task.py 60.5 KB
Newer Older
1
import abc
2
import ast
lintangsutawika's avatar
lintangsutawika committed
3
import logging
4
import random
5
import re
lintangsutawika's avatar
lintangsutawika committed
6
import uuid
7
from collections.abc import Callable
8
from copy import deepcopy
9
from dataclasses import asdict, dataclass
10
from inspect import getsource
11
12
13
14
15
16
17
18
19
20
21
22
from typing import (
    Any,
    Dict,
    Iterable,
    Iterator,
    List,
    Literal,
    Mapping,
    Optional,
    Tuple,
    Union,
)
23
24
25

import datasets
import numpy as np
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

lintangsutawika's avatar
lintangsutawika committed
54

lintangsutawika's avatar
lintangsutawika committed
55
56
@dataclass
class GroupConfig(dict):
lintangsutawika's avatar
lintangsutawika committed
57
58
59
    group: Optional[str] = None
    group_alias: Optional[str] = None
    task: Optional[Union[str, list]] = None
lintangsutawika's avatar
lintangsutawika committed
60
61
62
    aggregate_metric: Optional[str] = False
    aggregate_fn: Optional[str] = "mean"
    weight_by_size: Optional[str] = False
lintangsutawika's avatar
lintangsutawika committed
63
    metric_alias: Optional[str] = None
64
    version: Optional[str] = 0
lintangsutawika's avatar
lintangsutawika committed
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107

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

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

    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()):
            if v is None:
                cfg_dict.pop(k)
            elif callable(v):
                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
108
109
110
111
112
class ConfigurableGroup(abc.ABC):
    def __init__(
        self,
        config: Optional[dict] = None,
    ) -> None:
lintangsutawika's avatar
lintangsutawika committed
113
114
        # Create a unique identifier ID
        self._task_id = str(uuid.uuid1())
lintangsutawika's avatar
lintangsutawika committed
115
116
117
118
119
        self._config = GroupConfig(**config)

    @property
    def group(self):
        return self._config.group
120

lintangsutawika's avatar
lintangsutawika committed
121
122
123
    @property
    def group_alias(self):
        return self._config.group_alias
124
125
126
127
128

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

lintangsutawika's avatar
lintangsutawika committed
129
130
131
132
    @property
    def config(self):
        return self._config.to_dict()

lintangsutawika's avatar
lintangsutawika committed
133
134
135
136
    @property
    def task_id(self) -> Any:
        return self._task_id

lintangsutawika's avatar
lintangsutawika committed
137
138
    def __repr__(self):
        return (
139
            f"ConfigurableGroup(group={self.group}," f"group_alias={self.group_alias})"
lintangsutawika's avatar
lintangsutawika committed
140
141
        )

142

143
144
@dataclass
class TaskConfig(dict):
145
    # task naming/registry
146
147
    task: Optional[str] = None
    task_alias: Optional[str] = None
lintangsutawika's avatar
lintangsutawika committed
148
    tag: Optional[Union[str, list]] = None
149
150
    group: Optional[Union[str, list]] = None
    group_alias: Optional[Union[str, list]] = None
151
152
153
    # HF dataset options.
    # which dataset to use,
    # and what splits for what purpose
154
155
156
157
158
159
160
161
162
    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 (?)
163
164
    # formatting / prompting options.
    # see docs/advanced_task_guide.md for more info
165
166
167
168
169
170
    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
171
    description: str = ""
172
173
    target_delimiter: str = " "
    fewshot_delimiter: str = "\n\n"
174
    fewshot_config: Optional[dict] = None
175
    # runtime configuration options
176
    num_fewshot: Optional[int] = None
177
    # scoring options
178
179
180
    metric_list: Optional[list] = None
    output_type: OutputType = "generate_until"
    generation_kwargs: Optional[dict] = None
181
    repeats: int = 1
182
    filter_list: Optional[Union[str, list]] = None
183
    should_decontaminate: bool = False
184
185
186
187
    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
188

Ethan Smith's avatar
Ethan Smith committed
189
    def __post_init__(self) -> None:
Lintang Sutawika's avatar
Lintang Sutawika committed
190
        if self.generation_kwargs is not None:
191
            if self.output_type != "generate_until":
192
                eval_logger.warning(
193
                    f"[{self.task}] passed `generation_kwargs`, but not using `output_type: generate_until`!"
Lintang Sutawika's avatar
Lintang Sutawika committed
194
195
196
197
198
199
200
201
                )

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

            if "until" not in self.generation_kwargs:
202
                self.generation_kwargs["until"] = [self.fewshot_delimiter]
Lintang Sutawika's avatar
Lintang Sutawika committed
203
        else:
204
            if self.output_type == "generate_until":
Lintang Sutawika's avatar
Lintang Sutawika committed
205
206
                # ensure that we greedily generate in absence of explicit arguments otherwise
                self.generation_kwargs = {
207
208
209
210
211
                    "until": (
                        None
                        if self.fewshot_delimiter is None
                        else [self.fewshot_delimiter]
                    ),
Lintang Sutawika's avatar
Lintang Sutawika committed
212
213
                    "do_sample": False,
                }
214

215
216
217
    def __getitem__(self, item):
        return getattr(self, item)

218
219
220
    def __setitem__(self, item, value):
        return setattr(self, item, value)

221
    def to_dict(self, keep_callable: bool = False) -> dict:
222
223
        """dumps the current config as a dictionary object, as a printable format.
        null fields will not be printed.
haileyschoelkopf's avatar
haileyschoelkopf committed
224
        Used for dumping results alongside full task configuration
225

haileyschoelkopf's avatar
haileyschoelkopf committed
226
227
228
229
230
231
232
233
234
235
        :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)
236
237
238
239
240
241
242
243
244
245
            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
246
        return cfg_dict
247

248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
    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)

264
265
266
267
268
269
270
271
272
273
274

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)
    """

275
    VERSION: Optional[Union[int, str]] = None
276

277
278
    # The name of the `Task` benchmark as denoted in the HuggingFace datasets Hub
    # or a path to a custom `datasets` loading script.
279
    DATASET_PATH: Optional[str] = None
280
281

    # The name of a subset within `DATASET_PATH`.
282
    DATASET_NAME: Optional[str] = None
283

284
    OUTPUT_TYPE: Optional[OutputType] = None
lintangsutawika's avatar
lintangsutawika committed
285

286
287
    def __init__(
        self,
288
289
290
291
        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
292
    ) -> None:
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
        """
        :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)
315
316
317
        self._training_docs: Optional[list] = None
        self._fewshot_docs: Optional[list] = None
        self._instances: Optional[List[Instance]] = None
318

319
320
        # Create a unique identifier ID
        self._task_id = str(uuid.uuid1())
321
        self._config: TaskConfig = TaskConfig({**config}) if config else TaskConfig()
322

lintangsutawika's avatar
lintangsutawika committed
323
        self._filters = [build_filter_ensemble("none", [["take_first", None]])]
324
325
326
        self.fewshot_rnd: Optional[
            random.Random
        ] = None  # purposely induce errors in case of improper usage
327

328
329
330
331
332
333
    def download(
        self,
        data_dir: Optional[str] = None,
        cache_dir: Optional[str] = None,
        download_mode=None,
    ) -> None:
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
        """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.
        """
358
359
360
361
362
363
364
        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,
        )
365

366
    @property
367
    def config(self) -> TaskConfig:
368
369
370
        """Returns the TaskConfig associated with this class."""
        return self._config

371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
    @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

386
    def training_docs(self) -> Iterable:
387
388
389
390
391
392
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

393
    def validation_docs(self) -> Iterable:
394
395
396
397
398
399
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

400
    def test_docs(self) -> Iterable:
401
402
403
404
405
406
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

407
    def fewshot_docs(self) -> Iterable:
408
409
410
411
412
413
414
415
416
        """
        :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
417
            eval_logger.warning(
418
                f"[Task: {self.config.task}] has_training_docs and has_validation_docs are False"
419
                ", using test_docs as fewshot_docs but this is not recommended."
lintangsutawika's avatar
lintangsutawika committed
420
            )
421
422
            return self.test_docs()

423
    def _process_doc(self, doc: dict) -> dict:
424
425
426
427
428
429
430
431
432
        """
        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
433

434
    @property
435
    def instances(self) -> List[Instance]:
436
437
438
439
440
441
442
443
444
445
446
        """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)

447
448
    def doc_to_decontamination_query(self, doc):
        raise NotImplementedError(
449
450
451
452
453
454
455
456
457
458
459
            "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

460
461
    def build_all_requests(
        self,
462
        *,
463
464
465
466
467
468
        limit=None,
        rank=None,
        world_size=None,
        cache_requests=False,
        rewrite_requests_cache=False,
    ) -> None:
469
        """Build a set of Instances for a task, and store them in task.instances"""
470
471
472
473

        # used with caching
        og_limit = limit

474
        cache_key = f"requests-{self._config.task}-{self.config.num_fewshot}shot-rank{rank}-world_size{world_size}"
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489

        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
490
        eval_logger.info(f"Building contexts for {self.config.task} on rank {rank}...")
491

492
        instances = []
493
494
495
496
497
498
499
500
501
502

        # 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(
503
            self.doc_iterator(rank=rank, limit=limit, world_size=world_size)
504
505
506
507
508
509
510
        )

        num_docs = len(doc_id_docs)

        for doc_id, doc in tqdm(
            doc_id_docs,
            total=num_docs,
lintangsutawika's avatar
lintangsutawika committed
511
        ):
512
            # sample fewshot context #TODO: need to offset doc_id by rank now!
513
            fewshot_ctx = self.fewshot_context(
514
                doc,
515
                0 if self.config.num_fewshot is None else self.config.num_fewshot,
516
            )
517

518
            # TODO: we should override self.config.repeats if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
519
520
521
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
522
                metadata=(self.config["task"], doc_id, self.config.repeats),
lintangsutawika's avatar
lintangsutawika committed
523
            )
524
525
526
527

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

528
529
530
531
532
533
534
535
536
537
538
539
540
            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
541

542
543
        if len(self._instances) == 0:
            raise ValueError("task.build_requests() did not find any docs!")
544

545
546
547
        if cache_requests and (not cached_instances or rewrite_requests_cache):
            save_to_cache(file_name=cache_key, obj=instances)

548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
    @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
564
            The number of times each instance in a dataset is inferred on. Defaults to 1,
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
            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

600
601
602
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

haileyschoelkopf's avatar
haileyschoelkopf committed
603
604
605
606
607
608
609
610
611
612
    @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))

613
    @utils.positional_deprecated
lintangsutawika's avatar
lintangsutawika committed
614
    def fewshot_context(
615
616
617
        self,
        doc,
        num_fewshot,
618
        rnd=None,
619
        description=None,
lintangsutawika's avatar
lintangsutawika committed
620
    ):
621
622
623
624
625
626
627
        """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
628
629
630
631
632
        :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.
633
634
635
        :returns: str
            The fewshot context.
        """
636
        if rnd is None:
637
638
639
640
641
642
            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
643

644
        description = description if description else ""
645
646

        if num_fewshot == 0:
lintangsutawika's avatar
lintangsutawika committed
647
            labeled_examples = ""
648
        else:
lintangsutawika's avatar
lintangsutawika committed
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
            # 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
673
            )
674
675

        example = self.doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
676
        return description + labeled_examples + example
677

678
    def apply_filters(self) -> Optional[List[Instance]]:
Baber Abbasi's avatar
Baber Abbasi committed
679
        """Iterates over FilterEnsembles and applies them to instances"""
lintangsutawika's avatar
lintangsutawika committed
680
681
        if hasattr(self, "_filters"):
            for f in self._filters:
682
                f.apply(self._instances)
lintangsutawika's avatar
lintangsutawika committed
683
684
685
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances
686

baberabb's avatar
baberabb committed
687
    def dump_config(self) -> dict:
Baber Abbasi's avatar
Baber Abbasi committed
688
        """Returns the config as a dictionary."""
689
        # TODO: this should only return the overrides applied to a non-YAML task's configuration.
690
        # (num_fewshot)
691
        return self.config.to_dict()
692

Baber Abbasi's avatar
Baber Abbasi committed
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
    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)

733
734
735
736
737
    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

738
739
740
741
742
743
744
    @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:
745
746
747
            raise ValueError(
                f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"
            )
748
749
750
751
752
753
754
755
756
757
758
759
760

    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

761
762
763
    @property
    def task_id(self) -> Any:
        return self._task_id
764

765

766
class ConfigurableTask(Task):
767
    VERSION = "Yaml"
768
    OUTPUT_TYPE = None
769
    CONFIG = None
770
771

    def __init__(
772
773
774
775
776
        self,
        data_dir=None,
        cache_dir=None,
        download_mode=None,
        config: Optional[dict] = None,
Ethan Smith's avatar
Ethan Smith committed
777
    ) -> None:  # TODO no super() call here
lintangsutawika's avatar
lintangsutawika committed
778
779
780
        # Create a unique identifier ID
        self._task_id = str(uuid.uuid1())

781
        # Get pre-configured attributes
782
        self._config = self.CONFIG
783

784
        # Use new configurations if there was no preconfiguration
785
        if self.config is None:
786
            self._config = TaskConfig(**config)
787
788
        # Overwrite configs
        else:
lintangsutawika's avatar
lintangsutawika committed
789
            if config is not None:
790
                self._config.__dict__.update(config)
791

792
        if self.config is None:
lintangsutawika's avatar
lintangsutawika committed
793
794
795
            raise ValueError(
                "Must pass a config to ConfigurableTask, either in cls.CONFIG or `config` kwarg"
            )
796

797
798
799
800
        if isinstance(self.config.metadata, dict):
            if "version" in self.config.metadata:
                self.VERSION = self.config.metadata["version"]

801
        if self.config.output_type is not None:
802
803
804
805
            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)}'"
                )
806
            self.OUTPUT_TYPE = self.config.output_type
807

808
809
        if self.config.dataset_path is not None:
            self.DATASET_PATH = self.config.dataset_path
810

811
812
        if self.config.dataset_name is not None:
            self.DATASET_NAME = self.config.dataset_name
813

814
815
816
817
        self._metric_fn_list = {}
        self._metric_fn_kwargs = {}
        self._aggregation_list = {}
        self._higher_is_better = {}
818

819
        if self.config.metric_list is None:
820
            # TODO: handle this in TaskConfig.__post_init__ ?
821
822
            _metric_list = DEFAULT_METRIC_REGISTRY[self.config.output_type]

823
            for metric_name in _metric_list:
haileyschoelkopf's avatar
haileyschoelkopf committed
824
                self._metric_fn_list[metric_name] = get_metric(metric_name)
lintangsutawika's avatar
lintangsutawika committed
825
                self._metric_fn_kwargs[metric_name] = {}
826
827
828
                self._aggregation_list[metric_name] = get_metric_aggregation(
                    metric_name
                )
haileyschoelkopf's avatar
haileyschoelkopf committed
829
                self._higher_is_better[metric_name] = is_higher_better(metric_name)
830
        else:
831
            for metric_config in self.config.metric_list:
832
833
834
835
                if "metric" not in metric_config:
                    raise ValueError(
                        "'metric' key not provided for an entry in 'metric_list', must be specified!"
                    )
836
837
838
839
                metric_name = metric_config["metric"]
                kwargs = {
                    key: metric_config[key]
                    for key in metric_config
Chris's avatar
Chris committed
840
841
                    if key
                    not in ["metric", "aggregation", "higher_is_better", "hf_evaluate"]
842
                }
Chris's avatar
Chris committed
843
844
845
846
                hf_evaluate_metric = (
                    "hf_evaluate" in metric_config
                    and metric_config["hf_evaluate"] is True
                )
847

848
                if self.config.process_results is not None:
849
850
                    self._metric_fn_list[metric_name] = None
                    self._metric_fn_kwargs[metric_name] = {}
851
852
853
854
855
856
                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
857
858
859
                    self._metric_fn_list[metric_name] = get_metric(
                        metric_name, hf_evaluate_metric
                    )
860
                    self._metric_fn_kwargs[metric_name] = kwargs
lintangsutawika's avatar
lintangsutawika committed
861

862
                if "aggregation" in metric_config:
863
                    agg_name = metric_config["aggregation"]
864
                    if isinstance(agg_name, str):
haileyschoelkopf's avatar
haileyschoelkopf committed
865
                        self._aggregation_list[metric_name] = get_aggregation(agg_name)
866
                    elif callable(agg_name):  # noqa: E721
867
868
869
                        self._aggregation_list[metric_name] = metric_config[
                            "aggregation"
                        ]
870
                else:
871
                    INV_AGG_REGISTRY = {v: k for k, v in AGGREGATION_REGISTRY.items()}
lintangsutawika's avatar
lintangsutawika committed
872
                    metric_agg = get_metric_aggregation(metric_name)
873
                    eval_logger.warning(
874
                        f"[Task: {self.config.task}] metric {metric_name} is defined, but aggregation is not. "
875
876
                        f"using default "
                        f"aggregation={INV_AGG_REGISTRY[metric_agg]}"
877
                    )
878
                    self._aggregation_list[metric_name] = metric_agg
lintangsutawika's avatar
lintangsutawika committed
879

880
881
882
883
884
885
                if "higher_is_better" in metric_config:
                    self._higher_is_better[metric_name] = metric_config[
                        "higher_is_better"
                    ]
                else:
                    eval_logger.warning(
886
                        f"[Task: {self.config.task}] metric {metric_name} is defined, but higher_is_better is not. "
887
                        f"using default "
haileyschoelkopf's avatar
haileyschoelkopf committed
888
                        f"higher_is_better={is_higher_better(metric_name)}"
889
                    )
haileyschoelkopf's avatar
haileyschoelkopf committed
890
                    self._higher_is_better[metric_name] = is_higher_better(metric_name)
891

892
        self.download(self.config.dataset_kwargs)
893
894
895
        self._training_docs = None
        self._fewshot_docs = None

896
        if self.config.filter_list is not None:
lintangsutawika's avatar
lintangsutawika committed
897
            self._filters = []
898
            for filter_config in self.config.filter_list:
899
900
901
902
903
904
905
906
907
                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
908
                self._filters.append(filter_pipeline)
lintangsutawika's avatar
lintangsutawika committed
909
        else:
910
            self._filters = [build_filter_ensemble("none", [["take_first", None]])]
911

912
913
        if self.config.use_prompt is not None:
            eval_logger.info(f"loading prompt {self.config.use_prompt}")
914
            self.prompt = get_prompt(
915
                self.config.use_prompt, self.DATASET_PATH, self.DATASET_NAME
lintangsutawika's avatar
lintangsutawika committed
916
            )
917
918
919
        else:
            self.prompt = None

lintangsutawika's avatar
lintangsutawika committed
920
        if self.fewshot_docs() is not None:
921
922
923
924
            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
925
926
927
                self.config.fewshot_config.get("sampler", "default")
                if self.config.fewshot_config
                else "default"
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
            )
            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)}"
                )
944

945
        self.task_docs = self.eval_docs
946

947
        # Test One Doc
948
        self.features = list(self.task_docs.features.keys())
949
950
        self.multiple_input = 0
        self.multiple_target = 0
951
        test_doc = self.task_docs[0]
952
        test_text = self.doc_to_text(test_doc)
953
        test_target = self.doc_to_target(test_doc)
lintangsutawika's avatar
lintangsutawika committed
954

955
        if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
956
            test_choice = self.doc_to_choice(test_doc)
957
            if not isinstance(test_choice, list):
lintangsutawika's avatar
lintangsutawika committed
958
                eval_logger.error("doc_to_choice must return list")
959
960
            else:
                num_choice = len(test_choice)
961

962
            if isinstance(test_text, int):
963
                self.multiple_input = num_choice
964
965
        else:
            test_choice = None
966

967
        if isinstance(test_target, list):
968
            self.multiple_target = len(test_target)
969
        else:
970
            if (isinstance(test_target, int)) and (test_choice is not None):
lintangsutawika's avatar
lintangsutawika committed
971
                test_target = test_choice[test_target]
972
            else:
lintangsutawika's avatar
lintangsutawika committed
973
                test_target = str(test_target)
974

975
976
977
        if test_choice is not None:
            check_choices = test_choice
        else:
lintangsutawika's avatar
lintangsutawika committed
978
            check_choices = [test_target]
979
980
981
982
        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 = (
983
984
                    True
                    if self.config.target_delimiter.rstrip()
985
                    != self.config.target_delimiter
986
                    else False
987
                )
988

989
                if delimiter_has_whitespace and choice_has_whitespace:
990
991
                    eval_logger.debug(
                        f'Both target_delimiter "{self.config.target_delimiter}" and target choice: "{choice}" have whitespace'
992
993
                    )
                elif (not delimiter_has_whitespace) and (not choice_has_whitespace):
994
                    eval_logger.debug(
995
                        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'
996
997
                    )

998
    def download(self, dataset_kwargs: Optional[Dict[str, Any]] = None) -> None:
999
1000
1001
1002
1003
1004
        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
1005
    def has_training_docs(self) -> bool:
1006
        if self.config.training_split is not None:
1007
1008
1009
1010
            return True
        else:
            return False

baberabb's avatar
baberabb committed
1011
    def has_validation_docs(self) -> bool:
1012
        if self.config.validation_split is not None:
1013
1014
1015
1016
            return True
        else:
            return False

baberabb's avatar
baberabb committed
1017
    def has_test_docs(self) -> bool:
1018
        if self.config.test_split is not None:
1019
1020
1021
1022
            return True
        else:
            return False

baberabb's avatar
baberabb committed
1023
    def training_docs(self) -> datasets.Dataset:
1024
        if self.has_training_docs():
1025
1026
1027
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.training_split]
1028
                )
1029
            return self.dataset[self.config.training_split]
1030

baberabb's avatar
baberabb committed
1031
    def validation_docs(self) -> datasets.Dataset:
1032
        if self.has_validation_docs():
1033
1034
1035
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.validation_split]
1036
                )
1037
            return self.dataset[self.config.validation_split]
1038

baberabb's avatar
baberabb committed
1039
    def test_docs(self) -> datasets.Dataset:
1040
        if self.has_test_docs():
1041
1042
1043
            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]
1044

1045
    def fewshot_docs(self):
1046
        if self.config.fewshot_split is not None:
1047
1048
            if self.config.process_docs is not None:
                return self.config.process_docs(self.dataset[self.config.fewshot_split])
1049
            return self.dataset[self.config.fewshot_split]
1050
        else:
1051
            if (self.config.num_fewshot is not None) and (self.config.num_fewshot > 0):
1052
                eval_logger.warning(
lintangsutawika's avatar
lintangsutawika committed
1053
                    f"[Task: {self.config.task}] "
1054
1055
1056
1057
                    "num_fewshot > 0 but fewshot_split is None. "
                    "using preconfigured rule."
                )
            return super().fewshot_docs()
1058

lintangsutawika's avatar
lintangsutawika committed
1059
    @utils.positional_deprecated
1060
    def fewshot_context(self, doc: str, num_fewshot: int) -> str:
lintangsutawika's avatar
lintangsutawika committed
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
        """Returns a fewshot context string that is made up of a prepended description
        (if provided), the `num_fewshot` number of examples, and an appended prompt example.

        :param doc: str
            The document as returned from training_docs, validation_docs, or test_docs.
        :param num_fewshot: int
            The number of fewshot examples to provide in the returned context string.
        :returns: str
            The fewshot context.
        """
1071
1072
        if description := self.config.description:
            description = utils.apply_template(self.config.description, doc)
lintangsutawika's avatar
lintangsutawika committed
1073
1074
1075

        if num_fewshot == 0:
            # always prepend the (possibly empty) task description
1076
            labeled_examples = description
lintangsutawika's avatar
lintangsutawika committed
1077
        else:
1078
            labeled_examples = description + self.sampler.get_context(doc, num_fewshot)
lintangsutawika's avatar
lintangsutawika committed
1079
1080

        example = self.doc_to_text(doc)
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
        if self.multiple_input:
            return labeled_examples
        else:
            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
1094

1095
    def apply_filters(self):
Baber Abbasi's avatar
Baber Abbasi committed
1096
        """Iterates over FilterEnsembles and applies them to instances"""
1097
1098
        if hasattr(self, "_filters"):
            for f in self._filters:
1099
                f.apply(self._instances)
1100
1101
1102
1103
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances

1104
    def should_decontaminate(self):
1105
        return self.config.should_decontaminate
1106
1107

    def doc_to_decontamination_query(self, doc):
1108
        if self.config.should_decontaminate:
1109
1110
            if self.config.doc_to_decontamination_query is None:
                return self.doc_to_text(doc)
1111
            else:
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
                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
                        )
                    )
1123

1124
    def _process_doc(self, doc: dict) -> dict:
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
        """
        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):
1136
1137
        if self.prompt is not None:
            doc_to_text = self.prompt
1138
        else:
1139
            doc_to_text = self.config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
1140

1141
        if isinstance(doc_to_text, int):
1142
            return doc_to_text
1143
        elif isinstance(doc_to_text, str):
1144
            if doc_to_text in self.features:
1145
                # if self.config.doc_to_choice is not None:
1146
1147
                #     return self.doc_to_choice(doc)[doc[doc_to_text]]
                # else:
1148
1149
                return doc[doc_to_text]
            else:
lintangsutawika's avatar
lintangsutawika committed
1150
                text_string = utils.apply_template(doc_to_text, doc)
lintangsutawika's avatar
lintangsutawika committed
1151
                if text_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1152
1153
1154
                    return ast.literal_eval(text_string)
                else:
                    return text_string
1155
        elif callable(doc_to_text):
1156
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
1157
        # Used when applying a Promptsource template
1158
        elif hasattr(doc_to_text, "apply"):
1159
1160
1161
1162
1163
            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")
1164
                return self.config.fewshot_delimiter
1165
        else:
1166
            print(type(doc_to_text))
1167
            raise TypeError
1168

1169
    def doc_to_target(self, doc: Mapping) -> Union[int, str, list]:
1170
1171
        if self.prompt is not None:
            doc_to_target = self.prompt
1172
        else:
1173
            doc_to_target = self.config.doc_to_target
1174

1175
        if isinstance(doc_to_target, int):
1176
            return doc_to_target
1177
        elif isinstance(doc_to_target, str):
1178
            if doc_to_target in self.features:
1179
                # if self.config.doc_to_choice is not None:
1180
1181
1182
                #     return self.doc_to_choice(doc)[doc[doc_to_target]]
                # else:
                return doc[doc_to_target]
1183
            else:
lintangsutawika's avatar
lintangsutawika committed
1184
                target_string = utils.apply_template(doc_to_target, doc)
lintangsutawika's avatar
lintangsutawika committed
1185
                if target_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1186
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
1187
1188
1189
1190
1191
                elif (
                    len(target_string) >= 2
                    and (target_string[0] == "[")
                    and (target_string[-1] == "]")
                ):
baberabb's avatar
baberabb committed
1192
1193
1194
1195
                    try:
                        return ast.literal_eval(target_string)
                    except (SyntaxError, ValueError):
                        return target_string
lintangsutawika's avatar
lintangsutawika committed
1196
1197
                else:
                    return target_string
1198
        elif isinstance(doc_to_target, list):
1199
            return doc_to_target
1200
        elif callable(doc_to_target):
1201
            return doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
1202
        # Used when applying a Promptsource template
1203
        elif hasattr(doc_to_target, "apply"):
1204
            applied_prompt = doc_to_target.apply(doc)
1205
1206
1207
1208
            if len(applied_prompt) == 2:
                return applied_prompt[1]
            else:
                eval_logger.warning("Applied prompt returns empty string")
1209
                return self.config.fewshot_delimiter
1210
1211
        else:
            raise TypeError
1212

baberabb's avatar
baberabb committed
1213
    def doc_to_choice(self, doc: Any) -> List[str]:
1214
1215
        if self.prompt is not None:
            doc_to_choice = self.prompt
1216
        elif self.config.doc_to_choice is None:
1217
1218
            eval_logger.error("doc_to_choice was called but not set in config")
        else:
1219
            doc_to_choice = self.config.doc_to_choice
1220

1221
        if isinstance(doc_to_choice, str):
1222
1223
1224
1225
            if doc_to_choice in self.features:
                return doc[doc_to_choice]
            else:
                return ast.literal_eval(utils.apply_template(doc_to_choice, doc))
1226
        elif isinstance(doc_to_choice, list):
1227
            return doc_to_choice
1228
        elif isinstance(doc_to_choice, dict):
1229
1230
1231
1232
1233
1234
1235
            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
1236

baberabb's avatar
baberabb committed
1237
1238
1239
    def construct_requests(
        self, doc: dict, ctx: str, **kwargs
    ) -> Union[List[Instance], Instance]:
1240
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
1241
            arguments = (ctx, self.doc_to_target(doc))
1242
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
1243
            arguments = (self.doc_to_target(doc),)
1244
        elif self.OUTPUT_TYPE == "multiple_choice":
1245
            choices = self.doc_to_choice(doc)
1246
            target_delimiter = self.config.target_delimiter
1247
1248
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
1249
                cont = self.doc_to_target(doc)
1250
1251
1252
                arguments = [
                    (ctx + choice, f"{target_delimiter}{cont}") for choice in choices
                ]
1253
            else:
1254
                # Otherwise they are placed in the continuation
1255
                arguments = [(ctx, f"{target_delimiter}{cont}") for cont in choices]
1256

1257
            request_list = [
1258
1259
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1260
                    doc=doc,
1261
                    arguments=arg,
1262
                    idx=i,
1263
1264
                    **kwargs,
                )
1265
                for i, arg in enumerate(arguments)
1266
            ]
1267
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
1268
            if "acc_mutual_info" in self._metric_fn_list.keys():
1269
1270
1271
                # if we are calculating multiple choice accuracy
                # using mutual information instead of raw loglikelihood as metric, need unconditional lls.

lintangsutawika's avatar
lintangsutawika committed
1272
                # here mutual info refers to calculating
1273
1274
1275
1276
1277
1278
                # 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
1279
                            doc=doc,
1280
                            arguments=("", "{}".format(choice)),
1281
1282
1283
                            idx=i,
                            **kwargs,
                        )
lintangsutawika's avatar
lintangsutawika committed
1284
                        for i, choice in enumerate(choices)
1285
1286
1287
                    ]
                )
            return request_list
lintangsutawika's avatar
lintangsutawika committed
1288

1289
        elif self.OUTPUT_TYPE == "generate_until":
1290
            arguments = (ctx, deepcopy(self.config.generation_kwargs))
lintangsutawika's avatar
lintangsutawika committed
1291
1292

        return Instance(
lintangsutawika's avatar
lintangsutawika committed
1293
1294
            request_type=self.OUTPUT_TYPE, doc=doc, arguments=arguments, idx=0, **kwargs
        )
1295
1296

    def process_results(self, doc, results):
1297
1298
        if callable(self.config.process_results):
            return self.config.process_results(doc, results)
lintangsutawika's avatar
lintangsutawika committed
1299

1300
        result_dict = {}
1301
        use_metric = list(self._metric_fn_list.keys())
1302
1303
1304
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
1305
1306
1307
1308
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
1309
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
1310
            (loglikelihood,) = results
1311
1312
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
haileyschoelkopf's avatar
haileyschoelkopf committed
1313
            return {
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
                **(
                    {"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
1329
            }
1330
        elif self.OUTPUT_TYPE == "multiple_choice":
1331
            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
1332

1333
            # retrieve choices in List[str] form, to compute choice lengths, etc.
1334
            choices = self.doc_to_choice(doc)
1335
1336
            completion_len = np.array([float(len(i)) for i in choices])

1337
1338
            if (
                2 * len(choices) == len(lls)
1339
                and "acc_mutual_info" in self._metric_fn_list.keys()
1340
1341
1342
1343
            ):
                # then we are doing mutual info.
                # this stores the "dryrun" / unconditional answer loglikelihoods
                lls_unconditional = lls[1::2]
1344
1345
                if len(lls_unconditional) != len(choices):
                    raise ValueError
1346
1347
                # and this stores our "regular" conditional loglikelihoods
                lls = lls[::2]
1348

1349
1350
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
1351

1352
1353
            if self.multiple_input:
                gold = self.doc_to_text(doc)
1354
            else:
1355
                gold = self.doc_to_target(doc)
1356
1357

            gold_index_error = False
1358
            if isinstance(gold, list):
Lintang Sutawika's avatar
Lintang Sutawika committed
1359
1360
                gold = [i if i < len(choices) else -100 for i in gold]
                if -100 in gold:
1361
1362
                    gold_index_error = True
            else:
1363
                if isinstance(gold, int):
Lintang Sutawika's avatar
Lintang Sutawika committed
1364
                    gold = gold if gold < len(choices) else -100
1365
                elif isinstance(gold, str):
Lintang Sutawika's avatar
Lintang Sutawika committed
1366
                    gold = choices.index(gold) if gold in choices else -100
lintangsutawika's avatar
lintangsutawika committed
1367

Lintang Sutawika's avatar
Lintang Sutawika committed
1368
                if gold == -100:
1369
1370
1371
1372
                    gold_index_error = True

            if gold_index_error:
                eval_logger.warning(
lintangsutawika's avatar
lintangsutawika committed
1373
                    f"Label index was not in within range of available choices,"
1374
1375
                    f"Sample:\n\n{doc}\n\n"
                )
lintangsutawika's avatar
lintangsutawika committed
1376

1377
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
1378
1379
                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
1380
                exact_match = int(any([is_greedy[i] if i != -100 else 0 for i in gold]))
lintangsutawika's avatar
lintangsutawika committed
1381
1382
1383
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
1384
                # 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
1385
                exact_match = int(is_greedy[gold]) if gold != -100 else 0
1386

Lintang Sutawika's avatar
Lintang Sutawika committed
1387
1388
1389
1390
            prob_norm = utils.softmax(lls)

            # TODO use keyword arguments to the metric?
            # gold, pred, norm stuff, the original lls,
1391
            result_dict = {
1392
                **({"acc": acc} if "acc" in use_metric else {}),
1393
1394
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
1395
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
1396
                **({"exact_match": exact_match} if "exact_match" in use_metric else {}),
Lintang Sutawika's avatar
Lintang Sutawika committed
1397
1398
1399
1400
1401
                **(
                    {"brier_score": (gold, prob_norm)}
                    if "brier_score" in use_metric
                    else {}
                ),
1402
1403
            }

1404
            if "acc_mutual_info" in use_metric:
lintangsutawika's avatar
lintangsutawika committed
1405
1406
1407
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
1408
1409
1410
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

1411
        elif self.OUTPUT_TYPE == "generate_until":
1412
            gold = self.doc_to_target(doc)
Chris's avatar
Chris committed
1413
            result = results[0]
1414
            if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1415
                # If you set doc_to_choice,
lintangsutawika's avatar
lintangsutawika committed
1416
                # it assumes that doc_to_target returns a number.
1417
1418
                choices = self.doc_to_choice(doc)
                gold = choices[gold]
1419
1420
            # we expect multiple_targets to be a list.
            elif self.multiple_target:
baberabb's avatar
baberabb committed
1421
                gold = list(gold)
Chris's avatar
Chris committed
1422
1423
1424
            elif type(gold) != type(result):
                # cast gold to the same type as result
                gold = type(result)(gold)
1425

lintangsutawika's avatar
lintangsutawika committed
1426
            for metric in self._metric_fn_list.keys():
haileyschoelkopf's avatar
haileyschoelkopf committed
1427
1428
1429
1430
1431
                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
1432
1433
1434
1435
                    if not isinstance(gold, list):
                        # sometimes, a multiple_target dataset has exceptions where one doc has only one string answer
                        # print(gold)
                        gold = [gold]
1436
1437
1438
1439
1440
1441
1442
1443
                    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
1444
                    else:
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
                        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
1466
                else:
1467
                    try:
1468
                        result_score = self._metric_fn_list[metric](
1469
1470
                            references=[gold],
                            predictions=[result],
1471
                            **self._metric_fn_kwargs[metric],
1472
                        )
1473
                    except TypeError:  # needed for now in order to use a different interface between our own metrics and HF Evaluate metrics
1474
                        result_score = self._metric_fn_list[metric]([gold, result])
1475
1476
1477
1478
                    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
1479
        else:
lintangsutawika's avatar
lintangsutawika committed
1480
1481
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
1482
                "'loglikelihood', 'loglikelihood_rolling', 'generate_until' or 'multiple_choice'",
1483
            )
1484
1485
1486

        return result_dict

Baber Abbasi's avatar
Baber Abbasi committed
1487
    def aggregation(self) -> dict:
1488
1489
        return self._aggregation_list

Baber Abbasi's avatar
Baber Abbasi committed
1490
    def higher_is_better(self) -> dict:
haileyschoelkopf's avatar
haileyschoelkopf committed
1491
        return self._higher_is_better
1492

Baber Abbasi's avatar
Baber Abbasi committed
1493
1494
1495
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

lintangsutawika's avatar
lintangsutawika committed
1496
1497
1498
1499
    @property
    def task_id(self) -> Any:
        return self._task_id

1500
1501
1502
1503
1504
1505
1506
1507
    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)})"
        )

1508
1509

class MultipleChoiceTask(Task):
1510
    OUTPUT_TYPE = "loglikelihood"
1511

baberabb's avatar
baberabb committed
1512
    def doc_to_target(self, doc: dict) -> str:
1513
1514
        return " " + doc["choices"][doc["gold"]]

baberabb's avatar
baberabb committed
1515
    def construct_requests(self, doc: dict, ctx: str, **kwargs) -> List[Instance]:
1516
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1517
1518
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1519
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1520
                doc=doc,
1521
                arguments=(ctx, " {}".format(choice)),
1522
                idx=i,
1523
1524
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1525
1526
            for i, choice in enumerate(doc["choices"])
        ]
1527

1528
    def process_results(self, doc: dict, results: Iterable[Tuple[float, bool]]) -> dict:
lintangsutawika's avatar
lintangsutawika committed
1529
1530
1531
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
        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
1543
    def higher_is_better(self) -> dict:
1544
1545
1546
1547
1548
        return {
            "acc": True,
            "acc_norm": True,
        }

baberabb's avatar
baberabb committed
1549
    def aggregation(self) -> dict:
1550
1551
1552
1553
1554
1555
        return {
            "acc": mean,
            "acc_norm": mean,
        }


lintangsutawika's avatar
lintangsutawika committed
1556
class PerplexityTask(Task):
1557
1558
    OUTPUT_TYPE = "loglikelihood_rolling"

baberabb's avatar
baberabb committed
1559
    def has_training_docs(self) -> bool:
1560
1561
        return False

baberabb's avatar
baberabb committed
1562
    def fewshot_examples(self, k: int, rnd) -> List:
1563
1564
1565
1566
        if k != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1567
1568
        return []

baberabb's avatar
baberabb committed
1569
    def fewshot_context(self, doc: dict, num_fewshot: int) -> Literal[""]:
1570
1571
1572
1573
        if num_fewshot != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1574
1575
1576

        return ""

baberabb's avatar
baberabb committed
1577
    def higher_is_better(self) -> dict:
1578
1579
1580
1581
1582
1583
1584
1585
1586
        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
1587
    def doc_to_text(self, doc) -> str:
1588
1589
1590
1591
1592
        return ""

    def doc_to_target(self, doc):
        return doc

1593
1594
1595
    def construct_requests(self, doc: dict, ctx: Optional[str], **kwargs):
        if bool(ctx):
            raise ValueError
1596

lintangsutawika's avatar
lintangsutawika committed
1597
1598
1599
1600
1601
1602
1603
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1604

1605
    def process_results(self, doc: dict, results: Tuple[float]) -> dict:
1606
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1607
1608
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1609
1610
1611
1612
1613
1614
        return {
            "word_perplexity": (loglikelihood, words),
            "byte_perplexity": (loglikelihood, bytes_),
            "bits_per_byte": (loglikelihood, bytes_),
        }

baberabb's avatar
baberabb committed
1615
    def aggregation(self) -> dict:
1616
1617
1618
1619
1620
1621
1622
        return {
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
            "bits_per_byte": bits_per_byte,
        }

    @classmethod
baberabb's avatar
baberabb committed
1623
    def count_bytes(cls, doc) -> int:
1624
1625
1626
        return len(doc.encode("utf-8"))

    @classmethod
baberabb's avatar
baberabb committed
1627
    def count_words(cls, doc) -> int:
1628
1629
        """Downstream tasks with custom word boundaries should override this!"""
        return len(re.split(r"\s+", doc))