task.py 43.9 KB
Newer Older
1
import abc
2
from dataclasses import dataclass, field, asdict
3
4

import re
5
import ast
lintangsutawika's avatar
lintangsutawika committed
6
import yaml
7
8
9
import evaluate
import random
import itertools
10
import functools
11
from tqdm import tqdm
12
13
14
15

import datasets
import numpy as np

baberabb's avatar
baberabb committed
16
from typing import Union, List, Any, Tuple, Literal
17
from collections.abc import Callable
18

19
from lm_eval import utils
20
from lm_eval.api import samplers
haileyschoelkopf's avatar
haileyschoelkopf committed
21
from lm_eval.api.instance import Instance
lintangsutawika's avatar
lintangsutawika committed
22
from lm_eval.api.filter import FilterEnsemble
23
24
25
26

from lm_eval.logger import eval_logger
from lm_eval.prompts import get_prompt
from lm_eval.filters import build_filter_ensemble
lintangsutawika's avatar
lintangsutawika committed
27
28
29
30
from lm_eval.api.metrics import (
    mean,
    weighted_perplexity,
    bits_per_byte,
lintangsutawika's avatar
lintangsutawika committed
31
    metric_max_over_ground_truths,
lintangsutawika's avatar
lintangsutawika committed
32
33
)
from lm_eval.api.registry import (
haileyschoelkopf's avatar
haileyschoelkopf committed
34
35
36
37
    get_metric,
    get_aggregation,
    get_default_aggregation,
    is_higher_better,
38
39
    DEFAULT_METRIC_REGISTRY,
    OUTPUT_TYPE_REGISTRY,
lintangsutawika's avatar
lintangsutawika committed
40
41
    AGGREGATION_REGISTRY,
)
42

43
44
45
46
47
48
49
ALL_OUTPUT_TYPES = [
    "loglikelihood",
    "multiple_choice",
    "loglikelihood_rolling",
    "greedy_until",
]

50
51
52

@dataclass
class TaskConfig(dict):
53
    # task naming/registry
54
    task: str = None
55
    group: Union[str, list] = None
56
57
58
    # HF dataset options.
    # which dataset to use,
    # and what splits for what purpose
59
60
    dataset_path: str = None
    dataset_name: str = None
61
    dataset_kwargs: dict = None
62
63
64
    training_split: str = None
    validation_split: str = None
    test_split: str = None
lintangsutawika's avatar
lintangsutawika committed
65
    fewshot_split: str = None  # TODO: assert that this not None if num_fewshot > 0. (?) assert if this is same split as one evaling (?)
66
67
    # formatting / prompting options.
    # see docs/advanced_task_guide.md for more info
68
    process_docs: Callable = None
69
70
    doc_to_text: Union[Callable, str] = None
    doc_to_target: Union[Callable, str] = None
lintangsutawika's avatar
lintangsutawika committed
71
    doc_to_choice: Union[Callable, str, dict, list] = None
72
    gold_alias: Union[Callable, str] = None
lintangsutawika's avatar
lintangsutawika committed
73
    process_results: Union[Callable, str] = None
74
    use_prompt: str = None
75
    description: str = ""
76
77
    target_delimiter: str = " "
    fewshot_delimiter: str = "\n\n"
78
    # runtime configuration options
79
    num_fewshot: int = 0
80
    # scoring options
81
82
    metric_list: str = None
    output_type: str = "greedy_until"
83
    generation_kwargs: dict = None
84
    repeats: int = 1
lintangsutawika's avatar
lintangsutawika committed
85
    filter_list: Union[str, list] = None
86
87
    should_decontaminate: bool = False
    doc_to_decontamination_query: str = None
88

lintangsutawika's avatar
lintangsutawika committed
89
    metadata: str = None  # by default, not used in the code. allows for users to pass arbitrary info to tasks
90

91
    def __post_init__(self):
92

Lintang Sutawika's avatar
Lintang Sutawika committed
93
94
95
        if self.generation_kwargs is not None:
            if self.output_type != "greedy_until":
                eval_logger.warning(
96
                    "passed `generation_kwargs`, but not using `output_type: greedy_until`!"
Lintang Sutawika's avatar
Lintang Sutawika committed
97
                )
98
                assert self.output_type != "greedy_until"
Lintang Sutawika's avatar
Lintang Sutawika committed
99
100
101
102
103
104
105

            if "temperature" in self.generation_kwargs:
                self.generation_kwargs["temperature"] = float(
                    self.generation_kwargs["temperature"]
                )

            if "until" not in self.generation_kwargs:
106
                self.generation_kwargs["until"] = [self.fewshot_delimiter]
Lintang Sutawika's avatar
Lintang Sutawika committed
107
108
109
110
        else:
            if self.output_type == "greedy_until":
                # ensure that we greedily generate in absence of explicit arguments otherwise
                self.generation_kwargs = {
Lintang Sutawika's avatar
Lintang Sutawika committed
111
                    "until": None
112
113
                    if self.fewshot_delimiter is None
                    else [self.fewshot_delimiter],
Lintang Sutawika's avatar
Lintang Sutawika committed
114
115
116
                    "do_sample": False,
                    "temperature": 0.0,
                }
117

haileyschoelkopf's avatar
haileyschoelkopf committed
118
119
        # TODO: how to make TaskConfigs be de- and re-serializable, even when using the !function constructor?

120
121
122
    def __getitem__(self, item):
        return getattr(self, item)

123
124
125
    def __setitem__(self, item, value):
        return setattr(self, item, value)

126
    def to_dict(self):
127
128
        """dumps the current config as a dictionary object, as a printable format.
        null fields will not be printed.
haileyschoelkopf's avatar
haileyschoelkopf committed
129
        Used for dumping results alongside full task configuration
130

haileyschoelkopf's avatar
haileyschoelkopf committed
131
132
133
134
135
136
137
138
139
140
        :return: dict
            A printable dictionary version of the TaskConfig object.

        # TODO: should any default value in the TaskConfig not be printed?
        """
        cfg_dict = asdict(self)
        # remove values that are `None`
        for k, v in list(cfg_dict.items()):
            if v is None:
                cfg_dict.pop(k)
haileyschoelkopf's avatar
haileyschoelkopf committed
141
142
143
            elif isinstance(v, Callable):
                # TODO: this should handle Promptsource template objects as a separate case?
                cfg_dict[k] = str(v)
haileyschoelkopf's avatar
haileyschoelkopf committed
144
        return cfg_dict
145

146
147
148
149
150
151
152
153
154
155
156
157

class Task(abc.ABC):
    """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)
    """

    VERSION = None
158

159
160
161
162
163
164
165
166
    # The name of the `Task` benchmark as denoted in the HuggingFace datasets Hub
    # or a path to a custom `datasets` loading script.
    DATASET_PATH: str = None

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

    OUTPUT_TYPE: str = None
lintangsutawika's avatar
lintangsutawika committed
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
    def __init__(
        self,
        data_dir=None,
        cache_dir=None,
        download_mode=None,
        config=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)
        self._training_docs = None
        self._fewshot_docs = None
        self._instances = None

haileyschoelkopf's avatar
haileyschoelkopf committed
202
        self._config = TaskConfig(**config) if config else TaskConfig()
203
204
205

        if not hasattr(self, "_filters"):
            self._filters = []
lintangsutawika's avatar
lintangsutawika committed
206
            for name, components in self._config.get(
207
                "filters", [["none", [["take_first", None]]]]
lintangsutawika's avatar
lintangsutawika committed
208
            ):
209
210
211
                filter_pipeline = build_filter_ensemble(name, components)
                self._filters.append(filter_pipeline)

lintangsutawika's avatar
lintangsutawika committed
212
        self.sampler = samplers.Sampler(
213
214
            list(self.fewshot_docs()), self, rnd=random.Random(1234)
        )
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

    def download(self, data_dir=None, cache_dir=None, download_mode=None):
        """Downloads and returns the task dataset.
        Override this method to download the dataset from a custom API.

        :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.
        """
241
242
243
244
245
246
247
        self.dataset = datasets.load_dataset(
            path=self.DATASET_PATH,
            name=self.DATASET_NAME,
            data_dir=data_dir,
            cache_dir=cache_dir,
            download_mode=download_mode,
        )
248

249
250
251
252
253
    @property
    def config(self):
        """Returns the TaskConfig associated with this class."""
        return self._config

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
    @abc.abstractmethod
    def has_training_docs(self):
        """Whether the task has a training set"""
        pass

    @abc.abstractmethod
    def has_validation_docs(self):
        """Whether the task has a validation set"""
        pass

    @abc.abstractmethod
    def has_test_docs(self):
        """Whether the task has a test set"""
        pass

    def training_docs(self):
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

    def validation_docs(self):
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

    def test_docs(self):
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

290
291
292
293
294
295
296
297
298
299
    def fewshot_docs(self):
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        if self.has_training_docs():
            return self.training_docs()
        elif self.has_validation_docs():
            return self.validation_docs()
        else:
lintangsutawika's avatar
lintangsutawika committed
300
            eval_logger.warning(
301
                "has_training_docs and has_validation_docs are False"
302
                ", using test_docs as fewshot_docs but this is not recommended."
lintangsutawika's avatar
lintangsutawika committed
303
            )
304
305
            return self.test_docs()

306
307
308
309
310
311
312
313
314
315
    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.
        E.g. `map(self._process_doc, self.dataset["validation"])`

        :return: dict
            The processed version of the specified `doc`.
        """
        return doc
lintangsutawika's avatar
lintangsutawika committed
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
    @property
    def instances(self):
        """After calling `task.build_all_requests()`, tasks
        maintain a list of the dataset instances which will be evaluated.
        """
        return self._instances

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

        return rnd.sample(self._training_docs, k)

    def doc_to_decontamination_query(self, doc):
        print(
            "Override doc_to_decontamination_query with document specific decontamination query."
        )
        assert False

    @abc.abstractmethod
    def doc_to_text(self, doc):
        pass

    @abc.abstractmethod
    def doc_to_target(self, doc):
        pass

344
    def build_all_requests(self, limit=None, rank=None, world_size=None):
345
346
347
348
349
350
351
352
353
354
        """Build a set of Instances for a task, and store them in task.instances"""
        if self.has_test_docs():
            docs = self.test_docs()
        elif self.has_validation_docs():
            docs = self.validation_docs()
        else:
            assert (
                False
            ), f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"

355
        eval_logger.info(
356
            f"Building contexts for task '{self.config.task}' on rank {rank}..."
357
358
        )

359
        instances = []
360
361
        for doc_id, doc in utils.create_iterator(
            enumerate(docs), rank, world_size, limit
lintangsutawika's avatar
lintangsutawika committed
362
        ):
363
            # sample fewshot context #TODO: need to offset doc_id by rank now!
364
            fewshot_ctx = self.fewshot_context(
365
                doc,
366
                self.config.num_fewshot,
367
            )
368

369
            # TODO: we should override self.config.repeats if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
370
371
372
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
373
                metadata=(self.config["task"], doc_id, self.config.repeats),
lintangsutawika's avatar
lintangsutawika committed
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

            if not isinstance(inst, list):
                inst = [inst]

            instances.extend(inst)

        self._instances = instances
        assert len(self._instances) != 0, "task.build_requests() did not find any docs!"

    @abc.abstractmethod
    def construct_requests(self, doc, ctx, **kwargs):
        """Uses RequestFactory to construct Requests and returns an iterable of
        Requests which will be sent to the LM.

        :param doc:
            The document as returned from training_docs, validation_docs, or test_docs.
        :param ctx: str
            The context string, generated by fewshot_context. This includes the natural
            language description, as well as the few shot examples, and the question
            part of the document for `doc`.
        :param doc_idx: int
            The index of a document within `self.test_docs()` or `self.validation_docs()`,
            whichever is the main split used.
        :param repeats: int
        TODO: update this docstring
lintangsutawika's avatar
lintangsutawika committed
400
            The number of times each instance in a dataset is inferred on. Defaults to 1,
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
            can be increased for techniques like majority voting.
        """
        pass

    @abc.abstractmethod
    def process_results(self, doc, results):
        """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
        the metric for that one document

        :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.
        """
        pass

    @abc.abstractmethod
    def aggregation(self):
        """
        :returns: {str: [metric_score] -> float}
            A dictionary where keys are the names of submetrics and values are
            functions that aggregate a list of metric scores
        """
        pass

    @abc.abstractmethod
    def higher_is_better(self):
        """
        :returns: {str: bool}
            A dictionary where keys are the names of submetrics and values are
            whether a higher value of the submetric is better
        """
        pass

haileyschoelkopf's avatar
haileyschoelkopf committed
436
437
438
439
440
441
442
443
444
445
    @classmethod
    def count_bytes(cls, doc):
        """Used for byte-level perplexity metrics in rolling loglikelihood"""
        return len(doc.encode("utf-8"))

    @classmethod
    def count_words(cls, doc):
        """Downstream loglikelihood_rolling perplexity tasks with custom word boundaries should override this!"""
        return len(re.split(r"\s+", doc))

446
    @utils.positional_deprecated
447
    def fewshot_context(self, doc, num_fewshot):
448
449
450
451
452
453
454
455
456
457
458
459
        """Returns a fewshot context string that is made up of a prepended description
        (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.
        :returns: str
            The fewshot context.
        """

        if num_fewshot == 0:
460
            # always prepend the (possibly empty) task description
461
            labeled_examples = self.config.description
462
        else:
463
            labeled_examples = self.config.description + self.sampler.get_context(
lintangsutawika's avatar
lintangsutawika committed
464
465
                doc, num_fewshot
            )
466
467

        example = self.doc_to_text(doc)
468
469
470
471
        if type(example) == str:
            return labeled_examples + example
        elif type(example) == list:
            return [labeled_examples + ex for ex in example]
472
        elif type(example) == int:
473
            if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
474
475
476
477
                choices = self.doc_to_choice(doc)
                return labeled_examples + choices[example]
            else:
                return labeled_examples + str(example)
478
479
480

    def apply_filters(self):

lintangsutawika's avatar
lintangsutawika committed
481
482
483
484
485
486
        if hasattr(self, "_filters"):
            for f in self._filters:
                f.apply(self._instances)
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances
487

baberabb's avatar
baberabb committed
488
    def dump_config(self) -> dict:
489
        """Returns a dictionary representing the task's config.
490
491
492
493
494

        :returns: str
            The fewshot context.
        """
        # TODO: this should only return the overrides applied to a non-YAML task's configuration.
495
        # (num_fewshot)
496
        return self.config.to_dict()
497

498
499

class ConfigurableTask(Task):
500
    VERSION = "Yaml"
501
    OUTPUT_TYPE = None
502
    CONFIG = None
503
504
505

    def __init__(
        self, data_dir=None, cache_dir=None, download_mode=None, config: dict = None
baberabb's avatar
baberabb committed
506
    ):  # TODO no super() call here
507
        # Get pre-configured attributes
508
        self._config = self.CONFIG
509

510
        # Use new configurations if there was no preconfiguration
511
        if self.config is None:
512
            self._config = TaskConfig(**config)
513
514
        # Overwrite configs
        else:
lintangsutawika's avatar
lintangsutawika committed
515
            if config is not None:
516
                self._config.__dict__.update(config)
517

518
        if self.config is None:
lintangsutawika's avatar
lintangsutawika committed
519
520
521
            raise ValueError(
                "Must pass a config to ConfigurableTask, either in cls.CONFIG or `config` kwarg"
            )
522

523
524
525
        if self.config.output_type is not None:
            assert self.config.output_type in ALL_OUTPUT_TYPES
            self.OUTPUT_TYPE = self.config.output_type
526

527
528
        if self.config.dataset_path is not None:
            self.DATASET_PATH = self.config.dataset_path
529

530
531
        if self.config.dataset_name is not None:
            self.DATASET_NAME = self.config.dataset_name
532

533
534
535
536
        self._metric_fn_list = {}
        self._metric_fn_kwargs = {}
        self._aggregation_list = {}
        self._higher_is_better = {}
537

538
539
        _metric_list = DEFAULT_METRIC_REGISTRY[self.config.output_type]
        if self.config.metric_list is None:
540
            # TODO: handle this in TaskConfig.__post_init__ ?
541
            for metric_name in _metric_list:
haileyschoelkopf's avatar
haileyschoelkopf committed
542
543
                self._metric_fn_list[metric_name] = get_metric(metric_name)
                self._aggregation_list[metric_name] = get_default_aggregation(
544
                    metric_name
haileyschoelkopf's avatar
haileyschoelkopf committed
545
546
                )
                self._higher_is_better[metric_name] = is_higher_better(metric_name)
547
        else:
548
            for metric_config in self.config.metric_list:
549
550
551
552
553
554
555
                assert "metric" in metric_config
                metric_name = metric_config["metric"]
                kwargs = {
                    key: metric_config[key]
                    for key in metric_config
                    if key not in ["metric", "aggregation", "higher_is_better"]
                }
556

557
                if self.config.process_results is not None:
558
559
                    self._metric_fn_list[metric_name] = None
                    self._metric_fn_kwargs[metric_name] = {}
560
561
562
563
564
565
566
567
                elif callable(metric_name):
                    metric_fn = metric_name.__call__
                    metric_name = metric_name.__name__
                    self._metric_fn_list[metric_name] = metric_fn
                    self._metric_fn_kwargs[metric_name] = kwargs
                else:
                    self._metric_fn_list[metric_name] = get_metric(metric_name)
                    self._metric_fn_kwargs[metric_name] = kwargs
lintangsutawika's avatar
lintangsutawika committed
568

569
                if "aggregation" in metric_config:
570
                    agg_name = metric_config["aggregation"]
571
                    if type(agg_name) == str:
haileyschoelkopf's avatar
haileyschoelkopf committed
572
                        self._aggregation_list[metric_name] = get_aggregation(agg_name)
573
574
575
576
                    elif callable(agg_name):
                        self._aggregation_list[metric_name] = metric_config[
                            "aggregation"
                        ]
577
                else:
578
579

                    INV_AGG_REGISTRY = {v: k for k, v in AGGREGATION_REGISTRY.items()}
haileyschoelkopf's avatar
haileyschoelkopf committed
580
                    metric_agg = get_default_aggregation(metric_name)
581
                    eval_logger.warning(
582
583
584
                        f"metric {metric_name} is defined, but aggregation is not. "
                        f"using default "
                        f"aggregation={INV_AGG_REGISTRY[metric_agg]}"
585
                    )
586
                    self._aggregation_list[metric_name] = metric_agg
lintangsutawika's avatar
lintangsutawika committed
587

588
589
590
591
592
593
                if "higher_is_better" in metric_config:
                    self._higher_is_better[metric_name] = metric_config[
                        "higher_is_better"
                    ]
                else:
                    eval_logger.warning(
594
595
                        f"metric {metric_name} is defined, but higher_is_better is not. "
                        f"using default "
haileyschoelkopf's avatar
haileyschoelkopf committed
596
                        f"higher_is_better={is_higher_better(metric_name)}"
597
                    )
haileyschoelkopf's avatar
haileyschoelkopf committed
598
                    self._higher_is_better[metric_name] = is_higher_better(metric_name)
599

600
        self.download(self.config.dataset_kwargs)
601
602
603
        self._training_docs = None
        self._fewshot_docs = None

604
        if self.config.filter_list is not None:
lintangsutawika's avatar
lintangsutawika committed
605
            self._filters = []
606
            for filter_config in self.config.filter_list:
lintangsutawika's avatar
lintangsutawika committed
607
608
609
610
611
612
613
                for filter_pipeline in filter_config:
                    filter_name = filter_config["name"]
                    filter_functions = filter_config["filter"]
                    components = []
                    for function in filter_functions:
                        kwargs = {
                            key: function[key] for key in function if key != "function"
lintangsutawika's avatar
lintangsutawika committed
614
615
616
                        }
                        components.append([function["function"], kwargs])
                    filter_pipeline = build_filter_ensemble(filter_name, components)
lintangsutawika's avatar
lintangsutawika committed
617
                self._filters.append(filter_pipeline)
lintangsutawika's avatar
lintangsutawika committed
618
        else:
619
            self._filters = [build_filter_ensemble("none", [["take_first", None]])]
620

621
622
        if self.config.use_prompt is not None:
            eval_logger.info(f"loading prompt {self.config.use_prompt}")
623
            self.prompt = get_prompt(
624
                self.config.use_prompt, self.DATASET_PATH, self.DATASET_NAME
lintangsutawika's avatar
lintangsutawika committed
625
            )
626
627
628
        else:
            self.prompt = None

lintangsutawika's avatar
lintangsutawika committed
629
630
        if self.fewshot_docs() is not None:
            self.sampler = samplers.Sampler(
631
                list(self.fewshot_docs()), self, rnd=random.Random(1234)
632
            )
633

634
635
636
637
638
639
640
641
642
        if self.has_test_docs():
            docs = self.test_docs()
        elif self.has_validation_docs():
            docs = self.validation_docs()
        else:
            assert (
                False
            ), f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"

643
        # Test One Doc
644
645
646
        self.features = list(docs.features.keys())
        self.multiple_input = 0
        self.multiple_target = 0
647
648
        test_doc = docs[0]
        test_text = self.doc_to_text(test_doc)
649
        test_target = self.doc_to_target(test_doc)
lintangsutawika's avatar
lintangsutawika committed
650

651
        if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
652
653
654
            test_choice = self.doc_to_choice(test_doc)
            if type(test_choice) is not list:
                eval_logger.error("doc_to_choice must return list")
655
656
            else:
                num_choice = len(test_choice)
657

658
659
            if type(test_text) is int:
                self.multiple_input = num_choice
660
661
        else:
            test_choice = None
662

663
        if type(test_target) is list:
664
            self.multiple_target = len(test_target)
665
        else:
lintangsutawika's avatar
lintangsutawika committed
666
            if (type(test_target) is int) and (test_choice is not None):
667
668
669
                test_target = [self.doc_to_choice(test_target)[test_target]]
            else:
                test_target = [test_target]
670

671
672
673
674
675
676
677
678
        if test_choice is not None:
            check_choices = test_choice
        else:
            check_choices = test_target

        for choice in check_choices:
            choice_has_whitespace = True if " " in choice else False
            delimiter_has_whitespace = (
679
                True if " " in self.config.target_delimiter else False
680
681
682
683
684
685
686
            )

            if delimiter_has_whitespace and choice_has_whitespace:
                eval_logger.warning(
                    f'Both target_delimiter and target choice: "{choice}" have whitespace'
                )
            elif (not delimiter_has_whitespace) and (not choice_has_whitespace):
687
                eval_logger.warning(
688
                    f'Both target_delimiter and target choice: "{choice}" does not have whitespace, ignore if the language you are evaluating on does not require/use whitespace'
689
                )
690

691
692
693
694
695
696
697
698
    def download(self, dataset_kwargs=None):

        self.dataset = datasets.load_dataset(
            path=self.DATASET_PATH,
            name=self.DATASET_NAME,
            **dataset_kwargs if dataset_kwargs is not None else {},
        )

baberabb's avatar
baberabb committed
699
    def has_training_docs(self) -> bool:
700
        if self.config.training_split is not None:
701
702
703
704
            return True
        else:
            return False

baberabb's avatar
baberabb committed
705
    def has_validation_docs(self) -> bool:
706
        if self.config.validation_split is not None:
707
708
709
710
            return True
        else:
            return False

baberabb's avatar
baberabb committed
711
    def has_test_docs(self) -> bool:
712
        if self.config.test_split is not None:
713
714
715
716
            return True
        else:
            return False

baberabb's avatar
baberabb committed
717
    def training_docs(self) -> datasets.Dataset:
718
        if self.has_training_docs():
719
720
721
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.training_split]
722
                )
723
            return self.dataset[self.config.training_split]
724

baberabb's avatar
baberabb committed
725
    def validation_docs(self) -> datasets.Dataset:
726
        if self.has_validation_docs():
727
728
729
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.validation_split]
730
                )
731
            return self.dataset[self.config.validation_split]
732

baberabb's avatar
baberabb committed
733
    def test_docs(self) -> datasets.Dataset:
734
        if self.has_test_docs():
735
736
737
            if self.config.process_docs is not None:
                return self.config.process_docs(self.dataset[self.config.test_split])
            return self.dataset[self.config.test_split]
738

739
    def fewshot_docs(self):
740
741
        if self.config.fewshot_split is not None:
            return self.dataset[self.config.fewshot_split]
742
        else:
743
            if self.config.num_fewshot > 0:
744
                eval_logger.warning(
745
                    f"Task '{self.config.task}': "
746
747
748
749
                    "num_fewshot > 0 but fewshot_split is None. "
                    "using preconfigured rule."
                )
            return super().fewshot_docs()
750

751
    def should_decontaminate(self):
752
        return self.config.should_decontaminate
753
754

    def doc_to_decontamination_query(self, doc):
755
756
757
        if self.config.should_decontaminate:
            if self.config.doc_to_decontamination_query in self.features:
                return doc[self.config.doc_to_decontamination_query]
758
759
            else:
                return ast.literal_eval(
760
                    utils.apply_template(self.config.doc_to_decontamination_query, doc)
761
                )
762

763
764
765
766
767
768
769
770
771
772
773
774
    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.
        E.g. `map(self._process_doc, self.dataset["validation"])`

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

    def doc_to_text(self, doc):
775
776
777

        if self.prompt is not None:
            doc_to_text = self.prompt
778
        else:
779
            doc_to_text = self.config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
780

781
782
783
        if type(doc_to_text) == int:
            return doc_to_text
        elif type(doc_to_text) == str:
784
            if doc_to_text in self.features:
785
                # if self.config.doc_to_choice is not None:
786
787
                #     return self.doc_to_choice(doc)[doc[doc_to_text]]
                # else:
788
789
                return doc[doc_to_text]
            else:
lintangsutawika's avatar
lintangsutawika committed
790
791
792
793
794
                text_string = utils.apply_template(doc_to_text, doc)
                if text_string.isdigit():
                    return ast.literal_eval(text_string)
                else:
                    return text_string
795
        elif callable(doc_to_text):
796
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
797
        # Used when applying a Promptsource template
798
        elif hasattr(doc_to_text, "apply"):
799
800
801
802
803
            applied_prompt = doc_to_text.apply(doc)
            if len(applied_prompt) == 2:
                return applied_prompt[0]
            else:
                eval_logger.warning("Applied prompt returns empty string")
804
                return self.config.fewshot_delimiter
805
        else:
806
            print(type(doc_to_text))
807
            raise TypeError
808

809
    def doc_to_target(self, doc: dict) -> Union[int, str, list]:
810
811
812

        if self.prompt is not None:
            doc_to_target = self.prompt
813
        else:
814
            doc_to_target = self.config.doc_to_target
815

816
817
818
        if type(doc_to_target) == int:
            return doc_to_target
        elif type(doc_to_target) == str:
819
            if doc_to_target in self.features:
820
                # if self.config.doc_to_choice is not None:
821
822
823
                #     return self.doc_to_choice(doc)[doc[doc_to_target]]
                # else:
                return doc[doc_to_target]
824
            else:
lintangsutawika's avatar
lintangsutawika committed
825
826
827
                target_string = utils.apply_template(doc_to_target, doc)
                if target_string.isdigit():
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
828
829
830
831
832
                elif (
                    len(target_string) >= 2
                    and (target_string[0] == "[")
                    and (target_string[-1] == "]")
                ):
833
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
834
835
                else:
                    return target_string
836
837
        elif type(doc_to_target) == list:
            return doc_to_target
838
        elif callable(doc_to_target):
839
            return doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
840
        # Used when applying a Promptsource template
841
        elif hasattr(doc_to_target, "apply"):
842
            applied_prompt = doc_to_target.apply(doc)
843
844
845
846
            if len(applied_prompt) == 2:
                return applied_prompt[1]
            else:
                eval_logger.warning("Applied prompt returns empty string")
847
                return self.config.fewshot_delimiter
848
849
        else:
            raise TypeError
850

baberabb's avatar
baberabb committed
851
    def doc_to_choice(self, doc: Any) -> List[str]:
852
853
854

        if self.prompt is not None:
            doc_to_choice = self.prompt
855
        elif self.config.doc_to_choice is None:
856
857
            eval_logger.error("doc_to_choice was called but not set in config")
        else:
858
            doc_to_choice = self.config.doc_to_choice
859
860
861
862
863
864
865
866
867
868
869
870
871

        if type(doc_to_choice) == str:
            return ast.literal_eval(utils.apply_template(doc_to_choice, doc))
        elif type(doc_to_choice) == list:
            return doc_to_choice
        elif type(doc_to_choice) == dict:
            return list(doc_to_choice.values())
        elif callable(doc_to_choice):
            return doc_to_choice(doc)
        elif hasattr(doc_to_choice, "get_answer_choices_list"):
            return doc_to_choice.get_answer_choices_list(doc)
        else:
            raise TypeError
872

873
    def gold_alias(self, doc):
874
875
876
877
878
        # returns a version of the gold target answer to a document,
        # which should be passed into metric for scoring as the ground truth.

        # in multiple_choice tasks, this should be castable to an int corresponding to the index
        # within the answer choices, while doc_to_target is the string version of {{answer_choices[gold]}}.
879
880
        if self.config.gold_alias is not None:
            doc_to_target = self.config.gold_alias
881
        else:
lintangsutawika's avatar
lintangsutawika committed
882
            return self.doc_to_target(doc)
883
884
885
886
887
888
889
890
891
892

        if type(doc_to_target) == str:
            return utils.apply_template(doc_to_target, doc)
        elif callable(doc_to_target):
            return doc_to_target(doc)
        elif hasattr(doc_to_target, "apply"):
            return doc_to_target.apply(doc)[1]
        else:
            raise TypeError

baberabb's avatar
baberabb committed
893
894
895
    def construct_requests(
        self, doc: dict, ctx: str, **kwargs
    ) -> Union[List[Instance], Instance]:
896

897
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
898
            arguments = (ctx, self.doc_to_target(doc))
899
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
900
            arguments = (self.doc_to_target(doc),)
901
        elif self.OUTPUT_TYPE == "multiple_choice":
902
903

            choices = self.doc_to_choice(doc)
904
            target_delimiter = self.config.target_delimiter
905
906
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
907
                cont = self.doc_to_target(doc)
908
                arguments = [(ctx, f"{target_delimiter}{cont}") for ctx in choices]
909
            else:
910
                # Otherwise they are placed in the continuation
911
                arguments = [(ctx, f"{target_delimiter}{cont}") for cont in choices]
912

913
            request_list = [
914
915
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
916
                    doc=doc,
917
                    arguments=arg,
918
                    idx=i,
919
920
                    **kwargs,
                )
921
                for i, arg in enumerate(arguments)
922
            ]
923
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
924
            if "acc_mutual_info" in self._metric_fn_list.keys():
925
926
927
                # if we are calculating multiple choice accuracy
                # using mutual information instead of raw loglikelihood as metric, need unconditional lls.

lintangsutawika's avatar
lintangsutawika committed
928
                # here mutual info refers to calculating
929
930
931
932
933
934
                # log(P(choice|ctx) / P(choice)) = log(P(choice|ctx)) - log(P(choice))
                # in other words normalizing by subtracting the unconditional logprob of each choice.
                request_list.extend(
                    [
                        Instance(
                            request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
935
                            doc=doc,
936
                            arguments=("", "{}".format(choice)),
937
938
939
                            idx=i,
                            **kwargs,
                        )
lintangsutawika's avatar
lintangsutawika committed
940
                        for i, choice in enumerate(choices)
941
942
943
                    ]
                )
            return request_list
lintangsutawika's avatar
lintangsutawika committed
944

945
        elif self.OUTPUT_TYPE == "greedy_until":
946
            arguments = (ctx, self.config.generation_kwargs)
lintangsutawika's avatar
lintangsutawika committed
947
948

        return Instance(
lintangsutawika's avatar
lintangsutawika committed
949
950
            request_type=self.OUTPUT_TYPE, doc=doc, arguments=arguments, idx=0, **kwargs
        )
951
952
953

    def process_results(self, doc, results):

954
955
        if callable(self.config.process_results):
            return self.config.process_results(doc, results)
lintangsutawika's avatar
lintangsutawika committed
956

957
        result_dict = {}
958
        use_metric = list(self._metric_fn_list.keys())
959
960
961
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
962
963
964
965
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
966
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
967
            (loglikelihood,) = results
968
969
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
haileyschoelkopf's avatar
haileyschoelkopf committed
970
            return {
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
                **(
                    {"word_perplexity": (loglikelihood, _words)}
                    if "word_perplexity" in use_metric
                    else {}
                ),
                **(
                    {"byte_perplexity": (loglikelihood, _bytes)}
                    if "byte_perplexity" in use_metric
                    else {}
                ),
                **(
                    {"bits_per_byte": (loglikelihood, _bytes)}
                    if "bits_per_byte" in use_metric
                    else {}
                ),
haileyschoelkopf's avatar
haileyschoelkopf committed
986
            }
987
        elif self.OUTPUT_TYPE == "multiple_choice":
988
989

            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
990

991
            # retrieve choices in List[str] form, to compute choice lengths, etc.
992
            choices = self.doc_to_choice(doc)
993
994
            completion_len = np.array([float(len(i)) for i in choices])

995
996
            if (
                2 * len(choices) == len(lls)
997
                and "acc_mutual_info" in self._metric_fn_list.keys()
998
999
1000
1001
1002
1003
1004
            ):
                # then we are doing mutual info.
                # this stores the "dryrun" / unconditional answer loglikelihoods
                lls_unconditional = lls[1::2]
                assert len(lls_unconditional) == len(choices)
                # and this stores our "regular" conditional loglikelihoods
                lls = lls[::2]
1005

1006
1007
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
1008

1009
1010
            if self.multiple_input:
                gold = self.doc_to_text(doc)
1011
            else:
1012
                gold = self.doc_to_target(doc)
1013
1014
                if type(gold) is str:
                    gold = choices.index(gold)
lintangsutawika's avatar
lintangsutawika committed
1015

1016
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
1017
1018
                acc = 1.0 if pred in gold else 0.0
                acc_norm = 1.0 if pred_norm in gold else 0.0
1019
                exact_match = int(any([is_greedy[i] for i in gold]))
lintangsutawika's avatar
lintangsutawika committed
1020
1021
1022
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
1023
1024
                # TODO: this gets score of 0 on arc_challenge for pythia-70m. need to test that this works properly
                exact_match = int(is_greedy[gold])
1025
1026

            result_dict = {
1027
                **({"acc": acc} if "acc" in use_metric else {}),
1028
1029
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
1030
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
1031
                **({"exact_match": exact_match} if "exact_match" in use_metric else {}),
1032
1033
            }

1034
            if "acc_mutual_info" in use_metric:
lintangsutawika's avatar
lintangsutawika committed
1035
1036
1037
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
1038
1039
1040
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

1041
1042
        elif self.OUTPUT_TYPE == "greedy_until":

1043
            gold = self.doc_to_target(doc)
1044
            if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1045
                # If you set doc_to_choice,
lintangsutawika's avatar
lintangsutawika committed
1046
                # it assumes that doc_to_target returns a number.
1047
1048
                choices = self.doc_to_choice(doc)
                gold = choices[gold]
lintangsutawika's avatar
lintangsutawika committed
1049
1050
            else:
                gold = str(gold)
1051

1052
            for key, result in zip(self._metric_fn_list.keys(), results):
haileyschoelkopf's avatar
haileyschoelkopf committed
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
                if self.multiple_target:
                    # in the case where we have multiple targets,
                    # return true if any are true
                    # TODO: this may break for multipLe_target, non zero-or-1 metrics
                    scores = []
                    for gold_option in gold:
                        res = self._metric_fn_list[key](
                            references=[gold_option],
                            predictions=[result],
                            **self._metric_fn_kwargs[key],
                        )
                        if isinstance(res, dict):
                            # TODO: this handles the case where HF evaluate returns a dict.
                            res = res[key]
                        scores.append(res)
                    if any(scores):
1069
                        result_score = 1.0
haileyschoelkopf's avatar
haileyschoelkopf committed
1070
                    else:
1071
                        result_score = 0.0
haileyschoelkopf's avatar
haileyschoelkopf committed
1072
                else:
1073
                    result_score = self._metric_fn_list[key](
haileyschoelkopf's avatar
haileyschoelkopf committed
1074
1075
1076
1077
                        references=[gold],
                        predictions=[result],
                        **self._metric_fn_kwargs[key],
                    )
1078

1079
1080
                if isinstance(result_score, dict):
                    result_dict.update(result_score)
haileyschoelkopf's avatar
haileyschoelkopf committed
1081
                else:
1082
                    result_dict[key] = result_score
1083
        else:
lintangsutawika's avatar
lintangsutawika committed
1084
1085
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
1086
                "'loglikelihood', 'loglikelihood_rolling', 'greedy_until' or 'multiple_choice'",
1087
            )
1088
1089
1090
1091
1092
1093
1094

        return result_dict

    def aggregation(self):
        return self._aggregation_list

    def higher_is_better(self):
haileyschoelkopf's avatar
haileyschoelkopf committed
1095
        return self._higher_is_better
1096
1097
1098
1099
1100


class MultipleChoiceTask(Task):
    OUTPUT_TYPE: str = "loglikelihood"

baberabb's avatar
baberabb committed
1101
    def doc_to_target(self, doc: dict) -> str:
1102
1103
        return " " + doc["choices"][doc["gold"]]

baberabb's avatar
baberabb committed
1104
    def construct_requests(self, doc: dict, ctx: str, **kwargs) -> List[Instance]:
1105
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1106
1107
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1108
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1109
                doc=doc,
1110
                arguments=(ctx, " {}".format(choice)),
1111
                idx=i,
1112
1113
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1114
1115
            for i, choice in enumerate(doc["choices"])
        ]
1116

baberabb's avatar
baberabb committed
1117
    def process_results(self, doc: dict, results: List[Tuple[float, bool]]) -> dict:
lintangsutawika's avatar
lintangsutawika committed
1118
1119
1120
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
        gold = doc["gold"]

        acc = 1.0 if np.argmax(results) == gold else 0.0
        completion_len = np.array([float(len(i)) for i in doc["choices"]])
        acc_norm = 1.0 if np.argmax(results / completion_len) == gold else 0.0

        return {
            "acc": acc,
            "acc_norm": acc_norm,
        }

baberabb's avatar
baberabb committed
1132
    def higher_is_better(self) -> dict:
1133
1134
1135
1136
1137
        return {
            "acc": True,
            "acc_norm": True,
        }

baberabb's avatar
baberabb committed
1138
    def aggregation(self) -> dict:
1139
1140
1141
1142
1143
1144
        return {
            "acc": mean,
            "acc_norm": mean,
        }


lintangsutawika's avatar
lintangsutawika committed
1145
class PerplexityTask(Task):
1146
1147
    OUTPUT_TYPE = "loglikelihood_rolling"

baberabb's avatar
baberabb committed
1148
    def has_training_docs(self) -> bool:
1149
1150
        return False

baberabb's avatar
baberabb committed
1151
    def fewshot_examples(self, k: int, rnd) -> List:
1152
1153
1154
        assert k == 0
        return []

baberabb's avatar
baberabb committed
1155
    def fewshot_context(self, doc: dict, num_fewshot: int) -> Literal[""]:
1156
1157
1158
1159
1160
1161
        assert (
            num_fewshot == 0
        ), "The number of fewshot examples must be 0 for perplexity tasks."

        return ""

baberabb's avatar
baberabb committed
1162
    def higher_is_better(self) -> dict:
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
        return {
            "word_perplexity": False,
            "byte_perplexity": False,
            "bits_per_byte": False,
        }

    def doc_to_decontamination_query(self, doc):
        return doc

    def doc_to_text(self, doc):
        return ""

    def doc_to_target(self, doc):
        return doc

baberabb's avatar
baberabb committed
1178
    def construct_requests(self, doc: dict, ctx: Union[str, None], **kwargs):
1179
1180
        assert not ctx

lintangsutawika's avatar
lintangsutawika committed
1181
1182
1183
1184
1185
1186
1187
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1188

baberabb's avatar
baberabb committed
1189
    def process_results(self, doc: dict, results: float) -> dict:
1190
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1191
1192
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1193
1194
1195
1196
1197
1198
        return {
            "word_perplexity": (loglikelihood, words),
            "byte_perplexity": (loglikelihood, bytes_),
            "bits_per_byte": (loglikelihood, bytes_),
        }

baberabb's avatar
baberabb committed
1199
    def aggregation(self) -> dict:
1200
1201
1202
1203
1204
1205
1206
        return {
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
            "bits_per_byte": bits_per_byte,
        }

    @classmethod
baberabb's avatar
baberabb committed
1207
    def count_bytes(cls, doc) -> int:
1208
1209
1210
        return len(doc.encode("utf-8"))

    @classmethod
baberabb's avatar
baberabb committed
1211
    def count_words(cls, doc) -> int:
1212
1213
        """Downstream tasks with custom word boundaries should override this!"""
        return len(re.split(r"\s+", doc))