task.py 43.4 KB
Newer Older
1
import abc
2
import ast
3
import functools
Herbie Bradley's avatar
Herbie Bradley committed
4
5
6
7
8
9
import itertools
import random
import re
from collections.abc import Callable
from dataclasses import asdict, dataclass, field
from typing import Any, List, Literal, Tuple, Union
10
11

import datasets
Herbie Bradley's avatar
Herbie Bradley committed
12
import evaluate
13
import numpy as np
Herbie Bradley's avatar
Herbie Bradley committed
14
15
16
import scipy.special as sp
import yaml
from tqdm import tqdm
17

18
from lm_eval import utils
19
from lm_eval.api import samplers
lintangsutawika's avatar
lintangsutawika committed
20
from lm_eval.api.filter import FilterEnsemble
Herbie Bradley's avatar
Herbie Bradley committed
21
from lm_eval.api.instance import Instance
lintangsutawika's avatar
lintangsutawika committed
22
23
from lm_eval.api.metrics import (
    bits_per_byte,
Herbie Bradley's avatar
Herbie Bradley committed
24
    mean,
lintangsutawika's avatar
lintangsutawika committed
25
    metric_max_over_ground_truths,
Herbie Bradley's avatar
Herbie Bradley committed
26
    weighted_perplexity,
lintangsutawika's avatar
lintangsutawika committed
27
28
)
from lm_eval.api.registry import (
Herbie Bradley's avatar
Herbie Bradley committed
29
30
31
    AGGREGATION_REGISTRY,
    DEFAULT_METRIC_REGISTRY,
    OUTPUT_TYPE_REGISTRY,
haileyschoelkopf's avatar
haileyschoelkopf committed
32
33
    get_aggregation,
    get_default_aggregation,
Herbie Bradley's avatar
Herbie Bradley committed
34
    get_metric,
haileyschoelkopf's avatar
haileyschoelkopf committed
35
    is_higher_better,
lintangsutawika's avatar
lintangsutawika committed
36
)
Herbie Bradley's avatar
Herbie Bradley committed
37
38
39
from lm_eval.filters import build_filter_ensemble
from lm_eval.logger import eval_logger
from lm_eval.prompts import get_prompt
40

41
42
43
44
45
46
47
ALL_OUTPUT_TYPES = [
    "loglikelihood",
    "multiple_choice",
    "loglikelihood_rolling",
    "greedy_until",
]

48
49
50

@dataclass
class TaskConfig(dict):
51
    # task naming/registry
52
    task: str = None
53
    group: Union[str, list] = None
54
55
56
    # HF dataset options.
    # which dataset to use,
    # and what splits for what purpose
57
58
    dataset_path: str = None
    dataset_name: str = None
59
    dataset_kwargs: dict = None
60
61
62
    training_split: str = None
    validation_split: str = None
    test_split: str = None
lintangsutawika's avatar
lintangsutawika committed
63
    fewshot_split: str = None  # TODO: assert that this not None if num_fewshot > 0. (?) assert if this is same split as one evaling (?)
64
65
    # formatting / prompting options.
    # see docs/advanced_task_guide.md for more info
66
    process_docs: Callable = None
67
68
    doc_to_text: Union[Callable, str] = None
    doc_to_target: Union[Callable, str] = None
lintangsutawika's avatar
lintangsutawika committed
69
    doc_to_choice: Union[Callable, str, dict, list] = None
70
    gold_alias: Union[Callable, str] = None
lintangsutawika's avatar
lintangsutawika committed
71
    process_results: Union[Callable, str] = None
72
    use_prompt: str = None
73
    description: str = ""
74
75
    target_delimiter: str = " "
    fewshot_delimiter: str = "\n\n"
76
    # runtime configuration options
77
    num_fewshot: int = 0
78
    # scoring options
79
80
    metric_list: str = None
    output_type: str = "greedy_until"
81
    generation_kwargs: dict = None
82
    repeats: int = 1
lintangsutawika's avatar
lintangsutawika committed
83
    filter_list: Union[str, list] = None
84
85
    should_decontaminate: bool = False
    doc_to_decontamination_query: str = None
86

lintangsutawika's avatar
lintangsutawika committed
87
    metadata: str = None  # by default, not used in the code. allows for users to pass arbitrary info to tasks
88

89
    def __post_init__(self):
Lintang Sutawika's avatar
Lintang Sutawika committed
90
91
92
        if self.generation_kwargs is not None:
            if self.output_type != "greedy_until":
                eval_logger.warning(
93
                    "passed `generation_kwargs`, but not using `output_type: greedy_until`!"
Lintang Sutawika's avatar
Lintang Sutawika committed
94
                )
95
                assert self.output_type != "greedy_until"
Lintang Sutawika's avatar
Lintang Sutawika committed
96
97
98
99
100
101
102

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

            if "until" not in self.generation_kwargs:
103
                self.generation_kwargs["until"] = [self.fewshot_delimiter]
Lintang Sutawika's avatar
Lintang Sutawika committed
104
105
106
107
        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
108
                    "until": None
109
110
                    if self.fewshot_delimiter is None
                    else [self.fewshot_delimiter],
Lintang Sutawika's avatar
Lintang Sutawika committed
111
112
113
                    "do_sample": False,
                    "temperature": 0.0,
                }
114

haileyschoelkopf's avatar
haileyschoelkopf committed
115
116
        # TODO: how to make TaskConfigs be de- and re-serializable, even when using the !function constructor?

117
118
119
    def __getitem__(self, item):
        return getattr(self, item)

120
121
122
    def __setitem__(self, item, value):
        return setattr(self, item, value)

123
    def to_dict(self):
124
125
        """dumps the current config as a dictionary object, as a printable format.
        null fields will not be printed.
haileyschoelkopf's avatar
haileyschoelkopf committed
126
        Used for dumping results alongside full task configuration
127

haileyschoelkopf's avatar
haileyschoelkopf committed
128
129
130
131
132
133
134
135
136
137
        :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
138
139
140
            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
141
        return cfg_dict
142

143
144
145
146
147
148
149
150
151
152
153
154

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
155

156
157
158
159
160
161
162
163
    # 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
164

165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
    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
199
        self._config = TaskConfig(**config) if config else TaskConfig()
200
201
202

        if not hasattr(self, "_filters"):
            self._filters = []
lintangsutawika's avatar
lintangsutawika committed
203
            for name, components in self._config.get(
204
                "filters", [["none", [["take_first", None]]]]
lintangsutawika's avatar
lintangsutawika committed
205
            ):
206
207
208
                filter_pipeline = build_filter_ensemble(name, components)
                self._filters.append(filter_pipeline)

lintangsutawika's avatar
lintangsutawika committed
209
        self.sampler = samplers.Sampler(
210
211
            list(self.fewshot_docs()), self, rnd=random.Random(1234)
        )
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237

    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.
        """
238
239
240
241
242
243
244
        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,
        )
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281

    @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 []

282
283
284
285
286
287
288
289
290
291
    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
292
            eval_logger.warning(
293
                "has_training_docs and has_validation_docs are False"
294
                ", using test_docs as fewshot_docs but this is not recommended."
lintangsutawika's avatar
lintangsutawika committed
295
            )
296
297
            return self.test_docs()

298
299
300
301
302
303
304
305
306
307
    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
308

309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
    @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

336
    def build_all_requests(self, limit=None, rank=None, world_size=None):
337
338
339
340
341
342
343
344
345
346
        """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!"

347
348
349
350
        eval_logger.info(
            f"Building contexts for task '{self._config.task}' on rank {rank}..."
        )

351
        instances = []
352
353
        for doc_id, doc in utils.create_iterator(
            enumerate(docs), rank, world_size, limit
lintangsutawika's avatar
lintangsutawika committed
354
        ):
355
            # sample fewshot context #TODO: need to offset doc_id by rank now!
356
            fewshot_ctx = self.fewshot_context(
357
358
                doc,
                self._config.num_fewshot,
359
            )
360

haileyschoelkopf's avatar
haileyschoelkopf committed
361
            # TODO: we should override self._config.repeats if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
362
363
364
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
365
                metadata=(self._config["task"], doc_id, self._config.repeats),
lintangsutawika's avatar
lintangsutawika committed
366
            )
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391

            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
392
            The number of times each instance in a dataset is inferred on. Defaults to 1,
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
            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
428
429
430
431
432
433
434
435
436
437
    @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))

438
    @utils.positional_deprecated
439
    def fewshot_context(self, doc, num_fewshot):
440
441
442
443
444
445
446
447
448
449
450
451
        """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:
452
453
            # always prepend the (possibly empty) task description
            labeled_examples = self._config.description
454
        else:
lintangsutawika's avatar
lintangsutawika committed
455
456
457
            labeled_examples = self._config.description + self.sampler.get_context(
                doc, num_fewshot
            )
458
459

        example = self.doc_to_text(doc)
460
461
462
463
        if type(example) == str:
            return labeled_examples + example
        elif type(example) == list:
            return [labeled_examples + ex for ex in example]
464
        elif type(example) == int:
lintangsutawika's avatar
lintangsutawika committed
465
466
467
468
469
            if self._config.doc_to_choice is not None:
                choices = self.doc_to_choice(doc)
                return labeled_examples + choices[example]
            else:
                return labeled_examples + str(example)
470
471

    def apply_filters(self):
lintangsutawika's avatar
lintangsutawika committed
472
473
474
475
476
477
        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
478

baberabb's avatar
baberabb committed
479
    def dump_config(self) -> dict:
480
        """Returns a dictionary representing the task's config.
481
482
483
484
485

        :returns: str
            The fewshot context.
        """
        # TODO: this should only return the overrides applied to a non-YAML task's configuration.
486
        # (num_fewshot)
487
488
        return self._config.to_dict()

489
490

class ConfigurableTask(Task):
491
    VERSION = "Yaml"
492
    OUTPUT_TYPE = None
493
    CONFIG = None
494
495
496

    def __init__(
        self, data_dir=None, cache_dir=None, download_mode=None, config: dict = None
baberabb's avatar
baberabb committed
497
    ):  # TODO no super() call here
498
        # Get pre-configured attributes
499
        self._config = self.CONFIG
500

501
502
        # Use new configurations if there was no preconfiguration
        if self._config is None:
503
            self._config = TaskConfig(**config)
504
505
        # Overwrite configs
        else:
lintangsutawika's avatar
lintangsutawika committed
506
            if config is not None:
507
                self._config.__dict__.update(config)
508

509
        if self._config is None:
lintangsutawika's avatar
lintangsutawika committed
510
511
512
            raise ValueError(
                "Must pass a config to ConfigurableTask, either in cls.CONFIG or `config` kwarg"
            )
513
514

        if self._config.output_type is not None:
515
            assert self._config.output_type in ALL_OUTPUT_TYPES
516
517
            self.OUTPUT_TYPE = self._config.output_type

518
519
520
521
522
523
        if self._config.dataset_path is not None:
            self.DATASET_PATH = self._config.dataset_path

        if self._config.dataset_name is not None:
            self.DATASET_NAME = self._config.dataset_name

524
525
526
527
        self._metric_fn_list = {}
        self._metric_fn_kwargs = {}
        self._aggregation_list = {}
        self._higher_is_better = {}
528

529
        _metric_list = DEFAULT_METRIC_REGISTRY[self._config.output_type]
530
        if self._config.metric_list is None:
531
            # TODO: handle this in TaskConfig.__post_init__ ?
532
            for metric_name in _metric_list:
haileyschoelkopf's avatar
haileyschoelkopf committed
533
534
                self._metric_fn_list[metric_name] = get_metric(metric_name)
                self._aggregation_list[metric_name] = get_default_aggregation(
535
                    metric_name
haileyschoelkopf's avatar
haileyschoelkopf committed
536
537
                )
                self._higher_is_better[metric_name] = is_higher_better(metric_name)
538
539
540
541
542
543
544
545
546
        else:
            for metric_config in self._config.metric_list:
                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"]
                }
547

548
                if self._config.process_results is not None:
549
550
                    self._metric_fn_list[metric_name] = None
                    self._metric_fn_kwargs[metric_name] = {}
551
552
553
554
555
556
557
558
                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
559

560
                if "aggregation" in metric_config:
561
                    agg_name = metric_config["aggregation"]
562
                    if type(agg_name) == str:
haileyschoelkopf's avatar
haileyschoelkopf committed
563
                        self._aggregation_list[metric_name] = get_aggregation(agg_name)
564
565
566
567
                    elif callable(agg_name):
                        self._aggregation_list[metric_name] = metric_config[
                            "aggregation"
                        ]
568
                else:
569
                    INV_AGG_REGISTRY = {v: k for k, v in AGGREGATION_REGISTRY.items()}
haileyschoelkopf's avatar
haileyschoelkopf committed
570
                    metric_agg = get_default_aggregation(metric_name)
571
                    eval_logger.warning(
572
573
574
                        f"metric {metric_name} is defined, but aggregation is not. "
                        f"using default "
                        f"aggregation={INV_AGG_REGISTRY[metric_agg]}"
575
                    )
576
                    self._aggregation_list[metric_name] = metric_agg
lintangsutawika's avatar
lintangsutawika committed
577

578
579
580
581
582
583
                if "higher_is_better" in metric_config:
                    self._higher_is_better[metric_name] = metric_config[
                        "higher_is_better"
                    ]
                else:
                    eval_logger.warning(
584
585
                        f"metric {metric_name} is defined, but higher_is_better is not. "
                        f"using default "
haileyschoelkopf's avatar
haileyschoelkopf committed
586
                        f"higher_is_better={is_higher_better(metric_name)}"
587
                    )
haileyschoelkopf's avatar
haileyschoelkopf committed
588
                    self._higher_is_better[metric_name] = is_higher_better(metric_name)
589

590
        self.download(self._config.dataset_kwargs)
591
592
593
        self._training_docs = None
        self._fewshot_docs = None

lintangsutawika's avatar
lintangsutawika committed
594
        if self._config.filter_list is not None:
lintangsutawika's avatar
lintangsutawika committed
595
            self._filters = []
lintangsutawika's avatar
lintangsutawika committed
596
597
598
599
600
601
602
603
            for filter_config in self._config.filter_list:
                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
604
605
606
                        }
                        components.append([function["function"], kwargs])
                    filter_pipeline = build_filter_ensemble(filter_name, components)
lintangsutawika's avatar
lintangsutawika committed
607
                self._filters.append(filter_pipeline)
lintangsutawika's avatar
lintangsutawika committed
608
        else:
609
            self._filters = [build_filter_ensemble("none", [["take_first", None]])]
610
611

        if self._config.use_prompt is not None:
lintangsutawika's avatar
lintangsutawika committed
612
            eval_logger.info(f"loading prompt {self._config.use_prompt}")
613
            self.prompt = get_prompt(
lintangsutawika's avatar
lintangsutawika committed
614
615
                self._config.use_prompt, self.DATASET_PATH, self.DATASET_NAME
            )
616
617
618
        else:
            self.prompt = None

lintangsutawika's avatar
lintangsutawika committed
619
620
        if self.fewshot_docs() is not None:
            self.sampler = samplers.Sampler(
621
                list(self.fewshot_docs()), self, rnd=random.Random(1234)
622
            )
623

624
625
626
627
628
629
630
631
632
        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!"

633
        # Test One Doc
634
635
636
        self.features = list(docs.features.keys())
        self.multiple_input = 0
        self.multiple_target = 0
637
638
        test_doc = docs[0]
        test_text = self.doc_to_text(test_doc)
639
        test_target = self.doc_to_target(test_doc)
lintangsutawika's avatar
lintangsutawika committed
640
641
642
643
644

        if self._config.doc_to_choice is not None:
            test_choice = self.doc_to_choice(test_doc)
            if type(test_choice) is not list:
                eval_logger.error("doc_to_choice must return list")
645
646
            else:
                num_choice = len(test_choice)
647

648
649
            if type(test_text) is int:
                self.multiple_input = num_choice
650

651
        if type(test_target) is list:
652
653
            self.multiple_target = len(test_target)

Herbie Bradley's avatar
Herbie Bradley committed
654
    def download(self, dataset_kwargs=None):
655
656
657
658
659
660
        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
661
    def has_training_docs(self) -> bool:
662
663
664
665
666
        if self._config.training_split is not None:
            return True
        else:
            return False

baberabb's avatar
baberabb committed
667
    def has_validation_docs(self) -> bool:
668
669
670
671
672
        if self._config.validation_split is not None:
            return True
        else:
            return False

baberabb's avatar
baberabb committed
673
    def has_test_docs(self) -> bool:
674
675
676
677
678
        if self._config.test_split is not None:
            return True
        else:
            return False

baberabb's avatar
baberabb committed
679
    def training_docs(self) -> datasets.Dataset:
680
        if self.has_training_docs():
681
            if self._config.process_docs is not None:
682
683
684
                return self._config.process_docs(
                    self.dataset[self._config.training_split]
                )
685
686
            return self.dataset[self._config.training_split]

baberabb's avatar
baberabb committed
687
    def validation_docs(self) -> datasets.Dataset:
688
        if self.has_validation_docs():
689
            if self._config.process_docs is not None:
690
691
692
                return self._config.process_docs(
                    self.dataset[self._config.validation_split]
                )
693
694
            return self.dataset[self._config.validation_split]

baberabb's avatar
baberabb committed
695
    def test_docs(self) -> datasets.Dataset:
696
        if self.has_test_docs():
697
            if self._config.process_docs is not None:
698
                return self._config.process_docs(self.dataset[self._config.test_split])
699
700
            return self.dataset[self._config.test_split]

701
    def fewshot_docs(self):
702
        if self._config.fewshot_split is not None:
703
            return self.dataset[self._config.fewshot_split]
704
705
706
        else:
            if self._config.num_fewshot > 0:
                eval_logger.warning(
haileyschoelkopf's avatar
haileyschoelkopf committed
707
                    f"Task '{self._config.task}': "
708
709
710
711
                    "num_fewshot > 0 but fewshot_split is None. "
                    "using preconfigured rule."
                )
            return super().fewshot_docs()
712

713
714
715
716
717
    def should_decontaminate(self):
        return self._config.should_decontaminate

    def doc_to_decontamination_query(self, doc):
        if self._config.should_decontaminate:
718
719
720
721
722
723
            if self._config.doc_to_decontamination_query in self.features:
                return doc[self._config.doc_to_decontamination_query]
            else:
                return ast.literal_eval(
                    utils.apply_template(self._config.doc_to_decontamination_query, doc)
                )
724

725
726
727
728
729
730
731
732
733
734
735
736
    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):
737
738
        if self.prompt is not None:
            doc_to_text = self.prompt
739
740
        else:
            doc_to_text = self._config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
741

742
743
744
        if type(doc_to_text) == int:
            return doc_to_text
        elif type(doc_to_text) == str:
745
            if doc_to_text in self.features:
746
747
748
                # if self._config.doc_to_choice is not None:
                #     return self.doc_to_choice(doc)[doc[doc_to_text]]
                # else:
749
750
                return doc[doc_to_text]
            else:
lintangsutawika's avatar
lintangsutawika committed
751
752
753
754
755
                text_string = utils.apply_template(doc_to_text, doc)
                if text_string.isdigit():
                    return ast.literal_eval(text_string)
                else:
                    return text_string
756
        elif callable(doc_to_text):
757
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
758
        # Used when applying a Promptsource template
759
        elif hasattr(doc_to_text, "apply"):
760
761
762
763
764
            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")
765
                return self._config.fewshot_delimiter
766
        else:
767
            print(type(doc_to_text))
768
            raise TypeError
769

770
    def doc_to_target(self, doc: dict) -> Union[int, str, list]:
771
772
        if self.prompt is not None:
            doc_to_target = self.prompt
773
774
775
        else:
            doc_to_target = self._config.doc_to_target

776
777
778
        if type(doc_to_target) == int:
            return doc_to_target
        elif type(doc_to_target) == str:
779
            if doc_to_target in self.features:
780
781
782
783
                # if self._config.doc_to_choice is not None:
                #     return self.doc_to_choice(doc)[doc[doc_to_target]]
                # else:
                return doc[doc_to_target]
784
            else:
lintangsutawika's avatar
lintangsutawika committed
785
786
787
                target_string = utils.apply_template(doc_to_target, doc)
                if target_string.isdigit():
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
788
789
790
791
792
                elif (
                    len(target_string) >= 2
                    and (target_string[0] == "[")
                    and (target_string[-1] == "]")
                ):
793
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
794
795
                else:
                    return target_string
796
797
        elif type(doc_to_target) == list:
            return doc_to_target
798
        elif callable(doc_to_target):
799
            return doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
800
        # Used when applying a Promptsource template
801
        elif hasattr(doc_to_target, "apply"):
802
            applied_prompt = doc_to_target.apply(doc)
803
804
805
806
            if len(applied_prompt) == 2:
                return applied_prompt[1]
            else:
                eval_logger.warning("Applied prompt returns empty string")
807
                return self._config.fewshot_delimiter
808
809
        else:
            raise TypeError
810

baberabb's avatar
baberabb committed
811
    def doc_to_choice(self, doc: Any) -> List[str]:
812
813
        if self.prompt is not None:
            doc_to_choice = self.prompt
lintangsutawika's avatar
lintangsutawika committed
814
        elif self._config.doc_to_choice is None:
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
            eval_logger.error("doc_to_choice was called but not set in config")
        else:
            doc_to_choice = self._config.doc_to_choice
        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
830

831
    def gold_alias(self, doc):
832
833
834
835
836
        # 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]}}.
lintangsutawika's avatar
lintangsutawika committed
837
        if self._config.gold_alias is not None:
838
839
            doc_to_target = self._config.gold_alias
        else:
lintangsutawika's avatar
lintangsutawika committed
840
            return self.doc_to_target(doc)
841
842
843
844
845
846
847
848
849
850

        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
851
852
853
    def construct_requests(
        self, doc: dict, ctx: str, **kwargs
    ) -> Union[List[Instance], Instance]:
854
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
855
            arguments = (ctx, self.doc_to_target(doc))
856
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
857
            arguments = (self.doc_to_target(doc),)
858
        elif self.OUTPUT_TYPE == "multiple_choice":
859
            choices = self.doc_to_choice(doc)
860
            target_delimiter = self._config.target_delimiter
861
862
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
863
                cont = self.doc_to_target(doc)
864
                arguments = [(ctx, f"{target_delimiter}{cont}") for ctx in choices]
865
            else:
866
                # Otherwise they are placed in the continuation
867
                arguments = [(ctx, f"{target_delimiter}{cont}") for cont in choices]
868

869
            request_list = [
870
871
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
872
                    doc=doc,
873
                    arguments=arg,
874
                    idx=i,
875
876
                    **kwargs,
                )
877
                for i, arg in enumerate(arguments)
878
            ]
879
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
880
            if "acc_mutual_info" in self._metric_fn_list.keys():
881
882
883
                # if we are calculating multiple choice accuracy
                # using mutual information instead of raw loglikelihood as metric, need unconditional lls.

lintangsutawika's avatar
lintangsutawika committed
884
                # here mutual info refers to calculating
885
886
887
888
889
890
                # 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
891
                            doc=doc,
892
                            arguments=("", "{}".format(choice)),
893
894
895
                            idx=i,
                            **kwargs,
                        )
lintangsutawika's avatar
lintangsutawika committed
896
                        for i, choice in enumerate(choices)
897
898
899
                    ]
                )
            return request_list
lintangsutawika's avatar
lintangsutawika committed
900

901
        elif self.OUTPUT_TYPE == "greedy_until":
902
            arguments = (ctx, self._config.generation_kwargs)
lintangsutawika's avatar
lintangsutawika committed
903
904

        return Instance(
lintangsutawika's avatar
lintangsutawika committed
905
906
            request_type=self.OUTPUT_TYPE, doc=doc, arguments=arguments, idx=0, **kwargs
        )
907
908

    def process_results(self, doc, results):
lintangsutawika's avatar
lintangsutawika committed
909
910
        if callable(self._config.process_results):
            return self._config.process_results(doc, results)
lintangsutawika's avatar
lintangsutawika committed
911

912
        result_dict = {}
913
        use_metric = list(self._metric_fn_list.keys())
914
915
916
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
917
918
919
920
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
921
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
922
            (loglikelihood,) = results
923
924
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
haileyschoelkopf's avatar
haileyschoelkopf committed
925
            return {
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
                **(
                    {"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
941
            }
942
        elif self.OUTPUT_TYPE == "multiple_choice":
943
            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
944

945
            # retrieve choices in List[str] form, to compute choice lengths, etc.
946
            choices = self.doc_to_choice(doc)
947
948
            completion_len = np.array([float(len(i)) for i in choices])

949
            if 2 * len(choices) == len(lls) and "acc_mutual_info" in use_metric:
950
951
952
953
954
955
                # 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]
956

957
958
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
959

960
961
            if self.multiple_input:
                gold = self.doc_to_text(doc)
962
            else:
963
                gold = self.doc_to_target(doc)
964
965
                if type(gold) is str:
                    gold = choices.index(gold)
966
967
968

            if "ece" in use_metric:
                # Convert lls from log-probabilities to normalized probabilities
969
970
                norm_probs: list[float] = np.exp(lls - sp.logsumexp(lls)).tolist()
                calib_scores: list[float] = [0.0] * len(choices)
971
972
973
974
975
                if isinstance(gold, list):
                    for g in gold:
                        calib_scores[g] = 1.0
                else:
                    calib_scores[gold] = 1.0
976
                calibration_probs: dict[str, list[float]] = {
977
978
979
980
                    "probs": norm_probs,
                    "scores": calib_scores,
                }

981
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
982
983
                acc = 1.0 if pred in gold else 0.0
                acc_norm = 1.0 if pred_norm in gold else 0.0
984
                exact_match = int(any([is_greedy[i] for i in gold]))
lintangsutawika's avatar
lintangsutawika committed
985
986
987
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
988
989
                # 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])
990
991

            result_dict = {
992
                **({"acc": acc} if "acc" in use_metric else {}),
993
994
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
995
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
996
                **({"exact_match": exact_match} if "exact_match" in use_metric else {}),
997
                **({"ece": calibration_probs} if "ece" in use_metric else {}),
998
999
            }

1000
            if "acc_mutual_info" in use_metric:
lintangsutawika's avatar
lintangsutawika committed
1001
1002
1003
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
1004
1005
1006
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

1007
        elif self.OUTPUT_TYPE == "greedy_until":
1008
            gold = self.doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
1009
            if self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1010
                # If you set doc_to_choice,
lintangsutawika's avatar
lintangsutawika committed
1011
                # it assumes that doc_to_target returns a number.
1012
1013
                choices = self.doc_to_choice(doc)
                gold = choices[gold]
lintangsutawika's avatar
lintangsutawika committed
1014
1015
            else:
                gold = str(gold)
1016

1017
            for key, result in zip(self._metric_fn_list.keys(), results):
haileyschoelkopf's avatar
haileyschoelkopf committed
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
                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):
1034
                        result_score = 1.0
haileyschoelkopf's avatar
haileyschoelkopf committed
1035
                    else:
1036
                        result_score = 0.0
haileyschoelkopf's avatar
haileyschoelkopf committed
1037
                else:
1038
                    result_score = self._metric_fn_list[key](
haileyschoelkopf's avatar
haileyschoelkopf committed
1039
1040
1041
1042
                        references=[gold],
                        predictions=[result],
                        **self._metric_fn_kwargs[key],
                    )
1043

1044
1045
                if isinstance(result_score, dict):
                    result_dict.update(result_score)
haileyschoelkopf's avatar
haileyschoelkopf committed
1046
                else:
1047
                    result_dict[key] = result_score
1048
        else:
lintangsutawika's avatar
lintangsutawika committed
1049
1050
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
1051
                "'loglikelihood', 'loglikelihood_rolling', 'greedy_until' or 'multiple_choice'",
1052
            )
1053
1054
1055
1056
1057
1058
1059

        return result_dict

    def aggregation(self):
        return self._aggregation_list

    def higher_is_better(self):
haileyschoelkopf's avatar
haileyschoelkopf committed
1060
        return self._higher_is_better
1061
1062
1063
1064
1065


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

baberabb's avatar
baberabb committed
1066
    def doc_to_target(self, doc: dict) -> str:
1067
1068
        return " " + doc["choices"][doc["gold"]]

baberabb's avatar
baberabb committed
1069
    def construct_requests(self, doc: dict, ctx: str, **kwargs) -> List[Instance]:
1070
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1071
1072
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1073
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1074
                doc=doc,
1075
                arguments=(ctx, " {}".format(choice)),
1076
                idx=i,
1077
1078
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1079
1080
            for i, choice in enumerate(doc["choices"])
        ]
1081

baberabb's avatar
baberabb committed
1082
    def process_results(self, doc: dict, results: List[Tuple[float, bool]]) -> dict:
lintangsutawika's avatar
lintangsutawika committed
1083
1084
1085
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
        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
1097
    def higher_is_better(self) -> dict:
1098
1099
1100
1101
1102
        return {
            "acc": True,
            "acc_norm": True,
        }

baberabb's avatar
baberabb committed
1103
    def aggregation(self) -> dict:
1104
1105
1106
1107
1108
1109
        return {
            "acc": mean,
            "acc_norm": mean,
        }


lintangsutawika's avatar
lintangsutawika committed
1110
class PerplexityTask(Task):
1111
1112
    OUTPUT_TYPE = "loglikelihood_rolling"

baberabb's avatar
baberabb committed
1113
    def has_training_docs(self) -> bool:
1114
1115
        return False

baberabb's avatar
baberabb committed
1116
    def fewshot_examples(self, k: int, rnd) -> List:
1117
1118
1119
        assert k == 0
        return []

baberabb's avatar
baberabb committed
1120
    def fewshot_context(self, doc: dict, num_fewshot: int) -> Literal[""]:
1121
1122
1123
1124
1125
1126
        assert (
            num_fewshot == 0
        ), "The number of fewshot examples must be 0 for perplexity tasks."

        return ""

baberabb's avatar
baberabb committed
1127
    def higher_is_better(self) -> dict:
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
        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
1143
    def construct_requests(self, doc: dict, ctx: Union[str, None], **kwargs):
1144
1145
        assert not ctx

lintangsutawika's avatar
lintangsutawika committed
1146
1147
1148
1149
1150
1151
1152
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1153

baberabb's avatar
baberabb committed
1154
    def process_results(self, doc: dict, results: float) -> dict:
1155
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1156
1157
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1158
1159
1160
1161
1162
1163
        return {
            "word_perplexity": (loglikelihood, words),
            "byte_perplexity": (loglikelihood, bytes_),
            "bits_per_byte": (loglikelihood, bytes_),
        }

baberabb's avatar
baberabb committed
1164
    def aggregation(self) -> dict:
1165
1166
1167
1168
1169
1170
1171
        return {
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
            "bits_per_byte": bits_per_byte,
        }

    @classmethod
baberabb's avatar
baberabb committed
1172
    def count_bytes(cls, doc) -> int:
1173
1174
1175
        return len(doc.encode("utf-8"))

    @classmethod
baberabb's avatar
baberabb committed
1176
    def count_words(cls, doc) -> int:
1177
1178
        """Downstream tasks with custom word boundaries should override this!"""
        return len(re.split(r"\s+", doc))