries should override this!"""
Leo Gao's avatar
Leo Gao committed
836
        return len(re.split(r"\s+", doc))
Leo Gao's avatar
Leo Gao committed
837

Jason Phang's avatar
Jason Phang committed
838

Leo Gao's avatar
Leo Gao committed
839
840
def hash_args(attr, args):
    dat = json.dumps([attr] + list(args))
Fabrizio Milo's avatar
Fabrizio Milo committed
841
    return hashlib.sha256(dat.encode("utf-8")).hexdigest()
Leo Gao's avatar
Leo Gao committed
842
843


Leo Gao's avatar
Leo Gao committed
844
845
class CacheHook:
    def __init__(self, cachinglm):
Fabrizio Milo's avatar
Fabrizio Milo committed
846
        if cachinglm is None:
Leo Gao's avatar
Leo Gao committed
847
848
849
850
            self.dbdict = None
            return

        self.dbdict = cachinglm.dbdict
Fabrizio Milo's avatar
Fabrizio Milo committed
851

Leo Gao's avatar
Leo Gao committed
852
853
854
855
856
857
858
    def add_partial(self, attr, req, res):
        if self.dbdict is None:
            return
        hsh = hash_args(attr, req)
        self.dbdict[hsh] = res


Leo Gao's avatar
Leo Gao committed
859
860
class CachingLM:
    def __init__(self, lm, cache_db):
861
862
863
864
865
866
867
        """LM wrapper that returns cached results if they exist, and uses the underlying LM if not.

        :param lm: LM
            Underlying LM
        :param cache_db: str
            Path to cache db
        """
Leo Gao's avatar
Leo Gao committed
868
869
        self.lm = lm
        self.cache_db = cache_db
870
871
        if os.path.dirname(cache_db):
            os.makedirs(os.path.dirname(cache_db), exist_ok=True)
Leo Gao's avatar
Leo Gao committed
872
873
        self.dbdict = SqliteDict(cache_db, autocommit=True)

Leo Gao's avatar
Leo Gao committed
874
875
876
        # add hook to lm
        lm.set_cache_hook(self.get_cache_hook())

Leo Gao's avatar
Leo Gao committed
877
    def __getattr__(self, attr):
gk's avatar
gk committed
878
879
880
881
        lm_attr = getattr(self.lm, attr)
        if not callable(lm_attr):
            return lm_attr

Leo Gao's avatar
Leo Gao committed
882
883
884
        def fn(requests):
            res = []
            remaining_reqs = []
Fabrizio Milo's avatar
Fabrizio Milo committed
885

Leo Gao's avatar
Leo Gao committed
886
887
            # figure out which ones are cached and which ones are new
            for req in requests:
Leo Gao's avatar
Leo Gao committed
888
                hsh = hash_args(attr, req)
Leo Gao's avatar
Leo Gao committed
889
890
891
892
893
894
895
896
897
                if hsh in self.dbdict:
                    ob = self.dbdict[hsh]

                    assert ob is not None

                    res.append(ob)
                else:
                    res.append(None)
                    remaining_reqs.append(req)
Fabrizio Milo's avatar
Fabrizio Milo committed
898

899
            # actually run the LM on the requests that do not have cached results
Leo Gao's avatar
Leo Gao committed
900
901
902
903
904
            rem_res = getattr(self.lm, attr)(remaining_reqs)

            # stick the new ones back into the list and also cache any of the new ones
            resptr = 0
            for req, r in zip(remaining_reqs, rem_res):
905
906
                while res[resptr] is not None:
                    resptr += 1
Leo Gao's avatar
Leo Gao committed
907
908
909
910

                res[resptr] = r

                # caching
Leo Gao's avatar
Leo Gao committed
911
                hsh = hash_args(attr, req)
Leo Gao's avatar
Leo Gao committed
912
                self.dbdict[hsh] = r
Leo Gao's avatar
Leo Gao committed
913
            self.dbdict.commit()
Leo Gao's avatar
Leo Gao committed
914
915

            return res
Fabrizio Milo's avatar
Fabrizio Milo committed
916

Leo Gao's avatar
Leo Gao committed
917
        return fn
Fabrizio Milo's avatar
Fabrizio Milo committed
918

Leo Gao's avatar
Leo Gao committed
919
920
    def get_cache_hook(self):
        return CacheHook(self)
Leo Gao's avatar
Leo Gao committed
921

Jason Phang's avatar
Jason Phang committed
922

923
REQUEST_RETURN_LENGTHS = {
Fabrizio Milo's avatar
Fabrizio Milo committed
924
925
926
    "loglikelihood": 2,
    "greedy_until": None,
    "loglikelihood_rolling": None,
927
928
929
}


930
class Request:
Leo Gao's avatar
Leo Gao committed
931
932
    def __init__(self, request_type, args, index=None):
        if request_type not in REQUEST_RETURN_LENGTHS.keys():
Fabrizio Milo's avatar
Fabrizio Milo committed
933
934
935
            raise NotImplementedError(
                "The request type {} is not implemented!".format(request_type)
            )
Leo Gao's avatar
Leo Gao committed
936

Leo Gao's avatar
Leo Gao committed
937
        self.request_type = request_type
938
939
        self.args = args
        self.index = index
Fabrizio Milo's avatar
Fabrizio Milo committed
940

941
    def __iter__(self):
Leo Gao's avatar
Leo Gao committed
942
        if REQUEST_RETURN_LENGTHS[self.request_type] is None:
Fabrizio Milo's avatar
Fabrizio Milo committed
943
            raise IndexError("This request type does not return multiple arguments!")
Leo Gao's avatar
Leo Gao committed
944
945
        for i in range(REQUEST_RETURN_LENGTHS[self.request_type]):
            yield Request(self.request_type, self.args, i)
Fabrizio Milo's avatar
Fabrizio Milo committed
946

947
    def __getitem__(self, i):
Leo Gao's avatar
Leo Gao committed
948
        if REQUEST_RETURN_LENGTHS[self.request_type] is None:
Fabrizio Milo's avatar
Fabrizio Milo committed
949
            raise IndexError("This request type does not return multiple arguments!")
Leo Gao's avatar
Leo Gao committed
950
        return Request(self.request_type, self.args, i)
Fabrizio Milo's avatar
Fabrizio Milo committed
951

Leo Gao's avatar
Leo Gao committed
952
    def __eq__(self, other):
Fabrizio Milo's avatar
Fabrizio Milo committed
953
954
955
956
957
        return (
            self.request_type == other.request_type
            and self.args == other.args
            and self.index == other.index
        )
Leo Gao's avatar
Leo Gao committed
958

Leo Gao's avatar
Leo Gao committed
959
    def __repr__(self):
Leo Gao's avatar
Leo Gao committed
960
        return f"Req_{self.request_type}{self.args}[{self.index}]\n"
961

Jason Phang's avatar
Jason Phang committed
962

Leo Gao's avatar
Leo Gao committed
963
964
class RequestFactory:
    def __getattr__(self, attr):
Leo Gao's avatar
Update  
Leo Gao committed
965
966
        def fn(*args):
            return Request(attr, args)
Fabrizio Milo's avatar
Fabrizio Milo committed
967

Leo Gao's avatar
Leo Gao committed
968
969
970
971
        return fn


rf = RequestFactory()