task.py 38.8 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
ALL_OUTPUT_TYPES = [
    "loglikelihood",
    "multiple_choice",
    "loglikelihood_rolling",
    "greedy_until",
47
    "winograd_schema"
48
49
]

50
51
52
53

@dataclass
class TaskConfig(dict):

54
    task: str = None
55
56
    group: Union[str, list] = None

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
67
    doc_to_text: Union[Callable, str] = None
    doc_to_target: Union[Callable, str] = None
68
    use_prompt: str = None
69
    description: str = ""
70
71
    target_delimiter: str = " "
    fewshot_delimiter: str = "\n\n"
72

73
74
    num_fewshot: int = 0
    batch_size: int = 1
75
76
    repeats: int = 1

77
    metric_list: str = None
lintangsutawika's avatar
lintangsutawika committed
78
    gold_alias: Union[Callable, str] = None
79
    create_choices: Union[Callable, str] = None
80
    output_type: str = "greedy_until"
81
    generation_kwargs: dict = None
lintangsutawika's avatar
lintangsutawika committed
82
    filter_list: Union[str, list] = None
83
84
    should_decontaminate: bool = False
    doc_to_decontamination_query: str = None
85

lintangsutawika's avatar
lintangsutawika committed
86
    metadata: str = None  # by default, not used in the code. allows for users to pass arbitrary info to tasks
87

88
89
90
91
    def __post_init__(self):
        # allow user-specified aliases so that users can
        # force prompt-compatibility for some prompt regardless of
        # field names in prompt
92
93
94
        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
95

96
97
            if type(self.doc_to_target) == str:
                self.doc_to_target = self.template_aliases + self.doc_to_target
98

99
            if type(self.gold_alias) == str:
lintangsutawika's avatar
lintangsutawika committed
100
                self.gold_alias = self.template_aliases + self.gold_alias
101

haileyschoelkopf's avatar
haileyschoelkopf committed
102
        if self.generation_kwargs:
103
104
105
            assert (
                self.output_type == "greedy_until"
            ), "passed `generation_kwargs`, but not using a generation request type!"
haileyschoelkopf's avatar
haileyschoelkopf committed
106
        elif self.output_type == "greedy_until":
107
108
            # ensure that we greedily generate in absence of explicit arguments otherwise
            self.generation_kwargs = {"do_sample": False, "temperature": 0.0}
109

haileyschoelkopf's avatar
haileyschoelkopf committed
110
111
        # TODO: how to make TaskConfigs be de- and re-serializable, even when using the !function constructor?

112
113
114
    def __getitem__(self, item):
        return getattr(self, item)

115
    def to_dict(self):
116
117
        """dumps the current config as a dictionary object, as a printable format.
        null fields will not be printed.
haileyschoelkopf's avatar
haileyschoelkopf committed
118
        Used for dumping results alongside full task configuration
119

haileyschoelkopf's avatar
haileyschoelkopf committed
120
121
122
123
124
125
126
127
128
129
        :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)
haileyschoelkopf's avatar
haileyschoelkopf committed
130
131
132
            elif isinstance(v, Callable):
                # TODO: this should handle Promptsource template objects as a separate case?
                cfg_dict[k] = str(v)
haileyschoelkopf's avatar
haileyschoelkopf committed
133
        return cfg_dict
134

135
136
137
138
139
140
141
142
143
144
145
146

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
147

148
149
150
151
152
153
154
155
    # 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
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
184
185
186
187
188
189
190
    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
191
        self._config = TaskConfig(**config) if config else TaskConfig()
192
193
194

        if not hasattr(self, "_filters"):
            self._filters = []
lintangsutawika's avatar
lintangsutawika committed
195
            for name, components in self._config.get(
196
                "filters", [["none", [["take_first", None]]]]
lintangsutawika's avatar
lintangsutawika committed
197
            ):
198
199
200
                filter_pipeline = build_filter_ensemble(name, components)
                self._filters.append(filter_pipeline)

lintangsutawika's avatar
lintangsutawika committed
201
202
203
        self.sampler = samplers.Sampler(
            list(self.fewshot_docs()), self, rnd=random.Random()
        )  # TODO: pass the correct docs in here
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229

    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.
        """
230
231
232
233
234
235
236
        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,
        )
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
267
268
269
270
271
272
273

    @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 []

274
275
276
277
278
279
280
281
282
283
    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
284
            eval_logger.warning(
285
                "has_training_docs and has_validation_docs are False"
286
                ", using test_docs but this is not recommended."
lintangsutawika's avatar
lintangsutawika committed
287
            )
288
289
            return self.test_docs()

290
291
292
293
294
295
296
297
298
299
    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
300
301
302
303
304
305
306
307
308
309
    
    def create_choices(self, doc):
        if self._config.create_choices is None:
            return ast.literal_eval(
                    utils.apply_template(
                        self._config.template_aliases + "{{answer_choices}}", doc
                        )
                    )
        else:
            return self._config.create_choices(doc)
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337

    @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

338
    def build_all_requests(self, limit=None, rank=None, world_size=None):
339
340
341
342
343
344
345
346
347
348
349
        """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 = []
350
351
        for doc_id, doc in utils.create_iterator(
            enumerate(docs), rank, world_size, limit
lintangsutawika's avatar
lintangsutawika committed
352
        ):
353
            # sample fewshot context #TODO: need to offset doc_id by rank now!
354
355
356
            fewshot_ctx = self.fewshot_context(
                doc, self._config.num_fewshot, rnd=random.Random()
            )
357
            # TODO: we should override this if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
358
359
360
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
361
                metadata=(self._config["task"], doc_id, self._config.repeats),
lintangsutawika's avatar
lintangsutawika committed
362
            )
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387

            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
388
            The number of times each instance in a dataset is inferred on. Defaults to 1,
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
            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
424
425
426
427
428
429
430
431
432
433
    @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))

434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
    @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:
454
455
            # always prepend the (possibly empty) task description
            labeled_examples = self._config.description
456
        else:
lintangsutawika's avatar
lintangsutawika committed
457
458
459
            labeled_examples = self._config.description + self.sampler.get_context(
                doc, num_fewshot
            )
460
461
462
463
464
465

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

    def apply_filters(self):

lintangsutawika's avatar
lintangsutawika committed
466
467
468
469
470
471
        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
472

473
    def dump_config(self):
474
        """Returns a dictionary representing the task's config.
475
476
477
478
479
480
481
482

        :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()

483
484
485

class ConfigurableTask(Task):

486
    VERSION = "Yaml"
487
    OUTPUT_TYPE = None
488
    CONFIG = None
489
490
491
492

    def __init__(
        self, data_dir=None, cache_dir=None, download_mode=None, config: dict = None
    ):
493
        # Get pre-configured attributes
494
        self._config = self.CONFIG
495

496
497
        # Use new configurations if there was no preconfiguration
        if self._config is None:
498
            self._config = TaskConfig(**config)
499
500
        # Overwrite configs
        else:
lintangsutawika's avatar
lintangsutawika committed
501
            if config is not None:
502
                self._config.__dict__.update(config)
503

504
        if self._config is None:
lintangsutawika's avatar
lintangsutawika committed
505
506
507
            raise ValueError(
                "Must pass a config to ConfigurableTask, either in cls.CONFIG or `config` kwarg"
            )
508
509

        if self._config.output_type is not None:
510
            assert self._config.output_type in ALL_OUTPUT_TYPES
511
512
            self.OUTPUT_TYPE = self._config.output_type

513
514
515
516
517
518
        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

519
520
521
522
        self._metric_fn_list = {}
        self._metric_fn_kwargs = {}
        self._aggregation_list = {}
        self._higher_is_better = {}
523

524
        _metric_list = DEFAULT_METRIC_REGISTRY[self._config.output_type]
525
        if self._config.metric_list is None:
526
            # TODO: handle this in TaskConfig.__post_init__ ?
527
528
            for metric_name in _metric_list:
                self._metric_fn_list[metric_name] = METRIC_REGISTRY[metric_name]
lintangsutawika's avatar
lintangsutawika committed
529
530
531
                self._aggregation_list[metric_name] = DEFAULT_AGGREGATION_REGISTRY[
                    metric_name
                ]
532
533
534
                self._higher_is_better[metric_name] = HIGHER_IS_BETTER_REGISTRY[
                    metric_name
                ]
535
536
537
538
539
540
541
542
543
        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"]
                }
544
                try:
545
                    self._metric_fn_list[metric_name] = METRIC_REGISTRY[metric_name]
546
                except Exception:
547
548
549
550
551
552
553
554
555
556
557
558
559
560
                    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
561

562
                if "aggregation" in metric_config:
563
                    agg_name = metric_config["aggregation"]
564
565
566
567
568
569
570
571
                    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"
                        ]
572
                else:
573
574
575

                    INV_AGG_REGISTRY = {v: k for k, v in AGGREGATION_REGISTRY.items()}
                    metric_agg = DEFAULT_AGGREGATION_REGISTRY[metric_name]
576
                    eval_logger.warning(
577
578
579
                        f"metric {metric_name} is defined, but aggregation is not. "
                        f"using default "
                        f"aggregation={INV_AGG_REGISTRY[metric_agg]}"
580
                    )
581
                    self._aggregation_list[metric_name] = metric_agg
lintangsutawika's avatar
lintangsutawika committed
582

583
584
585
586
587
588
                if "higher_is_better" in metric_config:
                    self._higher_is_better[metric_name] = metric_config[
                        "higher_is_better"
                    ]
                else:
                    eval_logger.warning(
589
590
591
                        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]}"
592
                    )
593
594
                    self._higher_is_better[metric_name] = HIGHER_IS_BETTER_REGISTRY[
                        metric_name
lintangsutawika's avatar
lintangsutawika committed
595
                    ]
596

597
        self.download(self._config.dataset_kwargs)
598
599
600
        self._training_docs = None
        self._fewshot_docs = None

lintangsutawika's avatar
lintangsutawika committed
601
        if self._config.filter_list is not None:
lintangsutawika's avatar
lintangsutawika committed
602
            self._filters = []
lintangsutawika's avatar
lintangsutawika committed
603
604
605
606
607
608
609
610
            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
611
612
613
                        }
                        components.append([function["function"], kwargs])
                    filter_pipeline = build_filter_ensemble(filter_name, components)
lintangsutawika's avatar
lintangsutawika committed
614
                self._filters.append(filter_pipeline)
lintangsutawika's avatar
lintangsutawika committed
615
        else:
616
            self._filters = [build_filter_ensemble("none", [["take_first", None]])]
617
618

        if self._config.use_prompt is not None:
lintangsutawika's avatar
lintangsutawika committed
619
            eval_logger.info(f"loading prompt {self._config.use_prompt}")
620
            self.prompt = get_prompt(
lintangsutawika's avatar
lintangsutawika committed
621
622
                self._config.use_prompt, self.DATASET_PATH, self.DATASET_NAME
            )
623
624
625
        else:
            self.prompt = None

lintangsutawika's avatar
lintangsutawika committed
626
627
628
        if self.fewshot_docs() is not None:
            self.sampler = samplers.Sampler(
                list(self.fewshot_docs()), self, rnd=random.Random()
629
            )
630

631
632
633
634
635
636
637
638
    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 {},
        )

639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
    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]

669
    def fewshot_docs(self):
670
        if self._config.fewshot_split is not None:
671
            return self.dataset[self._config.fewshot_split]
672
673
674
        else:
            if self._config.num_fewshot > 0:
                eval_logger.warning(
haileyschoelkopf's avatar
haileyschoelkopf committed
675
                    f"Task '{self._config.task}': "
676
677
678
679
                    "num_fewshot > 0 but fewshot_split is None. "
                    "using preconfigured rule."
                )
            return super().fewshot_docs()
680

681
682
683
684
685
686
687
    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)

688
689
690
691
692
693
694
695
696
697
698
699
    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):
700
701
702

        if self.prompt is not None:
            doc_to_text = self.prompt
703
704
        else:
            doc_to_text = self._config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
705

706
707
        if type(doc_to_text) == str:
            return utils.apply_template(doc_to_text, doc)
708
        elif callable(doc_to_text):
709
710
711
            return doc_to_text(doc)
        if hasattr(doc_to_text, "apply"):
            return doc_to_text.apply(doc)[0]
712
        else:
713
            print(type(doc_to_text))
714
            raise TypeError
715
716

    def doc_to_target(self, doc):
717
718
719

        if self.prompt is not None:
            doc_to_target = self.prompt
720
721
722
        else:
            doc_to_target = self._config.doc_to_target

723
724
        if type(doc_to_target) == str:
            return utils.apply_template(doc_to_target, doc)
725
        elif callable(doc_to_target):
726
727
728
            return doc_to_target(doc)
        elif hasattr(doc_to_target, "apply"):
            return doc_to_target.apply(doc)[1]
729
730
        else:
            raise TypeError
731

732
    def gold_alias(self, doc):
733
734
735
736
737
        # returns a version of the gold target answer to a document,
        # which should be passed into metric for scoring as the ground truth.

        # in multiple_choice tasks, this should be castable to an int corresponding to the index
        # within the answer choices, while doc_to_target is the string version of {{answer_choices[gold]}}.
lintangsutawika's avatar
lintangsutawika committed
738
        if self._config.gold_alias is not None:
739
740
            doc_to_target = self._config.gold_alias
        else:
lintangsutawika's avatar
lintangsutawika committed
741
            return self.doc_to_target(doc)
742
743
744
745
746
747
748
749
750
751

        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

752
753
    def construct_requests(self, doc, ctx, **kwargs):

754
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
755
            arguments = (ctx, self.doc_to_target(doc))
756
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
757
            arguments = (self.doc_to_target(doc),)
758
        elif self.OUTPUT_TYPE == "multiple_choice":
759
760
            # 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?
761
762
            choices = self.create_choices(doc)
            
763
            request_list = [
764
765
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
766
                    doc=doc,
767
                    arguments=(ctx, " {}".format(choice)),
768
                    idx=i,
769
770
                    **kwargs,
                )
lintangsutawika's avatar
lintangsutawika committed
771
                for i, choice in enumerate(choices)
772
            ]
773
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
774
            if "acc_mutual_info" in self._metric_fn_list.keys():
775
776
777
                # if we are calculating multiple choice accuracy
                # using mutual information instead of raw loglikelihood as metric, need unconditional lls.

lintangsutawika's avatar
lintangsutawika committed
778
                # here mutual info refers to calculating
779
780
781
782
783
784
                # 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
785
                            doc=doc,
786
                            arguments=("", "{}".format(choice)),
787
788
789
                            idx=i,
                            **kwargs,
                        )
lintangsutawika's avatar
lintangsutawika committed
790
                        for i, choice in enumerate(choices)
791
792
793
                    ]
                )
            return request_list
lintangsutawika's avatar
lintangsutawika committed
794

795
        elif self.OUTPUT_TYPE == "greedy_until":
796
            arguments = (ctx, self._config.generation_kwargs)
lintangsutawika's avatar
lintangsutawika committed
797

798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
        elif self.OUTPUT_TYPE == "winograd_schema":
            # similar to multiple_choice task type except each request contains
            # multiple differing contexts with the same continuation

            contexts = self.create_choices(doc)
            choice = self.doc_to_target(doc)
            
            request_list = [
                Instance(
                    request_type="loglikelihood",
                    doc=doc,
                    arguments=(context, " {}".format(choice)),
                    idx=i,
                    **kwargs,
                )
                for i, context in enumerate(contexts)
            ]
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
            if "acc_mutual_info" in self._metric_fn_list.keys():
                # if we are calculating multiple choice accuracy
                # using mutual information instead of raw loglikelihood as metric, need unconditional lls.

                # here mutual info refers to calculating
                # 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",
                            doc=doc,
                            arguments=("", "{}".format(choice)),
                            idx=i,
                            **kwargs,
                        )
                        for i, choice in enumerate(choices)
                    ]
                )
            return request_list

lintangsutawika's avatar
lintangsutawika committed
837
        return Instance(
lintangsutawika's avatar
lintangsutawika committed
838
839
            request_type=self.OUTPUT_TYPE, doc=doc, arguments=arguments, idx=0, **kwargs
        )
840
841
842

    def process_results(self, doc, results):

lintangsutawika's avatar
lintangsutawika committed
843
844
845
        # if callable(self._config.process_results):
        #     return self._config.process_results(doc, results)

846
        result_dict = {}
847
        use_metric = list(self._metric_fn_list.keys())
848
849
850
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
851
852
853
854
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
855
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
856
            (loglikelihood,) = results
857
858
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
haileyschoelkopf's avatar
haileyschoelkopf committed
859
            return {
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
                **(
                    {"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
875
            }
876
        elif self.OUTPUT_TYPE == "multiple_choice":
877
878

            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
879
880
881
882
883
            if self._config.gold_alias is not None:
                gold = int(self.gold_alias(doc))
            else:
                gold = int(self.doc_to_target(doc))

884
            pred = np.argmax(lls)
885
            # retrieve choices in List[str] form, to compute choice lengths, etc.
886
            choices = self.create_choices(doc)
887
888
            if (
                2 * len(choices) == len(lls)
889
                and "acc_mutual_info" in self._metric_fn_list.keys()
890
891
892
893
894
895
896
            ):
                # 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]
897
898

            acc = 1.0 if np.argmax(lls) == gold else 0.0
899
900
            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
901
902

            result_dict = {
903
                **({"acc": acc} if "acc" in use_metric else {}),
haileyschoelkopf's avatar
haileyschoelkopf committed
904
905
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
906
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
907
908
909
            }

            # TODO: set which normalization metrics should be reported, and calculate them
910
            if "exact_match" in self._metric_fn_list.keys():
911
                # TODO: this gets score of 0 on arc_challenge for pythia-70m. need to test that this works properly
lintangsutawika's avatar
lintangsutawika committed
912
                is_greedy = is_greedy[gold]  # take value for the gold answer
913
914
                result_dict["exact_match"] = int(is_greedy)

915
            if "acc_mutual_info" in use_metric:
lintangsutawika's avatar
lintangsutawika committed
916
917
918
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
919
920
921
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
        elif self.OUTPUT_TYPE == "winograd_schema":

            lls, is_greedy = zip(*results)
            if self._config.gold_alias is not None:
                gold = int(self.gold_alias(doc))
            else:
                gold = int(self.doc_to_target(doc))

            pred = np.argmax(lls)
            acc = 1.0 if np.argmax(lls) == gold else 0.0

            result_dict = {
                **({"acc": acc} if "acc" in use_metric else {}),
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
            }

940
941
942
        elif self.OUTPUT_TYPE == "greedy_until":

            if self._config.gold_alias is not None:
943
                gold = self.gold_alias(doc)
944
945
946
            else:
                gold = self.doc_to_target(doc)

947
948
            for key, result in zip(self._metric_fn_list.keys(), results):
                _dict = self._metric_fn_list[key].compute(
haileyschoelkopf's avatar
haileyschoelkopf committed
949
950
951
                    references=[gold],
                    predictions=[result],
                    **self._metric_fn_kwargs[key],
952
                )
953

lintangsutawika's avatar
lintangsutawika committed
954
                result_dict = {**result_dict, **_dict}
955
        else:
lintangsutawika's avatar
lintangsutawika committed
956
957
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
958
                "'loglikelihood', 'loglikelihood_rolling', 'greedy_until', 'multiple_choice' or 'winograd_schema' ",
959
            )
960
961
962
963
964
965
966

        return result_dict

    def aggregation(self):
        return self._aggregation_list

    def higher_is_better(self):
haileyschoelkopf's avatar
haileyschoelkopf committed
967
        return self._higher_is_better
968
969
970
971
972
973
974
975
976
977


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):
978
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
979
980
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
981
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
982
                doc=doc,
983
                arguments=(ctx, " {}".format(choice)),
984
                idx=i,
985
986
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
987
988
            for i, choice in enumerate(doc["choices"])
        ]
989
990

    def process_results(self, doc, results):
lintangsutawika's avatar
lintangsutawika committed
991
992
993
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
        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
1018
class PerplexityTask(Task):
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028

    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
1029
    def fewshot_context(self, doc, num_fewshot, rnd=None):
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
        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
1058
1059
1060
1061
1062
1063
1064
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1065
1066
1067

    def process_results(self, doc, results):
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1068
1069
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
        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))