"docs/source/en/api/schedulers/unipc.md" did not exist on "856dad57bb7a9ee13af4a08492e524b0a145a2c5"
task.py 60.6 KB
Newer Older
Baber's avatar
cleanup  
Baber committed
1
2
from __future__ import annotations

3
import abc
4
import ast
lintangsutawika's avatar
lintangsutawika committed
5
import logging
6
import random
7
8
import re
from collections.abc import Callable
9
from copy import deepcopy
10
from functools import cached_property
Baber's avatar
Baber committed
11
from typing import TYPE_CHECKING, Any, Literal, overload
12
13
14

import datasets
import numpy as np
15
from tqdm import tqdm
Baber's avatar
Baber committed
16
from typing_extensions import deprecated
17
18

from lm_eval import utils
19
20
from lm_eval.api.instance import Instance, OutputType
from lm_eval.api.metrics import bits_per_byte, mean, weighted_perplexity
21
from lm_eval.api.utils import check_gold_index_error
22
from lm_eval.caching.cache import load_from_cache, save_to_cache
Baber's avatar
Baber committed
23
from lm_eval.config.metric import MetricConfig
Baber's avatar
Baber committed
24
from lm_eval.config.task import DataSet, TaskConfig
25
26
from lm_eval.filters import build_filter_ensemble

27

28
29
30
31
ALL_OUTPUT_TYPES = [
    "loglikelihood",
    "multiple_choice",
    "loglikelihood_rolling",
32
    "generate_until",
33
34
]

Baber's avatar
cleanup  
Baber committed
35
if TYPE_CHECKING:
Baber's avatar
Baber committed
36
    pass
Baber's avatar
cleanup  
Baber committed
37
38


Lintang Sutawika's avatar
Lintang Sutawika committed
39
eval_logger = logging.getLogger(__name__)
40

lintangsutawika's avatar
lintangsutawika committed
41

42
43
44
45
46
47
48
49
50
51
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)
    """

Baber's avatar
cleanup  
Baber committed
52
    VERSION: int | str | None = None
53

54
55
    # The name of the `Task` benchmark as denoted in the HuggingFace datasets Hub
    # or a path to a custom `datasets` loading script.
Baber's avatar
cleanup  
Baber committed
56
    DATASET_PATH: str | None = None
57
58

    # The name of a subset within `DATASET_PATH`.
Baber's avatar
cleanup  
Baber committed
59
    DATASET_NAME: str | None = None
60

Baber's avatar
cleanup  
Baber committed
61
    OUTPUT_TYPE: OutputType | None = None
lintangsutawika's avatar
lintangsutawika committed
62

63
64
    def __init__(
        self,
Baber's avatar
cleanup  
Baber committed
65
66
67
68
        data_dir: str | None = None,
        cache_dir: str | None = None,
        download_mode: datasets.DownloadMode | None = None,
        config: Mapping | None = None,  # Union[dict, TaskConfig]
Ethan Smith's avatar
Ethan Smith committed
69
    ) -> None:
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
        """
        :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
            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)
Baber's avatar
cleanup  
Baber committed
92
93
94
        self._training_docs: list | None = None
        self._fewshot_docs: list | None = None
        self._instances: list[Instance] | None = None
95

96
        self._config: TaskConfig = TaskConfig.from_yaml({**config})
97

98
        self._filters = [build_filter_ensemble("none", [("take_first", None)])]
Baber's avatar
cleanup  
Baber committed
99
        self.fewshot_rnd: random.Random | None = (
100
101
            None  # purposely induce errors in case of improper usage
        )
102

103
104
    def download(
        self,
Baber's avatar
cleanup  
Baber committed
105
106
        data_dir: str | None = None,
        cache_dir: str | None = None,
107
108
        download_mode=None,
    ) -> None:
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
        """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.
        """
Baber's avatar
Baber committed
133
        assert self.DATASET_PATH is not None, "DATASET_PATH must be set in Task class"
134
135
136
137
138
139
140
        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,
        )
141

142
    @property
143
    def config(self) -> TaskConfig:
144
145
146
        """Returns the TaskConfig associated with this class."""
        return self._config

147
    @property
Baber's avatar
Baber committed
148
    def has_training_docs(self) -> bool:
149
        """Whether the task has a training set"""
Baber's avatar
Baber committed
150
        raise NotImplementedError
151

152
    @property
Baber's avatar
Baber committed
153
    def has_validation_docs(self) -> bool:
154
        """Whether the task has a validation set"""
Baber's avatar
Baber committed
155
        raise NotImplementedError
156

157
    @property
Baber's avatar
Baber committed
158
    def has_test_docs(self) -> bool:
159
        """Whether the task has a test set"""
Baber's avatar
Baber committed
160
        raise NotImplementedError
161

Baber's avatar
Baber committed
162
    def training_docs(self) -> DataSet | None:
163
164
165
166
167
168
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

Baber's avatar
Baber committed
169
    def validation_docs(self) -> DataSet | None:
170
171
172
173
174
175
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

Baber's avatar
Baber committed
176
    def test_docs(self) -> DataSet | None:
177
178
179
180
181
182
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

Baber's avatar
Baber committed
183
    def fewshot_docs(self) -> DataSet | None:
184
185
186
187
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
188
        if self.has_training_docs:
189
            return self.training_docs()
190
        elif self.has_validation_docs:
191
192
            return self.validation_docs()
        else:
Baber's avatar
Baber committed
193
            if self.config.num_fewshot and self.config.num_fewshot > 0:
Baber Abbasi's avatar
Baber Abbasi committed
194
195
196
197
                eval_logger.warning(
                    f"[Task: {self.config.task}] has_training_docs and has_validation_docs are False"
                    ", using test_docs as fewshot_docs but this is not recommended."
                )
198
199
            return self.test_docs()

200
    def _process_doc(self, doc: dict) -> dict:
201
202
203
204
205
206
207
208
209
        """
        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
210

211
    @property
Baber's avatar
cleanup  
Baber committed
212
    def instances(self) -> list[Instance]:
213
214
215
216
217
        """After calling `task.build_all_requests()`, tasks
        maintain a list of the dataset instances which will be evaluated.
        """
        return self._instances

218
    def fewshot_examples(self, k: int, rnd) -> Iterable[dict]:
219
220
221
222
223
        if self._training_docs is None:
            self._training_docs = list(self.training_docs())

        return rnd.sample(self._training_docs, k)

Baber's avatar
cleanup  
Baber committed
224
    def doc_to_decontamination_query(self, doc: dict):
225
        raise NotImplementedError(
226
227
228
229
            "Override doc_to_decontamination_query with document specific decontamination query."
        )

    @abc.abstractmethod
Baber's avatar
cleanup  
Baber committed
230
    def doc_to_text(self, doc: dict) -> str:
231
232
233
        pass

    @abc.abstractmethod
Baber's avatar
cleanup  
Baber committed
234
    def doc_to_target(self, doc: dict) -> str | int:
235
236
        pass

237
    # not an abstractmethod because not every language-only task has to implement this
Baber's avatar
cleanup  
Baber committed
238
    def doc_to_image(self, doc: dict):
239
240
        raise NotImplementedError

Baber's avatar
cleanup  
Baber committed
241
    def doc_to_audio(self, doc: dict):
242
243
        raise NotImplementedError

Baber's avatar
cleanup  
Baber committed
244
    def doc_to_prefix(self, doc: dict) -> str:
Baber Abbasi's avatar
Baber Abbasi committed
245
246
        return ""

247
248
    def build_all_requests(
        self,
249
        *,
Baber's avatar
cleanup  
Baber committed
250
251
        limit: int | None = None,
        samples: list[int] | None = None,
252
253
254
255
        rank: int = 0,
        world_size: int = 1,
        cache_requests: bool = False,
        rewrite_requests_cache: bool = False,
Baber's avatar
cleanup  
Baber committed
256
        system_instruction: str | None = None,
257
258
        apply_chat_template: bool = False,
        fewshot_as_multiturn: bool = False,
Baber's avatar
cleanup  
Baber committed
259
        chat_template: Callable | None = None,
260
        tokenizer_name: str = "",
261
    ) -> None:
262
        """Build a set of Instances for a task, and store them in task.instances"""
263
264
265
266

        # used with caching
        og_limit = limit

267
        cache_key = f"requests-{self._config.task}-{self.config.num_fewshot}shot-rank{rank}-world_size{world_size}"
KonradSzafer's avatar
KonradSzafer committed
268
269
270
271
272
273
274
        cache_key += "-chat_template" if apply_chat_template else ""
        cache_key += "-fewshot_as_multiturn" if fewshot_as_multiturn else ""
        cache_key += (
            f"-system_prompt_hash{utils.hash_string(system_instruction)}"
            if system_instruction is not None
            else ""
        )
275
        cache_key += f"-tokenizer{tokenizer_name}"
276

Baber Abbasi's avatar
Baber Abbasi committed
277
        cached_instances = load_from_cache(file_name=cache_key, cache=cache_requests)
278
279
280
281
282
283
284
285
286
287
288
289
290

        if cache_requests and cached_instances and not rewrite_requests_cache:
            cached_instances = cached_instances[:limit]

            flattened_instances = [
                instance
                for instance_group in cached_instances
                for instance in instance_group
            ]

            self._instances = flattened_instances
            return

Baber Abbasi's avatar
Baber Abbasi committed
291
        eval_logger.info(f"Building contexts for {self.config.task} on rank {rank}...")
292

293
        instances = []
294
295
296
297
298
299
300
301
302
303

        # process all documents when caching is specified for simplicity
        if (
            cache_requests
            and (not cached_instances or rewrite_requests_cache)
            and limit is not None
        ):
            limit = None

        doc_id_docs = list(
304
305
306
            self.doc_iterator(
                rank=rank, limit=limit, samples=samples, world_size=world_size
            )
307
308
309
310
311
312
313
        )

        num_docs = len(doc_id_docs)

        for doc_id, doc in tqdm(
            doc_id_docs,
            total=num_docs,
lintangsutawika's avatar
lintangsutawika committed
314
        ):
315
            # sample fewshot context #TODO: need to offset doc_id by rank now!
316
            fewshot_ctx = self.fewshot_context(
317
                doc,
318
319
320
321
322
323
324
                num_fewshot=0
                if self.config.num_fewshot is None
                else self.config.num_fewshot,
                system_instruction=system_instruction,
                apply_chat_template=apply_chat_template,
                fewshot_as_multiturn=fewshot_as_multiturn,
                chat_template=chat_template,
Baber Abbasi's avatar
Baber Abbasi committed
325
                gen_prefix=self.doc_to_prefix(doc),
326
            )
327

328
            # TODO: we should override self.config.repeats if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
329
330
331
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
Baber's avatar
Baber committed
332
                metadata=(self.config.task, doc_id, self.config.repeats),
333
                apply_chat_template=apply_chat_template,
334
                chat_template=chat_template,
lintangsutawika's avatar
lintangsutawika committed
335
            )
336
337
338
339

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

340
341
342
343
344
345
346
347
348
349
350
351
352
            instances.append(inst)

        # now flatten, this is to allow slicing to work with pickles

        sliced_instances = instances[:og_limit]

        flattened_instances = [
            instance
            for instance_group in sliced_instances
            for instance in instance_group
        ]

        self._instances = flattened_instances
353

354
355
        if len(self._instances) == 0:
            raise ValueError("task.build_requests() did not find any docs!")
356

357
358
359
        if cache_requests and (not cached_instances or rewrite_requests_cache):
            save_to_cache(file_name=cache_key, obj=instances)

360
    @abc.abstractmethod
Baber's avatar
cleanup  
Baber committed
361
    def construct_requests(self, doc: dict, ctx: list[dict] | str, **kwargs):
362
363
364
365
366
367
368
369
370
371
372
373
374
375
        """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
376
            The number of times each instance in a dataset is inferred on. Defaults to 1,
377
378
379
380
381
            can be increased for techniques like majority voting.
        """
        pass

    @abc.abstractmethod
382
    def process_results(self, doc: dict, results: list) -> dict[str, Any]:
383
384
385
386
387
388
389
390
391
        """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.
        """
392
        raise NotImplementedError
393

Baber's avatar
Baber committed
394
    @deprecated("not used anymore")
395
396
397
398
399
400
    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
        """
Baber's avatar
cleanup  
Baber committed
401
        return True
402

Baber's avatar
Baber committed
403
    @deprecated("not used anymore")
404
405
406
407
408
409
    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
        """
Baber's avatar
cleanup  
Baber committed
410
        return True
411

412
413
414
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

haileyschoelkopf's avatar
haileyschoelkopf committed
415
    @classmethod
Baber's avatar
Baber committed
416
    def count_bytes(cls, doc: str) -> int:
haileyschoelkopf's avatar
haileyschoelkopf committed
417
418
419
420
        """Used for byte-level perplexity metrics in rolling loglikelihood"""
        return len(doc.encode("utf-8"))

    @classmethod
Baber's avatar
Baber committed
421
    def count_words(cls, doc: str) -> int:
haileyschoelkopf's avatar
haileyschoelkopf committed
422
423
424
        """Downstream loglikelihood_rolling perplexity tasks with custom word boundaries should override this!"""
        return len(re.split(r"\s+", doc))

425
    @utils.positional_deprecated
Baber Abbasi's avatar
Baber Abbasi committed
426
    def fewshot_context(self, doc, num_fewshot, rnd=None, description=None, **kwargs):
427
428
429
430
431
432
433
        """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.
lintangsutawika's avatar
lintangsutawika committed
434
435
436
437
438
        :param rnd: random.Random
            The pseudo-random number generator used to randomly sample examples.
            WARNING: This is currently a required arg although it's optionalized with a default `None`.
        :param description: str
            The task's description that will be prepended to the fewshot examples.
439
440
441
        :returns: str
            The fewshot context.
        """
442
        if rnd is None:
443
444
445
446
447
448
            if self.fewshot_rnd is not None:
                rnd = self.fewshot_rnd
            else:
                raise ValueError(
                    "A `random.Random` generator argument must be provided to `rnd`"
                )
lintangsutawika's avatar
lintangsutawika committed
449

450
        description = description if description else ""
451
452

        if num_fewshot == 0:
lintangsutawika's avatar
lintangsutawika committed
453
            labeled_examples = ""
454
        else:
lintangsutawika's avatar
lintangsutawika committed
455
            # for sets with no training docs, draw from other set *but ensure no overlap with current doc*
456
            if self.has_training_docs:
lintangsutawika's avatar
lintangsutawika committed
457
458
459
460
461
                fewshotex = self.fewshot_examples(k=num_fewshot, rnd=rnd)
            else:
                if self._fewshot_docs is None:
                    self._fewshot_docs = list(
                        self.validation_docs()
462
                        if self.has_validation_docs
lintangsutawika's avatar
lintangsutawika committed
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
                        else self.test_docs()
                    )

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

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

            labeled_examples = (
                "\n\n".join(
                    [
                        self.doc_to_text(doc) + self.doc_to_target(doc)
                        for doc in fewshotex
                    ]
                )
                + "\n\n"
lintangsutawika's avatar
lintangsutawika committed
479
            )
480
481

        example = self.doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
482
        return description + labeled_examples + example
483

Baber's avatar
cleanup  
Baber committed
484
    def apply_filters(self) -> list[Instance] | None:
Baber Abbasi's avatar
Baber Abbasi committed
485
        """Iterates over FilterEnsembles and applies them to instances"""
Baber's avatar
cleanup  
Baber committed
486
        if hasattr(self, "_filters") and self._instances:
lintangsutawika's avatar
lintangsutawika committed
487
            for f in self._filters:
488
                f.apply(self._instances)
lintangsutawika's avatar
lintangsutawika committed
489
        else:
Baber's avatar
cleanup  
Baber committed
490
491
492
            eval_logger.warning(
                "No filter defined or no instances, passing through instances"
            )
lintangsutawika's avatar
lintangsutawika committed
493
            return self._instances
494

baberabb's avatar
baberabb committed
495
    def dump_config(self) -> dict:
Baber Abbasi's avatar
Baber Abbasi committed
496
        """Returns the config as a dictionary."""
497
        # TODO: this should only return the overrides applied to a non-YAML task's configuration.
498
        # (num_fewshot)
499
        return self.config.to_dict()
500

Baber Abbasi's avatar
Baber Abbasi committed
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
    def set_config(self, key: str, value: Any, update: bool = False) -> None:
        """Set or update the configuration for a given key."""
        if update:
            current_value = getattr(self._config, key, {})
            if not isinstance(current_value, dict):
                raise TypeError(
                    f"Expected a dict for key '{key}', got {type(current_value).__name__} instead."
                )
            current_value.update(value)
        else:
            setattr(self._config, key, value)

    def override_metric(self, metric_name: str) -> None:
        """
        Override the default metrics used for evaluation with custom metrics.

        Parameters:
        - metric_name (str): The name of the custom metric to override. Should be registered in api.metrics.
        """
Baber's avatar
Baber committed
520
521
522
523
524
        # if not isinstance(self, ConfigurableTask):
        #     self.process_results = lambda x, y: {metric_name: get_metric(metric_name)}
        #     self.aggregation = lambda: {
        #         metric_name: get_metric_aggregation(metric_name)
        #     }
Baber's avatar
Baber committed
525
526
        self._config.metric_list = [MetricConfig(name=metric_name)]
        self._config.process_results = lambda *args: {"bypass": 0}
Baber Abbasi's avatar
Baber Abbasi committed
527

Baber's avatar
cleanup  
Baber committed
528
    def set_fewshot_seed(self, seed: int | None = None) -> None:
529
530
531
532
        self.fewshot_rnd = random.Random(seed)
        if hasattr(self, "sampler"):
            self.sampler.rnd = self.fewshot_rnd

533
    @property
Baber's avatar
cleanup  
Baber committed
534
    def eval_docs(self) -> datasets.Dataset | Iterable[dict]:
535
        if self.has_test_docs:
536
            return self.test_docs()
537
        elif self.has_validation_docs:
538
539
            return self.validation_docs()
        else:
540
541
542
            raise ValueError(
                f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"
            )
543
544

    def doc_iterator(
545
546
547
        self,
        *,
        rank: int = 0,
Baber's avatar
cleanup  
Baber committed
548
        limit: int | None = None,
549
        world_size: int = 1,
Baber's avatar
cleanup  
Baber committed
550
551
        samples: list[int] | None = None,
    ) -> Iterator[tuple[int, Any]]:
552
553
        if samples:
            n = len(self.eval_docs)
Baber's avatar
cleanup  
Baber committed
554
            assert all(e < n for e in samples), (
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
                f"Elements of --samples should be in the interval [0,k-1] where k is the number of total examples. In this case, k={n}."
            )
            eval_logger.info(
                f"{self.config.task}: Evaluating on {len(samples)} examples"
            )
            doc_iterator = utils.create_iterator(
                enumerate(x for i, x in enumerate(self.eval_docs) if i in samples),
                rank=int(rank),
                limit=None,  # limit does not matter here since we are selecting samples directly
                world_size=int(world_size),
            )
        else:
            limit = int(limit) if limit else None
            doc_iterator = utils.create_iterator(
                enumerate(self.eval_docs),
                rank=int(rank),
                limit=limit,
                world_size=int(world_size),
            )
574
575
        return doc_iterator

576
577

class ConfigurableTask(Task):
578
    VERSION = "Yaml"
579
    OUTPUT_TYPE = None
580
    CONFIG = None
581
582

    def __init__(
583
584
585
586
        self,
        data_dir=None,
        cache_dir=None,
        download_mode=None,
Baber's avatar
Baber committed
587
        config: Mapping[str, Any] | None = None,
Baber's avatar
Baber committed
588
    ) -> None:
589
        # Get pre-configured attributes
590
        self._config = self.CONFIG
591

592
        # Use new configurations if there was no preconfiguration
593
        if self.config is None:
594
            self._config = TaskConfig.from_yaml(config)
595
596
        # Overwrite configs
        else:
lintangsutawika's avatar
lintangsutawika committed
597
            if config is not None:
598
                self._config.__dict__.update(config)
599

600
        if self.config is None:
lintangsutawika's avatar
lintangsutawika committed
601
602
603
            raise ValueError(
                "Must pass a config to ConfigurableTask, either in cls.CONFIG or `config` kwarg"
            )
604

Baber's avatar
cleanup  
Baber committed
605
606
        if isinstance(self.config.metadata, dict) and "version" in self.config.metadata:
            self.VERSION = self.config.metadata["version"]
607

608
        if self.config.output_type is not None:
609
610
611
612
            if self.config.output_type not in ALL_OUTPUT_TYPES:
                raise ValueError(
                    f"Got invalid output_type '{self.config.output_type}', must be in '{','.join(ALL_OUTPUT_TYPES)}'"
                )
613
            self.OUTPUT_TYPE = self.config.output_type
614

615
616
617
618
        if self.config.doc_to_image is not None:
            # mark the task as requiring multimodality.
            self.MULTIMODAL = True

619
620
621
622
        if self.config.doc_to_audio:
            # mark the task as requiring multimodality.
            self.MULTIMODAL = True

Hojin Lee's avatar
Hojin Lee committed
623
624
625
        if self.config.unsafe_code is not False:
            self.UNSAFE_CODE = True

626
627
        if self.config.dataset_path is not None:
            self.DATASET_PATH = self.config.dataset_path
628

629
630
        if self.config.dataset_name is not None:
            self.DATASET_NAME = self.config.dataset_name
631

632
        # self.metric_list: list[MetricConfig] = self.config.get_metrics
633

634
        self.download(self.config.dataset_kwargs)
635
636
637
        self._training_docs = None
        self._fewshot_docs = None

Baber's avatar
Baber committed
638
        self._filters = self.config.get_filters
Baber's avatar
Baber committed
639

Baber's avatar
Baber committed
640
641
642
643
644
645
646
        # if self.config.use_prompt is not None:
        #     eval_logger.info(f"loading prompt {self.config.use_prompt}")
        #     self.prompt = get_prompt(
        #         self.config.use_prompt, self.DATASET_PATH, self.DATASET_NAME
        #     )
        # else:
        #     self.prompt = None
647

648
649
650
651
        if (
            self.config.fewshot_cfg.num_fewshot() > 0
            and self.fewshot_docs() is not None
        ):
Baber's avatar
Baber committed
652
653
654
            self.fewshot_rnd = random.Random()
            self.sampler = self.config.fewshot_cfg.init_sampler(
                list(self.fewshot_docs()), self, rnd=self.fewshot_rnd
655
            )
656
        self.task_docs = self.eval_docs
657

Baber's avatar
Baber committed
658
659
660
661
662
663
664
665
666
667
668
669
        # for name, fn in self.config._fn.items():
        #     if hasattr(self, name):
        #         setattr(
        #             self,
        #             name,
        #             types.MethodType(
        #                 lambda self, *args, _fn=fn, **kwargs: _fn(*args, **kwargs),
        #                 self,
        #             ),
        #         )

        self.runtime_checks(self.task_docs[0])
670

Baber Abbasi's avatar
Baber Abbasi committed
671
    def download(
Baber's avatar
cleanup  
Baber committed
672
        self, dataset_kwargs:dict[str, Any] | None = None, **kwargs
Baber Abbasi's avatar
Baber Abbasi committed
673
    ) -> None:
Baber Abbasi's avatar
Baber Abbasi committed
674
675
        from packaging.version import parse as vparse

676
677
678
679
        self.config.dataset_kwargs, self.config.metadata = (
            self.config.dataset_kwargs or {},
            self.config.metadata or {},
        )
Baber Abbasi's avatar
Baber Abbasi committed
680
681
        if dataset_kwargs and vparse(datasets.__version__) >= vparse("4.0.0"):
            dataset_kwargs.pop("trust_remote_code", None)
682
        if isinstance(df := self.config.custom_dataset, Callable):
Baber Abbasi's avatar
Baber Abbasi committed
683
684
685
686
            eval_logger.warning(
                f"{self.config.task}: Custom kwargs can be passed to `--metadata` in console (as json string) or to the TaskManager."
                + "\nFor example --metadata='{\"max_seq_lengths\":[4096, 8192]}'. For details see task Readme."
            )
687
            self.dataset = df(**(self.config.dataset_kwargs | self.config.metadata))
Baber Abbasi's avatar
Baber Abbasi committed
688
        else:
Baber's avatar
Baber committed
689
690
691
            assert self.config.dataset_path is not None, (
                "dataset_path must be set in TaskConfig"
            )
Baber Abbasi's avatar
Baber Abbasi committed
692
            self.dataset = datasets.load_dataset(
693
694
695
                path=self.config.dataset_path,
                name=self.config.dataset_name,
                **self.config.dataset_kwargs,
Baber Abbasi's avatar
Baber Abbasi committed
696
            )
697

698
    @cached_property
baberabb's avatar
baberabb committed
699
    def has_training_docs(self) -> bool:
Baber's avatar
cleanup  
Baber committed
700
        return self.config.training_split is not None
701

702
    @cached_property
baberabb's avatar
baberabb committed
703
    def has_validation_docs(self) -> bool:
Baber's avatar
cleanup  
Baber committed
704
        return self.config.validation_split is not None
705

706
    @cached_property
baberabb's avatar
baberabb committed
707
    def has_test_docs(self) -> bool:
Baber's avatar
cleanup  
Baber committed
708
        return self.config.test_split is not None
709

Baber's avatar
Baber committed
710
    def training_docs(self) -> DataSet | None:
711
        if self.has_training_docs:
712
713
714
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.training_split]
715
                )
716
            return self.dataset[self.config.training_split]
717

Baber's avatar
Baber committed
718
    def validation_docs(self) -> DataSet | None:
719
        if self.has_validation_docs:
720
721
722
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.validation_split]
723
                )
724
            return self.dataset[self.config.validation_split]
725

Baber's avatar
Baber committed
726
    def test_docs(self) -> DataSet | None:
727
        if self.has_test_docs:
728
729
730
            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]
731

732
    def fewshot_docs(self):
Baber's avatar
Baber committed
733
734
735
736
737
738
        docs = self.config.fewshot_cfg.get_docs(self.dataset)

        if docs is not None:
            return docs

        # Fallback to parent implementation
Baber's avatar
cleanup  
Baber committed
739
740
741
742
743
744
745
746
747
748
        if (
            (_num_fewshot := self.config.num_fewshot)
            and isinstance(_num_fewshot, int)
            and _num_fewshot > 0
        ):
            eval_logger.warning(
                f"[Task: {self.config.task}] "
                "num_fewshot > 0 but no fewshot source configured. "
                "Using preconfigured rule."
            )
Baber's avatar
Baber committed
749
750

        return super().fewshot_docs()
751

KonradSzafer's avatar
KonradSzafer committed
752
753
    @staticmethod
    def append_target_question(
Baber's avatar
cleanup  
Baber committed
754
        labeled_examples: list[dict[str, str]],
KonradSzafer's avatar
KonradSzafer committed
755
756
        question: str,
        fewshot_as_multiturn: bool = False,
Baber's avatar
cleanup  
Baber committed
757
        gen_prefix: str | None = None,
KonradSzafer's avatar
KonradSzafer committed
758
759
760
761
762
763
764
765
    ) -> None:
        """Adds a target question to the labeled examples list.
        If fewshot_as_multiturn is True, or labeled_examples is empty, or the last entry is a system turn, appends the question as a new user entry.
        Otherwise, it is appended to the last user entry, ensuring that the conversation alternates between the user and the assistant.
        """
        if not fewshot_as_multiturn:
            # if no messages or last message is system, append as new user entry
            if len(labeled_examples) == 0 or labeled_examples[-1]["role"] == "system":
766
                labeled_examples.append({"role": "user", "content": question})
KonradSzafer's avatar
KonradSzafer committed
767
768
            # if last message is user, append to it to avoid two user messages in a row
            else:
769
                labeled_examples[-1]["content"] += question
KonradSzafer's avatar
KonradSzafer committed
770
771
        else:
            # if fewshot_as_multiturn is True, append as next user entry (last is always assistant)
772
            labeled_examples.append({"role": "user", "content": question})
Baber Abbasi's avatar
Baber Abbasi committed
773
774
        if gen_prefix:
            labeled_examples.append({"role": "assistant", "content": gen_prefix})
KonradSzafer's avatar
KonradSzafer committed
775

lintangsutawika's avatar
lintangsutawika committed
776
    @utils.positional_deprecated
KonradSzafer's avatar
KonradSzafer committed
777
778
    def fewshot_context(
        self,
Baber Abbasi's avatar
Baber Abbasi committed
779
        doc: dict,
KonradSzafer's avatar
KonradSzafer committed
780
        num_fewshot: int,
Baber's avatar
cleanup  
Baber committed
781
        system_instruction: str | None = None,
KonradSzafer's avatar
KonradSzafer committed
782
783
        apply_chat_template: bool = False,
        fewshot_as_multiturn: bool = False,
Baber's avatar
cleanup  
Baber committed
784
785
786
        chat_template: Callable | None = None,
        gen_prefix: str | None = None,
    ) -> str | list[str] | None:
lintangsutawika's avatar
lintangsutawika committed
787
788
789
790
791
792
793
        """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.
KonradSzafer's avatar
KonradSzafer committed
794
795
796
797
798
799
        :param  system_instruction: str
            System instruction to be applied to the prompt.
        :param apply_chat_template: bool
            Whether to apply the chat template to the fewshot context.
        :param fewshot_as_multiturn: bool
            Whether to provide the fewshot examples as a multiturn conversation or a single user turn.
800
801
        :param chat_template:
            callable (from lm.apply_chat_template) that takes in a list[Dict] chat transcript and renders it into a string.
802
803
        :param gen_prefix:
            String to append after the <|assistant|> token.
lintangsutawika's avatar
lintangsutawika committed
804
805
806
        :returns: str
            The fewshot context.
        """
Baber's avatar
cleanup  
Baber committed
807
        labeled_examples = [] if apply_chat_template else ""
KonradSzafer's avatar
KonradSzafer committed
808
809

        # get task description
810
811
        if description := self.config.description:
            description = utils.apply_template(self.config.description, doc)
lintangsutawika's avatar
lintangsutawika committed
812

KonradSzafer's avatar
KonradSzafer committed
813
814
815
816
817
818
819
820
821
        # create system prompt based on the provided system instruction and description
        if system_instruction is not None and description:
            system_prompt = (
                f"{system_instruction}{self.sampler.fewshot_delimiter}{description}"
            )
        elif system_instruction is not None:
            system_prompt = system_instruction
        elif description:
            system_prompt = description
lintangsutawika's avatar
lintangsutawika committed
822
        else:
KonradSzafer's avatar
KonradSzafer committed
823
824
825
826
827
828
829
830
831
832
833
834
835
            system_prompt = ""

        # add system prompt if specified
        if system_prompt:
            if apply_chat_template:
                labeled_examples.append({"role": "system", "content": system_prompt})
            else:
                labeled_examples = system_prompt
        # if few-shot - append examples after the system prompt
        if num_fewshot > 0:
            if apply_chat_template:
                labeled_examples.extend(
                    self.sampler.get_chat_context(
Baber Abbasi's avatar
Baber Abbasi committed
836
837
838
                        doc,
                        num_fewshot,
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
839
                        gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
840
841
842
                    )
                )
            else:
Baber Abbasi's avatar
Baber Abbasi committed
843
                labeled_examples += self.sampler.get_context(
Baber Abbasi's avatar
Baber Abbasi committed
844
                    doc, num_fewshot, gen_prefix=gen_prefix
Baber Abbasi's avatar
Baber Abbasi committed
845
                )
lintangsutawika's avatar
lintangsutawika committed
846
847

        example = self.doc_to_text(doc)
KonradSzafer's avatar
KonradSzafer committed
848
849
        if apply_chat_template:
            if self.multiple_input:
Baber Abbasi's avatar
Baber Abbasi committed
850
                # TODO: append prefill?
851
852
                if not labeled_examples:
                    return ""
853
                return chat_template(labeled_examples)
KonradSzafer's avatar
KonradSzafer committed
854
855
            if isinstance(example, str):
                self.append_target_question(
Baber Abbasi's avatar
Baber Abbasi committed
856
857
858
                    labeled_examples,
                    example,
                    fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
859
                    gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
860
861
862
863
864
865
866
                )
            # for loglikelihood create a list of questions with appended choices
            elif isinstance(example, list):
                labeled_examples_list = []
                # copy chat history for each example and append the answer
                for ex in example:
                    chat = deepcopy(labeled_examples)
Baber Abbasi's avatar
Baber Abbasi committed
867
868
869
870
                    self.append_target_question(
                        chat,
                        ex,
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
871
                        gen_prefix=gen_prefix,
Baber Abbasi's avatar
Baber Abbasi committed
872
873
874
875
876
                    )
                    # TODO: append prefill?
                    labeled_examples_list.append(
                        chat_template(
                            chat,
Baber's avatar
cleanup  
Baber committed
877
                            add_generation_prompt=not gen_prefix,
Baber Abbasi's avatar
Baber Abbasi committed
878
879
                        )
                    )
KonradSzafer's avatar
KonradSzafer committed
880
881
882
883
884
885
                return labeled_examples_list
            # if example is an integer, append the choice or convert to string
            elif isinstance(example, int):
                if self.config.doc_to_choice is not None:
                    choices = self.doc_to_choice(doc)
                    self.append_target_question(
Baber Abbasi's avatar
Baber Abbasi committed
886
887
888
                        labeled_examples,
                        choices[example],
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
889
                        gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
890
891
892
                    )
                else:
                    self.append_target_question(
Baber Abbasi's avatar
Baber Abbasi committed
893
894
895
                        labeled_examples,
                        str(example),
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
896
                        gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
897
898
                    )
                # return lm.apply_chat_template(labeled_examples)
Baber Abbasi's avatar
Baber Abbasi committed
899
900
            return chat_template(
                labeled_examples,
Baber's avatar
cleanup  
Baber committed
901
                add_generation_prompt=not gen_prefix,
Baber Abbasi's avatar
Baber Abbasi committed
902
            )
903
        else:
Baber Abbasi's avatar
Baber Abbasi committed
904
            prefix = (
Baber Abbasi's avatar
Baber Abbasi committed
905
906
                self.config.target_delimiter + gen_prefix
                if gen_prefix is not None
Baber Abbasi's avatar
Baber Abbasi committed
907
908
                else ""
            )
KonradSzafer's avatar
KonradSzafer committed
909
910
            if self.multiple_input:
                return labeled_examples
911
            if isinstance(example, str):
Baber Abbasi's avatar
Baber Abbasi committed
912
                return labeled_examples + example + prefix
913
            elif isinstance(example, list):
Baber Abbasi's avatar
Baber Abbasi committed
914
                return [labeled_examples + ex + prefix for ex in example]
915
916
917
            elif isinstance(example, int):
                if self.config.doc_to_choice is not None:
                    choices = self.doc_to_choice(doc)
Baber Abbasi's avatar
Baber Abbasi committed
918
                    return labeled_examples + choices[example] + prefix
919
                else:
Baber Abbasi's avatar
Baber Abbasi committed
920
                    return labeled_examples + str(example) + prefix
lintangsutawika's avatar
lintangsutawika committed
921

Baber's avatar
cleanup  
Baber committed
922
    def apply_filters(self) -> list[Instance] | None:
Baber Abbasi's avatar
Baber Abbasi committed
923
        """Iterates over FilterEnsembles and applies them to instances"""
924
        if hasattr(self, "_filters") and self._instances:
925
            for f in self._filters:
926
                f.ensemble.apply(self._instances)
927
        else:
928
929
930
            eval_logger.warning(
                "No filter defined or instances found. Passing through instances"
            )
931
932
            return self._instances

933
    def should_decontaminate(self):
934
        return self.config.should_decontaminate
935

Baber Abbasi's avatar
Baber Abbasi committed
936
    def doc_to_decontamination_query(self, doc: dict):
937
        if self.config.should_decontaminate:
938
939
            if self.config.doc_to_decontamination_query is None:
                return self.doc_to_text(doc)
940
            else:
941
942
943
944
945
946
947
948
949
950
951
                doc_to_decontamination_query = self.config.doc_to_decontamination_query
                if doc_to_decontamination_query in self.features:
                    return doc[doc_to_decontamination_query]
                elif callable(doc_to_decontamination_query):
                    return doc_to_decontamination_query(doc)
                else:
                    return ast.literal_eval(
                        utils.apply_template(
                            self.config.doc_to_decontamination_query, doc
                        )
                    )
952

953
    def _process_doc(self, doc: dict) -> dict:
954
955
956
957
958
959
960
961
962
963
        """
        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

Baber's avatar
Baber committed
964
965
966
967
968
969
970
971
972
973
974
975
    @overload
    def doc_to_text(self, doc: dict, doc_to_text: None = None) -> str | int: ...

    @overload
    def doc_to_text(self, doc: dict, doc_to_text: int) -> int: ...

    @overload
    def doc_to_text(self, doc: dict, doc_to_text: str) -> str: ...

    @overload
    def doc_to_text(self, doc: dict, doc_to_text: Callable[..., str]) -> str: ...

976
977
    def doc_to_text(
        self, doc: dict, doc_to_text: int | str | Callable[..., str] | None = None
Baber's avatar
Baber committed
978
    ) -> str | int:
Baber's avatar
Baber committed
979
980
        # if self.prompt is not None:
        #     doc_to_text = self.prompt
981
        doc_to_text = doc_to_text or self.config.doc_to_text
Baber's avatar
Baber committed
982
983
        if callable(doc_to_text):
            return doc_to_text(doc)
984
985
        if doc_to_text in doc:
            return doc[doc_to_text]
986
        elif isinstance(doc_to_text, str):
987
988
989
            text_string = utils.apply_template(doc_to_text, doc)
            if text_string.isdigit() and self.config.doc_to_choice is not None:
                return ast.literal_eval(text_string)
990
            else:
991
992
993
                return text_string
        elif isinstance(doc_to_text, int):
            return doc_to_text
lintangsutawika's avatar
lintangsutawika committed
994
        # Used when applying a Promptsource template
Baber's avatar
Baber committed
995
996
997
998
999
1000
1001
        # elif hasattr(doc_to_text, "apply"):
        #     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")
        #         return self.config.fewshot_delimiter
1002
        else:
1003
            print(type(doc_to_text))
1004
            raise TypeError
1005

Baber's avatar
Baber committed
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
    @overload
    def doc_to_target(
        self, doc: dict, doc_to_target: None = None
    ) -> int | str | list[int]: ...

    @overload
    def doc_to_target(self, doc: dict, doc_to_target: int) -> int: ...

    @overload
    def doc_to_target(self, doc: dict, doc_to_target: str) -> int | str | list[int]: ...

    @overload
    def doc_to_target(self, doc: dict, doc_to_target: list) -> list[int]: ...

    @overload
    def doc_to_target(
        self, doc: dict, doc_to_target: Callable[..., int | str | list[int]]
    ) -> int | str | list[int]: ...

Baber's avatar
cleanup  
Baber committed
1025
    def doc_to_target(self, doc: dict, doc_to_target=None) -> int | str | list[int]:
Baber's avatar
Baber committed
1026
1027
        # if self.prompt is not None:
        #     doc_to_target = self.prompt
1028
        doc_to_target = doc_to_target or self.config.doc_to_target
Baber's avatar
Baber committed
1029
1030
        if callable(doc_to_target):
            doc_to_target(doc)
1031
1032
        if doc_to_target in doc:
            return doc[doc_to_target]
1033
        elif isinstance(doc_to_target, str):
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
            target_string = utils.apply_template(doc_to_target, doc)
            if target_string.isdigit() and self.config.doc_to_choice is not None:
                return ast.literal_eval(target_string)
            # elif (
            #     len(target_string) >= 2
            #     and (target_string[0] == "[")
            #     and (target_string[-1] == "]")
            # ):
            #     try:
            #         return ast.literal_eval(target_string)
            #     except (SyntaxError, ValueError):
            #         return target_string
1046
            else:
1047
1048
1049
                return target_string

        elif isinstance(doc_to_target, (int, list)):
1050
            return doc_to_target
1051
1052
1053
1054
        # elif isinstance(doc_to_target, list):
        #     return doc_to_target
        # elif callable(doc_to_target):
        #     return doc_to_target(doc)
Baber's avatar
Baber committed
1055
1056
1057
1058
1059
1060
1061
1062
        # # Used when applying a Promptsource template
        # elif hasattr(doc_to_target, "apply"):
        #     applied_prompt = doc_to_target.apply(doc)
        #     if len(applied_prompt) == 2:
        #         return applied_prompt[1]
        #     else:
        #         eval_logger.warning("Applied prompt returns empty string")
        #         return self.config.fewshot_delimiter
1063
1064
        else:
            raise TypeError
1065

Baber's avatar
Baber committed
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
    @overload
    def doc_to_choice(self, doc: dict, doc_to_choice: None = None) -> list[str]: ...

    @overload
    def doc_to_choice(self, doc: dict, doc_to_choice: str) -> list[str]: ...

    @overload
    def doc_to_choice(self, doc: dict, doc_to_choice: list) -> list[str]: ...

    @overload
    def doc_to_choice(self, doc: dict, doc_to_choice: dict) -> list[str]: ...

    @overload
    def doc_to_choice(
        self, doc: dict, doc_to_choice: Callable[..., list[str]]
    ) -> list[str]: ...

Baber's avatar
cleanup  
Baber committed
1083
    def doc_to_choice(
Baber's avatar
Baber committed
1084
1085
        self,
        doc: dict,
Baber's avatar
cleanup  
Baber committed
1086
1087
        doc_to_choice: str | list | dict | Callable[..., list[str]] | None = None,
    ) -> list[str]:
Baber's avatar
Baber committed
1088
1089
1090
        # if self.prompt is not None:
        #     doc_to_choice = self.prompt
        if doc_to_choice is not None:
Yu Shi Jie's avatar
Yu Shi Jie committed
1091
            doc_to_choice = doc_to_choice
1092
        elif self.config.doc_to_choice is None:
1093
            eval_logger.error("doc_to_choice was called but not set in config")
Baber's avatar
Baber committed
1094
            doc_to_choice = None
1095
        else:
1096
            doc_to_choice = self.config.doc_to_choice
1097

1098
        if isinstance(doc_to_choice, str):
1099
            if doc_to_choice in doc:
1100
1101
1102
                return doc[doc_to_choice]
            else:
                return ast.literal_eval(utils.apply_template(doc_to_choice, doc))
1103
        elif isinstance(doc_to_choice, list):
1104
            return doc_to_choice
1105
1106
        # elif isinstance(doc_to_choice, dict):
        #     return list(doc_to_choice.values())
Baber's avatar
Baber committed
1107
1108
        # elif hasattr(doc_to_choice, "get_answer_choices_list"):
        #     return doc_to_choice.get_answer_choices_list(doc)
1109
1110
        else:
            raise TypeError
1111

Baber's avatar
Baber committed
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
    @overload
    def doc_to_image(self, doc: dict, doc_to_image: None = None) -> None: ...

    @overload
    def doc_to_image(self, doc: dict, doc_to_image: list) -> list: ...

    @overload
    def doc_to_image(self, doc: dict, doc_to_image: str) -> int | str | None: ...

    @overload
    def doc_to_image(self, doc: dict, doc_to_image: Callable[..., Any]) -> Any: ...

Baber's avatar
cleanup  
Baber committed
1124
    def doc_to_image(self, doc: dict, doc_to_image=None) -> int | str | list | None:
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
        if doc_to_image is not None:
            doc_to_image = doc_to_image
        elif self.config.doc_to_image is not None:
            doc_to_image = self.config.doc_to_image
        else:
            return None

        if isinstance(doc_to_image, list):
            image_feature = [
                self.doc_to_image(doc, feature) for feature in doc_to_image
            ]
            return [feature for feature in image_feature if feature is not None]
        elif isinstance(doc_to_image, str):
            if doc_to_image in self.features:
                return doc[doc_to_image]
            else:
                return ast.literal_eval(utils.apply_template(doc_to_image, doc))
        elif callable(doc_to_image):
            return doc_to_image(doc)
        else:
            return None

Baber's avatar
Baber committed
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
    @overload
    def doc_to_audio(self, doc: Any, doc_to_audio: None = None) -> None: ...

    @overload
    def doc_to_audio(self, doc: Any, doc_to_audio: list) -> list: ...

    @overload
    def doc_to_audio(self, doc: Any, doc_to_audio: str) -> int | str | None: ...

    @overload
    def doc_to_audio(self, doc: Any, doc_to_audio: Callable[..., Any]) -> Any: ...

Baber's avatar
cleanup  
Baber committed
1159
    def doc_to_audio(self, doc: Any, doc_to_audio=None) -> int | str | list | None:
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
        if doc_to_audio is not None:
            doc_to_audio = doc_to_audio
        elif self.config.doc_to_audio is not None:
            doc_to_audio = self.config.doc_to_audio
        else:
            return None

        if isinstance(doc_to_audio, list):
            audio_feature = [
                self.doc_to_audio(doc, feature) for feature in doc_to_audio
            ]
            return [feature for feature in audio_feature if feature is not None]
        elif isinstance(doc_to_audio, str):
            if doc_to_audio in self.features:
                return doc[doc_to_audio]
            else:
                return ast.literal_eval(utils.apply_template(doc_to_audio, doc))
        elif callable(doc_to_audio):
            return doc_to_audio(doc)
        else:
            return None

Baber's avatar
cleanup  
Baber committed
1182
    def doc_to_prefix(self, doc: dict) -> str | None:
Baber Abbasi's avatar
Baber Abbasi committed
1183
        if (gen_prefix := self.config.gen_prefix) is not None:
1184
            if gen_prefix in doc:
Baber Abbasi's avatar
Baber Abbasi committed
1185
1186
1187
1188
1189
                return doc[gen_prefix]
            else:
                return utils.apply_template(gen_prefix, doc)
        return None

baberabb's avatar
baberabb committed
1190
1191
    def construct_requests(
        self, doc: dict, ctx: str, **kwargs
Baber's avatar
cleanup  
Baber committed
1192
    ) -> list[Instance] | Instance:
1193
        apply_chat_template = kwargs.pop("apply_chat_template", False)
1194
        chat_template: Callable | None = kwargs.pop("chat_template", None)
1195

1196
1197
        aux_arguments = None

1198
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
1199
            arguments = (ctx, self.doc_to_target(doc))
1200
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
1201
            arguments = (self.doc_to_target(doc),)
1202
        elif self.OUTPUT_TYPE == "multiple_choice":
1203
            choices = self.doc_to_choice(doc)
1204
            target_delimiter = self.config.target_delimiter
1205
1206
            if apply_chat_template:
                target_delimiter = ""
1207
1208
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
1209
                # apply chat_template to choices if apply_chat_template
1210
                cont = self.doc_to_target(doc)
1211

1212
                arguments = [
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
                    (
                        ctx
                        + (
                            chat_template([{"role": "user", "content": choice}])
                            if apply_chat_template
                            else choice
                        ),
                        f"{target_delimiter}{cont}",
                    )
                    for choice in choices
1223
                ]
1224
            else:
1225
                # Otherwise they are placed in the continuation
1226
                arguments = [(ctx, f"{target_delimiter}{cont}") for cont in choices]
1227

1228
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
1229
            if "acc_mutual_info" in [m.metric_name for m in self.config._metric_list]:
1230
1231
1232
1233
1234
1235
                # if we are calculating multiple choice accuracy
                # using mutual information instead of raw loglikelihood as metric, need unconditional lls.

                # here mutual info refers to calculating
                # 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.
1236
1237
1238
1239
                # TODO: should these be strided? will have to modify the processing in process_results if so
                aux_arguments = [
                    ("", f"{target_delimiter}{choice}") for choice in choices
                ]
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254

                arguments.extend(aux_arguments)

        elif self.OUTPUT_TYPE == "generate_until":
            arguments = (ctx, deepcopy(self.config.generation_kwargs))

        multimodal_arg = {}
        if (
            self.config.doc_to_image
        ):  # TODO: ensure that non-multimodal tasks aren't getting visual args
            multimodal_arg = {
                **multimodal_arg,
                **{"visual": self.doc_to_image(doc)},
            }

1255
1256
1257
1258
1259
1260
1261
1262
        if (
            self.config.doc_to_audio
        ):  # TODO: ensure that non-multimodal tasks aren't getting audio args
            multimodal_arg = {
                **multimodal_arg,
                **{"audio": self.doc_to_audio(doc)},
            }

1263
1264
1265
1266
1267
1268
1269
        if bool(multimodal_arg):
            if isinstance(arguments, list):
                arguments = [arg + (multimodal_arg,) for arg in arguments]
            else:
                arguments = arguments + (multimodal_arg,)

        if self.OUTPUT_TYPE == "multiple_choice":
1270
            request_list = [
1271
1272
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1273
                    doc=doc,
1274
                    arguments=arg,
1275
                    idx=i,
1276
1277
                    **kwargs,
                )
1278
                for i, arg in enumerate(arguments)
1279
            ]
1280
1281

            return request_list
lintangsutawika's avatar
lintangsutawika committed
1282

lintangsutawika's avatar
lintangsutawika committed
1283
        return Instance(
1284
1285
1286
1287
1288
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=arguments,
            idx=0,
            **kwargs,
lintangsutawika's avatar
lintangsutawika committed
1289
        )
1290

1291
    def process_results(self, doc: dict, results: list) -> dict[str, Any]:
Baber's avatar
Baber committed
1292
1293
        if callable(self.config.process_results):
            return self.config.process_results(doc, results)
1294
        result_dict = {}
1295
        use_metric = list(m.metric_name for m in self.config._metric_list)
1296
1297
1298
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
1299
1300
1301
1302
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
1303
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
1304
1305
1306
1307
1308
1309
            (loglikelihood, *_) = results
            assert isinstance(_target := self.doc_to_target(doc), str), (
                "Require target to be a string for loglikelihood_rolling"
            )
            _words = self.count_words(_target)
            _bytes = self.count_bytes(_target)
haileyschoelkopf's avatar
haileyschoelkopf committed
1310
            return {
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
                **(
                    {"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
1326
            }
1327
        elif self.OUTPUT_TYPE == "multiple_choice":
1328
            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
1329

Baber's avatar
cleanup  
Baber committed
1330
            # retrieve choices in list[str] form, to compute choice lengths, etc.
1331
            choices = self.doc_to_choice(doc)
1332
1333
            completion_len = np.array([float(len(i)) for i in choices])

Baber's avatar
Baber committed
1334
            if 2 * len(choices) == len(lls) and "acc_mutual_info" in use_metric:
1335
1336
                # then we are doing mutual info.
                # this stores the "dryrun" / unconditional answer loglikelihoods
1337
1338
                # as we extend the args list with unconditional ("", continuation) pairs
                lls_unconditional = lls[len(choices) :]
1339
1340
                if len(lls_unconditional) != len(choices):
                    raise ValueError
1341
                # and this stores our "regular" conditional loglikelihoods
1342
                lls = lls[: len(choices)]
Baber's avatar
Baber committed
1343
1344
            else:
                lls_unconditional = None
1345

1346
1347
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
1348

1349
1350
            if self.multiple_input:
                gold = self.doc_to_text(doc)
1351
            else:
1352
                gold = self.doc_to_target(doc)
1353

1354
            gold, gold_index_error = check_gold_index_error(choices, gold)
1355
1356
1357

            if gold_index_error:
                eval_logger.warning(
lintangsutawika's avatar
lintangsutawika committed
1358
                    f"Label index was not in within range of available choices,"
1359
1360
                    f"Sample:\n\n{doc}\n\n"
                )
lintangsutawika's avatar
lintangsutawika committed
1361

1362
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
1363
1364
                acc = 1.0 if pred in gold else 0.0
                acc_norm = 1.0 if pred_norm in gold else 0.0
Baber's avatar
cleanup  
Baber committed
1365
                exact_match = int(any(is_greedy[i] if i != -100 else 0 for i in gold))
lintangsutawika's avatar
lintangsutawika committed
1366
1367
1368
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
1369
                # TODO: this gets score of 0 on arc_challenge for pythia-70m. need to test that this works properly
Lintang Sutawika's avatar
Lintang Sutawika committed
1370
                exact_match = int(is_greedy[gold]) if gold != -100 else 0
1371

Lintang Sutawika's avatar
Lintang Sutawika committed
1372
1373
1374
1375
            prob_norm = utils.softmax(lls)

            # TODO use keyword arguments to the metric?
            # gold, pred, norm stuff, the original lls,
1376
            result_dict = {
1377
                **({"acc": acc} if "acc" in use_metric else {}),
1378
1379
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
1380
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
1381
                **({"exact_match": exact_match} if "exact_match" in use_metric else {}),
Lintang Sutawika's avatar
Lintang Sutawika committed
1382
1383
1384
1385
1386
                **(
                    {"brier_score": (gold, prob_norm)}
                    if "brier_score" in use_metric
                    else {}
                ),
1387
1388
            }

1389
            if "acc_mutual_info" in use_metric:
Baber's avatar
Baber committed
1390
1391
1392
                assert lls_unconditional is not None, (
                    "lls_unconditional should not be None if acc_mutual_info is in use_metric"
                )
lintangsutawika's avatar
lintangsutawika committed
1393
1394
1395
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
1396
1397
1398
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

1399
        elif self.OUTPUT_TYPE == "generate_until":
1400
            gold = self.doc_to_target(doc)
Chris's avatar
Chris committed
1401
            result = results[0]
Baber's avatar
Baber committed
1402
            for metric in self.config._metric_list:
1403
                try:
Baber's avatar
Baber committed
1404
                    result_score = metric.fn(
1405
1406
                        references=[gold] if not isinstance(gold, list) else gold,
                        predictions=[result],
Baber's avatar
Baber committed
1407
                        **metric.kwargs,
1408
1409
                    )
                except TypeError:  # needed for now in order to use a different interface between our own metrics and HF Evaluate metrics
Baber's avatar
Baber committed
1410
                    result_score = metric.fn([gold, result])
1411
1412
1413
1414
1415
1416
                if isinstance(result_score, dict):
                    # TODO: this handles the case where HF evaluate returns a dict.
                    # This allows for multiple metrics to be returned from the same function
                    for k, v in result_score.items():
                        result_dict[k] = v
                else:
Baber's avatar
Baber committed
1417
                    result_dict[metric.name] = result_score
1418
        else:
lintangsutawika's avatar
lintangsutawika committed
1419
1420
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
1421
                "'loglikelihood', 'loglikelihood_rolling', 'generate_until' or 'multiple_choice'",
1422
            )
1423
1424
1425

        return result_dict

Baber Abbasi's avatar
Baber Abbasi committed
1426
    def aggregation(self) -> dict:
1427
        return {k.name: k.aggregation_fn for k in self.config._metric_list}
1428

Baber Abbasi's avatar
Baber Abbasi committed
1429
    def higher_is_better(self) -> dict:
1430
        return {k.name: k.higher_is_better for k in self.config._metric_list}
1431

Baber Abbasi's avatar
Baber Abbasi committed
1432
1433
1434
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

Lintang Sutawika's avatar
Lintang Sutawika committed
1435
    @property
Baber's avatar
cleanup  
Baber committed
1436
    def task_name(self) -> str | None:
Lintang Sutawika's avatar
Lintang Sutawika committed
1437
1438
        return getattr(self.config, "task", None)

Baber's avatar
Baber committed
1439
1440
1441
1442
    def runtime_checks(self, test_doc):
        # Test One Doc
        self.features: list[str] = list(self.task_docs.features.keys())
        self.multiple_target = 0
Baber's avatar
Baber committed
1443
        self.multiple_input = 0
Baber's avatar
Baber committed
1444
1445
1446
1447
1448
1449
1450
        test_text = self.doc_to_text(test_doc)
        test_target = self.doc_to_target(test_doc)

        if self.config.doc_to_choice is not None:
            test_choice = self.doc_to_choice(test_doc)
            if not isinstance(test_choice, list):
                eval_logger.error("doc_to_choice must return list")
Baber's avatar
Baber committed
1451
1452
1453
1454
1455
1456
1457
            else:
                num_choice = len(test_choice)

            if isinstance(test_text, int):
                eval_logger.debug(
                    "doc_to_text returned an int. Assuming multiple inputs."
                )
Baber's avatar
Baber committed
1458
1459
1460
1461
1462

            if isinstance(test_text, int):
                eval_logger.debug(
                    "doc_to_text returned an int. Assuming multiple inputs."
                )
Baber's avatar
Baber committed
1463
                self.multiple_input = num_choice
Baber's avatar
Baber committed
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
        else:
            test_choice = None

        if isinstance(test_target, list):
            eval_logger.debug(
                "doc_to_target returned a list. Assuming multiple targets."
            )
            self.multiple_target = len(test_target)
        else:
            if (isinstance(test_target, int)) and (test_choice is not None):
                test_target = test_choice[test_target]
            else:
                test_target = str(test_target)

        check_choices = test_choice if test_choice is not None else [test_target]
        if self.config.doc_to_choice is not None:
            for choice in check_choices:
                choice_has_whitespace = choice[0].isspace()
                delimiter_has_whitespace = (
                    self.config.target_delimiter.rstrip()
                    != self.config.target_delimiter
                )

                if delimiter_has_whitespace and choice_has_whitespace:
                    eval_logger.debug(
                        f'Both target_delimiter "{self.config.target_delimiter}" and target choice: "{choice}" have whitespace'
                    )
                elif (not delimiter_has_whitespace) and (not choice_has_whitespace):
                    eval_logger.debug(
                        f'Both target_delimiter "{self.config.target_delimiter}" and target choice: "{choice}" do not have whitespace, ignore if the language you are evaluating on does not require/use whitespace'
                    )

1496
1497
1498
1499
1500
    def __repr__(self):
        return (
            f"ConfigurableTask(task_name={getattr(self.config, 'task', None)},"
            f"output_type={self.OUTPUT_TYPE},"
            f"num_fewshot={getattr(self.config, 'num_fewshot', None)},"
Baber Abbasi's avatar
Baber Abbasi committed
1501
            f"num_samples={len(self.eval_docs)})"
1502
1503
        )

1504
1505

class MultipleChoiceTask(Task):
1506
    OUTPUT_TYPE = "loglikelihood"
1507

baberabb's avatar
baberabb committed
1508
    def doc_to_target(self, doc: dict) -> str:
1509
1510
        return " " + doc["choices"][doc["gold"]]

Baber's avatar
cleanup  
Baber committed
1511
    def construct_requests(self, doc: dict, ctx: str, **kwargs) -> list[Instance]:
1512
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1513
1514
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1515
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1516
                doc=doc,
Baber's avatar
Baber committed
1517
                arguments=(ctx, f" {choice}"),
1518
                idx=i,
1519
1520
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1521
1522
            for i, choice in enumerate(doc["choices"])
        ]
1523

Baber's avatar
cleanup  
Baber committed
1524
    def process_results(self, doc: dict, results: Iterable[tuple[float, bool]]) -> dict:
lintangsutawika's avatar
lintangsutawika committed
1525
1526
1527
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
        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
1539
    def higher_is_better(self) -> dict:
1540
1541
1542
1543
1544
        return {
            "acc": True,
            "acc_norm": True,
        }

baberabb's avatar
baberabb committed
1545
    def aggregation(self) -> dict:
1546
1547
1548
1549
1550
1551
        return {
            "acc": mean,
            "acc_norm": mean,
        }


lintangsutawika's avatar
lintangsutawika committed
1552
class PerplexityTask(Task):
1553
1554
    OUTPUT_TYPE = "loglikelihood_rolling"

baberabb's avatar
baberabb committed
1555
    def has_training_docs(self) -> bool:
1556
1557
        return False

Baber's avatar
cleanup  
Baber committed
1558
    def fewshot_examples(self, k: int, rnd) -> list:
1559
1560
1561
1562
        if k != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1563
1564
        return []

baberabb's avatar
baberabb committed
1565
    def fewshot_context(self, doc: dict, num_fewshot: int) -> Literal[""]:
1566
1567
1568
1569
        if num_fewshot != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1570
1571
1572

        return ""

baberabb's avatar
baberabb committed
1573
    def higher_is_better(self) -> dict:
1574
1575
1576
1577
1578
1579
1580
1581
1582
        return {
            "word_perplexity": False,
            "byte_perplexity": False,
            "bits_per_byte": False,
        }

    def doc_to_decontamination_query(self, doc):
        return doc

Ethan Smith's avatar
Ethan Smith committed
1583
    def doc_to_text(self, doc) -> str:
1584
1585
1586
1587
1588
        return ""

    def doc_to_target(self, doc):
        return doc

Baber's avatar
cleanup  
Baber committed
1589
    def construct_requests(self, doc: dict, ctx: str | None, **kwargs):
1590
1591
        if bool(ctx):
            raise ValueError
1592

lintangsutawika's avatar
lintangsutawika committed
1593
1594
1595
1596
1597
1598
1599
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1600

Baber's avatar
cleanup  
Baber committed
1601
    def process_results(self, doc: dict, results: tuple[float]) -> dict:
1602
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1603
1604
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1605
1606
1607
1608
1609
1610
        return {
            "word_perplexity": (loglikelihood, words),
            "byte_perplexity": (loglikelihood, bytes_),
            "bits_per_byte": (loglikelihood, bytes_),
        }

baberabb's avatar
baberabb committed
1611
    def aggregation(self) -> dict:
1612
1613
1614
1615
1616
1617
1618
        return {
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
            "bits_per_byte": bits_per_byte,
        }

    @classmethod
baberabb's avatar
baberabb committed
1619
    def count_bytes(cls, doc) -> int:
1620
1621
1622
        return len(doc.encode("utf-8"))

    @classmethod
baberabb's avatar
baberabb committed
1623
    def count_words(cls, doc) -> int:
1624
        """Downstream tasks with custom word boundaries should override this!"""
Lintang Sutawika's avatar
Lintang Sutawika committed
1625
        return len(re.split(r"\s+", doc))