task.py 56.4 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
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
25
from tqdm import tqdm
26
27

from lm_eval import utils
28
from lm_eval.api import samplers
29
30
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
31
from lm_eval.api.registry import (
32
33
    AGGREGATION_REGISTRY,
    DEFAULT_METRIC_REGISTRY,
haileyschoelkopf's avatar
haileyschoelkopf committed
34
    get_aggregation,
35
    get_metric,
36
    get_metric_aggregation,
haileyschoelkopf's avatar
haileyschoelkopf committed
37
    is_higher_better,
lintangsutawika's avatar
lintangsutawika committed
38
)
39
from lm_eval.caching.cache import load_from_cache, save_to_cache
40
41
42
from lm_eval.filters import build_filter_ensemble
from lm_eval.prompts import get_prompt

43

44
45
46
47
ALL_OUTPUT_TYPES = [
    "loglikelihood",
    "multiple_choice",
    "loglikelihood_rolling",
48
    "generate_until",
49
50
]

51
eval_logger = logging.getLogger("lm-eval")
52

lintangsutawika's avatar
lintangsutawika committed
53

54
55
@dataclass
class TaskConfig(dict):
56
    # task naming/registry
57
58
59
60
    task: Optional[str] = None
    task_alias: Optional[str] = None
    group: Optional[Union[str, list]] = None
    group_alias: Optional[Union[str, list]] = None
61
62
63
    # HF dataset options.
    # which dataset to use,
    # and what splits for what purpose
64
65
66
67
68
69
70
71
72
    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 (?)
73
74
    # formatting / prompting options.
    # see docs/advanced_task_guide.md for more info
75
76
77
78
79
80
    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
81
    description: str = ""
82
83
    target_delimiter: str = " "
    fewshot_delimiter: str = "\n\n"
84
    fewshot_config: Optional[dict] = None
85
    # runtime configuration options
86
    num_fewshot: Optional[int] = None
87
    # scoring options
88
89
90
    metric_list: Optional[list] = None
    output_type: OutputType = "generate_until"
    generation_kwargs: Optional[dict] = None
91
    repeats: int = 1
92
    filter_list: Optional[Union[str, list]] = None
93
    should_decontaminate: bool = False
94
95
96
97
    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
98

Ethan Smith's avatar
Ethan Smith committed
99
    def __post_init__(self) -> None:
Lintang Sutawika's avatar
Lintang Sutawika committed
100
        if self.generation_kwargs is not None:
101
            if self.output_type != "generate_until":
102
                raise ValueError(
103
                    f"[{self.task}] passed `generation_kwargs`, but not using `output_type: generate_until`!"
Lintang Sutawika's avatar
Lintang Sutawika committed
104
105
106
107
108
109
110
111
                )

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

            if "until" not in self.generation_kwargs:
112
                self.generation_kwargs["until"] = [self.fewshot_delimiter]
Lintang Sutawika's avatar
Lintang Sutawika committed
113
        else:
114
            if self.output_type == "generate_until":
Lintang Sutawika's avatar
Lintang Sutawika committed
115
116
                # ensure that we greedily generate in absence of explicit arguments otherwise
                self.generation_kwargs = {
117
118
119
120
121
                    "until": (
                        None
                        if self.fewshot_delimiter is None
                        else [self.fewshot_delimiter]
                    ),
Lintang Sutawika's avatar
Lintang Sutawika committed
122
123
                    "do_sample": False,
                }
124

125
126
127
    def __getitem__(self, item):
        return getattr(self, item)

128
129
130
    def __setitem__(self, item, value):
        return setattr(self, item, value)

131
    def to_dict(self, keep_callable: bool = False) -> dict:
132
133
        """dumps the current config as a dictionary object, as a printable format.
        null fields will not be printed.
haileyschoelkopf's avatar
haileyschoelkopf committed
134
        Used for dumping results alongside full task configuration
135

haileyschoelkopf's avatar
haileyschoelkopf committed
136
137
138
139
140
141
142
143
144
145
        :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)
146
147
148
149
150
151
152
153
154
155
            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
156
        return cfg_dict
157

158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
    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)

174
175
176
177
178
179
180
181
182
183
184

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)
    """

185
    VERSION: Optional[Union[int, str]] = None
186

187
188
    # The name of the `Task` benchmark as denoted in the HuggingFace datasets Hub
    # or a path to a custom `datasets` loading script.
189
    DATASET_PATH: Optional[str] = None
190
191

    # The name of a subset within `DATASET_PATH`.
192
    DATASET_NAME: Optional[str] = None
193

194
    OUTPUT_TYPE: Optional[OutputType] = None
lintangsutawika's avatar
lintangsutawika committed
195

196
197
    def __init__(
        self,
198
199
200
201
        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
202
    ) -> None:
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
        """
        :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)
225
226
227
        self._training_docs: Optional[list] = None
        self._fewshot_docs: Optional[list] = None
        self._instances: Optional[List[Instance]] = None
228

229
        self._config: TaskConfig = TaskConfig({**config}) if config else TaskConfig()
230

lintangsutawika's avatar
lintangsutawika committed
231
        self._filters = [build_filter_ensemble("none", [["take_first", None]])]
232

233
234
235
236
237
238
    def download(
        self,
        data_dir: Optional[str] = None,
        cache_dir: Optional[str] = None,
        download_mode=None,
    ) -> None:
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
        """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.
        """
263
264
265
266
267
268
269
        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,
        )
270

271
    @property
272
    def config(self) -> TaskConfig:
273
274
275
        """Returns the TaskConfig associated with this class."""
        return self._config

276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
    @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

291
    def training_docs(self) -> Iterable:
292
293
294
295
296
297
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

298
    def validation_docs(self) -> Iterable:
299
300
301
302
303
304
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

305
    def test_docs(self) -> Iterable:
306
307
308
309
310
311
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

312
    def fewshot_docs(self) -> Iterable:
313
314
315
316
317
318
319
320
321
        """
        :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
322
            eval_logger.warning(
323
                f"[Task: {self.config.task}] has_training_docs and has_validation_docs are False"
324
                ", using test_docs as fewshot_docs but this is not recommended."
lintangsutawika's avatar
lintangsutawika committed
325
            )
326
327
            return self.test_docs()

328
    def _process_doc(self, doc: dict) -> dict:
329
330
331
332
333
334
335
336
337
        """
        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
338

339
    @property
340
    def instances(self) -> List[Instance]:
341
342
343
344
345
346
347
348
349
350
351
        """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)

352
353
    def doc_to_decontamination_query(self, doc):
        raise NotImplementedError(
354
355
356
357
358
359
360
361
362
363
364
            "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

365
366
    def build_all_requests(
        self,
367
        *,
368
369
370
371
372
373
        limit=None,
        rank=None,
        world_size=None,
        cache_requests=False,
        rewrite_requests_cache=False,
    ) -> None:
374
        """Build a set of Instances for a task, and store them in task.instances"""
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394

        # used with caching
        og_limit = limit

        cache_key = f"requests-{self._config.task}"

        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
395
        eval_logger.info(f"Building contexts for {self.config.task} on rank {rank}...")
396

397
        instances = []
398
399
400
401
402
403
404
405
406
407

        # 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(
408
            self.doc_iterator(rank=rank, limit=limit, world_size=world_size)
409
410
411
412
413
414
415
        )

        num_docs = len(doc_id_docs)

        for doc_id, doc in tqdm(
            doc_id_docs,
            total=num_docs,
lintangsutawika's avatar
lintangsutawika committed
416
        ):
417
            # sample fewshot context #TODO: need to offset doc_id by rank now!
418
            fewshot_ctx = self.fewshot_context(
419
                doc,
420
                0 if self.config.num_fewshot is None else self.config.num_fewshot,
421
            )
422

423
            # TODO: we should override self.config.repeats if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
424
425
426
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
427
                metadata=(self.config["task"], doc_id, self.config.repeats),
lintangsutawika's avatar
lintangsutawika committed
428
            )
429
430
431
432

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

433
434
435
436
437
438
439
440
441
442
443
444
445
            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
446

447
448
        if len(self._instances) == 0:
            raise ValueError("task.build_requests() did not find any docs!")
449

450
451
452
        if cache_requests and (not cached_instances or rewrite_requests_cache):
            save_to_cache(file_name=cache_key, obj=instances)

453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
    @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
469
            The number of times each instance in a dataset is inferred on. Defaults to 1,
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
            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

505
506
507
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

haileyschoelkopf's avatar
haileyschoelkopf committed
508
509
510
511
512
513
514
515
516
517
    @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))

518
    @utils.positional_deprecated
lintangsutawika's avatar
lintangsutawika committed
519
    def fewshot_context(
520
521
522
523
524
        self,
        doc,
        num_fewshot,
        rnd=random.Random(1234),
        description=None,
lintangsutawika's avatar
lintangsutawika committed
525
    ):
526
527
528
529
530
531
532
        """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
533
534
535
536
537
        :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.
538
539
540
        :returns: str
            The fewshot context.
        """
541
542
543
544
        if rnd is None:
            raise ValueError(
                "A `random.Random` generator argument must be provided to `rnd`"
            )
lintangsutawika's avatar
lintangsutawika committed
545

546
        description = description if description else ""
547
548

        if num_fewshot == 0:
lintangsutawika's avatar
lintangsutawika committed
549
            labeled_examples = ""
550
        else:
lintangsutawika's avatar
lintangsutawika committed
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
            # 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
575
            )
576
577

        example = self.doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
578
        return description + labeled_examples + example
579

580
    def apply_filters(self) -> Optional[List[Instance]]:
Baber Abbasi's avatar
Baber Abbasi committed
581
        """Iterates over FilterEnsembles and applies them to instances"""
lintangsutawika's avatar
lintangsutawika committed
582
583
        if hasattr(self, "_filters"):
            for f in self._filters:
584
                f.apply(self._instances)
lintangsutawika's avatar
lintangsutawika committed
585
586
587
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances
588

baberabb's avatar
baberabb committed
589
    def dump_config(self) -> dict:
Baber Abbasi's avatar
Baber Abbasi committed
590
        """Returns the config as a dictionary."""
591
        # TODO: this should only return the overrides applied to a non-YAML task's configuration.
592
        # (num_fewshot)
593
        return self.config.to_dict()
594

Baber Abbasi's avatar
Baber Abbasi committed
595
596
597
598
599
600
601
602
603
604
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
    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)

635
636
637
638
639
640
641
    @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:
642
643
644
            raise ValueError(
                f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"
            )
645
646
647
648
649
650
651
652
653
654
655
656
657

    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

658
659

class ConfigurableTask(Task):
660
    VERSION = "Yaml"
661
    OUTPUT_TYPE = None
662
    CONFIG = None
663
664

    def __init__(
665
666
667
668
669
        self,
        data_dir=None,
        cache_dir=None,
        download_mode=None,
        config: Optional[dict] = None,
Ethan Smith's avatar
Ethan Smith committed
670
    ) -> None:  # TODO no super() call here
671
        # Get pre-configured attributes
672
        self._config = self.CONFIG
673

674
        # Use new configurations if there was no preconfiguration
675
        if self.config is None:
676
            self._config = TaskConfig(**config)
677
678
        # Overwrite configs
        else:
lintangsutawika's avatar
lintangsutawika committed
679
            if config is not None:
680
                self._config.__dict__.update(config)
681

682
        if self.config is None:
lintangsutawika's avatar
lintangsutawika committed
683
684
685
            raise ValueError(
                "Must pass a config to ConfigurableTask, either in cls.CONFIG or `config` kwarg"
            )
686

687
688
689
690
        if isinstance(self.config.metadata, dict):
            if "version" in self.config.metadata:
                self.VERSION = self.config.metadata["version"]

691
        if self.config.output_type is not None:
692
693
694
695
            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)}'"
                )
696
            self.OUTPUT_TYPE = self.config.output_type
697

698
699
        if self.config.dataset_path is not None:
            self.DATASET_PATH = self.config.dataset_path
700

701
702
        if self.config.dataset_name is not None:
            self.DATASET_NAME = self.config.dataset_name
703

704
705
706
707
        self._metric_fn_list = {}
        self._metric_fn_kwargs = {}
        self._aggregation_list = {}
        self._higher_is_better = {}
708

709
        if self.config.metric_list is None:
710
            # TODO: handle this in TaskConfig.__post_init__ ?
711
712
            _metric_list = DEFAULT_METRIC_REGISTRY[self.config.output_type]

713
            for metric_name in _metric_list:
haileyschoelkopf's avatar
haileyschoelkopf committed
714
                self._metric_fn_list[metric_name] = get_metric(metric_name)
lintangsutawika's avatar
lintangsutawika committed
715
                self._metric_fn_kwargs[metric_name] = {}
716
717
718
                self._aggregation_list[metric_name] = get_metric_aggregation(
                    metric_name
                )
haileyschoelkopf's avatar
haileyschoelkopf committed
719
                self._higher_is_better[metric_name] = is_higher_better(metric_name)
720
        else:
721
            for metric_config in self.config.metric_list:
722
723
724
725
                if "metric" not in metric_config:
                    raise ValueError(
                        "'metric' key not provided for an entry in 'metric_list', must be specified!"
                    )
726
727
728
729
                metric_name = metric_config["metric"]
                kwargs = {
                    key: metric_config[key]
                    for key in metric_config
Chris's avatar
Chris committed
730
731
                    if key
                    not in ["metric", "aggregation", "higher_is_better", "hf_evaluate"]
732
                }
Chris's avatar
Chris committed
733
734
735
736
                hf_evaluate_metric = (
                    "hf_evaluate" in metric_config
                    and metric_config["hf_evaluate"] is True
                )
737

738
                if self.config.process_results is not None:
739
740
                    self._metric_fn_list[metric_name] = None
                    self._metric_fn_kwargs[metric_name] = {}
741
742
743
744
745
746
                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
747
748
749
                    self._metric_fn_list[metric_name] = get_metric(
                        metric_name, hf_evaluate_metric
                    )
750
                    self._metric_fn_kwargs[metric_name] = kwargs
lintangsutawika's avatar
lintangsutawika committed
751

752
                if "aggregation" in metric_config:
753
                    agg_name = metric_config["aggregation"]
754
                    if isinstance(agg_name, str):
haileyschoelkopf's avatar
haileyschoelkopf committed
755
                        self._aggregation_list[metric_name] = get_aggregation(agg_name)
756
                    elif callable(agg_name):  # noqa: E721
757
758
759
                        self._aggregation_list[metric_name] = metric_config[
                            "aggregation"
                        ]
760
                else:
761
                    INV_AGG_REGISTRY = {v: k for k, v in AGGREGATION_REGISTRY.items()}
lintangsutawika's avatar
lintangsutawika committed
762
                    metric_agg = get_metric_aggregation(metric_name)
763
                    eval_logger.warning(
764
                        f"[Task: {self.config.task}] metric {metric_name} is defined, but aggregation is not. "
765
766
                        f"using default "
                        f"aggregation={INV_AGG_REGISTRY[metric_agg]}"
767
                    )
768
                    self._aggregation_list[metric_name] = metric_agg
lintangsutawika's avatar
lintangsutawika committed
769

770
771
772
773
774
775
                if "higher_is_better" in metric_config:
                    self._higher_is_better[metric_name] = metric_config[
                        "higher_is_better"
                    ]
                else:
                    eval_logger.warning(
776
                        f"[Task: {self.config.task}] metric {metric_name} is defined, but higher_is_better is not. "
777
                        f"using default "
haileyschoelkopf's avatar
haileyschoelkopf committed
778
                        f"higher_is_better={is_higher_better(metric_name)}"
779
                    )
haileyschoelkopf's avatar
haileyschoelkopf committed
780
                    self._higher_is_better[metric_name] = is_higher_better(metric_name)
781

782
        self.download(self.config.dataset_kwargs)
783
784
785
        self._training_docs = None
        self._fewshot_docs = None

786
        if self.config.filter_list is not None:
lintangsutawika's avatar
lintangsutawika committed
787
            self._filters = []
788
            for filter_config in self.config.filter_list:
789
790
791
792
793
794
795
796
797
                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
798
                self._filters.append(filter_pipeline)
lintangsutawika's avatar
lintangsutawika committed
799
        else:
800
            self._filters = [build_filter_ensemble("none", [["take_first", None]])]
801

802
803
        if self.config.use_prompt is not None:
            eval_logger.info(f"loading prompt {self.config.use_prompt}")
804
            self.prompt = get_prompt(
805
                self.config.use_prompt, self.DATASET_PATH, self.DATASET_NAME
lintangsutawika's avatar
lintangsutawika committed
806
            )
807
808
809
        else:
            self.prompt = None

lintangsutawika's avatar
lintangsutawika committed
810
        if self.fewshot_docs() is not None:
haileyschoelkopf's avatar
haileyschoelkopf committed
811
            self.sampler = samplers.get_sampler(
haileyschoelkopf's avatar
haileyschoelkopf committed
812
813
814
                self.config.fewshot_config.get("sampler", "default")
                if self.config.fewshot_config
                else "default"
haileyschoelkopf's avatar
haileyschoelkopf committed
815
            )(list(self.fewshot_docs()), self, rnd=random.Random(1234))
816

817
        self.task_docs = self.eval_docs
818

819
        # Test One Doc
820
        self.features = list(self.task_docs.features.keys())
821
822
        self.multiple_input = 0
        self.multiple_target = 0
823
        test_doc = self.task_docs[0]
824
        test_text = self.doc_to_text(test_doc)
825
        test_target = self.doc_to_target(test_doc)
lintangsutawika's avatar
lintangsutawika committed
826

827
        if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
828
            test_choice = self.doc_to_choice(test_doc)
829
            if not isinstance(test_choice, list):
lintangsutawika's avatar
lintangsutawika committed
830
                eval_logger.error("doc_to_choice must return list")
831
832
            else:
                num_choice = len(test_choice)
833

834
            if isinstance(test_text, int):
835
                self.multiple_input = num_choice
836
837
        else:
            test_choice = None
838

839
        if isinstance(test_target, list):
840
            self.multiple_target = len(test_target)
841
        else:
842
            if (isinstance(test_target, int)) and (test_choice is not None):
lintangsutawika's avatar
lintangsutawika committed
843
                test_target = test_choice[test_target]
844
            else:
lintangsutawika's avatar
lintangsutawika committed
845
                test_target = str(test_target)
846

847
848
849
        if test_choice is not None:
            check_choices = test_choice
        else:
lintangsutawika's avatar
lintangsutawika committed
850
            check_choices = [test_target]
851
852
853
854
        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 = (
855
856
                    True
                    if self.config.target_delimiter.rstrip()
857
                    != self.config.target_delimiter
858
                    else False
859
                )
860

861
                if delimiter_has_whitespace and choice_has_whitespace:
862
863
                    eval_logger.debug(
                        f'Both target_delimiter "{self.config.target_delimiter}" and target choice: "{choice}" have whitespace'
864
865
                    )
                elif (not delimiter_has_whitespace) and (not choice_has_whitespace):
866
                    eval_logger.debug(
867
                        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'
868
869
                    )

870
    def download(self, dataset_kwargs: Optional[Dict[str, Any]] = None) -> None:
871
872
873
874
875
876
        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
877
    def has_training_docs(self) -> bool:
878
        if self.config.training_split is not None:
879
880
881
882
            return True
        else:
            return False

baberabb's avatar
baberabb committed
883
    def has_validation_docs(self) -> bool:
884
        if self.config.validation_split is not None:
885
886
887
888
            return True
        else:
            return False

baberabb's avatar
baberabb committed
889
    def has_test_docs(self) -> bool:
890
        if self.config.test_split is not None:
891
892
893
894
            return True
        else:
            return False

baberabb's avatar
baberabb committed
895
    def training_docs(self) -> datasets.Dataset:
896
        if self.has_training_docs():
897
898
899
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.training_split]
900
                )
901
            return self.dataset[self.config.training_split]
902

baberabb's avatar
baberabb committed
903
    def validation_docs(self) -> datasets.Dataset:
904
        if self.has_validation_docs():
905
906
907
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.validation_split]
908
                )
909
            return self.dataset[self.config.validation_split]
910

baberabb's avatar
baberabb committed
911
    def test_docs(self) -> datasets.Dataset:
912
        if self.has_test_docs():
913
914
915
            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]
916

917
    def fewshot_docs(self):
918
        if self.config.fewshot_split is not None:
919
920
            if self.config.process_docs is not None:
                return self.config.process_docs(self.dataset[self.config.fewshot_split])
921
            return self.dataset[self.config.fewshot_split]
922
        else:
923
            if (self.config.num_fewshot is not None) and (self.config.num_fewshot > 0):
924
                eval_logger.warning(
925
                    f"Task '{self.config.task}': "
926
927
928
929
                    "num_fewshot > 0 but fewshot_split is None. "
                    "using preconfigured rule."
                )
            return super().fewshot_docs()
930

lintangsutawika's avatar
lintangsutawika committed
931
    @utils.positional_deprecated
932
    def fewshot_context(self, doc: str, num_fewshot: int) -> str:
lintangsutawika's avatar
lintangsutawika committed
933
934
935
936
937
938
939
940
941
942
        """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.
        """
943
944
        if description := self.config.description:
            description = utils.apply_template(self.config.description, doc)
lintangsutawika's avatar
lintangsutawika committed
945
946
947

        if num_fewshot == 0:
            # always prepend the (possibly empty) task description
948
            labeled_examples = description
lintangsutawika's avatar
lintangsutawika committed
949
        else:
950
            labeled_examples = description + self.sampler.get_context(doc, num_fewshot)
lintangsutawika's avatar
lintangsutawika committed
951
952

        example = self.doc_to_text(doc)
953
954
955
956
957
958
959
960
961
962
963
964
965
        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
966

967
    def apply_filters(self):
Baber Abbasi's avatar
Baber Abbasi committed
968
        """Iterates over FilterEnsembles and applies them to instances"""
969
970
        if hasattr(self, "_filters"):
            for f in self._filters:
971
                f.apply(self._instances)
972
973
974
975
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances

976
    def should_decontaminate(self):
977
        return self.config.should_decontaminate
978
979

    def doc_to_decontamination_query(self, doc):
980
        if self.config.should_decontaminate:
981
982
            if self.config.doc_to_decontamination_query is None:
                return self.doc_to_text(doc)
983
            else:
984
985
986
987
988
989
990
991
992
993
994
                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
                        )
                    )
995

996
    def _process_doc(self, doc: dict) -> dict:
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
        """
        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):
1008
1009
        if self.prompt is not None:
            doc_to_text = self.prompt
1010
        else:
1011
            doc_to_text = self.config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
1012

1013
        if isinstance(doc_to_text, int):
1014
            return doc_to_text
1015
        elif isinstance(doc_to_text, str):
1016
            if doc_to_text in self.features:
1017
                # if self.config.doc_to_choice is not None:
1018
1019
                #     return self.doc_to_choice(doc)[doc[doc_to_text]]
                # else:
1020
1021
                return doc[doc_to_text]
            else:
lintangsutawika's avatar
lintangsutawika committed
1022
                text_string = utils.apply_template(doc_to_text, doc)
lintangsutawika's avatar
lintangsutawika committed
1023
                if text_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1024
1025
1026
                    return ast.literal_eval(text_string)
                else:
                    return text_string
1027
        elif callable(doc_to_text):
1028
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
1029
        # Used when applying a Promptsource template
1030
        elif hasattr(doc_to_text, "apply"):
1031
1032
1033
1034
1035
            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")
1036
                return self.config.fewshot_delimiter
1037
        else:
1038
            print(type(doc_to_text))
1039
            raise TypeError
1040

1041
    def doc_to_target(self, doc: Mapping) -> Union[int, str, list]:
1042
1043
        if self.prompt is not None:
            doc_to_target = self.prompt
1044
        else:
1045
            doc_to_target = self.config.doc_to_target
1046

1047
        if isinstance(doc_to_target, int):
1048
            return doc_to_target
1049
        elif isinstance(doc_to_target, str):
1050
            if doc_to_target in self.features:
1051
                # if self.config.doc_to_choice is not None:
1052
1053
1054
                #     return self.doc_to_choice(doc)[doc[doc_to_target]]
                # else:
                return doc[doc_to_target]
1055
            else:
lintangsutawika's avatar
lintangsutawika committed
1056
                target_string = utils.apply_template(doc_to_target, doc)
lintangsutawika's avatar
lintangsutawika committed
1057
                if target_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1058
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
1059
1060
1061
1062
1063
                elif (
                    len(target_string) >= 2
                    and (target_string[0] == "[")
                    and (target_string[-1] == "]")
                ):
baberabb's avatar
baberabb committed
1064
1065
1066
1067
                    try:
                        return ast.literal_eval(target_string)
                    except (SyntaxError, ValueError):
                        return target_string
lintangsutawika's avatar
lintangsutawika committed
1068
1069
                else:
                    return target_string
1070
        elif isinstance(doc_to_target, list):
1071
            return doc_to_target
1072
        elif callable(doc_to_target):
1073
            return doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
1074
        # Used when applying a Promptsource template
1075
        elif hasattr(doc_to_target, "apply"):
1076
            applied_prompt = doc_to_target.apply(doc)
1077
1078
1079
1080
            if len(applied_prompt) == 2:
                return applied_prompt[1]
            else:
                eval_logger.warning("Applied prompt returns empty string")
1081
                return self.config.fewshot_delimiter
1082
1083
        else:
            raise TypeError
1084

baberabb's avatar
baberabb committed
1085
    def doc_to_choice(self, doc: Any) -> List[str]:
1086
1087
        if self.prompt is not None:
            doc_to_choice = self.prompt
1088
        elif self.config.doc_to_choice is None:
1089
1090
            eval_logger.error("doc_to_choice was called but not set in config")
        else:
1091
            doc_to_choice = self.config.doc_to_choice
1092

1093
        if isinstance(doc_to_choice, str):
1094
1095
1096
1097
            if doc_to_choice in self.features:
                return doc[doc_to_choice]
            else:
                return ast.literal_eval(utils.apply_template(doc_to_choice, doc))
1098
        elif isinstance(doc_to_choice, list):
1099
            return doc_to_choice
1100
        elif isinstance(doc_to_choice, dict):
1101
1102
1103
1104
1105
1106
1107
            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
1108

baberabb's avatar
baberabb committed
1109
1110
1111
    def construct_requests(
        self, doc: dict, ctx: str, **kwargs
    ) -> Union[List[Instance], Instance]:
1112
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
1113
            arguments = (ctx, self.doc_to_target(doc))
1114
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
1115
            arguments = (self.doc_to_target(doc),)
1116
        elif self.OUTPUT_TYPE == "multiple_choice":
1117
            choices = self.doc_to_choice(doc)
1118
            target_delimiter = self.config.target_delimiter
1119
1120
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
1121
                cont = self.doc_to_target(doc)
1122
1123
1124
                arguments = [
                    (ctx + choice, f"{target_delimiter}{cont}") for choice in choices
                ]
1125
            else:
1126
                # Otherwise they are placed in the continuation
1127
                arguments = [(ctx, f"{target_delimiter}{cont}") for cont in choices]
1128

1129
            request_list = [
1130
1131
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1132
                    doc=doc,
1133
                    arguments=arg,
1134
                    idx=i,
1135
1136
                    **kwargs,
                )
1137
                for i, arg in enumerate(arguments)
1138
            ]
1139
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
1140
            if "acc_mutual_info" in self._metric_fn_list.keys():
1141
1142
1143
                # if we are calculating multiple choice accuracy
                # using mutual information instead of raw loglikelihood as metric, need unconditional lls.

lintangsutawika's avatar
lintangsutawika committed
1144
                # here mutual info refers to calculating
1145
1146
1147
1148
1149
1150
                # 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
1151
                            doc=doc,
1152
                            arguments=("", "{}".format(choice)),
1153
1154
1155
                            idx=i,
                            **kwargs,
                        )
lintangsutawika's avatar
lintangsutawika committed
1156
                        for i, choice in enumerate(choices)
1157
1158
1159
                    ]
                )
            return request_list
lintangsutawika's avatar
lintangsutawika committed
1160

1161
        elif self.OUTPUT_TYPE == "generate_until":
1162
            arguments = (ctx, deepcopy(self.config.generation_kwargs))
lintangsutawika's avatar
lintangsutawika committed
1163
1164

        return Instance(
lintangsutawika's avatar
lintangsutawika committed
1165
1166
            request_type=self.OUTPUT_TYPE, doc=doc, arguments=arguments, idx=0, **kwargs
        )
1167
1168

    def process_results(self, doc, results):
1169
1170
        if callable(self.config.process_results):
            return self.config.process_results(doc, results)
lintangsutawika's avatar
lintangsutawika committed
1171

1172
        result_dict = {}
1173
        use_metric = list(self._metric_fn_list.keys())
1174
1175
1176
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
1177
1178
1179
1180
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
1181
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
1182
            (loglikelihood,) = results
1183
1184
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
haileyschoelkopf's avatar
haileyschoelkopf committed
1185
            return {
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
                **(
                    {"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
1201
            }
1202
        elif self.OUTPUT_TYPE == "multiple_choice":
1203
            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
1204

1205
            # retrieve choices in List[str] form, to compute choice lengths, etc.
1206
            choices = self.doc_to_choice(doc)
1207
1208
            completion_len = np.array([float(len(i)) for i in choices])

1209
1210
            if (
                2 * len(choices) == len(lls)
1211
                and "acc_mutual_info" in self._metric_fn_list.keys()
1212
1213
1214
1215
            ):
                # then we are doing mutual info.
                # this stores the "dryrun" / unconditional answer loglikelihoods
                lls_unconditional = lls[1::2]
1216
1217
                if len(lls_unconditional) != len(choices):
                    raise ValueError
1218
1219
                # and this stores our "regular" conditional loglikelihoods
                lls = lls[::2]
1220

1221
1222
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
1223

1224
1225
            if self.multiple_input:
                gold = self.doc_to_text(doc)
1226
            else:
1227
                gold = self.doc_to_target(doc)
1228
1229

            gold_index_error = False
1230
            if isinstance(gold, list):
Lintang Sutawika's avatar
Lintang Sutawika committed
1231
1232
                gold = [i if i < len(choices) else -100 for i in gold]
                if -100 in gold:
1233
1234
                    gold_index_error = True
            else:
1235
                if isinstance(gold, int):
Lintang Sutawika's avatar
Lintang Sutawika committed
1236
                    gold = gold if gold < len(choices) else -100
1237
                elif isinstance(gold, str):
Lintang Sutawika's avatar
Lintang Sutawika committed
1238
                    gold = choices.index(gold) if gold in choices else -100
lintangsutawika's avatar
lintangsutawika committed
1239

Lintang Sutawika's avatar
Lintang Sutawika committed
1240
                if gold == -100:
1241
1242
1243
1244
                    gold_index_error = True

            if gold_index_error:
                eval_logger.warning(
lintangsutawika's avatar
lintangsutawika committed
1245
                    f"Label index was not in within range of available choices,"
1246
1247
                    f"Sample:\n\n{doc}\n\n"
                )
lintangsutawika's avatar
lintangsutawika committed
1248

1249
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
1250
1251
                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
1252
                exact_match = int(any([is_greedy[i] if i != -100 else 0 for i in gold]))
lintangsutawika's avatar
lintangsutawika committed
1253
1254
1255
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
1256
                # 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
1257
                exact_match = int(is_greedy[gold]) if gold != -100 else 0
1258

Lintang Sutawika's avatar
Lintang Sutawika committed
1259
1260
1261
1262
            prob_norm = utils.softmax(lls)

            # TODO use keyword arguments to the metric?
            # gold, pred, norm stuff, the original lls,
1263
            result_dict = {
1264
                **({"acc": acc} if "acc" in use_metric else {}),
1265
1266
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
1267
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
1268
                **({"exact_match": exact_match} if "exact_match" in use_metric else {}),
Lintang Sutawika's avatar
Lintang Sutawika committed
1269
1270
1271
1272
1273
                **(
                    {"brier_score": (gold, prob_norm)}
                    if "brier_score" in use_metric
                    else {}
                ),
1274
1275
            }

1276
            if "acc_mutual_info" in use_metric:
lintangsutawika's avatar
lintangsutawika committed
1277
1278
1279
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
1280
1281
1282
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

1283
        elif self.OUTPUT_TYPE == "generate_until":
1284
            gold = self.doc_to_target(doc)
Chris's avatar
Chris committed
1285
            result = results[0]
1286
            if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1287
                # If you set doc_to_choice,
lintangsutawika's avatar
lintangsutawika committed
1288
                # it assumes that doc_to_target returns a number.
1289
1290
                choices = self.doc_to_choice(doc)
                gold = choices[gold]
1291
1292
            # we expect multiple_targets to be a list.
            elif self.multiple_target:
baberabb's avatar
baberabb committed
1293
                gold = list(gold)
Chris's avatar
Chris committed
1294
1295
1296
            elif type(gold) != type(result):
                # cast gold to the same type as result
                gold = type(result)(gold)
1297

lintangsutawika's avatar
lintangsutawika committed
1298
            for metric in self._metric_fn_list.keys():
haileyschoelkopf's avatar
haileyschoelkopf committed
1299
1300
1301
1302
1303
                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
1304
1305
1306
1307
                    if not isinstance(gold, list):
                        # sometimes, a multiple_target dataset has exceptions where one doc has only one string answer
                        # print(gold)
                        gold = [gold]
1308
1309
1310
1311
1312
1313
1314
1315
                    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
1316
                    else:
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
                        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
1338
                else:
1339
                    try:
1340
                        result_score = self._metric_fn_list[metric](
1341
1342
                            references=[gold],
                            predictions=[result],
1343
                            **self._metric_fn_kwargs[metric],
1344
                        )
1345
                    except TypeError:  # needed for now in order to use a different interface between our own metrics and HF Evaluate metrics
1346
                        result_score = self._metric_fn_list[metric]([gold, result])
1347
1348
1349
1350
                    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
1351
        else:
lintangsutawika's avatar
lintangsutawika committed
1352
1353
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
1354
                "'loglikelihood', 'loglikelihood_rolling', 'generate_until' or 'multiple_choice'",
1355
            )
1356
1357
1358

        return result_dict

Baber Abbasi's avatar
Baber Abbasi committed
1359
    def aggregation(self) -> dict:
1360
1361
        return self._aggregation_list

Baber Abbasi's avatar
Baber Abbasi committed
1362
    def higher_is_better(self) -> dict:
haileyschoelkopf's avatar
haileyschoelkopf committed
1363
        return self._higher_is_better
1364

Baber Abbasi's avatar
Baber Abbasi committed
1365
1366
1367
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

1368
1369
1370
1371
1372
1373
1374
1375
1376
    def __repr__(self):
        return (
            f"ConfigurableTask(task_name={getattr(self.config, 'task', None)},"
            f"group_name={getattr(self.config, 'group', None)},"
            f"output_type={self.OUTPUT_TYPE},"
            f"num_fewshot={getattr(self.config, 'num_fewshot', None)},"
            f"num_samples={len(self.eval_docs)})"
        )

1377
1378

class MultipleChoiceTask(Task):
1379
    OUTPUT_TYPE = "loglikelihood"
1380

baberabb's avatar
baberabb committed
1381
    def doc_to_target(self, doc: dict) -> str:
1382
1383
        return " " + doc["choices"][doc["gold"]]

baberabb's avatar
baberabb committed
1384
    def construct_requests(self, doc: dict, ctx: str, **kwargs) -> List[Instance]:
1385
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1386
1387
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1388
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1389
                doc=doc,
1390
                arguments=(ctx, " {}".format(choice)),
1391
                idx=i,
1392
1393
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1394
1395
            for i, choice in enumerate(doc["choices"])
        ]
1396

1397
    def process_results(self, doc: dict, results: Iterable[Tuple[float, bool]]) -> dict:
lintangsutawika's avatar
lintangsutawika committed
1398
1399
1400
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
        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
1412
    def higher_is_better(self) -> dict:
1413
1414
1415
1416
1417
        return {
            "acc": True,
            "acc_norm": True,
        }

baberabb's avatar
baberabb committed
1418
    def aggregation(self) -> dict:
1419
1420
1421
1422
1423
1424
        return {
            "acc": mean,
            "acc_norm": mean,
        }


lintangsutawika's avatar
lintangsutawika committed
1425
class PerplexityTask(Task):
1426
1427
    OUTPUT_TYPE = "loglikelihood_rolling"

baberabb's avatar
baberabb committed
1428
    def has_training_docs(self) -> bool:
1429
1430
        return False

baberabb's avatar
baberabb committed
1431
    def fewshot_examples(self, k: int, rnd) -> List:
1432
1433
1434
1435
        if k != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1436
1437
        return []

baberabb's avatar
baberabb committed
1438
    def fewshot_context(self, doc: dict, num_fewshot: int) -> Literal[""]:
1439
1440
1441
1442
        if num_fewshot != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1443
1444
1445

        return ""

baberabb's avatar
baberabb committed
1446
    def higher_is_better(self) -> dict:
1447
1448
1449
1450
1451
1452
1453
1454
1455
        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
1456
    def doc_to_text(self, doc) -> str:
1457
1458
1459
1460
1461
        return ""

    def doc_to_target(self, doc):
        return doc

1462
1463
1464
    def construct_requests(self, doc: dict, ctx: Optional[str], **kwargs):
        if bool(ctx):
            raise ValueError
1465

lintangsutawika's avatar
lintangsutawika committed
1466
1467
1468
1469
1470
1471
1472
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1473

1474
    def process_results(self, doc: dict, results: Tuple[float]) -> dict:
1475
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1476
1477
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1478
1479
1480
1481
1482
1483
        return {
            "word_perplexity": (loglikelihood, words),
            "byte_perplexity": (loglikelihood, bytes_),
            "bits_per_byte": (loglikelihood, bytes_),
        }

baberabb's avatar
baberabb committed
1484
    def aggregation(self) -> dict:
1485
1486
1487
1488
1489
1490
1491
        return {
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
            "bits_per_byte": bits_per_byte,
        }

    @classmethod
baberabb's avatar
baberabb committed
1492
    def count_bytes(cls, doc) -> int:
1493
1494
1495
        return len(doc.encode("utf-8"))

    @classmethod
baberabb's avatar
baberabb committed
1496
    def count_words(cls, doc) -> int:
1497
1498
        """Downstream tasks with custom word boundaries should override this!"""
        return len(re.split(r"\s+", doc))