task.py 36.8 KB
Newer Older
1
import abc
2
from dataclasses import dataclass, field, asdict
3
4

import re
5
import ast
lintangsutawika's avatar
lintangsutawika committed
6
import yaml
7
8
9
import evaluate
import random
import itertools
10
import functools
11
12
13
14

import datasets
import numpy as np

15
16
from typing import Union
from collections.abc import Callable
17

18
from lm_eval import utils
19
from lm_eval.api import samplers
haileyschoelkopf's avatar
haileyschoelkopf committed
20
from lm_eval.api.instance import Instance
lintangsutawika's avatar
lintangsutawika committed
21
from lm_eval.api.filter import FilterEnsemble
22
23
24
25

from lm_eval.logger import eval_logger
from lm_eval.prompts import get_prompt
from lm_eval.filters import build_filter_ensemble
lintangsutawika's avatar
lintangsutawika committed
26
27
28
29
30
31
32
33
from lm_eval.api.metrics import (
    # get_metric,
    # get_aggregation,
    mean,
    weighted_perplexity,
    bits_per_byte,
)
from lm_eval.api.registry import (
lintangsutawika's avatar
lintangsutawika committed
34
    METRIC_REGISTRY,
35
36
    DEFAULT_METRIC_REGISTRY,
    OUTPUT_TYPE_REGISTRY,
lintangsutawika's avatar
lintangsutawika committed
37
38
    AGGREGATION_REGISTRY,
    HIGHER_IS_BETTER_REGISTRY,
39
    DEFAULT_AGGREGATION_REGISTRY,
lintangsutawika's avatar
lintangsutawika committed
40
)
41

42
43
44
45
46
47
48
ALL_OUTPUT_TYPES = [
    "loglikelihood",
    "multiple_choice",
    "loglikelihood_rolling",
    "greedy_until",
]

49
50
51
52

@dataclass
class TaskConfig(dict):

53
    task: str = None
54
    group: Union[str, list] = None
lintangsutawika's avatar
lintangsutawika committed
55
    reference: str = None
56

57
58
    dataset_path: str = None
    dataset_name: str = None
59
    dataset_kwargs: dict = None
60
61
62
    training_split: str = None
    validation_split: str = None
    test_split: str = None
lintangsutawika's avatar
lintangsutawika committed
63
    fewshot_split: str = None  # TODO: assert that this not None if num_fewshot > 0. (?) assert if this is same split as one evaling (?)
64

65
    template_aliases: str = None
66
    aliases: Union[str, list] = None
67
68
    doc_to_text: Union[Callable, str] = None
    doc_to_target: Union[Callable, str] = None
69
    use_prompt: str = None
70

71
72
    num_fewshot: int = 0
    batch_size: int = 1
73
74
    repeats: int = 1

75
    metric_list: str = None
lintangsutawika's avatar
lintangsutawika committed
76
    gold_alias: Union[Callable, str] = None
77
    output_type: str = "greedy_until"
78
    generation_kwargs: dict = None
lintangsutawika's avatar
lintangsutawika committed
79
80
    target_delimiter: str = "\n\n"
    sample_delimiter: str = " "
lintangsutawika's avatar
lintangsutawika committed
81
    filter_list: Union[str, list] = None
82
83
    should_decontaminate: bool = False
    doc_to_decontamination_query: str = None
84

lintangsutawika's avatar
lintangsutawika committed
85
    metadata: str = None  # by default, not used in the code. allows for users to pass arbitrary info to tasks
86

87
88
89
90
    def __post_init__(self):
        # allow user-specified aliases so that users can
        # force prompt-compatibility for some prompt regardless of
        # field names in prompt
91
92
93
        if self.template_aliases is not None:
            if type(self.doc_to_text) == str:
                self.doc_to_text = self.template_aliases + self.doc_to_text
94

95
96
            if type(self.doc_to_target) == str:
                self.doc_to_target = self.template_aliases + self.doc_to_target
97

98
            if type(self.gold_alias) == str:
lintangsutawika's avatar
lintangsutawika committed
99
                self.gold_alias = self.template_aliases + self.gold_alias
100

101
        if self.generation_kwargs or self.output_type == "greedy_until":
102
103
104
            assert (
                self.output_type == "greedy_until"
            ), "passed `generation_kwargs`, but not using a generation request type!"
105
106
            # ensure that we greedily generate in absence of explicit arguments otherwise
            self.generation_kwargs = {"do_sample": False, "temperature": 0.0}
107

108
109
110
    def __getitem__(self, item):
        return getattr(self, item)

111
    def to_dict(self):
112
113
        """dumps the current config as a dictionary object, as a printable format.
        null fields will not be printed.
haileyschoelkopf's avatar
haileyschoelkopf committed
114
        Used for dumping results alongside full task configuration
115

haileyschoelkopf's avatar
haileyschoelkopf committed
116
117
118
119
120
121
122
123
124
125
126
        :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)
        return cfg_dict
127

128
129
130
131
132
133
134
135
136
137
138
139

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
140

141
142
143
144
145
146
147
148
    # 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
149

150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
    def __init__(
        self,
        data_dir=None,
        cache_dir=None,
        download_mode=None,
        config=None,
    ):
        """
        :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.
        """
        self.download(data_dir, cache_dir, download_mode)
        self._training_docs = None
        self._fewshot_docs = None
        self._instances = None

haileyschoelkopf's avatar
haileyschoelkopf committed
184
        self._config = TaskConfig(**config) if config else TaskConfig()
185
186
187

        if not hasattr(self, "_filters"):
            self._filters = []
lintangsutawika's avatar
lintangsutawika committed
188
            for name, components in self._config.get(
189
                "filters", [["none", [["take_first", None]]]]
lintangsutawika's avatar
lintangsutawika committed
190
            ):
191
192
193
                filter_pipeline = build_filter_ensemble(name, components)
                self._filters.append(filter_pipeline)

lintangsutawika's avatar
lintangsutawika committed
194
195
196
        self.sampler = samplers.Sampler(
            list(self.fewshot_docs()), self, rnd=random.Random()
        )  # TODO: pass the correct docs in here
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222

    def download(self, data_dir=None, cache_dir=None, download_mode=None):
        """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.
        """
223
224
225
226
227
228
229
        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,
        )
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266

    @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 []

267
268
269
270
271
272
273
274
275
276
    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
277
            eval_logger.warning(
278
                "has_training_docs and has_validation_docs are False"
279
                ", using test_docs but this is not recommended."
lintangsutawika's avatar
lintangsutawika committed
280
            )
281
282
            return self.test_docs()

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
318
319
320
    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

    @property
    def instances(self):
        """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)

    def doc_to_decontamination_query(self, doc):
        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

321
    def build_all_requests(self, limit=None, rank=None, world_size=None):
322
323
324
325
326
327
328
329
330
331
332
        """Build a set of Instances for a task, and store them in task.instances"""
        if self.has_test_docs():
            docs = self.test_docs()
        elif self.has_validation_docs():
            docs = self.validation_docs()
        else:
            assert (
                False
            ), f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"

        instances = []
333
334
        for doc_id, doc in utils.create_iterator(
            enumerate(docs), rank, world_size, limit
lintangsutawika's avatar
lintangsutawika committed
335
        ):
336
            # sample fewshot context #TODO: need to offset doc_id by rank now!
337
338
339
340
            fewshot_ctx = self.fewshot_context(
                doc, self._config.num_fewshot, rnd=random.Random()
            )
            # TODO: hardcoded for now: # of runs on each input to be 2. # TODO: we should override this if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
341
342
343
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
344
                metadata=(self._config["task"], doc_id, self._config.repeats),
lintangsutawika's avatar
lintangsutawika committed
345
            )
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370

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

            instances.extend(inst)

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

    @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
371
            The number of times each instance in a dataset is inferred on. Defaults to 1,
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
            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

haileyschoelkopf's avatar
haileyschoelkopf committed
407
408
409
410
411
412
413
414
415
416
    @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))

417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
    @utils.positional_deprecated
    def fewshot_context(self, doc, num_fewshot, rnd=None):
        """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.
        :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`.
        :returns: str
            The fewshot context.
        """
        assert (
            rnd is not None
        ), "A `random.Random` generator argument must be provided to `rnd`"

        if num_fewshot == 0:
            labeled_examples = ""
        else:
Lintang Sutawika's avatar
Lintang Sutawika committed
439
            labeled_examples = self.sampler.get_context(doc, num_fewshot)
440
441

            # for sets with no training docs, draw from other set *but ensure no overlap with current doc*
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
            # 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"
            # )
466
467
468
469
470
471

        example = self.doc_to_text(doc)
        return labeled_examples + example

    def apply_filters(self):

lintangsutawika's avatar
lintangsutawika committed
472
473
474
475
476
477
        if hasattr(self, "_filters"):
            for f in self._filters:
                f.apply(self._instances)
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances
478

479
    def dump_config(self):
480
        """Returns a dictionary representing the task's config.
481
482
483
484
485
486
487
488

        :returns: str
            The fewshot context.
        """
        # TODO: this should only return the overrides applied to a non-YAML task's configuration.
        # (batch size, num_fewshot)
        return self._config.to_dict()

489
490
491

class ConfigurableTask(Task):

492
    VERSION = "Yaml"
493
    OUTPUT_TYPE = None
494
    CONFIG = None
495
496
497
498

    def __init__(
        self, data_dir=None, cache_dir=None, download_mode=None, config: dict = None
    ):
499
        # Get pre-configured attributes
500
        self._config = self.CONFIG
501

502
503
        # Use new configurations if there was no preconfiguration
        if self._config is None:
504
            self._config = TaskConfig(**config)
505
506
        # Overwrite configs
        else:
lintangsutawika's avatar
lintangsutawika committed
507
            if config is not None:
508
                self._config.__dict__.update(config)
509

510
        if self._config is None:
lintangsutawika's avatar
lintangsutawika committed
511
512
513
            raise ValueError(
                "Must pass a config to ConfigurableTask, either in cls.CONFIG or `config` kwarg"
            )
514
515

        if self._config.output_type is not None:
516
            assert self._config.output_type in ALL_OUTPUT_TYPES
517
518
            self.OUTPUT_TYPE = self._config.output_type

519
520
521
522
523
524
        if self._config.dataset_path is not None:
            self.DATASET_PATH = self._config.dataset_path

        if self._config.dataset_name is not None:
            self.DATASET_NAME = self._config.dataset_name

525
526
527
528
        self._metric_fn_list = {}
        self._metric_fn_kwargs = {}
        self._aggregation_list = {}
        self._higher_is_better = {}
529

530
        _metric_list = DEFAULT_METRIC_REGISTRY[self._config.output_type]
531
        if self._config.metric_list is None:
532
            # TODO: handle this in TaskConfig.__post_init__ ?
533
534
            for metric_name in _metric_list:
                self._metric_fn_list[metric_name] = METRIC_REGISTRY[metric_name]
lintangsutawika's avatar
lintangsutawika committed
535
536
537
                self._aggregation_list[metric_name] = DEFAULT_AGGREGATION_REGISTRY[
                    metric_name
                ]
538
539
540
                self._higher_is_better[metric_name] = HIGHER_IS_BETTER_REGISTRY[
                    metric_name
                ]
541
542
543
544
545
546
547
548
549
        else:
            for metric_config in self._config.metric_list:
                assert "metric" in metric_config
                metric_name = metric_config["metric"]
                kwargs = {
                    key: metric_config[key]
                    for key in metric_config
                    if key not in ["metric", "aggregation", "higher_is_better"]
                }
550
                try:
551
                    self._metric_fn_list[metric_name] = METRIC_REGISTRY[metric_name]
552
                except Exception:
553
554
555
556
557
558
559
560
561
562
563
564
565
566
                    eval_logger.warning(
                        f"Metric {metric_name} not found, "
                        "Searching from https://huggingface.co/evaluate-metric"
                    )
                    try:
                        metric_object = evaluate.load(metric_name)
                        self._metric_fn_list[metric_name] = metric_object
                        self._metric_fn_kwargs[metric_name] = kwargs

                    except Exception:
                        raise Warning(
                            "{} not found in the evaluate library!".format(metric_name),
                            "Please check https://huggingface.co/evaluate-metric",
                        )
lintangsutawika's avatar
lintangsutawika committed
567

568
                if "aggregation" in metric_config:
569
                    agg_name = metric_config["aggregation"]
570
571
572
573
574
575
576
577
                    if type(agg_name) == str:
                        self._aggregation_list[metric_name] = AGGREGATION_REGISTRY[
                            agg_name
                        ]
                    elif callable(agg_name):
                        self._aggregation_list[metric_name] = metric_config[
                            "aggregation"
                        ]
578
                else:
579
580
581

                    INV_AGG_REGISTRY = {v: k for k, v in AGGREGATION_REGISTRY.items()}
                    metric_agg = DEFAULT_AGGREGATION_REGISTRY[metric_name]
582
                    eval_logger.warning(
583
584
585
                        f"metric {metric_name} is defined, but aggregation is not. "
                        f"using default "
                        f"aggregation={INV_AGG_REGISTRY[metric_agg]}"
586
                    )
587
                    self._aggregation_list[metric_name] = metric_agg
lintangsutawika's avatar
lintangsutawika committed
588

589
590
591
592
593
594
                if "higher_is_better" in metric_config:
                    self._higher_is_better[metric_name] = metric_config[
                        "higher_is_better"
                    ]
                else:
                    eval_logger.warning(
595
596
597
                        f"metric {metric_name} is defined, but higher_is_better is not. "
                        f"using default "
                        f"higher_is_better={HIGHER_IS_BETTER_REGISTRY[metric_name]}"
598
                    )
599
600
                    self._higher_is_better[metric_name] = HIGHER_IS_BETTER_REGISTRY[
                        metric_name
lintangsutawika's avatar
lintangsutawika committed
601
                    ]
602

603
        self.download(self._config.dataset_kwargs)
604
605
606
        self._training_docs = None
        self._fewshot_docs = None

lintangsutawika's avatar
lintangsutawika committed
607
        if self._config.filter_list is not None:
lintangsutawika's avatar
lintangsutawika committed
608
            self._filters = []
lintangsutawika's avatar
lintangsutawika committed
609
610
611
612
613
614
615
616
            for filter_config in self._config.filter_list:
                for filter_pipeline in filter_config:
                    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"
lintangsutawika's avatar
lintangsutawika committed
617
618
619
                        }
                        components.append([function["function"], kwargs])
                    filter_pipeline = build_filter_ensemble(filter_name, components)
lintangsutawika's avatar
lintangsutawika committed
620
                self._filters.append(filter_pipeline)
lintangsutawika's avatar
lintangsutawika committed
621
        else:
622
            self._filters = [build_filter_ensemble("none", [["take_first", None]])]
623
624

        if self._config.use_prompt is not None:
lintangsutawika's avatar
lintangsutawika committed
625
            eval_logger.info(f"loading prompt {self._config.use_prompt}")
626
            self.prompt = get_prompt(
lintangsutawika's avatar
lintangsutawika committed
627
628
                self._config.use_prompt, self.DATASET_PATH, self.DATASET_NAME
            )
629
630
631
        else:
            self.prompt = None

lintangsutawika's avatar
lintangsutawika committed
632
633
634
        if self.fewshot_docs() is not None:
            self.sampler = samplers.Sampler(
                list(self.fewshot_docs()), self, rnd=random.Random()
635
            )
636

637
638
639
640
641
642
643
644
    def download(self, dataset_kwargs=None):

        self.dataset = datasets.load_dataset(
            path=self.DATASET_PATH,
            name=self.DATASET_NAME,
            **dataset_kwargs if dataset_kwargs is not None else {},
        )

645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
    def has_training_docs(self):
        if self._config.training_split is not None:
            return True
        else:
            return False

    def has_validation_docs(self):
        if self._config.validation_split is not None:
            return True
        else:
            return False

    def has_test_docs(self):
        if self._config.test_split is not None:
            return True
        else:
            return False

    def training_docs(self):
        if self._config.training_split is not None:
            return self.dataset[self._config.training_split]

    def validation_docs(self):
        if self._config.validation_split is not None:
            return self.dataset[self._config.validation_split]

    def test_docs(self):
        if self._config.test_split is not None:
            return self.dataset[self._config.test_split]

675
    def fewshot_docs(self):
676
        if self._config.fewshot_split is not None:
677
            return self.dataset[self._config.fewshot_split]
678
679
680
        else:
            if self._config.num_fewshot > 0:
                eval_logger.warning(
haileyschoelkopf's avatar
haileyschoelkopf committed
681
                    f"Task '{self._config.task}': "
682
683
684
685
                    "num_fewshot > 0 but fewshot_split is None. "
                    "using preconfigured rule."
                )
            return super().fewshot_docs()
686

687
688
689
690
691
692
693
    def should_decontaminate(self):
        return self._config.should_decontaminate

    def doc_to_decontamination_query(self, doc):
        if self._config.should_decontaminate:
            return utils.apply_template(self._config.doc_to_decontamination_query, doc)

694
695
696
697
698
699
700
701
702
703
704
705
    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):
706
707
708

        if self.prompt is not None:
            doc_to_text = self.prompt
709
710
        else:
            doc_to_text = self._config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
711

712
713
        if type(doc_to_text) == str:
            return utils.apply_template(doc_to_text, doc)
714
        elif callable(doc_to_text):
715
716
717
            return doc_to_text(doc)
        if hasattr(doc_to_text, "apply"):
            return doc_to_text.apply(doc)[0]
718
        else:
719
            print(type(doc_to_text))
720
            raise TypeError
721
722

    def doc_to_target(self, doc):
723
724
725

        if self.prompt is not None:
            doc_to_target = self.prompt
726
727
728
        else:
            doc_to_target = self._config.doc_to_target

729
730
        if type(doc_to_target) == str:
            return utils.apply_template(doc_to_target, doc)
731
        elif callable(doc_to_target):
732
733
734
            return doc_to_target(doc)
        elif hasattr(doc_to_target, "apply"):
            return doc_to_target.apply(doc)[1]
735
736
        else:
            raise TypeError
737

738
    def gold_alias(self, doc):
739
        # TODO: reevaluate if we need this. implemented to have a
740
        # processed version of answer to put into gsm8k exact_match scoring as ref.
lintangsutawika's avatar
lintangsutawika committed
741
        if self._config.gold_alias is not None:
742
743
            doc_to_target = self._config.gold_alias
        else:
lintangsutawika's avatar
lintangsutawika committed
744
745
            # doc_to_target = self._config.doc_to_target
            return self.doc_to_target(doc)
746
747
748
749
750
751
752
753
754
755

        if type(doc_to_target) == str:
            return utils.apply_template(doc_to_target, doc)
        elif callable(doc_to_target):
            return doc_to_target(doc)
        elif hasattr(doc_to_target, "apply"):
            return doc_to_target.apply(doc)[1]
        else:
            raise TypeError

756
757
    def construct_requests(self, doc, ctx, **kwargs):

758
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
759
            arguments = (ctx, self.doc_to_target(doc))
760
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
761
            arguments = (self.doc_to_target(doc),)
762
        elif self.OUTPUT_TYPE == "multiple_choice":
763
764
            # we pass the user-defined answer_choices var (in aliases) and translate the result to a Python list.
            # TODO: any cleaner way to do this?
lintangsutawika's avatar
lintangsutawika committed
765
766
767
768
769
            choices = ast.literal_eval(
                utils.apply_template(
                    self._config.template_aliases + "{{answer_choices}}", doc
                )
            )
770
            request_list = [
771
772
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
773
                    doc=doc,
774
                    arguments=(ctx, " {}".format(choice)),
775
                    idx=i,
776
777
                    **kwargs,
                )
lintangsutawika's avatar
lintangsutawika committed
778
                for i, choice in enumerate(choices)
779
            ]
780
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
781
            if "acc_mutual_info" in self._metric_fn_list.keys():
782
783
784
                # if we are calculating multiple choice accuracy
                # using mutual information instead of raw loglikelihood as metric, need unconditional lls.

lintangsutawika's avatar
lintangsutawika committed
785
                # here mutual info refers to calculating
786
787
788
789
790
791
                # 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
792
                            doc=doc,
793
794
795
796
                            arguments=("", "{}".format(choice)),
                            idx=i,
                            **kwargs,
                        )
lintangsutawika's avatar
lintangsutawika committed
797
                        for i, choice in enumerate(choices)
798
799
800
                    ]
                )
            return request_list
lintangsutawika's avatar
lintangsutawika committed
801

802
        elif self.OUTPUT_TYPE == "greedy_until":
803
            arguments = (ctx, self._config.generation_kwargs)
lintangsutawika's avatar
lintangsutawika committed
804
805

        return Instance(
lintangsutawika's avatar
lintangsutawika committed
806
807
            request_type=self.OUTPUT_TYPE, doc=doc, arguments=arguments, idx=0, **kwargs
        )
808
809
810

    def process_results(self, doc, results):

lintangsutawika's avatar
lintangsutawika committed
811
812
813
        # if callable(self._config.process_results):
        #     return self._config.process_results(doc, results)

814
        result_dict = {}
815
        use_metric = list(self._metric_fn_list.keys())
816
817
818
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
819
820
821
822
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
823
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
824
            (loglikelihood,) = results
825
826
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
haileyschoelkopf's avatar
haileyschoelkopf committed
827
            return {
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
                **(
                    {"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
843
            }
844
        elif self.OUTPUT_TYPE == "multiple_choice":
845
846

            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
847
848
849
850
851
            if self._config.gold_alias is not None:
                gold = int(self.gold_alias(doc))
            else:
                gold = int(self.doc_to_target(doc))

852
            pred = np.argmax(lls)
853
            # retrieve choices in List[str] form, to compute choice lengths, etc.
lintangsutawika's avatar
lintangsutawika committed
854
855
856
857
858
            choices = ast.literal_eval(
                utils.apply_template(
                    self._config.template_aliases + "{{answer_choices}}", doc
                )
            )
859
860
            if (
                2 * len(choices) == len(lls)
861
                and "acc_mutual_info" in self._metric_fn_list.keys()
862
863
864
865
866
867
868
            ):
                # 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]
869
870

            acc = 1.0 if np.argmax(lls) == gold else 0.0
871
872
            completion_len = np.array([float(len(i)) for i in choices])
            acc_norm = 1.0 if np.argmax(lls / completion_len) == gold else 0.0
873
874

            result_dict = {
875
                **({"acc": acc} if "acc" in use_metric else {}),
haileyschoelkopf's avatar
haileyschoelkopf committed
876
877
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
878
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
879
880
881
            }

            # TODO: set which normalization metrics should be reported, and calculate them
882
            if "exact_match" in self._metric_fn_list.keys():
883
                # TODO: this gets score of 0 on arc_challenge for pythia-70m. need to test that this works properly
lintangsutawika's avatar
lintangsutawika committed
884
                is_greedy = is_greedy[gold]  # take value for the gold answer
885
886
                result_dict["exact_match"] = int(is_greedy)

887
            if "acc_mutual_info" in use_metric:
lintangsutawika's avatar
lintangsutawika committed
888
889
890
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
891
892
893
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

894
895
896
        elif self.OUTPUT_TYPE == "greedy_until":

            if self._config.gold_alias is not None:
897
                gold = self.gold_alias(doc)
898
899
900
            else:
                gold = self.doc_to_target(doc)

901
902
            for key, result in zip(self._metric_fn_list.keys(), results):
                _dict = self._metric_fn_list[key].compute(
lintangsutawika's avatar
lintangsutawika committed
903
                    references=[gold], predictions=[result], **self._metric_kwargs[key]
904
                )
905

lintangsutawika's avatar
lintangsutawika committed
906
                result_dict = {**result_dict, **_dict}
907
        else:
lintangsutawika's avatar
lintangsutawika committed
908
909
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
910
                "'loglikelihood', 'loglikelihood_rolling', 'greedy_until', or 'multiple_choice'",
911
            )
912
913
914
915
916
917
918

        return result_dict

    def aggregation(self):
        return self._aggregation_list

    def higher_is_better(self):
haileyschoelkopf's avatar
haileyschoelkopf committed
919
        return self._higher_is_better
920
921
922
923
924
925
926
927
928
929


class MultipleChoiceTask(Task):

    OUTPUT_TYPE: str = "loglikelihood"

    def doc_to_target(self, doc):
        return " " + doc["choices"][doc["gold"]]

    def construct_requests(self, doc, ctx, **kwargs):
930
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
931
932
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
933
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
934
                doc=doc,
935
                arguments=(ctx, " {}".format(choice)),
936
                idx=i,
937
938
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
939
940
            for i, choice in enumerate(doc["choices"])
        ]
941
942

    def process_results(self, doc, results):
lintangsutawika's avatar
lintangsutawika committed
943
944
945
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
        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,
        }

    def higher_is_better(self):
        return {
            "acc": True,
            "acc_norm": True,
        }

    def aggregation(self):
        return {
            "acc": mean,
            "acc_norm": mean,
        }


lintangsutawika's avatar
lintangsutawika committed
970
class PerplexityTask(Task):
971
972
973
974
975
976
977
978
979
980

    OUTPUT_TYPE = "loglikelihood_rolling"

    def has_training_docs(self):
        return False

    def fewshot_examples(self, k, rnd):
        assert k == 0
        return []

lintangsutawika's avatar
lintangsutawika committed
981
    def fewshot_context(self, doc, num_fewshot, rnd=None):
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
        assert (
            num_fewshot == 0
        ), "The number of fewshot examples must be 0 for perplexity tasks."
        assert (
            rnd is not None
        ), "A `random.Random` generator argument must be provided to `rnd`."

        return ""

    def higher_is_better(self):
        return {
            "word_perplexity": False,
            "byte_perplexity": False,
            "bits_per_byte": False,
        }

    def doc_to_decontamination_query(self, doc):
        return doc

    def doc_to_text(self, doc):
        return ""

    def doc_to_target(self, doc):
        return doc

    def construct_requests(self, doc, ctx, **kwargs):
        assert not ctx

lintangsutawika's avatar
lintangsutawika committed
1010
1011
1012
1013
1014
1015
1016
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1017
1018
1019

    def process_results(self, doc, results):
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1020
1021
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
        return {
            "word_perplexity": (loglikelihood, words),
            "byte_perplexity": (loglikelihood, bytes_),
            "bits_per_byte": (loglikelihood, bytes_),
        }

    def aggregation(self):
        return {
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
            "bits_per_byte": bits_per_byte,
        }

    @classmethod
    def count_bytes(cls, doc):
        return len(doc.encode("utf-8"))

    @classmethod
    def count_words(cls, doc):
        """Downstream tasks with custom word boundaries should override this!"""
        return len(re.split(r"\s+", doc))