task.py 57.7 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
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
                eval_logger.warning(
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
        self.fewshot_rnd: Optional[
            random.Random
        ] = None  # purposely induce errors in case of improper usage
235

236
237
238
239
240
241
    def download(
        self,
        data_dir: Optional[str] = None,
        cache_dir: Optional[str] = None,
        download_mode=None,
    ) -> None:
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
        """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.
        """
266
267
268
269
270
271
272
        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,
        )
273

274
    @property
275
    def config(self) -> TaskConfig:
276
277
278
        """Returns the TaskConfig associated with this class."""
        return self._config

279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
    @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

294
    def training_docs(self) -> Iterable:
295
296
297
298
299
300
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

301
    def validation_docs(self) -> Iterable:
302
303
304
305
306
307
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

308
    def test_docs(self) -> Iterable:
309
310
311
312
313
314
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

315
    def fewshot_docs(self) -> Iterable:
316
317
318
319
320
321
322
323
324
        """
        :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
325
            eval_logger.warning(
326
                f"[Task: {self.config.task}] has_training_docs and has_validation_docs are False"
327
                ", using test_docs as fewshot_docs but this is not recommended."
lintangsutawika's avatar
lintangsutawika committed
328
            )
329
330
            return self.test_docs()

331
    def _process_doc(self, doc: dict) -> dict:
332
333
334
335
336
337
338
339
340
        """
        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
341

342
    @property
343
    def instances(self) -> List[Instance]:
344
345
346
347
348
349
350
351
352
353
354
        """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)

355
356
    def doc_to_decontamination_query(self, doc):
        raise NotImplementedError(
357
358
359
360
361
362
363
364
365
366
367
            "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

368
369
    def build_all_requests(
        self,
370
        *,
371
372
373
374
375
376
        limit=None,
        rank=None,
        world_size=None,
        cache_requests=False,
        rewrite_requests_cache=False,
    ) -> None:
377
        """Build a set of Instances for a task, and store them in task.instances"""
378
379
380
381

        # used with caching
        og_limit = limit

382
        cache_key = f"requests-{self._config.task}-{self.config.num_fewshot}shot-rank{rank}-world_size{world_size}"
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397

        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
398
        eval_logger.info(f"Building contexts for {self.config.task} on rank {rank}...")
399

400
        instances = []
401
402
403
404
405
406
407
408
409
410

        # 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(
411
            self.doc_iterator(rank=rank, limit=limit, world_size=world_size)
412
413
414
415
416
417
418
        )

        num_docs = len(doc_id_docs)

        for doc_id, doc in tqdm(
            doc_id_docs,
            total=num_docs,
lintangsutawika's avatar
lintangsutawika committed
419
        ):
420
            # sample fewshot context #TODO: need to offset doc_id by rank now!
421
            fewshot_ctx = self.fewshot_context(
422
                doc,
423
                0 if self.config.num_fewshot is None else self.config.num_fewshot,
424
            )
425

426
            # TODO: we should override self.config.repeats if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
427
428
429
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
430
                metadata=(self.config["task"], doc_id, self.config.repeats),
lintangsutawika's avatar
lintangsutawika committed
431
            )
432
433
434
435

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

436
437
438
439
440
441
442
443
444
445
446
447
448
            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
449

450
451
        if len(self._instances) == 0:
            raise ValueError("task.build_requests() did not find any docs!")
452

453
454
455
        if cache_requests and (not cached_instances or rewrite_requests_cache):
            save_to_cache(file_name=cache_key, obj=instances)

456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
    @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
472
            The number of times each instance in a dataset is inferred on. Defaults to 1,
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
505
506
507
            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

508
509
510
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

haileyschoelkopf's avatar
haileyschoelkopf committed
511
512
513
514
515
516
517
518
519
520
    @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))

521
    @utils.positional_deprecated
lintangsutawika's avatar
lintangsutawika committed
522
    def fewshot_context(
523
524
525
        self,
        doc,
        num_fewshot,
526
        rnd=None,
527
        description=None,
lintangsutawika's avatar
lintangsutawika committed
528
    ):
529
530
531
532
533
534
535
        """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
536
537
538
539
540
        :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.
541
542
543
        :returns: str
            The fewshot context.
        """
544
        if rnd is None:
545
546
547
548
549
550
            if self.fewshot_rnd is not None:
                rnd = self.fewshot_rnd
            else:
                raise ValueError(
                    "A `random.Random` generator argument must be provided to `rnd`"
                )
lintangsutawika's avatar
lintangsutawika committed
551

552
        description = description if description else ""
553
554

        if num_fewshot == 0:
lintangsutawika's avatar
lintangsutawika committed
555
            labeled_examples = ""
556
        else:
lintangsutawika's avatar
lintangsutawika committed
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
            # 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
581
            )
582
583

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

586
    def apply_filters(self) -> Optional[List[Instance]]:
Baber Abbasi's avatar
Baber Abbasi committed
587
        """Iterates over FilterEnsembles and applies them to instances"""
lintangsutawika's avatar
lintangsutawika committed
588
589
        if hasattr(self, "_filters"):
            for f in self._filters:
590
                f.apply(self._instances)
lintangsutawika's avatar
lintangsutawika committed
591
592
593
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances
594

baberabb's avatar
baberabb committed
595
    def dump_config(self) -> dict:
Baber Abbasi's avatar
Baber Abbasi committed
596
        """Returns the config as a dictionary."""
597
        # TODO: this should only return the overrides applied to a non-YAML task's configuration.
598
        # (num_fewshot)
599
        return self.config.to_dict()
600

Baber Abbasi's avatar
Baber Abbasi committed
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
640
    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)

641
642
643
644
645
    def set_fewshot_seed(self, seed: Optional[int] = None) -> None:
        self.fewshot_rnd = random.Random(seed)
        if hasattr(self, "sampler"):
            self.sampler.rnd = self.fewshot_rnd

646
647
648
649
650
651
652
    @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:
653
654
655
            raise ValueError(
                f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"
            )
656
657
658
659
660
661
662
663
664
665
666
667
668

    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

669
670

class ConfigurableTask(Task):
671
    VERSION = "Yaml"
672
    OUTPUT_TYPE = None
673
    CONFIG = None
674
675

    def __init__(
676
677
678
679
680
        self,
        data_dir=None,
        cache_dir=None,
        download_mode=None,
        config: Optional[dict] = None,
Ethan Smith's avatar
Ethan Smith committed
681
    ) -> None:  # TODO no super() call here
682
        # Get pre-configured attributes
683
        self._config = self.CONFIG
684

685
        # Use new configurations if there was no preconfiguration
686
        if self.config is None:
687
            self._config = TaskConfig(**config)
688
689
        # Overwrite configs
        else:
lintangsutawika's avatar
lintangsutawika committed
690
            if config is not None:
691
                self._config.__dict__.update(config)
692

693
        if self.config is None:
lintangsutawika's avatar
lintangsutawika committed
694
695
696
            raise ValueError(
                "Must pass a config to ConfigurableTask, either in cls.CONFIG or `config` kwarg"
            )
697

698
699
700
701
        if isinstance(self.config.metadata, dict):
            if "version" in self.config.metadata:
                self.VERSION = self.config.metadata["version"]

702
        if self.config.output_type is not None:
703
704
705
706
            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)}'"
                )
707
            self.OUTPUT_TYPE = self.config.output_type
708

709
710
        if self.config.dataset_path is not None:
            self.DATASET_PATH = self.config.dataset_path
711

712
713
        if self.config.dataset_name is not None:
            self.DATASET_NAME = self.config.dataset_name
714

715
716
717
718
        self._metric_fn_list = {}
        self._metric_fn_kwargs = {}
        self._aggregation_list = {}
        self._higher_is_better = {}
719

720
        if self.config.metric_list is None:
721
            # TODO: handle this in TaskConfig.__post_init__ ?
722
723
            _metric_list = DEFAULT_METRIC_REGISTRY[self.config.output_type]

724
            for metric_name in _metric_list:
haileyschoelkopf's avatar
haileyschoelkopf committed
725
                self._metric_fn_list[metric_name] = get_metric(metric_name)
lintangsutawika's avatar
lintangsutawika committed
726
                self._metric_fn_kwargs[metric_name] = {}
727
728
729
                self._aggregation_list[metric_name] = get_metric_aggregation(
                    metric_name
                )
haileyschoelkopf's avatar
haileyschoelkopf committed
730
                self._higher_is_better[metric_name] = is_higher_better(metric_name)
731
        else:
732
            for metric_config in self.config.metric_list:
733
734
735
736
                if "metric" not in metric_config:
                    raise ValueError(
                        "'metric' key not provided for an entry in 'metric_list', must be specified!"
                    )
737
738
739
740
                metric_name = metric_config["metric"]
                kwargs = {
                    key: metric_config[key]
                    for key in metric_config
Chris's avatar
Chris committed
741
742
                    if key
                    not in ["metric", "aggregation", "higher_is_better", "hf_evaluate"]
743
                }
Chris's avatar
Chris committed
744
745
746
747
                hf_evaluate_metric = (
                    "hf_evaluate" in metric_config
                    and metric_config["hf_evaluate"] is True
                )
748

749
                if self.config.process_results is not None:
750
751
                    self._metric_fn_list[metric_name] = None
                    self._metric_fn_kwargs[metric_name] = {}
752
753
754
755
756
757
                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
758
759
760
                    self._metric_fn_list[metric_name] = get_metric(
                        metric_name, hf_evaluate_metric
                    )
761
                    self._metric_fn_kwargs[metric_name] = kwargs
lintangsutawika's avatar
lintangsutawika committed
762

763
                if "aggregation" in metric_config:
764
                    agg_name = metric_config["aggregation"]
765
                    if isinstance(agg_name, str):
haileyschoelkopf's avatar
haileyschoelkopf committed
766
                        self._aggregation_list[metric_name] = get_aggregation(agg_name)
767
                    elif callable(agg_name):  # noqa: E721
768
769
770
                        self._aggregation_list[metric_name] = metric_config[
                            "aggregation"
                        ]
771
                else:
772
                    INV_AGG_REGISTRY = {v: k for k, v in AGGREGATION_REGISTRY.items()}
lintangsutawika's avatar
lintangsutawika committed
773
                    metric_agg = get_metric_aggregation(metric_name)
774
                    eval_logger.warning(
775
                        f"[Task: {self.config.task}] metric {metric_name} is defined, but aggregation is not. "
776
777
                        f"using default "
                        f"aggregation={INV_AGG_REGISTRY[metric_agg]}"
778
                    )
779
                    self._aggregation_list[metric_name] = metric_agg
lintangsutawika's avatar
lintangsutawika committed
780

781
782
783
784
785
786
                if "higher_is_better" in metric_config:
                    self._higher_is_better[metric_name] = metric_config[
                        "higher_is_better"
                    ]
                else:
                    eval_logger.warning(
787
                        f"[Task: {self.config.task}] metric {metric_name} is defined, but higher_is_better is not. "
788
                        f"using default "
haileyschoelkopf's avatar
haileyschoelkopf committed
789
                        f"higher_is_better={is_higher_better(metric_name)}"
790
                    )
haileyschoelkopf's avatar
haileyschoelkopf committed
791
                    self._higher_is_better[metric_name] = is_higher_better(metric_name)
792

793
        self.download(self.config.dataset_kwargs)
794
795
796
        self._training_docs = None
        self._fewshot_docs = None

797
        if self.config.filter_list is not None:
lintangsutawika's avatar
lintangsutawika committed
798
            self._filters = []
799
            for filter_config in self.config.filter_list:
800
801
802
803
804
805
806
807
808
                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
809
                self._filters.append(filter_pipeline)
lintangsutawika's avatar
lintangsutawika committed
810
        else:
811
            self._filters = [build_filter_ensemble("none", [["take_first", None]])]
812

813
814
        if self.config.use_prompt is not None:
            eval_logger.info(f"loading prompt {self.config.use_prompt}")
815
            self.prompt = get_prompt(
816
                self.config.use_prompt, self.DATASET_PATH, self.DATASET_NAME
lintangsutawika's avatar
lintangsutawika committed
817
            )
818
819
820
        else:
            self.prompt = None

lintangsutawika's avatar
lintangsutawika committed
821
        if self.fewshot_docs() is not None:
822
823
824
825
            self.fewshot_rnd = (
                random.Random()
            )  # setting with no seed, to be overridden at a later time
            config_sampler: Union[str, Callable] = (
haileyschoelkopf's avatar
haileyschoelkopf committed
826
827
828
                self.config.fewshot_config.get("sampler", "default")
                if self.config.fewshot_config
                else "default"
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
            )
            if isinstance(config_sampler, str):
                self.sampler = samplers.get_sampler(config_sampler)(
                    list(self.fewshot_docs()), self, rnd=self.fewshot_rnd
                )
            elif callable(config_sampler) and issubclass(
                config_sampler, samplers.ContextSampler
            ):
                self.sampler = config_sampler(
                    docs=list(self.fewshot_docs()), task=self, rnd=self.fewshot_rnd
                )
            else:
                raise TypeError(
                    f"fewshot_config.sampler should be a string or callable of ContextSampler type, "
                    f"not {type(config_sampler)}"
                )
845

846
        self.task_docs = self.eval_docs
847

848
        # Test One Doc
849
        self.features = list(self.task_docs.features.keys())
850
851
        self.multiple_input = 0
        self.multiple_target = 0
852
        test_doc = self.task_docs[0]
853
        test_text = self.doc_to_text(test_doc)
854
        test_target = self.doc_to_target(test_doc)
lintangsutawika's avatar
lintangsutawika committed
855

856
        if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
857
            test_choice = self.doc_to_choice(test_doc)
858
            if not isinstance(test_choice, list):
lintangsutawika's avatar
lintangsutawika committed
859
                eval_logger.error("doc_to_choice must return list")
860
861
            else:
                num_choice = len(test_choice)
862

863
            if isinstance(test_text, int):
864
                self.multiple_input = num_choice
865
866
        else:
            test_choice = None
867

868
        if isinstance(test_target, list):
869
            self.multiple_target = len(test_target)
870
        else:
871
            if (isinstance(test_target, int)) and (test_choice is not None):
lintangsutawika's avatar
lintangsutawika committed
872
                test_target = test_choice[test_target]
873
            else:
lintangsutawika's avatar
lintangsutawika committed
874
                test_target = str(test_target)
875

876
877
878
        if test_choice is not None:
            check_choices = test_choice
        else:
lintangsutawika's avatar
lintangsutawika committed
879
            check_choices = [test_target]
880
881
882
883
        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 = (
884
885
                    True
                    if self.config.target_delimiter.rstrip()
886
                    != self.config.target_delimiter
887
                    else False
888
                )
889

890
                if delimiter_has_whitespace and choice_has_whitespace:
891
892
                    eval_logger.debug(
                        f'Both target_delimiter "{self.config.target_delimiter}" and target choice: "{choice}" have whitespace'
893
894
                    )
                elif (not delimiter_has_whitespace) and (not choice_has_whitespace):
895
                    eval_logger.debug(
896
                        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'
897
898
                    )

899
    def download(self, dataset_kwargs: Optional[Dict[str, Any]] = None) -> None:
900
901
902
903
904
905
        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
906
    def has_training_docs(self) -> bool:
907
        if self.config.training_split is not None:
908
909
910
911
            return True
        else:
            return False

baberabb's avatar
baberabb committed
912
    def has_validation_docs(self) -> bool:
913
        if self.config.validation_split is not None:
914
915
916
917
            return True
        else:
            return False

baberabb's avatar
baberabb committed
918
    def has_test_docs(self) -> bool:
919
        if self.config.test_split is not None:
920
921
922
923
            return True
        else:
            return False

baberabb's avatar
baberabb committed
924
    def training_docs(self) -> datasets.Dataset:
925
        if self.has_training_docs():
926
927
928
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.training_split]
929
                )
930
            return self.dataset[self.config.training_split]
931

baberabb's avatar
baberabb committed
932
    def validation_docs(self) -> datasets.Dataset:
933
        if self.has_validation_docs():
934
935
936
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.validation_split]
937
                )
938
            return self.dataset[self.config.validation_split]
939

baberabb's avatar
baberabb committed
940
    def test_docs(self) -> datasets.Dataset:
941
        if self.has_test_docs():
942
943
944
            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]
945

946
    def fewshot_docs(self):
947
        if self.config.fewshot_split is not None:
948
949
            if self.config.process_docs is not None:
                return self.config.process_docs(self.dataset[self.config.fewshot_split])
950
            return self.dataset[self.config.fewshot_split]
951
        else:
952
            if (self.config.num_fewshot is not None) and (self.config.num_fewshot > 0):
953
                eval_logger.warning(
954
                    f"Task '{self.config.task}': "
955
956
957
958
                    "num_fewshot > 0 but fewshot_split is None. "
                    "using preconfigured rule."
                )
            return super().fewshot_docs()
959

lintangsutawika's avatar
lintangsutawika committed
960
    @utils.positional_deprecated
961
    def fewshot_context(self, doc: str, num_fewshot: int) -> str:
lintangsutawika's avatar
lintangsutawika committed
962
963
964
965
966
967
968
969
970
971
        """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.
        """
972
973
        if description := self.config.description:
            description = utils.apply_template(self.config.description, doc)
lintangsutawika's avatar
lintangsutawika committed
974
975
976

        if num_fewshot == 0:
            # always prepend the (possibly empty) task description
977
            labeled_examples = description
lintangsutawika's avatar
lintangsutawika committed
978
        else:
979
            labeled_examples = description + self.sampler.get_context(doc, num_fewshot)
lintangsutawika's avatar
lintangsutawika committed
980
981

        example = self.doc_to_text(doc)
982
983
984
985
986
987
988
989
990
991
992
993
994
        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
995

996
    def apply_filters(self):
Baber Abbasi's avatar
Baber Abbasi committed
997
        """Iterates over FilterEnsembles and applies them to instances"""
998
999
        if hasattr(self, "_filters"):
            for f in self._filters:
1000
                f.apply(self._instances)
1001
1002
1003
1004
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances

1005
    def should_decontaminate(self):
1006
        return self.config.should_decontaminate
1007
1008

    def doc_to_decontamination_query(self, doc):
1009
        if self.config.should_decontaminate:
1010
1011
            if self.config.doc_to_decontamination_query is None:
                return self.doc_to_text(doc)
1012
            else:
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
                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
                        )
                    )
1024

1025
    def _process_doc(self, doc: dict) -> dict:
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
        """
        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):
1037
1038
        if self.prompt is not None:
            doc_to_text = self.prompt
1039
        else:
1040
            doc_to_text = self.config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
1041

1042
        if isinstance(doc_to_text, int):
1043
            return doc_to_text
1044
        elif isinstance(doc_to_text, str):
1045
            if doc_to_text in self.features:
1046
                # if self.config.doc_to_choice is not None:
1047
1048
                #     return self.doc_to_choice(doc)[doc[doc_to_text]]
                # else:
1049
1050
                return doc[doc_to_text]
            else:
lintangsutawika's avatar
lintangsutawika committed
1051
                text_string = utils.apply_template(doc_to_text, doc)
lintangsutawika's avatar
lintangsutawika committed
1052
                if text_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1053
1054
1055
                    return ast.literal_eval(text_string)
                else:
                    return text_string
1056
        elif callable(doc_to_text):
1057
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
1058
        # Used when applying a Promptsource template
1059
        elif hasattr(doc_to_text, "apply"):
1060
1061
1062
1063
1064
            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")
1065
                return self.config.fewshot_delimiter
1066
        else:
1067
            print(type(doc_to_text))
1068
            raise TypeError
1069

1070
    def doc_to_target(self, doc: Mapping) -> Union[int, str, list]:
1071
1072
        if self.prompt is not None:
            doc_to_target = self.prompt
1073
        else:
1074
            doc_to_target = self.config.doc_to_target
1075

1076
        if isinstance(doc_to_target, int):
1077
            return doc_to_target
1078
        elif isinstance(doc_to_target, str):
1079
            if doc_to_target in self.features:
1080
                # if self.config.doc_to_choice is not None:
1081
1082
1083
                #     return self.doc_to_choice(doc)[doc[doc_to_target]]
                # else:
                return doc[doc_to_target]
1084
            else:
lintangsutawika's avatar
lintangsutawika committed
1085
                target_string = utils.apply_template(doc_to_target, doc)
lintangsutawika's avatar
lintangsutawika committed
1086
                if target_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1087
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
1088
1089
1090
1091
1092
                elif (
                    len(target_string) >= 2
                    and (target_string[0] == "[")
                    and (target_string[-1] == "]")
                ):
baberabb's avatar
baberabb committed
1093
1094
1095
1096
                    try:
                        return ast.literal_eval(target_string)
                    except (SyntaxError, ValueError):
                        return target_string
lintangsutawika's avatar
lintangsutawika committed
1097
1098
                else:
                    return target_string
1099
        elif isinstance(doc_to_target, list):
1100
            return doc_to_target
1101
        elif callable(doc_to_target):
1102
            return doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
1103
        # Used when applying a Promptsource template
1104
        elif hasattr(doc_to_target, "apply"):
1105
            applied_prompt = doc_to_target.apply(doc)
1106
1107
1108
1109
            if len(applied_prompt) == 2:
                return applied_prompt[1]
            else:
                eval_logger.warning("Applied prompt returns empty string")
1110
                return self.config.fewshot_delimiter
1111
1112
        else:
            raise TypeError
1113

baberabb's avatar
baberabb committed
1114
    def doc_to_choice(self, doc: Any) -> List[str]:
1115
1116
        if self.prompt is not None:
            doc_to_choice = self.prompt
1117
        elif self.config.doc_to_choice is None:
1118
1119
            eval_logger.error("doc_to_choice was called but not set in config")
        else:
1120
            doc_to_choice = self.config.doc_to_choice
1121

1122
        if isinstance(doc_to_choice, str):
1123
1124
1125
1126
            if doc_to_choice in self.features:
                return doc[doc_to_choice]
            else:
                return ast.literal_eval(utils.apply_template(doc_to_choice, doc))
1127
        elif isinstance(doc_to_choice, list):
1128
            return doc_to_choice
1129
        elif isinstance(doc_to_choice, dict):
1130
1131
1132
1133
1134
1135
1136
            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
1137

baberabb's avatar
baberabb committed
1138
1139
1140
    def construct_requests(
        self, doc: dict, ctx: str, **kwargs
    ) -> Union[List[Instance], Instance]:
1141
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
1142
            arguments = (ctx, self.doc_to_target(doc))
1143
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
1144
            arguments = (self.doc_to_target(doc),)
1145
        elif self.OUTPUT_TYPE == "multiple_choice":
1146
            choices = self.doc_to_choice(doc)
1147
            target_delimiter = self.config.target_delimiter
1148
1149
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
1150
                cont = self.doc_to_target(doc)
1151
1152
1153
                arguments = [
                    (ctx + choice, f"{target_delimiter}{cont}") for choice in choices
                ]
1154
            else:
1155
                # Otherwise they are placed in the continuation
1156
                arguments = [(ctx, f"{target_delimiter}{cont}") for cont in choices]
1157

1158
            request_list = [
1159
1160
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1161
                    doc=doc,
1162
                    arguments=arg,
1163
                    idx=i,
1164
1165
                    **kwargs,
                )
1166
                for i, arg in enumerate(arguments)
1167
            ]
1168
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
1169
            if "acc_mutual_info" in self._metric_fn_list.keys():
1170
1171
1172
                # if we are calculating multiple choice accuracy
                # using mutual information instead of raw loglikelihood as metric, need unconditional lls.

lintangsutawika's avatar
lintangsutawika committed
1173
                # here mutual info refers to calculating
1174
1175
1176
1177
1178
1179
                # 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
1180
                            doc=doc,
1181
                            arguments=("", "{}".format(choice)),
1182
1183
1184
                            idx=i,
                            **kwargs,
                        )
lintangsutawika's avatar
lintangsutawika committed
1185
                        for i, choice in enumerate(choices)
1186
1187
1188
                    ]
                )
            return request_list
lintangsutawika's avatar
lintangsutawika committed
1189

1190
        elif self.OUTPUT_TYPE == "generate_until":
1191
            arguments = (ctx, deepcopy(self.config.generation_kwargs))
lintangsutawika's avatar
lintangsutawika committed
1192
1193

        return Instance(
lintangsutawika's avatar
lintangsutawika committed
1194
1195
            request_type=self.OUTPUT_TYPE, doc=doc, arguments=arguments, idx=0, **kwargs
        )
1196
1197

    def process_results(self, doc, results):
1198
1199
        if callable(self.config.process_results):
            return self.config.process_results(doc, results)
lintangsutawika's avatar
lintangsutawika committed
1200

1201
        result_dict = {}
1202
        use_metric = list(self._metric_fn_list.keys())
1203
1204
1205
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
1206
1207
1208
1209
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
1210
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
1211
            (loglikelihood,) = results
1212
1213
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
haileyschoelkopf's avatar
haileyschoelkopf committed
1214
            return {
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
                **(
                    {"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
1230
            }
1231
        elif self.OUTPUT_TYPE == "multiple_choice":
1232
            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
1233

1234
            # retrieve choices in List[str] form, to compute choice lengths, etc.
1235
            choices = self.doc_to_choice(doc)
1236
1237
            completion_len = np.array([float(len(i)) for i in choices])

1238
1239
            if (
                2 * len(choices) == len(lls)
1240
                and "acc_mutual_info" in self._metric_fn_list.keys()
1241
1242
1243
1244
            ):
                # then we are doing mutual info.
                # this stores the "dryrun" / unconditional answer loglikelihoods
                lls_unconditional = lls[1::2]
1245
1246
                if len(lls_unconditional) != len(choices):
                    raise ValueError
1247
1248
                # and this stores our "regular" conditional loglikelihoods
                lls = lls[::2]
1249

1250
1251
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
1252

1253
1254
            if self.multiple_input:
                gold = self.doc_to_text(doc)
1255
            else:
1256
                gold = self.doc_to_target(doc)
1257
1258

            gold_index_error = False
1259
            if isinstance(gold, list):
Lintang Sutawika's avatar
Lintang Sutawika committed
1260
1261
                gold = [i if i < len(choices) else -100 for i in gold]
                if -100 in gold:
1262
1263
                    gold_index_error = True
            else:
1264
                if isinstance(gold, int):
Lintang Sutawika's avatar
Lintang Sutawika committed
1265
                    gold = gold if gold < len(choices) else -100
1266
                elif isinstance(gold, str):
Lintang Sutawika's avatar
Lintang Sutawika committed
1267
                    gold = choices.index(gold) if gold in choices else -100
lintangsutawika's avatar
lintangsutawika committed
1268

Lintang Sutawika's avatar
Lintang Sutawika committed
1269
                if gold == -100:
1270
1271
1272
1273
                    gold_index_error = True

            if gold_index_error:
                eval_logger.warning(
lintangsutawika's avatar
lintangsutawika committed
1274
                    f"Label index was not in within range of available choices,"
1275
1276
                    f"Sample:\n\n{doc}\n\n"
                )
lintangsutawika's avatar
lintangsutawika committed
1277

1278
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
1279
1280
                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
1281
                exact_match = int(any([is_greedy[i] if i != -100 else 0 for i in gold]))
lintangsutawika's avatar
lintangsutawika committed
1282
1283
1284
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
1285
                # 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
1286
                exact_match = int(is_greedy[gold]) if gold != -100 else 0
1287

Lintang Sutawika's avatar
Lintang Sutawika committed
1288
1289
1290
1291
            prob_norm = utils.softmax(lls)

            # TODO use keyword arguments to the metric?
            # gold, pred, norm stuff, the original lls,
1292
            result_dict = {
1293
                **({"acc": acc} if "acc" in use_metric else {}),
1294
1295
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
1296
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
1297
                **({"exact_match": exact_match} if "exact_match" in use_metric else {}),
Lintang Sutawika's avatar
Lintang Sutawika committed
1298
1299
1300
1301
1302
                **(
                    {"brier_score": (gold, prob_norm)}
                    if "brier_score" in use_metric
                    else {}
                ),
1303
1304
            }

1305
            if "acc_mutual_info" in use_metric:
lintangsutawika's avatar
lintangsutawika committed
1306
1307
1308
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
1309
1310
1311
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

1312
        elif self.OUTPUT_TYPE == "generate_until":
1313
            gold = self.doc_to_target(doc)
Chris's avatar
Chris committed
1314
            result = results[0]
1315
            if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1316
                # If you set doc_to_choice,
lintangsutawika's avatar
lintangsutawika committed
1317
                # it assumes that doc_to_target returns a number.
1318
1319
                choices = self.doc_to_choice(doc)
                gold = choices[gold]
1320
1321
            # we expect multiple_targets to be a list.
            elif self.multiple_target:
baberabb's avatar
baberabb committed
1322
                gold = list(gold)
Chris's avatar
Chris committed
1323
1324
1325
            elif type(gold) != type(result):
                # cast gold to the same type as result
                gold = type(result)(gold)
1326

lintangsutawika's avatar
lintangsutawika committed
1327
            for metric in self._metric_fn_list.keys():
haileyschoelkopf's avatar
haileyschoelkopf committed
1328
1329
1330
1331
1332
                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
1333
1334
1335
1336
                    if not isinstance(gold, list):
                        # sometimes, a multiple_target dataset has exceptions where one doc has only one string answer
                        # print(gold)
                        gold = [gold]
1337
1338
1339
1340
1341
1342
1343
1344
                    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
1345
                    else:
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
                        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
1367
                else:
1368
                    try:
1369
                        result_score = self._metric_fn_list[metric](
1370
1371
                            references=[gold],
                            predictions=[result],
1372
                            **self._metric_fn_kwargs[metric],
1373
                        )
JessicaOjo's avatar
JessicaOjo committed
1374
                    except TypeError as error:  # needed for now in order to use a different interface between our own metrics and HF Evaluate metrics
1375
                        result_score = self._metric_fn_list[metric]([gold, result])
1376
1377
1378
1379
                    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
1380
        else:
lintangsutawika's avatar
lintangsutawika committed
1381
1382
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
1383
                "'loglikelihood', 'loglikelihood_rolling', 'generate_until' or 'multiple_choice'",
1384
            )
1385
1386
1387

        return result_dict

Baber Abbasi's avatar
Baber Abbasi committed
1388
    def aggregation(self) -> dict:
1389
1390
        return self._aggregation_list

Baber Abbasi's avatar
Baber Abbasi committed
1391
    def higher_is_better(self) -> dict:
haileyschoelkopf's avatar
haileyschoelkopf committed
1392
        return self._higher_is_better
1393

Baber Abbasi's avatar
Baber Abbasi committed
1394
1395
1396
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

1397
1398
1399
1400
1401
1402
1403
1404
1405
    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)})"
        )

1406
1407

class MultipleChoiceTask(Task):
1408
    OUTPUT_TYPE = "loglikelihood"
1409

baberabb's avatar
baberabb committed
1410
    def doc_to_target(self, doc: dict) -> str:
1411
1412
        return " " + doc["choices"][doc["gold"]]

baberabb's avatar
baberabb committed
1413
    def construct_requests(self, doc: dict, ctx: str, **kwargs) -> List[Instance]:
1414
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1415
1416
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1417
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1418
                doc=doc,
1419
                arguments=(ctx, " {}".format(choice)),
1420
                idx=i,
1421
1422
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1423
1424
            for i, choice in enumerate(doc["choices"])
        ]
1425

1426
    def process_results(self, doc: dict, results: Iterable[Tuple[float, bool]]) -> dict:
lintangsutawika's avatar
lintangsutawika committed
1427
1428
1429
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
        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
1441
    def higher_is_better(self) -> dict:
1442
1443
1444
1445
1446
        return {
            "acc": True,
            "acc_norm": True,
        }

baberabb's avatar
baberabb committed
1447
    def aggregation(self) -> dict:
1448
1449
1450
1451
1452
1453
        return {
            "acc": mean,
            "acc_norm": mean,
        }


lintangsutawika's avatar
lintangsutawika committed
1454
class PerplexityTask(Task):
1455
1456
    OUTPUT_TYPE = "loglikelihood_rolling"

baberabb's avatar
baberabb committed
1457
    def has_training_docs(self) -> bool:
1458
1459
        return False

baberabb's avatar
baberabb committed
1460
    def fewshot_examples(self, k: int, rnd) -> List:
1461
1462
1463
1464
        if k != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1465
1466
        return []

baberabb's avatar
baberabb committed
1467
    def fewshot_context(self, doc: dict, num_fewshot: int) -> Literal[""]:
1468
1469
1470
1471
        if num_fewshot != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1472
1473
1474

        return ""

baberabb's avatar
baberabb committed
1475
    def higher_is_better(self) -> dict:
1476
1477
1478
1479
1480
1481
1482
1483
1484
        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
1485
    def doc_to_text(self, doc) -> str:
1486
1487
1488
1489
1490
        return ""

    def doc_to_target(self, doc):
        return doc

1491
1492
1493
    def construct_requests(self, doc: dict, ctx: Optional[str], **kwargs):
        if bool(ctx):
            raise ValueError
1494

lintangsutawika's avatar
lintangsutawika committed
1495
1496
1497
1498
1499
1500
1501
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1502

1503
    def process_results(self, doc: dict, results: Tuple[float]) -> dict:
1504
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1505
1506
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1507
1508
1509
1510
1511
1512
        return {
            "word_perplexity": (loglikelihood, words),
            "byte_perplexity": (loglikelihood, bytes_),
            "bits_per_byte": (loglikelihood, bytes_),
        }

baberabb's avatar
baberabb committed
1513
    def aggregation(self) -> dict:
1514
1515
1516
1517
1518
1519
1520
        return {
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
            "bits_per_byte": bits_per_byte,
        }

    @classmethod
baberabb's avatar
baberabb committed
1521
    def count_bytes(cls, doc) -> int:
1522
1523
1524
        return len(doc.encode("utf-8"))

    @classmethod
baberabb's avatar
baberabb committed
1525
    def count_words(cls, doc) -> int:
1526
1527
        """Downstream tasks with custom word boundaries should override this!"""
        return len(re.split(r"\s+", doc))