task.py 55.7 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
from typing import Any, Iterator, List, Literal, Tuple, Union
11
12
13

import datasets
import numpy as np
14
from tqdm import tqdm
15
16

from lm_eval import utils
17
from lm_eval.api import samplers
haileyschoelkopf's avatar
haileyschoelkopf committed
18
from lm_eval.api.instance import Instance
lintangsutawika's avatar
lintangsutawika committed
19
from lm_eval.api.metrics import (
20
    bits_per_byte,
lintangsutawika's avatar
lintangsutawika committed
21
22
23
24
    mean,
    weighted_perplexity,
)
from lm_eval.api.registry import (
25
26
    AGGREGATION_REGISTRY,
    DEFAULT_METRIC_REGISTRY,
haileyschoelkopf's avatar
haileyschoelkopf committed
27
    get_aggregation,
28
    get_metric,
29
    get_metric_aggregation,
haileyschoelkopf's avatar
haileyschoelkopf committed
30
    is_higher_better,
lintangsutawika's avatar
lintangsutawika committed
31
)
32
from lm_eval.caching.cache import load_from_cache, save_to_cache
33
34
35
from lm_eval.filters import build_filter_ensemble
from lm_eval.prompts import get_prompt

36

37
38
39
40
ALL_OUTPUT_TYPES = [
    "loglikelihood",
    "multiple_choice",
    "loglikelihood_rolling",
41
    "generate_until",
42
43
]

44
eval_logger = logging.getLogger("lm-eval")
45

lintangsutawika's avatar
lintangsutawika committed
46

lintangsutawika's avatar
lintangsutawika committed
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
@dataclass
class GroupConfig(dict):
    group: str = None
    task: Union[str, list] = None
    weight_by_size: bool = False

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

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

    def to_dict(self):
        return asdict(self)

lintangsutawika's avatar
lintangsutawika committed
62

63
64
@dataclass
class TaskConfig(dict):
65
    # task naming/registry
66
    task: str = None
lintangsutawika's avatar
lintangsutawika committed
67
    task_alias: str = None
68
    group: Union[str, list] = None
lintangsutawika's avatar
lintangsutawika committed
69
    group_alias: Union[str, list] = None
70
71
72
    # HF dataset options.
    # which dataset to use,
    # and what splits for what purpose
73
74
    dataset_path: str = None
    dataset_name: str = None
75
    dataset_kwargs: dict = None
76
77
78
    training_split: str = None
    validation_split: str = None
    test_split: str = None
lintangsutawika's avatar
lintangsutawika committed
79
    fewshot_split: str = None  # TODO: assert that this not None if num_fewshot > 0. (?) assert if this is same split as one evaling (?)
80
81
    # formatting / prompting options.
    # see docs/advanced_task_guide.md for more info
82
    process_docs: Callable = None
83
84
    doc_to_text: Union[Callable, str] = None
    doc_to_target: Union[Callable, str] = None
lintangsutawika's avatar
lintangsutawika committed
85
    doc_to_choice: Union[Callable, str, dict, list] = None
lintangsutawika's avatar
lintangsutawika committed
86
    process_results: Union[Callable, str] = None
87
    use_prompt: str = None
88
    description: str = ""
89
90
    target_delimiter: str = " "
    fewshot_delimiter: str = "\n\n"
haileyschoelkopf's avatar
haileyschoelkopf committed
91
    fewshot_config: dict = None
92
    # runtime configuration options
93
    num_fewshot: int = None
94
    # scoring options
95
    metric_list: list = None
96
97
98
99
100
101
    output_type: Literal[
        "loglikelihood",
        "loglikelihood_rolling",
        "generate_until",
        "multiple_choice",
    ] = "generate_until"
102
    generation_kwargs: dict = None
103
    repeats: int = 1
lintangsutawika's avatar
lintangsutawika committed
104
    filter_list: Union[str, list] = None
105
106
    should_decontaminate: bool = False
    doc_to_decontamination_query: str = None
107
    metadata: dict = None  # by default, not used in the code. allows for users to pass arbitrary info to tasks
108

Ethan Smith's avatar
Ethan Smith committed
109
    def __post_init__(self) -> None:
Lintang Sutawika's avatar
Lintang Sutawika committed
110
        if self.generation_kwargs is not None:
111
            if self.output_type != "generate_until":
Lintang Sutawika's avatar
Lintang Sutawika committed
112
                eval_logger.warning(
113
                    f"[{self.task}] passed `generation_kwargs`, but not using `output_type: generate_until`!"
Lintang Sutawika's avatar
Lintang Sutawika committed
114
                )
115
                assert self.output_type != "generate_until"
Lintang Sutawika's avatar
Lintang Sutawika committed
116
117
118
119
120
121
122

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

            if "until" not in self.generation_kwargs:
123
                self.generation_kwargs["until"] = [self.fewshot_delimiter]
Lintang Sutawika's avatar
Lintang Sutawika committed
124
        else:
125
            if self.output_type == "generate_until":
Lintang Sutawika's avatar
Lintang Sutawika committed
126
127
                # ensure that we greedily generate in absence of explicit arguments otherwise
                self.generation_kwargs = {
128
129
130
131
132
                    "until": (
                        None
                        if self.fewshot_delimiter is None
                        else [self.fewshot_delimiter]
                    ),
Lintang Sutawika's avatar
Lintang Sutawika committed
133
134
                    "do_sample": False,
                }
135

136
137
138
    def __getitem__(self, item):
        return getattr(self, item)

139
140
141
    def __setitem__(self, item, value):
        return setattr(self, item, value)

142
    def to_dict(self, keep_callable: bool = False) -> dict:
143
144
        """dumps the current config as a dictionary object, as a printable format.
        null fields will not be printed.
haileyschoelkopf's avatar
haileyschoelkopf committed
145
        Used for dumping results alongside full task configuration
146

haileyschoelkopf's avatar
haileyschoelkopf committed
147
148
149
150
151
152
153
154
155
156
        :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)
157
158
159
160
161
162
163
164
165
166
            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
167
        return cfg_dict
168

169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
    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)

185
186
187
188
189
190
191
192
193
194
195
196

class Task(abc.ABC):
    """A task represents an entire benchmark including its dataset, problems,
    answers, and evaluation methods. See BoolQ for a simple example implementation

    A `doc` can be any python object which represents one instance of evaluation.
    This is usually a dictionary e.g.
        {"question": ..., "answer": ...} or
        {"question": ..., question, answer)
    """

    VERSION = None
197

198
199
200
201
202
203
204
205
    # The name of the `Task` benchmark as denoted in the HuggingFace datasets Hub
    # or a path to a custom `datasets` loading script.
    DATASET_PATH: str = None

    # The name of a subset within `DATASET_PATH`.
    DATASET_NAME: str = None

    OUTPUT_TYPE: str = None
lintangsutawika's avatar
lintangsutawika committed
206

207
208
209
210
211
212
    def __init__(
        self,
        data_dir=None,
        cache_dir=None,
        download_mode=None,
        config=None,
Ethan Smith's avatar
Ethan Smith committed
213
    ) -> None:
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
        """
        :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)
        self._training_docs = None
        self._fewshot_docs = None
        self._instances = None

lintangsutawika's avatar
lintangsutawika committed
240
        self._config = TaskConfig({**config}) if config else TaskConfig()
241

lintangsutawika's avatar
lintangsutawika committed
242
        self._filters = [build_filter_ensemble("none", [["take_first", None]])]
243

Ethan Smith's avatar
Ethan Smith committed
244
    def download(self, data_dir=None, cache_dir=None, download_mode=None) -> None:
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
        """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.
        """
269
270
271
272
273
274
275
        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,
        )
276

277
278
279
280
281
    @property
    def config(self):
        """Returns the TaskConfig associated with this class."""
        return self._config

282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
    @abc.abstractmethod
    def has_training_docs(self):
        """Whether the task has a training set"""
        pass

    @abc.abstractmethod
    def has_validation_docs(self):
        """Whether the task has a validation set"""
        pass

    @abc.abstractmethod
    def has_test_docs(self):
        """Whether the task has a test set"""
        pass

    def training_docs(self):
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

    def validation_docs(self):
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

    def test_docs(self):
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

318
319
320
321
322
323
324
325
326
327
    def fewshot_docs(self):
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        if self.has_training_docs():
            return self.training_docs()
        elif self.has_validation_docs():
            return self.validation_docs()
        else:
lintangsutawika's avatar
lintangsutawika committed
328
            eval_logger.warning(
329
                f"[Task: {self.config.task}] has_training_docs and has_validation_docs are False"
330
                ", using test_docs as fewshot_docs but this is not recommended."
lintangsutawika's avatar
lintangsutawika committed
331
            )
332
333
            return self.test_docs()

334
335
336
337
338
339
340
341
342
343
    def _process_doc(self, doc):
        """
        Override this to process (detokenize, strip, replace, etc.) individual
        documents. This can be used in a map over documents of a data split.
        E.g. `map(self._process_doc, self.dataset["validation"])`

        :return: dict
            The processed version of the specified `doc`.
        """
        return doc
lintangsutawika's avatar
lintangsutawika committed
344

345
    @property
346
    def instances(self) -> List[Instance]:
347
348
349
350
351
352
353
354
355
356
357
        """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)

Ethan Smith's avatar
Ethan Smith committed
358
    def doc_to_decontamination_query(self, doc) -> None:
359
360
361
362
363
364
365
366
367
368
369
370
371
        print(
            "Override doc_to_decontamination_query with document specific decontamination query."
        )
        assert False

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

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

372
373
    def build_all_requests(
        self,
374
        *,
375
376
377
378
379
380
        limit=None,
        rank=None,
        world_size=None,
        cache_requests=False,
        rewrite_requests_cache=False,
    ) -> None:
381
        """Build a set of Instances for a task, and store them in task.instances"""
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400

        # 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
401

Baber Abbasi's avatar
Baber Abbasi committed
402
        eval_logger.info(f"Building contexts for {self.config.task} on rank {rank}...")
403

404
        instances = []
405
406
407
408
409
410
411
412
413
414

        # 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(
415
            self.doc_iterator(rank=rank, limit=limit, world_size=world_size)
416
417
418
419
420
421
422
        )

        num_docs = len(doc_id_docs)

        for doc_id, doc in tqdm(
            doc_id_docs,
            total=num_docs,
lintangsutawika's avatar
lintangsutawika committed
423
        ):
424
            # sample fewshot context #TODO: need to offset doc_id by rank now!
425
            fewshot_ctx = self.fewshot_context(
426
                doc,
427
                0 if self.config.num_fewshot is None else self.config.num_fewshot,
428
            )
429

430
            # TODO: we should override self.config.repeats if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
431
432
433
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
434
                metadata=(self.config["task"], doc_id, self.config.repeats),
lintangsutawika's avatar
lintangsutawika committed
435
            )
436
437
438
439

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

440
441
442
443
444
445
446
447
448
449
450
451
452
            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
453
454
455

        assert len(self._instances) != 0, "task.build_requests() did not find any docs!"

456
457
458
        if cache_requests and (not cached_instances or rewrite_requests_cache):
            save_to_cache(file_name=cache_key, obj=instances)

459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
    @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
475
            The number of times each instance in a dataset is inferred on. Defaults to 1,
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
505
506
507
508
509
510
            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

511
512
513
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

haileyschoelkopf's avatar
haileyschoelkopf committed
514
515
516
517
518
519
520
521
522
523
    @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))

524
    @utils.positional_deprecated
lintangsutawika's avatar
lintangsutawika committed
525
    def fewshot_context(
526
527
528
529
530
        self,
        doc,
        num_fewshot,
        rnd=random.Random(1234),
        description=None,
lintangsutawika's avatar
lintangsutawika committed
531
    ):
532
533
534
535
536
537
538
        """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
539
540
541
542
543
        :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.
544
545
546
        :returns: str
            The fewshot context.
        """
lintangsutawika's avatar
lintangsutawika committed
547
548
549
550
        assert (
            rnd is not None
        ), "A `random.Random` generator argument must be provided to `rnd`"

551
        description = description if description else ""
552
553

        if num_fewshot == 0:
lintangsutawika's avatar
lintangsutawika committed
554
            labeled_examples = ""
555
        else:
lintangsutawika's avatar
lintangsutawika committed
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
            # 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
580
            )
581
582

        example = self.doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
583
        return description + labeled_examples + example
584
585

    def apply_filters(self):
Baber Abbasi's avatar
Baber Abbasi committed
586
        """Iterates over FilterEnsembles and applies them to instances"""
lintangsutawika's avatar
lintangsutawika committed
587
588
        if hasattr(self, "_filters"):
            for f in self._filters:
589
                f.apply(self._instances)
lintangsutawika's avatar
lintangsutawika committed
590
591
592
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances
593

baberabb's avatar
baberabb committed
594
    def dump_config(self) -> dict:
Baber Abbasi's avatar
Baber Abbasi committed
595
        """Returns the config as a dictionary."""
596
        # TODO: this should only return the overrides applied to a non-YAML task's configuration.
597
        # (num_fewshot)
598
        return self.config.to_dict()
599

Baber Abbasi's avatar
Baber Abbasi committed
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
635
636
637
638
639
    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)

640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
    @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:
            assert False, f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"

    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

661
662

class ConfigurableTask(Task):
663
    VERSION = "Yaml"
664
    OUTPUT_TYPE = None
665
    CONFIG = None
666
667
668

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

673
        # Use new configurations if there was no preconfiguration
674
        if self.config is None:
675
            self._config = TaskConfig(**config)
676
677
        # Overwrite configs
        else:
lintangsutawika's avatar
lintangsutawika committed
678
            if config is not None:
679
                self._config.__dict__.update(config)
680

681
        if self.config is None:
lintangsutawika's avatar
lintangsutawika committed
682
683
684
            raise ValueError(
                "Must pass a config to ConfigurableTask, either in cls.CONFIG or `config` kwarg"
            )
685

686
687
688
689
        if isinstance(self.config.metadata, dict):
            if "version" in self.config.metadata:
                self.VERSION = self.config.metadata["version"]

690
691
692
        if self.config.output_type is not None:
            assert self.config.output_type in ALL_OUTPUT_TYPES
            self.OUTPUT_TYPE = self.config.output_type
693

694
695
        if self.config.dataset_path is not None:
            self.DATASET_PATH = self.config.dataset_path
696

697
698
        if self.config.dataset_name is not None:
            self.DATASET_NAME = self.config.dataset_name
699

700
701
702
703
        self._metric_fn_list = {}
        self._metric_fn_kwargs = {}
        self._aggregation_list = {}
        self._higher_is_better = {}
704

705
        if self.config.metric_list is None:
706
            # TODO: handle this in TaskConfig.__post_init__ ?
707
708
            _metric_list = DEFAULT_METRIC_REGISTRY[self.config.output_type]

709
            for metric_name in _metric_list:
haileyschoelkopf's avatar
haileyschoelkopf committed
710
                self._metric_fn_list[metric_name] = get_metric(metric_name)
lintangsutawika's avatar
lintangsutawika committed
711
                self._metric_fn_kwargs[metric_name] = {}
712
713
714
                self._aggregation_list[metric_name] = get_metric_aggregation(
                    metric_name
                )
haileyschoelkopf's avatar
haileyschoelkopf committed
715
                self._higher_is_better[metric_name] = is_higher_better(metric_name)
716
        else:
717
            for metric_config in self.config.metric_list:
718
719
720
721
722
                assert "metric" in metric_config
                metric_name = metric_config["metric"]
                kwargs = {
                    key: metric_config[key]
                    for key in metric_config
Chris's avatar
Chris committed
723
724
                    if key
                    not in ["metric", "aggregation", "higher_is_better", "hf_evaluate"]
725
                }
Chris's avatar
Chris committed
726
727
728
729
                hf_evaluate_metric = (
                    "hf_evaluate" in metric_config
                    and metric_config["hf_evaluate"] is True
                )
730

731
                if self.config.process_results is not None:
732
733
                    self._metric_fn_list[metric_name] = None
                    self._metric_fn_kwargs[metric_name] = {}
734
735
736
737
738
739
                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
740
741
742
                    self._metric_fn_list[metric_name] = get_metric(
                        metric_name, hf_evaluate_metric
                    )
743
                    self._metric_fn_kwargs[metric_name] = kwargs
lintangsutawika's avatar
lintangsutawika committed
744

745
                if "aggregation" in metric_config:
746
                    agg_name = metric_config["aggregation"]
747
                    if isinstance(agg_name, str):
haileyschoelkopf's avatar
haileyschoelkopf committed
748
                        self._aggregation_list[metric_name] = get_aggregation(agg_name)
749
                    elif callable(agg_name):  # noqa: E721
750
751
752
                        self._aggregation_list[metric_name] = metric_config[
                            "aggregation"
                        ]
753
                else:
754
                    INV_AGG_REGISTRY = {v: k for k, v in AGGREGATION_REGISTRY.items()}
lintangsutawika's avatar
lintangsutawika committed
755
                    metric_agg = get_metric_aggregation(metric_name)
756
                    eval_logger.warning(
757
                        f"[Task: {self.config.task}] metric {metric_name} is defined, but aggregation is not. "
758
759
                        f"using default "
                        f"aggregation={INV_AGG_REGISTRY[metric_agg]}"
760
                    )
761
                    self._aggregation_list[metric_name] = metric_agg
lintangsutawika's avatar
lintangsutawika committed
762

763
764
765
766
767
768
                if "higher_is_better" in metric_config:
                    self._higher_is_better[metric_name] = metric_config[
                        "higher_is_better"
                    ]
                else:
                    eval_logger.warning(
769
                        f"[Task: {self.config.task}] metric {metric_name} is defined, but higher_is_better is not. "
770
                        f"using default "
haileyschoelkopf's avatar
haileyschoelkopf committed
771
                        f"higher_is_better={is_higher_better(metric_name)}"
772
                    )
haileyschoelkopf's avatar
haileyschoelkopf committed
773
                    self._higher_is_better[metric_name] = is_higher_better(metric_name)
774

775
        self.download(self.config.dataset_kwargs)
776
777
778
        self._training_docs = None
        self._fewshot_docs = None

779
        if self.config.filter_list is not None:
lintangsutawika's avatar
lintangsutawika committed
780
            self._filters = []
781
            for filter_config in self.config.filter_list:
782
783
784
785
786
787
788
789
790
                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
791
                self._filters.append(filter_pipeline)
lintangsutawika's avatar
lintangsutawika committed
792
        else:
793
            self._filters = [build_filter_ensemble("none", [["take_first", None]])]
794

795
796
        if self.config.use_prompt is not None:
            eval_logger.info(f"loading prompt {self.config.use_prompt}")
797
            self.prompt = get_prompt(
798
                self.config.use_prompt, self.DATASET_PATH, self.DATASET_NAME
lintangsutawika's avatar
lintangsutawika committed
799
            )
800
801
802
        else:
            self.prompt = None

lintangsutawika's avatar
lintangsutawika committed
803
        if self.fewshot_docs() is not None:
haileyschoelkopf's avatar
haileyschoelkopf committed
804
            self.sampler = samplers.get_sampler(
haileyschoelkopf's avatar
haileyschoelkopf committed
805
806
807
                self.config.fewshot_config.get("sampler", "default")
                if self.config.fewshot_config
                else "default"
haileyschoelkopf's avatar
haileyschoelkopf committed
808
            )(list(self.fewshot_docs()), self, rnd=random.Random(1234))
809

810
        self.task_docs = self.eval_docs
811

812
        # Test One Doc
813
        self.features = list(self.task_docs.features.keys())
814
815
        self.multiple_input = 0
        self.multiple_target = 0
816
        test_doc = self.task_docs[0]
817
        test_text = self.doc_to_text(test_doc)
818
        test_target = self.doc_to_target(test_doc)
lintangsutawika's avatar
lintangsutawika committed
819

820
        if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
821
            test_choice = self.doc_to_choice(test_doc)
822
            if not isinstance(test_choice, list):
lintangsutawika's avatar
lintangsutawika committed
823
                eval_logger.error("doc_to_choice must return list")
824
825
            else:
                num_choice = len(test_choice)
826

827
            if isinstance(test_text, int):
828
                self.multiple_input = num_choice
829
830
        else:
            test_choice = None
831

832
        if isinstance(test_target, list):
833
            self.multiple_target = len(test_target)
834
        else:
835
            if (isinstance(test_target, int)) and (test_choice is not None):
lintangsutawika's avatar
lintangsutawika committed
836
                test_target = test_choice[test_target]
837
            else:
lintangsutawika's avatar
lintangsutawika committed
838
                test_target = str(test_target)
839

840
841
842
        if test_choice is not None:
            check_choices = test_choice
        else:
lintangsutawika's avatar
lintangsutawika committed
843
            check_choices = [test_target]
844
845
846
847
        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 = (
848
849
                    True
                    if self.config.target_delimiter.rstrip()
850
                    != self.config.target_delimiter
851
                    else False
852
                )
853

854
                if delimiter_has_whitespace and choice_has_whitespace:
855
856
                    eval_logger.debug(
                        f'Both target_delimiter "{self.config.target_delimiter}" and target choice: "{choice}" have whitespace'
857
858
                    )
                elif (not delimiter_has_whitespace) and (not choice_has_whitespace):
859
                    eval_logger.debug(
860
                        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'
861
862
                    )

Ethan Smith's avatar
Ethan Smith committed
863
    def download(self, dataset_kwargs=None) -> None:
864
865
866
867
868
869
        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
870
    def has_training_docs(self) -> bool:
871
        if self.config.training_split is not None:
872
873
874
875
            return True
        else:
            return False

baberabb's avatar
baberabb committed
876
    def has_validation_docs(self) -> bool:
877
        if self.config.validation_split is not None:
878
879
880
881
            return True
        else:
            return False

baberabb's avatar
baberabb committed
882
    def has_test_docs(self) -> bool:
883
        if self.config.test_split is not None:
884
885
886
887
            return True
        else:
            return False

baberabb's avatar
baberabb committed
888
    def training_docs(self) -> datasets.Dataset:
889
        if self.has_training_docs():
890
891
892
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.training_split]
893
                )
894
            return self.dataset[self.config.training_split]
895

baberabb's avatar
baberabb committed
896
    def validation_docs(self) -> datasets.Dataset:
897
        if self.has_validation_docs():
898
899
900
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.validation_split]
901
                )
902
            return self.dataset[self.config.validation_split]
903

baberabb's avatar
baberabb committed
904
    def test_docs(self) -> datasets.Dataset:
905
        if self.has_test_docs():
906
907
908
            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]
909

910
    def fewshot_docs(self):
911
        if self.config.fewshot_split is not None:
912
913
            if self.config.process_docs is not None:
                return self.config.process_docs(self.dataset[self.config.fewshot_split])
914
            return self.dataset[self.config.fewshot_split]
915
        else:
916
            if (self.config.num_fewshot is not None) and (self.config.num_fewshot > 0):
917
                eval_logger.warning(
918
                    f"Task '{self.config.task}': "
919
920
921
922
                    "num_fewshot > 0 but fewshot_split is None. "
                    "using preconfigured rule."
                )
            return super().fewshot_docs()
923

lintangsutawika's avatar
lintangsutawika committed
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
    @utils.positional_deprecated
    def fewshot_context(self, doc, num_fewshot):
        """Returns a fewshot context string that is made up of a prepended description
        (if provided), the `num_fewshot` number of examples, and an appended prompt example.

        :param doc: str
            The document as returned from training_docs, validation_docs, or test_docs.
        :param num_fewshot: int
            The number of fewshot examples to provide in the returned context string.
        :returns: str
            The fewshot context.
        """

        if num_fewshot == 0:
            # always prepend the (possibly empty) task description
            labeled_examples = self.config.description
        else:
            labeled_examples = self.config.description + self.sampler.get_context(
                doc, num_fewshot
            )

        example = self.doc_to_text(doc)
946
947
948
949
950
951
952
953
954
955
956
957
958
        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
959

960
    def apply_filters(self):
Baber Abbasi's avatar
Baber Abbasi committed
961
        """Iterates over FilterEnsembles and applies them to instances"""
962
963
        if hasattr(self, "_filters"):
            for f in self._filters:
964
                f.apply(self._instances)
965
966
967
968
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances

969
    def should_decontaminate(self):
970
        return self.config.should_decontaminate
971
972

    def doc_to_decontamination_query(self, doc):
973
        if self.config.should_decontaminate:
974
975
            if self.config.doc_to_decontamination_query is None:
                return self.doc_to_text(doc)
976
            else:
977
978
979
980
981
982
983
984
985
986
987
                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
                        )
                    )
988

989
990
991
992
993
994
995
996
997
998
999
1000
    def _process_doc(self, doc):
        """
        Override this to process (detokenize, strip, replace, etc.) individual
        documents. This can be used in a map over documents of a data split.
        E.g. `map(self._process_doc, self.dataset["validation"])`

        :return: dict
            The processed version of the specified `doc`.
        """
        return doc

    def doc_to_text(self, doc):
1001
1002
        if self.prompt is not None:
            doc_to_text = self.prompt
1003
        else:
1004
            doc_to_text = self.config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
1005

1006
        if isinstance(doc_to_text, int):
1007
            return doc_to_text
1008
        elif isinstance(doc_to_text, str):
1009
            if doc_to_text in self.features:
1010
                # if self.config.doc_to_choice is not None:
1011
1012
                #     return self.doc_to_choice(doc)[doc[doc_to_text]]
                # else:
1013
1014
                return doc[doc_to_text]
            else:
lintangsutawika's avatar
lintangsutawika committed
1015
                text_string = utils.apply_template(doc_to_text, doc)
lintangsutawika's avatar
lintangsutawika committed
1016
                if text_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1017
1018
1019
                    return ast.literal_eval(text_string)
                else:
                    return text_string
1020
        elif callable(doc_to_text):
1021
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
1022
        # Used when applying a Promptsource template
1023
        elif hasattr(doc_to_text, "apply"):
1024
1025
1026
1027
1028
            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")
1029
                return self.config.fewshot_delimiter
1030
        else:
1031
            print(type(doc_to_text))
1032
            raise TypeError
1033

1034
    def doc_to_target(self, doc: dict) -> Union[int, str, list]:
1035
1036
        if self.prompt is not None:
            doc_to_target = self.prompt
1037
        else:
1038
            doc_to_target = self.config.doc_to_target
1039

1040
        if isinstance(doc_to_target, int):
1041
            return doc_to_target
1042
        elif isinstance(doc_to_target, str):
1043
            if doc_to_target in self.features:
1044
                # if self.config.doc_to_choice is not None:
1045
1046
1047
                #     return self.doc_to_choice(doc)[doc[doc_to_target]]
                # else:
                return doc[doc_to_target]
1048
            else:
lintangsutawika's avatar
lintangsutawika committed
1049
                target_string = utils.apply_template(doc_to_target, doc)
lintangsutawika's avatar
lintangsutawika committed
1050
                if target_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1051
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
1052
1053
1054
1055
1056
                elif (
                    len(target_string) >= 2
                    and (target_string[0] == "[")
                    and (target_string[-1] == "]")
                ):
baberabb's avatar
baberabb committed
1057
1058
1059
1060
                    try:
                        return ast.literal_eval(target_string)
                    except (SyntaxError, ValueError):
                        return target_string
lintangsutawika's avatar
lintangsutawika committed
1061
1062
                else:
                    return target_string
1063
        elif isinstance(doc_to_target, list):
1064
            return doc_to_target
1065
        elif callable(doc_to_target):
1066
            return doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
1067
        # Used when applying a Promptsource template
1068
        elif hasattr(doc_to_target, "apply"):
1069
            applied_prompt = doc_to_target.apply(doc)
1070
1071
1072
1073
            if len(applied_prompt) == 2:
                return applied_prompt[1]
            else:
                eval_logger.warning("Applied prompt returns empty string")
1074
                return self.config.fewshot_delimiter
1075
1076
        else:
            raise TypeError
1077

baberabb's avatar
baberabb committed
1078
    def doc_to_choice(self, doc: Any) -> List[str]:
1079
1080
        if self.prompt is not None:
            doc_to_choice = self.prompt
1081
        elif self.config.doc_to_choice is None:
1082
1083
            eval_logger.error("doc_to_choice was called but not set in config")
        else:
1084
            doc_to_choice = self.config.doc_to_choice
1085

1086
        if isinstance(doc_to_choice, str):
1087
1088
1089
1090
            if doc_to_choice in self.features:
                return doc[doc_to_choice]
            else:
                return ast.literal_eval(utils.apply_template(doc_to_choice, doc))
1091
        elif isinstance(doc_to_choice, list):
1092
            return doc_to_choice
1093
        elif isinstance(doc_to_choice, dict):
1094
1095
1096
1097
1098
1099
1100
            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
1101

baberabb's avatar
baberabb committed
1102
1103
1104
    def construct_requests(
        self, doc: dict, ctx: str, **kwargs
    ) -> Union[List[Instance], Instance]:
1105
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
1106
            arguments = (ctx, self.doc_to_target(doc))
1107
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
1108
            arguments = (self.doc_to_target(doc),)
1109
        elif self.OUTPUT_TYPE == "multiple_choice":
1110
            choices = self.doc_to_choice(doc)
1111
            target_delimiter = self.config.target_delimiter
1112
1113
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
1114
                cont = self.doc_to_target(doc)
1115
1116
1117
                arguments = [
                    (ctx + choice, f"{target_delimiter}{cont}") for choice in choices
                ]
1118
            else:
1119
                # Otherwise they are placed in the continuation
1120
                arguments = [(ctx, f"{target_delimiter}{cont}") for cont in choices]
1121

1122
            request_list = [
1123
1124
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1125
                    doc=doc,
1126
                    arguments=arg,
1127
                    idx=i,
1128
1129
                    **kwargs,
                )
1130
                for i, arg in enumerate(arguments)
1131
            ]
1132
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
1133
            if "acc_mutual_info" in self._metric_fn_list.keys():
1134
1135
1136
                # if we are calculating multiple choice accuracy
                # using mutual information instead of raw loglikelihood as metric, need unconditional lls.

lintangsutawika's avatar
lintangsutawika committed
1137
                # here mutual info refers to calculating
1138
1139
1140
1141
1142
1143
                # 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
1144
                            doc=doc,
1145
                            arguments=("", "{}".format(choice)),
1146
1147
1148
                            idx=i,
                            **kwargs,
                        )
lintangsutawika's avatar
lintangsutawika committed
1149
                        for i, choice in enumerate(choices)
1150
1151
1152
                    ]
                )
            return request_list
lintangsutawika's avatar
lintangsutawika committed
1153

1154
        elif self.OUTPUT_TYPE == "generate_until":
1155
            arguments = (ctx, deepcopy(self.config.generation_kwargs))
lintangsutawika's avatar
lintangsutawika committed
1156
1157

        return Instance(
lintangsutawika's avatar
lintangsutawika committed
1158
1159
            request_type=self.OUTPUT_TYPE, doc=doc, arguments=arguments, idx=0, **kwargs
        )
1160
1161

    def process_results(self, doc, results):
1162
        if callable(self.config.process_results):
Lintang Sutawika's avatar
Lintang Sutawika committed
1163
            return self.config.process_results(doc, results)
lintangsutawika's avatar
lintangsutawika committed
1164

1165
        result_dict = {}
1166
        use_metric = list(self._metric_fn_list.keys())
1167
1168
1169
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
1170
1171
            prob_norm = np.exp(ll)

1172
1173
1174
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
1175
                **(
lintangsutawika's avatar
lintangsutawika committed
1176
                    {"brier_score": (0, [prob_norm])}  # Gold is Index 0
1177
1178
1179
                    if "brier_score" in use_metric
                    else {}
                ),
1180
            }
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
1216
1217
1218
            ):
                # then we are doing mutual info.
                # this stores the "dryrun" / unconditional answer loglikelihoods
                lls_unconditional = lls[1::2]
                assert len(lls_unconditional) == len(choices)
                # and this stores our "regular" conditional loglikelihoods
                lls = lls[::2]
1219

1220
1221
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
1222

1223
1224
            if self.multiple_input:
                gold = self.doc_to_text(doc)
1225
            else:
1226
                gold = self.doc_to_target(doc)
1227
1228

            gold_index_error = False
1229
            if isinstance(gold, list):
Lintang Sutawika's avatar
Lintang Sutawika committed
1230
1231
                gold = [i if i < len(choices) else -100 for i in gold]
                if -100 in gold:
1232
1233
                    gold_index_error = True
            else:
1234
                if isinstance(gold, int):
Lintang Sutawika's avatar
Lintang Sutawika committed
1235
                    gold = gold if gold < len(choices) else -100
1236
                elif isinstance(gold, str):
Lintang Sutawika's avatar
Lintang Sutawika committed
1237
                    gold = choices.index(gold) if gold in choices else -100
lintangsutawika's avatar
lintangsutawika committed
1238

Lintang Sutawika's avatar
Lintang Sutawika committed
1239
                if gold == -100:
1240
1241
1242
1243
                    gold_index_error = True

            if gold_index_error:
                eval_logger.warning(
lintangsutawika's avatar
lintangsutawika committed
1244
                    f"Label index was not in within range of available choices,"
1245
1246
                    f"Sample:\n\n{doc}\n\n"
                )
lintangsutawika's avatar
lintangsutawika committed
1247

1248
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
1249
1250
                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
1251
                exact_match = int(any([is_greedy[i] if i != -100 else 0 for i in gold]))
lintangsutawika's avatar
lintangsutawika committed
1252
1253
1254
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
1255
                # 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
1256
                exact_match = int(is_greedy[gold]) if gold != -100 else 0
1257

lintangsutawika's avatar
lintangsutawika committed
1258
            prob_norm = utils.softmax(lls)
lintangsutawika's avatar
lintangsutawika committed
1259

lintangsutawika's avatar
lintangsutawika committed
1260
            # TODO use keyword arguments to the metric?
lintangsutawika's avatar
format  
lintangsutawika committed
1261
            # gold, pred, norm stuff, the original lls,
1262
            result_dict = {
1263
                **({"acc": acc} if "acc" in use_metric else {}),
1264
1265
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
1266
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
1267
                **({"exact_match": exact_match} if "exact_match" in use_metric else {}),
lintangsutawika's avatar
format  
lintangsutawika committed
1268
                **(
lintangsutawika's avatar
lintangsutawika committed
1269
                    # {"brier_score": (gold, prob_norm)}
1270
                    {"brier_score": [np.eye(len(prob_norm))[gold], prob_norm]}
lintangsutawika's avatar
format  
lintangsutawika committed
1271
1272
1273
                    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
1379
1380

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

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

baberabb's avatar
baberabb committed
1397
    def process_results(self, doc: dict, results: List[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
        assert k == 0
        return []

baberabb's avatar
baberabb committed
1435
    def fewshot_context(self, doc: dict, num_fewshot: int) -> Literal[""]:
1436
1437
1438
1439
1440
1441
        assert (
            num_fewshot == 0
        ), "The number of fewshot examples must be 0 for perplexity tasks."

        return ""

baberabb's avatar
baberabb committed
1442
    def higher_is_better(self) -> dict:
1443
1444
1445
1446
1447
1448
1449
1450
1451
        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
1452
    def doc_to_text(self, doc) -> str:
1453
1454
1455
1456
1457
        return ""

    def doc_to_target(self, doc):
        return doc

baberabb's avatar
baberabb committed
1458
    def construct_requests(self, doc: dict, ctx: Union[str, None], **kwargs):
1459
1460
        assert not ctx

lintangsutawika's avatar
lintangsutawika committed
1461
1462
1463
1464
1465
1466
1467
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1468

baberabb's avatar
baberabb committed
1469
    def process_results(self, doc: dict, results: float) -> dict:
1470
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1471
1472
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1473
1474
1475
1476
1477
1478
        return {
            "word_perplexity": (loglikelihood, words),
            "byte_perplexity": (loglikelihood, bytes_),
            "bits_per_byte": (loglikelihood, bytes_),
        }

baberabb's avatar
baberabb committed
1479
    def aggregation(self) -> dict:
1480
1481
1482
1483
1484
1485
1486
        return {
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
            "bits_per_byte": bits_per_byte,
        }

    @classmethod
baberabb's avatar
baberabb committed
1487
    def count_bytes(cls, doc) -> int:
1488
1489
1490
        return len(doc.encode("utf-8"))

    @classmethod
baberabb's avatar
baberabb committed
1491
    def count_words(cls, doc) -> int:
1492
1493
        """Downstream tasks with custom word boundaries should override this!"""
        return len(re.split(r"\s+", doc))