task.py 67.1 KB
Newer Older
1
import abc
2
import ast
lintangsutawika's avatar
lintangsutawika committed
3
import logging
4
import random
5
6
import re
from collections.abc import Callable
7
from copy import deepcopy
8
from dataclasses import asdict, dataclass, field
9
from inspect import getsource
10
11
12
13
14
15
16
17
18
19
20
21
from typing import (
    Any,
    Dict,
    Iterable,
    Iterator,
    List,
    Literal,
    Mapping,
    Optional,
    Tuple,
    Union,
)
22
23
24

import datasets
import numpy as np
lintangsutawika's avatar
lintangsutawika committed
25
import shortuuid
26
from tqdm import tqdm
27
28

from lm_eval import utils
29
from lm_eval.api import samplers
30
31
from lm_eval.api.instance import Instance, OutputType
from lm_eval.api.metrics import bits_per_byte, mean, weighted_perplexity
lintangsutawika's avatar
lintangsutawika committed
32
from lm_eval.api.registry import (
33
34
    AGGREGATION_REGISTRY,
    DEFAULT_METRIC_REGISTRY,
haileyschoelkopf's avatar
haileyschoelkopf committed
35
    get_aggregation,
36
    get_metric,
37
    get_metric_aggregation,
haileyschoelkopf's avatar
haileyschoelkopf committed
38
    is_higher_better,
lintangsutawika's avatar
lintangsutawika committed
39
)
40
from lm_eval.caching.cache import load_from_cache, save_to_cache
41
42
43
from lm_eval.filters import build_filter_ensemble
from lm_eval.prompts import get_prompt

44

45
46
47
48
ALL_OUTPUT_TYPES = [
    "loglikelihood",
    "multiple_choice",
    "loglikelihood_rolling",
49
    "generate_until",
50
51
]

52
eval_logger = logging.getLogger("lm-eval")
53

54
55
56
57
58
59
60
61
62
63
64
@dataclass
class AggMetricConfig(dict):
    metric: Optional[str] = "acc"
    metric_alias: Optional[str] = "acc"
    aggregation: Optional[str] = "mean"
    weight_by_size: Optional[str] = False
    filter_list: Optional[Union[str,list]] = "none"

    def __post_init__(self):
        if isinstance(self.filter_list, str):
            self.filter_list = [self.filter_list]
lintangsutawika's avatar
lintangsutawika committed
65

lintangsutawika's avatar
lintangsutawika committed
66
67
@dataclass
class GroupConfig(dict):
lintangsutawika's avatar
lintangsutawika committed
68
69
70
    group: Optional[str] = None
    group_alias: Optional[str] = None
    task: Optional[Union[str, list]] = None
71
    tag_to_task: Optional[str] = False
72
    aggregate_metric: Optional[
73
74
        Union[List[AggMetricConfig], AggMetricConfig, dict]
    ] = None
lintangsutawika's avatar
lintangsutawika committed
75
76
77
    metadata: Optional[
        dict
    ] = None  # by default, not used in the code. allows for users to pass arbitrary info to tasks
lintangsutawika's avatar
lintangsutawika committed
78
79
80
81
82
83
84

    def __getitem__(self, item):
        return getattr(self, item)

    def __setitem__(self, item, value):
        return setattr(self, item, value)

85
86
87
88
89
90
91
92
93
94
    def __post_init__(self):
        if self.aggregate_metric_list is not None:
            if isinstance(self.aggregate_metric_list, dict):
                self.aggregate_metric_list = [self.aggregate_metric_list]

            self.aggregate_metric_list = [
                AggMetricConfig(**item) if isinstance(item, dict) else item
                for item in self.aggregate_metric_list
            ]

lintangsutawika's avatar
lintangsutawika committed
95
96
97
98
99
100
101
102
103
104
105
106
107
    def to_dict(self, keep_callable: bool = False) -> dict:
        """dumps the current config as a dictionary object, as a printable format.
        null fields will not be printed.
        Used for dumping results alongside full task configuration

        :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()):
lintangsutawika's avatar
lintangsutawika committed
108
            if callable(v):
lintangsutawika's avatar
lintangsutawika committed
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
                cfg_dict[k] = self.serialize_function(v, keep_callable=keep_callable)
        return cfg_dict

    def serialize_function(
        self, value: Union[Callable, str], keep_callable=False
    ) -> Union[Callable, str]:
        """Serializes a given function or string.

        If 'keep_callable' is True, the original callable is returned.
        Otherwise, attempts to return the source code of the callable using 'getsource'.
        """
        if keep_callable:
            return value
        else:
            try:
                return getsource(value)
            except (TypeError, OSError):
                return str(value)


lintangsutawika's avatar
lintangsutawika committed
129
130
131
132
133
class ConfigurableGroup(abc.ABC):
    def __init__(
        self,
        config: Optional[dict] = None,
    ) -> None:
lintangsutawika's avatar
lintangsutawika committed
134
        # Create a unique identifier ID
lintangsutawika's avatar
lintangsutawika committed
135
        self._task_id = shortuuid.uuid()[:8]
lintangsutawika's avatar
lintangsutawika committed
136
137
138
139
140
        self._config = GroupConfig(**config)

    @property
    def group(self):
        return self._config.group
141

lintangsutawika's avatar
lintangsutawika committed
142
143
144
    @property
    def group_alias(self):
        return self._config.group_alias
145
146
147
148
149

    @property
    def version(self):
        return self._config.version

lintangsutawika's avatar
lintangsutawika committed
150
151
152
153
    @property
    def config(self):
        return self._config.to_dict()

lintangsutawika's avatar
lintangsutawika committed
154
155
    @property
    def task_id(self) -> Any:
lintangsutawika's avatar
lintangsutawika committed
156
157
158
159
160
        return "-".join((self.group_name, self._task_id))

    @property
    def group_name(self) -> Any:
        return self._config.group
lintangsutawika's avatar
lintangsutawika committed
161

lintangsutawika's avatar
lintangsutawika committed
162
163
    def __repr__(self):
        return (
164
            f"ConfigurableGroup(group={self.group}," f"group_alias={self.group_alias})"
lintangsutawika's avatar
lintangsutawika committed
165
166
        )

167

168
169
@dataclass
class TaskConfig(dict):
170
    # task naming/registry
171
172
    task: Optional[str] = None
    task_alias: Optional[str] = None
lintangsutawika's avatar
lintangsutawika committed
173
    tag: Optional[Union[str, list]] = None
174
    group: Optional[Union[str, list]] = None
175
176
177
    # HF dataset options.
    # which dataset to use,
    # and what splits for what purpose
178
179
180
181
182
183
184
185
186
    dataset_path: Optional[str] = None
    dataset_name: Optional[str] = None
    dataset_kwargs: Optional[dict] = None
    training_split: Optional[str] = None
    validation_split: Optional[str] = None
    test_split: Optional[str] = None
    fewshot_split: Optional[
        str
    ] = None  # TODO: assert that this not None if num_fewshot > 0. (?) assert if this is same split as one evaling (?)
187
188
    # formatting / prompting options.
    # see docs/advanced_task_guide.md for more info
189
190
191
192
193
194
    process_docs: Optional[Callable] = None
    doc_to_text: Optional[Union[Callable, str]] = None
    doc_to_target: Optional[Union[Callable, str]] = None
    doc_to_choice: Optional[Union[Callable, str, dict, list]] = None
    process_results: Optional[Union[Callable, str]] = None
    use_prompt: Optional[str] = None
195
    description: str = ""
196
197
    target_delimiter: str = " "
    fewshot_delimiter: str = "\n\n"
198
    fewshot_config: Optional[dict] = None
199
    # runtime configuration options
200
    num_fewshot: Optional[int] = None
201
    # scoring options
202
203
204
    metric_list: Optional[list] = None
    output_type: OutputType = "generate_until"
    generation_kwargs: Optional[dict] = None
205
    repeats: int = 1
206
    filter_list: Optional[Union[str, list]] = None
207
    should_decontaminate: bool = False
208
209
210
211
    doc_to_decontamination_query: Optional[str] = None
    metadata: Optional[
        dict
    ] = None  # by default, not used in the code. allows for users to pass arbitrary info to tasks
212

Ethan Smith's avatar
Ethan Smith committed
213
    def __post_init__(self) -> None:
Lintang Sutawika's avatar
Lintang Sutawika committed
214
        if self.generation_kwargs is not None:
215
            if self.output_type != "generate_until":
216
                eval_logger.warning(
217
                    f"[{self.task}] passed `generation_kwargs`, but not using `output_type: generate_until`!"
Lintang Sutawika's avatar
Lintang Sutawika committed
218
219
220
221
222
223
224
225
                )

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

            if "until" not in self.generation_kwargs:
226
                self.generation_kwargs["until"] = [self.fewshot_delimiter]
Lintang Sutawika's avatar
Lintang Sutawika committed
227
        else:
228
            if self.output_type == "generate_until":
Lintang Sutawika's avatar
Lintang Sutawika committed
229
230
                # ensure that we greedily generate in absence of explicit arguments otherwise
                self.generation_kwargs = {
231
232
233
234
235
                    "until": (
                        None
                        if self.fewshot_delimiter is None
                        else [self.fewshot_delimiter]
                    ),
Lintang Sutawika's avatar
Lintang Sutawika committed
236
237
                    "do_sample": False,
                }
238

239
240
241
    def __getitem__(self, item):
        return getattr(self, item)

242
243
244
    def __setitem__(self, item, value):
        return setattr(self, item, value)

245
    def to_dict(self, keep_callable: bool = False) -> dict:
246
247
        """dumps the current config as a dictionary object, as a printable format.
        null fields will not be printed.
haileyschoelkopf's avatar
haileyschoelkopf committed
248
        Used for dumping results alongside full task configuration
249

haileyschoelkopf's avatar
haileyschoelkopf committed
250
251
252
253
254
255
256
257
258
259
        :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)
260
261
262
263
264
265
266
267
268
269
            elif k == "metric_list":
                for metric_dict in v:
                    for metric_key, metric_value in metric_dict.items():
                        if callable(metric_value):
                            metric_dict[metric_key] = self.serialize_function(
                                metric_value, keep_callable=keep_callable
                            )
                cfg_dict[k] = v
            elif callable(v):
                cfg_dict[k] = self.serialize_function(v, keep_callable=keep_callable)
haileyschoelkopf's avatar
haileyschoelkopf committed
270
        return cfg_dict
271

272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
    def serialize_function(
        self, value: Union[Callable, str], keep_callable=False
    ) -> Union[Callable, str]:
        """Serializes a given function or string.

        If 'keep_callable' is True, the original callable is returned.
        Otherwise, attempts to return the source code of the callable using 'getsource'.
        """
        if keep_callable:
            return value
        else:
            try:
                return getsource(value)
            except (TypeError, OSError):
                return str(value)

288
289
290
291
292
293
294
295
296
297
298

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)
    """

299
    VERSION: Optional[Union[int, str]] = None
300

301
302
    # The name of the `Task` benchmark as denoted in the HuggingFace datasets Hub
    # or a path to a custom `datasets` loading script.
303
    DATASET_PATH: Optional[str] = None
304
305

    # The name of a subset within `DATASET_PATH`.
306
    DATASET_NAME: Optional[str] = None
307

308
    OUTPUT_TYPE: Optional[OutputType] = None
lintangsutawika's avatar
lintangsutawika committed
309

310
311
    def __init__(
        self,
312
313
314
315
        data_dir: Optional[str] = None,
        cache_dir: Optional[str] = None,
        download_mode: Optional[datasets.DownloadMode] = None,
        config: Optional[Mapping] = None,  # Union[dict, TaskConfig]
Ethan Smith's avatar
Ethan Smith committed
316
    ) -> None:
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
        """
        :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)
339
340
341
        self._training_docs: Optional[list] = None
        self._fewshot_docs: Optional[list] = None
        self._instances: Optional[List[Instance]] = None
342

343
        # Create a unique identifier ID
lintangsutawika's avatar
lintangsutawika committed
344
        self._task_id = shortuuid.uuid()[:8]
345
        self._config: TaskConfig = TaskConfig({**config}) if config else TaskConfig()
346

lintangsutawika's avatar
lintangsutawika committed
347
        self._filters = [build_filter_ensemble("none", [["take_first", None]])]
348
349
350
        self.fewshot_rnd: Optional[
            random.Random
        ] = None  # purposely induce errors in case of improper usage
351

352
353
354
355
356
357
    def download(
        self,
        data_dir: Optional[str] = None,
        cache_dir: Optional[str] = None,
        download_mode=None,
    ) -> None:
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
        """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.
        """
382
383
384
385
386
387
388
        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,
        )
389

390
    @property
391
    def config(self) -> TaskConfig:
392
393
394
        """Returns the TaskConfig associated with this class."""
        return self._config

395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
    @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

410
    def training_docs(self) -> Iterable:
411
412
413
414
415
416
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

417
    def validation_docs(self) -> Iterable:
418
419
420
421
422
423
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

424
    def test_docs(self) -> Iterable:
425
426
427
428
429
430
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

431
    def fewshot_docs(self) -> Iterable:
432
433
434
435
436
437
438
439
440
        """
        :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
441
            eval_logger.warning(
442
                f"[Task: {self.config.task}] has_training_docs and has_validation_docs are False"
443
                ", using test_docs as fewshot_docs but this is not recommended."
lintangsutawika's avatar
lintangsutawika committed
444
            )
445
446
            return self.test_docs()

447
    def _process_doc(self, doc: dict) -> dict:
448
449
450
451
452
453
454
455
456
        """
        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
457

458
    @property
459
    def instances(self) -> List[Instance]:
460
461
462
463
464
465
466
467
468
469
470
        """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)

471
472
    def doc_to_decontamination_query(self, doc):
        raise NotImplementedError(
473
474
475
476
477
478
479
480
481
482
483
            "Override doc_to_decontamination_query with document specific decontamination query."
        )

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

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

484
485
    def build_all_requests(
        self,
486
        *,
487
488
489
490
491
        limit=None,
        rank=None,
        world_size=None,
        cache_requests=False,
        rewrite_requests_cache=False,
KonradSzafer's avatar
KonradSzafer committed
492
493
494
495
        system_instruction=None,
        apply_chat_template=False,
        fewshot_as_multiturn=False,
        lm=None,
496
    ) -> None:
497
        """Build a set of Instances for a task, and store them in task.instances"""
498
499
500
501

        # used with caching
        og_limit = limit

502
        cache_key = f"requests-{self._config.task}-{self.config.num_fewshot}shot-rank{rank}-world_size{world_size}"
KonradSzafer's avatar
KonradSzafer committed
503
504
505
506
507
508
509
510
        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 ""
        )
        cache_key += f"-tokenizer{lm.tokenizer_name}" if apply_chat_template else ""
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525

        cached_instances = load_from_cache(file_name=cache_key)

        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
526
        eval_logger.info(f"Building contexts for {self.config.task} on rank {rank}...")
527

528
        instances = []
529
530
531
532
533
534
535
536
537
538

        # 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(
539
            self.doc_iterator(rank=rank, limit=limit, world_size=world_size)
540
541
542
543
544
545
546
        )

        num_docs = len(doc_id_docs)

        for doc_id, doc in tqdm(
            doc_id_docs,
            total=num_docs,
lintangsutawika's avatar
lintangsutawika committed
547
        ):
548
            # sample fewshot context #TODO: need to offset doc_id by rank now!
549
            fewshot_ctx = self.fewshot_context(
550
                doc,
551
                0 if self.config.num_fewshot is None else self.config.num_fewshot,
KonradSzafer's avatar
KonradSzafer committed
552
553
554
555
                system_instruction,
                apply_chat_template,
                fewshot_as_multiturn,
                lm,
556
            )
557

558
            # TODO: we should override self.config.repeats if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
559
560
561
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
562
                metadata=(self.config["task"], doc_id, self.config.repeats),
lintangsutawika's avatar
lintangsutawika committed
563
            )
564
565
566
567

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

568
569
570
571
572
573
574
575
576
577
578
579
580
            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
581

582
583
        if len(self._instances) == 0:
            raise ValueError("task.build_requests() did not find any docs!")
584

585
586
587
        if cache_requests and (not cached_instances or rewrite_requests_cache):
            save_to_cache(file_name=cache_key, obj=instances)

588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
    @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
604
            The number of times each instance in a dataset is inferred on. Defaults to 1,
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
            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

640
641
642
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

haileyschoelkopf's avatar
haileyschoelkopf committed
643
644
645
646
647
648
649
650
651
652
    @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))

653
    @utils.positional_deprecated
lintangsutawika's avatar
lintangsutawika committed
654
    def fewshot_context(
655
656
657
        self,
        doc,
        num_fewshot,
658
        rnd=None,
659
        description=None,
lintangsutawika's avatar
lintangsutawika committed
660
    ):
661
662
663
664
665
666
667
        """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
668
669
670
671
672
        :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.
673
674
675
        :returns: str
            The fewshot context.
        """
676
        if rnd is None:
677
678
679
680
681
682
            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
683

684
        description = description if description else ""
685
686

        if num_fewshot == 0:
lintangsutawika's avatar
lintangsutawika committed
687
            labeled_examples = ""
688
        else:
lintangsutawika's avatar
lintangsutawika committed
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
            # for sets with no training docs, draw from other set *but ensure no overlap with current doc*
            if self.has_training_docs():
                fewshotex = self.fewshot_examples(k=num_fewshot, rnd=rnd)
            else:
                if self._fewshot_docs is None:
                    self._fewshot_docs = list(
                        self.validation_docs()
                        if self.has_validation_docs()
                        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
713
            )
714
715

        example = self.doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
716
        return description + labeled_examples + example
717

718
    def apply_filters(self) -> Optional[List[Instance]]:
Baber Abbasi's avatar
Baber Abbasi committed
719
        """Iterates over FilterEnsembles and applies them to instances"""
lintangsutawika's avatar
lintangsutawika committed
720
721
        if hasattr(self, "_filters"):
            for f in self._filters:
722
                f.apply(self._instances)
lintangsutawika's avatar
lintangsutawika committed
723
724
725
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances
726

baberabb's avatar
baberabb committed
727
    def dump_config(self) -> dict:
Baber Abbasi's avatar
Baber Abbasi committed
728
        """Returns the config as a dictionary."""
729
        # TODO: this should only return the overrides applied to a non-YAML task's configuration.
730
        # (num_fewshot)
731
        return self.config.to_dict()
732

Baber Abbasi's avatar
Baber Abbasi committed
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
    def set_config(self, key: str, value: Any, update: bool = False) -> None:
        """Set or update the configuration for a given key."""
        if key is None:
            raise ValueError("Key must be provided.")

        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.
        """
        (
            self._metric_fn_list,
            self._aggregation_list,
            self._metric_fn_kwargs,
            self._higher_is_better,
        ) = ({}, {}, {}, {})
        self._metric_fn_list[metric_name] = get_metric(metric_name)
        self._aggregation_list[metric_name] = get_metric_aggregation(metric_name)
        self._higher_is_better[metric_name] = is_higher_better(metric_name)
        self._metric_fn_kwargs[metric_name] = {}
        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)
            }
        setattr(self._config, "metric_list", [{"metric": metric_name}])
        setattr(self._config, "process_results", None)

773
774
775
776
777
    def set_fewshot_seed(self, seed: Optional[int] = None) -> None:
        self.fewshot_rnd = random.Random(seed)
        if hasattr(self, "sampler"):
            self.sampler.rnd = self.fewshot_rnd

778
779
780
781
782
783
784
    @property
    def eval_docs(self) -> Union[datasets.Dataset, List[dict]]:
        if self.has_test_docs():
            return self.test_docs()
        elif self.has_validation_docs():
            return self.validation_docs()
        else:
785
786
787
            raise ValueError(
                f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"
            )
788
789
790
791
792
793
794
795
796
797
798
799
800

    def doc_iterator(
        self, *, rank: int = 0, limit: Union[int, None] = None, world_size: int = 1
    ) -> Iterator[Tuple[int, Any]]:
        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),
        )
        return doc_iterator

801
802
803
    @property
    def task_id(self) -> Any:
        return self._task_id
804

805

806
class ConfigurableTask(Task):
807
    VERSION = "Yaml"
808
    OUTPUT_TYPE = None
809
    CONFIG = None
810
811

    def __init__(
812
813
814
815
816
        self,
        data_dir=None,
        cache_dir=None,
        download_mode=None,
        config: Optional[dict] = None,
Ethan Smith's avatar
Ethan Smith committed
817
    ) -> None:  # TODO no super() call here
lintangsutawika's avatar
lintangsutawika committed
818
        # Create a unique identifier ID
lintangsutawika's avatar
lintangsutawika committed
819
        self._task_id = shortuuid.uuid()[:8]
lintangsutawika's avatar
lintangsutawika committed
820

821
        # Get pre-configured attributes
822
        self._config = self.CONFIG
823

824
        # Use new configurations if there was no preconfiguration
825
        if self.config is None:
826
            self._config = TaskConfig(**config)
827
828
        # Overwrite configs
        else:
lintangsutawika's avatar
lintangsutawika committed
829
            if config is not None:
830
                self._config.__dict__.update(config)
831

832
        if self.config is None:
lintangsutawika's avatar
lintangsutawika committed
833
834
835
            raise ValueError(
                "Must pass a config to ConfigurableTask, either in cls.CONFIG or `config` kwarg"
            )
836

837
838
839
840
        if isinstance(self.config.metadata, dict):
            if "version" in self.config.metadata:
                self.VERSION = self.config.metadata["version"]

841
        if self.config.output_type is not None:
842
843
844
845
            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)}'"
                )
846
            self.OUTPUT_TYPE = self.config.output_type
847

848
849
        if self.config.dataset_path is not None:
            self.DATASET_PATH = self.config.dataset_path
850

851
852
        if self.config.dataset_name is not None:
            self.DATASET_NAME = self.config.dataset_name
853

854
855
856
857
        self._metric_fn_list = {}
        self._metric_fn_kwargs = {}
        self._aggregation_list = {}
        self._higher_is_better = {}
858

859
        if self.config.metric_list is None:
860
            # TODO: handle this in TaskConfig.__post_init__ ?
861
862
            _metric_list = DEFAULT_METRIC_REGISTRY[self.config.output_type]

863
            for metric_name in _metric_list:
haileyschoelkopf's avatar
haileyschoelkopf committed
864
                self._metric_fn_list[metric_name] = get_metric(metric_name)
lintangsutawika's avatar
lintangsutawika committed
865
                self._metric_fn_kwargs[metric_name] = {}
866
867
868
                self._aggregation_list[metric_name] = get_metric_aggregation(
                    metric_name
                )
haileyschoelkopf's avatar
haileyschoelkopf committed
869
                self._higher_is_better[metric_name] = is_higher_better(metric_name)
870
        else:
871
            for metric_config in self.config.metric_list:
872
873
874
875
                if "metric" not in metric_config:
                    raise ValueError(
                        "'metric' key not provided for an entry in 'metric_list', must be specified!"
                    )
876
877
878
879
                metric_name = metric_config["metric"]
                kwargs = {
                    key: metric_config[key]
                    for key in metric_config
Chris's avatar
Chris committed
880
881
                    if key
                    not in ["metric", "aggregation", "higher_is_better", "hf_evaluate"]
882
                }
Chris's avatar
Chris committed
883
884
885
886
                hf_evaluate_metric = (
                    "hf_evaluate" in metric_config
                    and metric_config["hf_evaluate"] is True
                )
887

888
                if self.config.process_results is not None:
889
890
                    self._metric_fn_list[metric_name] = None
                    self._metric_fn_kwargs[metric_name] = {}
891
892
893
894
895
896
                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:
Chris's avatar
Chris committed
897
898
899
                    self._metric_fn_list[metric_name] = get_metric(
                        metric_name, hf_evaluate_metric
                    )
900
                    self._metric_fn_kwargs[metric_name] = kwargs
lintangsutawika's avatar
lintangsutawika committed
901

902
                if "aggregation" in metric_config:
903
                    agg_name = metric_config["aggregation"]
904
                    if isinstance(agg_name, str):
haileyschoelkopf's avatar
haileyschoelkopf committed
905
                        self._aggregation_list[metric_name] = get_aggregation(agg_name)
906
                    elif callable(agg_name):  # noqa: E721
907
908
909
                        self._aggregation_list[metric_name] = metric_config[
                            "aggregation"
                        ]
910
                else:
911
                    INV_AGG_REGISTRY = {v: k for k, v in AGGREGATION_REGISTRY.items()}
lintangsutawika's avatar
lintangsutawika committed
912
                    metric_agg = get_metric_aggregation(metric_name)
913
                    eval_logger.warning(
914
                        f"[Task: {self.config.task}] metric {metric_name} is defined, but aggregation is not. "
915
916
                        f"using default "
                        f"aggregation={INV_AGG_REGISTRY[metric_agg]}"
917
                    )
918
                    self._aggregation_list[metric_name] = metric_agg
lintangsutawika's avatar
lintangsutawika committed
919

920
921
922
923
924
925
                if "higher_is_better" in metric_config:
                    self._higher_is_better[metric_name] = metric_config[
                        "higher_is_better"
                    ]
                else:
                    eval_logger.warning(
926
                        f"[Task: {self.config.task}] metric {metric_name} is defined, but higher_is_better is not. "
927
                        f"using default "
haileyschoelkopf's avatar
haileyschoelkopf committed
928
                        f"higher_is_better={is_higher_better(metric_name)}"
929
                    )
haileyschoelkopf's avatar
haileyschoelkopf committed
930
                    self._higher_is_better[metric_name] = is_higher_better(metric_name)
931

932
        self.download(self.config.dataset_kwargs)
933
934
935
        self._training_docs = None
        self._fewshot_docs = None

936
        if self.config.filter_list is not None:
lintangsutawika's avatar
lintangsutawika committed
937
            self._filters = []
938
            for filter_config in self.config.filter_list:
939
940
941
942
943
944
945
946
947
                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"
                    }
                    components.append([function["function"], kwargs])
                filter_pipeline = build_filter_ensemble(filter_name, components)
lintangsutawika's avatar
lintangsutawika committed
948
                self._filters.append(filter_pipeline)
lintangsutawika's avatar
lintangsutawika committed
949
        else:
950
            self._filters = [build_filter_ensemble("none", [["take_first", None]])]
951

952
953
        if self.config.use_prompt is not None:
            eval_logger.info(f"loading prompt {self.config.use_prompt}")
954
            self.prompt = get_prompt(
955
                self.config.use_prompt, self.DATASET_PATH, self.DATASET_NAME
lintangsutawika's avatar
lintangsutawika committed
956
            )
957
958
959
        else:
            self.prompt = None

lintangsutawika's avatar
lintangsutawika committed
960
        if self.fewshot_docs() is not None:
961
962
963
964
            self.fewshot_rnd = (
                random.Random()
            )  # setting with no seed, to be overridden at a later time
            config_sampler: Union[str, Callable] = (
haileyschoelkopf's avatar
haileyschoelkopf committed
965
966
967
                self.config.fewshot_config.get("sampler", "default")
                if self.config.fewshot_config
                else "default"
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
            )
            if isinstance(config_sampler, str):
                self.sampler = samplers.get_sampler(config_sampler)(
                    list(self.fewshot_docs()), self, rnd=self.fewshot_rnd
                )
            elif callable(config_sampler) and issubclass(
                config_sampler, samplers.ContextSampler
            ):
                self.sampler = config_sampler(
                    docs=list(self.fewshot_docs()), task=self, rnd=self.fewshot_rnd
                )
            else:
                raise TypeError(
                    f"fewshot_config.sampler should be a string or callable of ContextSampler type, "
                    f"not {type(config_sampler)}"
                )
984

985
        self.task_docs = self.eval_docs
986

987
        # Test One Doc
988
        self.features = list(self.task_docs.features.keys())
989
990
        self.multiple_input = 0
        self.multiple_target = 0
991
        test_doc = self.task_docs[0]
992
        test_text = self.doc_to_text(test_doc)
993
        test_target = self.doc_to_target(test_doc)
lintangsutawika's avatar
lintangsutawika committed
994

995
        if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
996
            test_choice = self.doc_to_choice(test_doc)
997
            if not isinstance(test_choice, list):
lintangsutawika's avatar
lintangsutawika committed
998
                eval_logger.error("doc_to_choice must return list")
999
1000
            else:
                num_choice = len(test_choice)
1001

1002
            if isinstance(test_text, int):
1003
                self.multiple_input = num_choice
1004
1005
        else:
            test_choice = None
1006

1007
        if isinstance(test_target, list):
1008
            self.multiple_target = len(test_target)
1009
        else:
1010
            if (isinstance(test_target, int)) and (test_choice is not None):
lintangsutawika's avatar
lintangsutawika committed
1011
                test_target = test_choice[test_target]
1012
            else:
lintangsutawika's avatar
lintangsutawika committed
1013
                test_target = str(test_target)
1014

1015
1016
1017
        if test_choice is not None:
            check_choices = test_choice
        else:
lintangsutawika's avatar
lintangsutawika committed
1018
            check_choices = [test_target]
1019
1020
1021
1022
        if self.config.doc_to_choice is not None:
            for choice in check_choices:
                choice_has_whitespace = True if choice[0].isspace() else False
                delimiter_has_whitespace = (
1023
1024
                    True
                    if self.config.target_delimiter.rstrip()
1025
                    != self.config.target_delimiter
1026
                    else False
1027
                )
1028

1029
                if delimiter_has_whitespace and choice_has_whitespace:
1030
1031
                    eval_logger.debug(
                        f'Both target_delimiter "{self.config.target_delimiter}" and target choice: "{choice}" have whitespace'
1032
1033
                    )
                elif (not delimiter_has_whitespace) and (not choice_has_whitespace):
1034
                    eval_logger.debug(
1035
                        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'
1036
1037
                    )

1038
    def download(self, dataset_kwargs: Optional[Dict[str, Any]] = None) -> None:
1039
1040
1041
1042
1043
1044
        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
1045
    def has_training_docs(self) -> bool:
1046
        if self.config.training_split is not None:
1047
1048
1049
1050
            return True
        else:
            return False

baberabb's avatar
baberabb committed
1051
    def has_validation_docs(self) -> bool:
1052
        if self.config.validation_split is not None:
1053
1054
1055
1056
            return True
        else:
            return False

baberabb's avatar
baberabb committed
1057
    def has_test_docs(self) -> bool:
1058
        if self.config.test_split is not None:
1059
1060
1061
1062
            return True
        else:
            return False

baberabb's avatar
baberabb committed
1063
    def training_docs(self) -> datasets.Dataset:
1064
        if self.has_training_docs():
1065
1066
1067
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.training_split]
1068
                )
1069
            return self.dataset[self.config.training_split]
1070

baberabb's avatar
baberabb committed
1071
    def validation_docs(self) -> datasets.Dataset:
1072
        if self.has_validation_docs():
1073
1074
1075
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.validation_split]
1076
                )
1077
            return self.dataset[self.config.validation_split]
1078

baberabb's avatar
baberabb committed
1079
    def test_docs(self) -> datasets.Dataset:
1080
        if self.has_test_docs():
1081
1082
1083
            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]
1084

1085
    def fewshot_docs(self):
1086
        if self.config.fewshot_split is not None:
1087
1088
            if self.config.process_docs is not None:
                return self.config.process_docs(self.dataset[self.config.fewshot_split])
1089
            return self.dataset[self.config.fewshot_split]
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
        elif (
            self.config.fewshot_config is not None
            and self.config.fewshot_config.get("samples", None) is not None
        ):
            if isinstance(self.config.fewshot_config["samples"], list):
                return self.config.fewshot_config["samples"]
            elif callable(self.config.fewshot_config["samples"]):
                return self.config.fewshot_config["samples"]()
            else:
                raise Exception(
                    "`fewshot_config['samples']` was incorrectly defined in the configuration. It should be either a list of samples as a dict, or function returning this list."
                )
1102
        else:
1103
            if (self.config.num_fewshot is not None) and (self.config.num_fewshot > 0):
1104
                eval_logger.warning(
lintangsutawika's avatar
lintangsutawika committed
1105
                    f"[Task: {self.config.task}] "
1106
1107
1108
1109
                    "num_fewshot > 0 but fewshot_split is None. "
                    "using preconfigured rule."
                )
            return super().fewshot_docs()
1110

KonradSzafer's avatar
KonradSzafer committed
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
    @staticmethod
    def append_target_question(
        labeled_examples: List[Dict[str, str]],
        question: str,
        fewshot_as_multiturn: bool = False,
    ) -> 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":
                labeled_examples.append({"role": "user", "content": question})
            # if last message is user, append to it to avoid two user messages in a row
            else:
                labeled_examples[-1]["content"] += question
        else:
            # if fewshot_as_multiturn is True, append as next user entry (last is always assistant)
            labeled_examples.append({"role": "user", "content": question})

lintangsutawika's avatar
lintangsutawika committed
1132
    @utils.positional_deprecated
KonradSzafer's avatar
KonradSzafer committed
1133
1134
1135
1136
1137
1138
1139
1140
1141
    def fewshot_context(
        self,
        doc: str,
        num_fewshot: int,
        system_instruction: Optional[str] = None,
        apply_chat_template: bool = False,
        fewshot_as_multiturn: bool = False,
        lm=None,
    ) -> str:
lintangsutawika's avatar
lintangsutawika committed
1142
1143
1144
1145
1146
1147
1148
        """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
1149
1150
1151
1152
1153
1154
1155
1156
        :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.
        :param lm:
            Language model with definition of the tokenizer/function to use for applying the chat template.
lintangsutawika's avatar
lintangsutawika committed
1157
1158
1159
        :returns: str
            The fewshot context.
        """
KonradSzafer's avatar
KonradSzafer committed
1160
1161
1162
1163
1164
1165
1166

        if apply_chat_template:
            labeled_examples = []
        else:
            labeled_examples = ""

        # get task description
1167
1168
        if description := self.config.description:
            description = utils.apply_template(self.config.description, doc)
lintangsutawika's avatar
lintangsutawika committed
1169

KonradSzafer's avatar
KonradSzafer committed
1170
1171
1172
1173
1174
1175
1176
1177
1178
        # 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
1179
        else:
KonradSzafer's avatar
KonradSzafer committed
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
            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(
                        doc, num_fewshot, fewshot_as_multiturn
                    )
                )
            else:
                labeled_examples += self.sampler.get_context(doc, num_fewshot)
lintangsutawika's avatar
lintangsutawika committed
1199
1200

        example = self.doc_to_text(doc)
KonradSzafer's avatar
KonradSzafer committed
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
        if apply_chat_template:
            if self.multiple_input:
                return lm.apply_chat_template(labeled_examples)
            if isinstance(example, str):
                self.append_target_question(
                    labeled_examples, example, fewshot_as_multiturn
                )
            # 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)
                    self.append_target_question(chat, ex, fewshot_as_multiturn)
                    labeled_examples_list.append(lm.apply_chat_template(chat))
                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(
                        labeled_examples, choices[example], fewshot_as_multiturn
                    )
                else:
                    self.append_target_question(
                        labeled_examples, str(example), fewshot_as_multiturn
                    )
                # return lm.apply_chat_template(labeled_examples)
            return lm.apply_chat_template(labeled_examples)
1230
        else:
KonradSzafer's avatar
KonradSzafer committed
1231
1232
            if self.multiple_input:
                return labeled_examples
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
            if isinstance(example, str):
                return labeled_examples + example
            elif isinstance(example, list):
                return [labeled_examples + ex for ex in example]
            elif isinstance(example, int):
                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)
lintangsutawika's avatar
lintangsutawika committed
1243

1244
    def apply_filters(self):
Baber Abbasi's avatar
Baber Abbasi committed
1245
        """Iterates over FilterEnsembles and applies them to instances"""
1246
1247
        if hasattr(self, "_filters"):
            for f in self._filters:
1248
                f.apply(self._instances)
1249
1250
1251
1252
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances

1253
    def should_decontaminate(self):
1254
        return self.config.should_decontaminate
1255
1256

    def doc_to_decontamination_query(self, doc):
1257
        if self.config.should_decontaminate:
1258
1259
            if self.config.doc_to_decontamination_query is None:
                return self.doc_to_text(doc)
1260
            else:
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
                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
                        )
                    )
1272

1273
    def _process_doc(self, doc: dict) -> dict:
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
        """
        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):
1285
1286
        if self.prompt is not None:
            doc_to_text = self.prompt
1287
        else:
1288
            doc_to_text = self.config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
1289

1290
        if isinstance(doc_to_text, int):
1291
            return doc_to_text
1292
        elif isinstance(doc_to_text, str):
1293
            if doc_to_text in self.features:
1294
                # if self.config.doc_to_choice is not None:
1295
1296
                #     return self.doc_to_choice(doc)[doc[doc_to_text]]
                # else:
1297
1298
                return doc[doc_to_text]
            else:
lintangsutawika's avatar
lintangsutawika committed
1299
                text_string = utils.apply_template(doc_to_text, doc)
lintangsutawika's avatar
lintangsutawika committed
1300
                if text_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1301
1302
1303
                    return ast.literal_eval(text_string)
                else:
                    return text_string
1304
        elif callable(doc_to_text):
1305
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
1306
        # Used when applying a Promptsource template
1307
        elif hasattr(doc_to_text, "apply"):
1308
1309
1310
1311
1312
            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")
1313
                return self.config.fewshot_delimiter
1314
        else:
1315
            print(type(doc_to_text))
1316
            raise TypeError
1317

1318
    def doc_to_target(self, doc: Mapping) -> Union[int, str, list]:
1319
1320
        if self.prompt is not None:
            doc_to_target = self.prompt
1321
        else:
1322
            doc_to_target = self.config.doc_to_target
1323

1324
        if isinstance(doc_to_target, int):
1325
            return doc_to_target
1326
        elif isinstance(doc_to_target, str):
1327
            if doc_to_target in self.features:
1328
                # if self.config.doc_to_choice is not None:
1329
1330
1331
                #     return self.doc_to_choice(doc)[doc[doc_to_target]]
                # else:
                return doc[doc_to_target]
1332
            else:
lintangsutawika's avatar
lintangsutawika committed
1333
                target_string = utils.apply_template(doc_to_target, doc)
lintangsutawika's avatar
lintangsutawika committed
1334
                if target_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1335
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
1336
1337
1338
1339
1340
                elif (
                    len(target_string) >= 2
                    and (target_string[0] == "[")
                    and (target_string[-1] == "]")
                ):
baberabb's avatar
baberabb committed
1341
1342
1343
1344
                    try:
                        return ast.literal_eval(target_string)
                    except (SyntaxError, ValueError):
                        return target_string
lintangsutawika's avatar
lintangsutawika committed
1345
1346
                else:
                    return target_string
1347
        elif isinstance(doc_to_target, list):
1348
            return doc_to_target
1349
        elif callable(doc_to_target):
1350
            return doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
1351
        # Used when applying a Promptsource template
1352
        elif hasattr(doc_to_target, "apply"):
1353
            applied_prompt = doc_to_target.apply(doc)
1354
1355
1356
1357
            if len(applied_prompt) == 2:
                return applied_prompt[1]
            else:
                eval_logger.warning("Applied prompt returns empty string")
1358
                return self.config.fewshot_delimiter
1359
1360
        else:
            raise TypeError
1361

baberabb's avatar
baberabb committed
1362
    def doc_to_choice(self, doc: Any) -> List[str]:
1363
1364
        if self.prompt is not None:
            doc_to_choice = self.prompt
1365
        elif self.config.doc_to_choice is None:
1366
1367
            eval_logger.error("doc_to_choice was called but not set in config")
        else:
1368
            doc_to_choice = self.config.doc_to_choice
1369

1370
        if isinstance(doc_to_choice, str):
1371
1372
1373
1374
            if doc_to_choice in self.features:
                return doc[doc_to_choice]
            else:
                return ast.literal_eval(utils.apply_template(doc_to_choice, doc))
1375
        elif isinstance(doc_to_choice, list):
1376
            return doc_to_choice
1377
        elif isinstance(doc_to_choice, dict):
1378
1379
1380
1381
1382
1383
1384
            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
1385

baberabb's avatar
baberabb committed
1386
1387
1388
    def construct_requests(
        self, doc: dict, ctx: str, **kwargs
    ) -> Union[List[Instance], Instance]:
1389
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
1390
            arguments = (ctx, self.doc_to_target(doc))
1391
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
1392
            arguments = (self.doc_to_target(doc),)
1393
        elif self.OUTPUT_TYPE == "multiple_choice":
1394
            choices = self.doc_to_choice(doc)
1395
            target_delimiter = self.config.target_delimiter
1396
1397
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
1398
                cont = self.doc_to_target(doc)
1399
1400
1401
                arguments = [
                    (ctx + choice, f"{target_delimiter}{cont}") for choice in choices
                ]
1402
            else:
1403
                # Otherwise they are placed in the continuation
1404
                arguments = [(ctx, f"{target_delimiter}{cont}") for cont in choices]
1405

1406
            request_list = [
1407
1408
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1409
                    doc=doc,
1410
                    arguments=arg,
1411
                    idx=i,
1412
1413
                    **kwargs,
                )
1414
                for i, arg in enumerate(arguments)
1415
            ]
1416
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
1417
            if "acc_mutual_info" in self._metric_fn_list.keys():
1418
1419
1420
                # if we are calculating multiple choice accuracy
                # using mutual information instead of raw loglikelihood as metric, need unconditional lls.

lintangsutawika's avatar
lintangsutawika committed
1421
                # here mutual info refers to calculating
1422
1423
1424
1425
1426
1427
                # 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
1428
                            doc=doc,
1429
                            arguments=("", "{}".format(choice)),
1430
1431
1432
                            idx=i,
                            **kwargs,
                        )
lintangsutawika's avatar
lintangsutawika committed
1433
                        for i, choice in enumerate(choices)
1434
1435
1436
                    ]
                )
            return request_list
lintangsutawika's avatar
lintangsutawika committed
1437

1438
        elif self.OUTPUT_TYPE == "generate_until":
1439
            arguments = (ctx, deepcopy(self.config.generation_kwargs))
lintangsutawika's avatar
lintangsutawika committed
1440
1441

        return Instance(
lintangsutawika's avatar
lintangsutawika committed
1442
1443
            request_type=self.OUTPUT_TYPE, doc=doc, arguments=arguments, idx=0, **kwargs
        )
1444
1445

    def process_results(self, doc, results):
1446
1447
        if callable(self.config.process_results):
            return self.config.process_results(doc, results)
lintangsutawika's avatar
lintangsutawika committed
1448

1449
        result_dict = {}
1450
        use_metric = list(self._metric_fn_list.keys())
1451
1452
1453
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
1454
1455
1456
1457
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
1458
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
1459
            (loglikelihood,) = results
1460
1461
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
haileyschoelkopf's avatar
haileyschoelkopf committed
1462
            return {
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
                **(
                    {"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
1478
            }
1479
        elif self.OUTPUT_TYPE == "multiple_choice":
1480
            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
1481

1482
            # retrieve choices in List[str] form, to compute choice lengths, etc.
1483
            choices = self.doc_to_choice(doc)
1484
1485
            completion_len = np.array([float(len(i)) for i in choices])

1486
1487
            if (
                2 * len(choices) == len(lls)
1488
                and "acc_mutual_info" in self._metric_fn_list.keys()
1489
1490
1491
1492
            ):
                # then we are doing mutual info.
                # this stores the "dryrun" / unconditional answer loglikelihoods
                lls_unconditional = lls[1::2]
1493
1494
                if len(lls_unconditional) != len(choices):
                    raise ValueError
1495
1496
                # and this stores our "regular" conditional loglikelihoods
                lls = lls[::2]
1497

1498
1499
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
1500

1501
1502
            if self.multiple_input:
                gold = self.doc_to_text(doc)
1503
            else:
1504
                gold = self.doc_to_target(doc)
1505
1506

            gold_index_error = False
1507
            if isinstance(gold, list):
Lintang Sutawika's avatar
Lintang Sutawika committed
1508
1509
                gold = [i if i < len(choices) else -100 for i in gold]
                if -100 in gold:
1510
1511
                    gold_index_error = True
            else:
1512
                if isinstance(gold, int):
Lintang Sutawika's avatar
Lintang Sutawika committed
1513
                    gold = gold if gold < len(choices) else -100
1514
                elif isinstance(gold, str):
Lintang Sutawika's avatar
Lintang Sutawika committed
1515
                    gold = choices.index(gold) if gold in choices else -100
lintangsutawika's avatar
lintangsutawika committed
1516

Lintang Sutawika's avatar
Lintang Sutawika committed
1517
                if gold == -100:
1518
1519
1520
1521
                    gold_index_error = True

            if gold_index_error:
                eval_logger.warning(
lintangsutawika's avatar
lintangsutawika committed
1522
                    f"Label index was not in within range of available choices,"
1523
1524
                    f"Sample:\n\n{doc}\n\n"
                )
lintangsutawika's avatar
lintangsutawika committed
1525

1526
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
1527
1528
                acc = 1.0 if pred in gold else 0.0
                acc_norm = 1.0 if pred_norm in gold else 0.0
Lintang Sutawika's avatar
Lintang Sutawika committed
1529
                exact_match = int(any([is_greedy[i] if i != -100 else 0 for i in gold]))
lintangsutawika's avatar
lintangsutawika committed
1530
1531
1532
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
1533
                # 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
1534
                exact_match = int(is_greedy[gold]) if gold != -100 else 0
1535

Lintang Sutawika's avatar
Lintang Sutawika committed
1536
1537
1538
1539
            prob_norm = utils.softmax(lls)

            # TODO use keyword arguments to the metric?
            # gold, pred, norm stuff, the original lls,
1540
            result_dict = {
1541
                **({"acc": acc} if "acc" in use_metric else {}),
1542
1543
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
1544
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
1545
                **({"exact_match": exact_match} if "exact_match" in use_metric else {}),
Lintang Sutawika's avatar
Lintang Sutawika committed
1546
1547
1548
1549
1550
                **(
                    {"brier_score": (gold, prob_norm)}
                    if "brier_score" in use_metric
                    else {}
                ),
1551
1552
            }

1553
            if "acc_mutual_info" in use_metric:
lintangsutawika's avatar
lintangsutawika committed
1554
1555
1556
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
1557
1558
1559
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

1560
        elif self.OUTPUT_TYPE == "generate_until":
1561
            gold = self.doc_to_target(doc)
Chris's avatar
Chris committed
1562
            result = results[0]
1563
            if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1564
                # If you set doc_to_choice,
lintangsutawika's avatar
lintangsutawika committed
1565
                # it assumes that doc_to_target returns a number.
1566
1567
                choices = self.doc_to_choice(doc)
                gold = choices[gold]
1568
1569
            # we expect multiple_targets to be a list.
            elif self.multiple_target:
baberabb's avatar
baberabb committed
1570
                gold = list(gold)
Chris's avatar
Chris committed
1571
1572
1573
            elif type(gold) != type(result):
                # cast gold to the same type as result
                gold = type(result)(gold)
1574

lintangsutawika's avatar
lintangsutawika committed
1575
            for metric in self._metric_fn_list.keys():
haileyschoelkopf's avatar
haileyschoelkopf committed
1576
1577
1578
1579
1580
                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 = []
haileyschoelkopf's avatar
haileyschoelkopf committed
1581
1582
1583
1584
                    if not isinstance(gold, list):
                        # sometimes, a multiple_target dataset has exceptions where one doc has only one string answer
                        # print(gold)
                        gold = [gold]
1585
1586
1587
1588
1589
1590
1591
1592
                    if metric == "exact_match":
                        result = [result for _ in range(len(gold))]
                        scores = self._metric_fn_list[metric](
                            references=gold,
                            predictions=result,
                            **self._metric_fn_kwargs[metric],
                        )[metric]
                        result_score = 1.0 if scores > 0.0 else 0.0
haileyschoelkopf's avatar
haileyschoelkopf committed
1593
                    else:
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
                        for gold_option in gold:
                            try:
                                result_score = self._metric_fn_list[metric](
                                    references=[gold_option],
                                    predictions=[result],
                                    **self._metric_fn_kwargs[metric],
                                )
                            except (
                                TypeError
                            ):  # TODO: this is hacky and I don't want to do it
                                result_score = self._metric_fn_list[metric](
                                    [gold_option, result]
                                )
                            if isinstance(result_score, dict):
                                # TODO: this handles the case where HF evaluate returns a dict.
                                result_score = result_score[metric]
                            scores.append(result_score)
                        if any(scores):
                            result_score = 1.0
                        else:
                            result_score = 0.0
haileyschoelkopf's avatar
haileyschoelkopf committed
1615
                else:
1616
                    try:
1617
                        result_score = self._metric_fn_list[metric](
1618
1619
                            references=[gold],
                            predictions=[result],
1620
                            **self._metric_fn_kwargs[metric],
1621
                        )
1622
                    except TypeError:  # needed for now in order to use a different interface between our own metrics and HF Evaluate metrics
1623
                        result_score = self._metric_fn_list[metric]([gold, result])
1624
1625
1626
1627
                    if isinstance(result_score, dict):
                        # TODO: this handles the case where HF evaluate returns a dict.
                        result_score = result_score[metric]
                result_dict[metric] = result_score
1628
        else:
lintangsutawika's avatar
lintangsutawika committed
1629
1630
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
1631
                "'loglikelihood', 'loglikelihood_rolling', 'generate_until' or 'multiple_choice'",
1632
            )
1633
1634
1635

        return result_dict

Baber Abbasi's avatar
Baber Abbasi committed
1636
    def aggregation(self) -> dict:
1637
1638
        return self._aggregation_list

Baber Abbasi's avatar
Baber Abbasi committed
1639
    def higher_is_better(self) -> dict:
haileyschoelkopf's avatar
haileyschoelkopf committed
1640
        return self._higher_is_better
1641

Baber Abbasi's avatar
Baber Abbasi committed
1642
1643
1644
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

lintangsutawika's avatar
lintangsutawika committed
1645
1646
    @property
    def task_id(self) -> Any:
lintangsutawika's avatar
lintangsutawika committed
1647
1648
1649
1650
1651
        return "-".join((self.task_name, self._task_id))

    @property
    def task_name(self) -> Any:
        return getattr(self.config, "task", None)
lintangsutawika's avatar
lintangsutawika committed
1652

1653
1654
1655
1656
1657
1658
1659
1660
    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)},"
            f"num_samples={len(self.eval_docs)})"
        )

1661
1662

class MultipleChoiceTask(Task):
1663
    OUTPUT_TYPE = "loglikelihood"
1664

baberabb's avatar
baberabb committed
1665
    def doc_to_target(self, doc: dict) -> str:
1666
1667
        return " " + doc["choices"][doc["gold"]]

baberabb's avatar
baberabb committed
1668
    def construct_requests(self, doc: dict, ctx: str, **kwargs) -> List[Instance]:
1669
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1670
1671
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1672
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1673
                doc=doc,
1674
                arguments=(ctx, " {}".format(choice)),
1675
                idx=i,
1676
1677
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1678
1679
            for i, choice in enumerate(doc["choices"])
        ]
1680

1681
    def process_results(self, doc: dict, results: Iterable[Tuple[float, bool]]) -> dict:
lintangsutawika's avatar
lintangsutawika committed
1682
1683
1684
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
        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
1696
    def higher_is_better(self) -> dict:
1697
1698
1699
1700
1701
        return {
            "acc": True,
            "acc_norm": True,
        }

baberabb's avatar
baberabb committed
1702
    def aggregation(self) -> dict:
1703
1704
1705
1706
1707
1708
        return {
            "acc": mean,
            "acc_norm": mean,
        }


lintangsutawika's avatar
lintangsutawika committed
1709
class PerplexityTask(Task):
1710
1711
    OUTPUT_TYPE = "loglikelihood_rolling"

baberabb's avatar
baberabb committed
1712
    def has_training_docs(self) -> bool:
1713
1714
        return False

baberabb's avatar
baberabb committed
1715
    def fewshot_examples(self, k: int, rnd) -> List:
1716
1717
1718
1719
        if k != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1720
1721
        return []

baberabb's avatar
baberabb committed
1722
    def fewshot_context(self, doc: dict, num_fewshot: int) -> Literal[""]:
1723
1724
1725
1726
        if num_fewshot != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1727
1728
1729

        return ""

baberabb's avatar
baberabb committed
1730
    def higher_is_better(self) -> dict:
1731
1732
1733
1734
1735
1736
1737
1738
1739
        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
1740
    def doc_to_text(self, doc) -> str:
1741
1742
1743
1744
1745
        return ""

    def doc_to_target(self, doc):
        return doc

1746
1747
1748
    def construct_requests(self, doc: dict, ctx: Optional[str], **kwargs):
        if bool(ctx):
            raise ValueError
1749

lintangsutawika's avatar
lintangsutawika committed
1750
1751
1752
1753
1754
1755
1756
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1757

1758
    def process_results(self, doc: dict, results: Tuple[float]) -> dict:
1759
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1760
1761
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1762
1763
1764
1765
1766
1767
        return {
            "word_perplexity": (loglikelihood, words),
            "byte_perplexity": (loglikelihood, bytes_),
            "bits_per_byte": (loglikelihood, bytes_),
        }

baberabb's avatar
baberabb committed
1768
    def aggregation(self) -> dict:
1769
1770
1771
1772
1773
1774
1775
        return {
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
            "bits_per_byte": bits_per_byte,
        }

    @classmethod
baberabb's avatar
baberabb committed
1776
    def count_bytes(cls, doc) -> int:
1777
1778
1779
        return len(doc.encode("utf-8"))

    @classmethod
baberabb's avatar
baberabb committed
1780
    def count_words(cls, doc) -> int:
1781
1782
        """Downstream tasks with custom word boundaries should override this!"""
        return len(re.split(r"\s+", doc))