task.py 49.2 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
from tqdm import tqdm
12
13
14
15

import datasets
import numpy as np

baberabb's avatar
baberabb committed
16
from typing import Union, List, Any, Tuple, Literal
17
from collections.abc import Callable
18

19
from lm_eval import utils
20
from lm_eval.api import samplers
haileyschoelkopf's avatar
haileyschoelkopf committed
21
from lm_eval.api.instance import Instance
lintangsutawika's avatar
lintangsutawika committed
22
from lm_eval.api.filter import FilterEnsemble
23
24
25
26

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
27
28
29
30
from lm_eval.api.metrics import (
    mean,
    weighted_perplexity,
    bits_per_byte,
lintangsutawika's avatar
lintangsutawika committed
31
    metric_max_over_ground_truths,
lintangsutawika's avatar
lintangsutawika committed
32
33
)
from lm_eval.api.registry import (
haileyschoelkopf's avatar
haileyschoelkopf committed
34
35
    get_metric,
    get_aggregation,
36
    get_metric_aggregation,
haileyschoelkopf's avatar
haileyschoelkopf committed
37
    is_higher_better,
38
39
    DEFAULT_METRIC_REGISTRY,
    OUTPUT_TYPE_REGISTRY,
lintangsutawika's avatar
lintangsutawika committed
40
41
    AGGREGATION_REGISTRY,
)
42

43
44
45
46
ALL_OUTPUT_TYPES = [
    "loglikelihood",
    "multiple_choice",
    "loglikelihood_rolling",
47
    "generate_until",
48
49
]

50
51
52

@dataclass
class TaskConfig(dict):
53
    # task naming/registry
54
    task: str = None
lintangsutawika's avatar
lintangsutawika committed
55
    task_alias: str = None
56
    group: Union[str, list] = None
lintangsutawika's avatar
lintangsutawika committed
57
    group_alias: Union[str, list] = None
58
59
60
    # HF dataset options.
    # which dataset to use,
    # and what splits for what purpose
61
62
    dataset_path: str = None
    dataset_name: str = None
63
    dataset_kwargs: dict = None
64
65
66
    training_split: str = None
    validation_split: str = None
    test_split: str = None
lintangsutawika's avatar
lintangsutawika committed
67
    fewshot_split: str = None  # TODO: assert that this not None if num_fewshot > 0. (?) assert if this is same split as one evaling (?)
68
69
    # formatting / prompting options.
    # see docs/advanced_task_guide.md for more info
70
    process_docs: Callable = None
71
72
    doc_to_text: Union[Callable, str] = None
    doc_to_target: Union[Callable, str] = None
lintangsutawika's avatar
lintangsutawika committed
73
    doc_to_choice: Union[Callable, str, dict, list] = None
lintangsutawika's avatar
lintangsutawika committed
74
    process_results: Union[Callable, str] = None
75
    use_prompt: str = None
76
    description: str = ""
77
78
    target_delimiter: str = " "
    fewshot_delimiter: str = "\n\n"
haileyschoelkopf's avatar
haileyschoelkopf committed
79
    fewshot_config: dict = None
80
    # runtime configuration options
81
    num_fewshot: int = 0
82
    # scoring options
83
    metric_list: list = None
84
    output_type: str = "generate_until"
85
    generation_kwargs: dict = None
86
    repeats: int = 1
lintangsutawika's avatar
lintangsutawika committed
87
    filter_list: Union[str, list] = None
88
89
    should_decontaminate: bool = False
    doc_to_decontamination_query: str = None
90

lintangsutawika's avatar
lintangsutawika committed
91
    metadata: str = None  # by default, not used in the code. allows for users to pass arbitrary info to tasks
92

Ethan Smith's avatar
Ethan Smith committed
93
    def __post_init__(self) -> None:
lintangsutawika's avatar
lintangsutawika committed
94
        if self.dataset_path and ("." in self.dataset_path):
lintangsutawika's avatar
lintangsutawika committed
95
96
            import inspect
            from importlib import import_module
lintangsutawika's avatar
format  
lintangsutawika committed
97

lintangsutawika's avatar
lintangsutawika committed
98
            self.dataset_path = inspect.getfile(import_module(self.dataset_path))
99

Lintang Sutawika's avatar
Lintang Sutawika committed
100
        if self.generation_kwargs is not None:
101
            if self.output_type != "generate_until":
Lintang Sutawika's avatar
Lintang Sutawika committed
102
                eval_logger.warning(
103
                    f"[{self.task}] passed `generation_kwargs`, but not using `output_type: generate_until`!"
Lintang Sutawika's avatar
Lintang Sutawika committed
104
                )
105
                assert self.output_type != "generate_until"
Lintang Sutawika's avatar
Lintang Sutawika committed
106
107
108
109
110
111
112

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

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

haileyschoelkopf's avatar
haileyschoelkopf committed
124
125
        # TODO: how to make TaskConfigs be de- and re-serializable, even when using the !function constructor?

126
127
128
    def __getitem__(self, item):
        return getattr(self, item)

129
130
131
    def __setitem__(self, item, value):
        return setattr(self, item, value)

132
    def to_dict(self):
133
134
        """dumps the current config as a dictionary object, as a printable format.
        null fields will not be printed.
haileyschoelkopf's avatar
haileyschoelkopf committed
135
        Used for dumping results alongside full task configuration
136

haileyschoelkopf's avatar
haileyschoelkopf committed
137
138
139
140
141
142
143
144
145
146
        :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
147
148
149
            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
150
        return cfg_dict
151

152
153
154
155
156
157
158
159
160
161
162
163

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
164

165
166
167
168
169
170
171
172
    # 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
173

174
175
176
177
178
179
    def __init__(
        self,
        data_dir=None,
        cache_dir=None,
        download_mode=None,
        config=None,
Ethan Smith's avatar
Ethan Smith committed
180
    ) -> None:
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
        """
        :param data_dir: str
            Stores the path to a local folder containing the `Task`'s data files.
            Use this to specify the path to manually downloaded data (usually when
            the dataset is not publicly accessible).
        :param cache_dir: str
            The directory to read/write the `Task` dataset. This follows the
            HuggingFace `datasets` API with the default cache directory located at:
                `~/.cache/huggingface/datasets`
            NOTE: You can change the cache location globally for a given process
            to another directory:
                `export HF_DATASETS_CACHE="/path/to/another/directory"`
        :param download_mode: datasets.DownloadMode
            How to treat pre-existing `Task` downloads and data.
            - `datasets.DownloadMode.REUSE_DATASET_IF_EXISTS`
                Reuse download and reuse dataset.
            - `datasets.DownloadMode.REUSE_CACHE_IF_EXISTS`
                Reuse download with fresh dataset.
            - `datasets.DownloadMode.FORCE_REDOWNLOAD`
                Fresh download and fresh dataset.
        """
        self.download(data_dir, cache_dir, download_mode)
        self._training_docs = None
        self._fewshot_docs = None
        self._instances = None

lintangsutawika's avatar
lintangsutawika committed
207
        self._config = TaskConfig({**config}) if config else TaskConfig()
208

lintangsutawika's avatar
lintangsutawika committed
209
        self._filters = [build_filter_ensemble("none", [["take_first", None]])]
210

Ethan Smith's avatar
Ethan Smith committed
211
    def download(self, data_dir=None, cache_dir=None, download_mode=None) -> None:
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
        """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.
        """
236
237
238
239
240
241
242
        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,
        )
243

244
245
246
247
248
    @property
    def config(self):
        """Returns the TaskConfig associated with this class."""
        return self._config

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
274
275
276
277
278
279
280
281
282
283
284
    @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 []

285
286
287
288
289
290
291
292
293
294
    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
295
            eval_logger.warning(
296
                "has_training_docs and has_validation_docs are False"
297
                ", using test_docs as fewshot_docs but this is not recommended."
lintangsutawika's avatar
lintangsutawika committed
298
            )
299
300
            return self.test_docs()

301
302
303
304
305
306
307
308
309
310
    def _process_doc(self, doc):
        """
        Override this to process (detokenize, strip, replace, etc.) individual
        documents. This can be used in a map over documents of a data split.
        E.g. `map(self._process_doc, self.dataset["validation"])`

        :return: dict
            The processed version of the specified `doc`.
        """
        return doc
lintangsutawika's avatar
lintangsutawika committed
311

312
313
314
315
316
317
318
319
320
321
322
323
324
    @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)

Ethan Smith's avatar
Ethan Smith committed
325
    def doc_to_decontamination_query(self, doc) -> None:
326
327
328
329
330
331
332
333
334
335
336
337
338
        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

Ethan Smith's avatar
Ethan Smith committed
339
    def build_all_requests(self, limit=None, rank=None, world_size=None) -> None:
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!"

350
        eval_logger.info(f"Building contexts for task on rank {rank}...")
351

352
        instances = []
353
354
        for doc_id, doc in utils.create_iterator(
            enumerate(docs), rank, world_size, limit
lintangsutawika's avatar
lintangsutawika committed
355
        ):
356
            # sample fewshot context #TODO: need to offset doc_id by rank now!
357
            fewshot_ctx = self.fewshot_context(
358
                doc,
359
                self.config.num_fewshot,
360
            )
361

362
            # TODO: we should override self.config.repeats if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
363
364
365
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
366
                metadata=(self.config["task"], doc_id, self.config.repeats),
lintangsutawika's avatar
lintangsutawika committed
367
            )
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392

            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
393
            The number of times each instance in a dataset is inferred on. Defaults to 1,
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
424
425
426
427
428
            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
429
430
431
432
433
434
435
436
437
438
    @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))

439
    @utils.positional_deprecated
lintangsutawika's avatar
lintangsutawika committed
440
    def fewshot_context(
441
442
443
444
445
446
        self,
        doc,
        num_fewshot,
        provide_description=None,
        rnd=random.Random(1234),
        description=None,
lintangsutawika's avatar
lintangsutawika committed
447
    ):
448
449
450
451
452
453
454
        """Returns a fewshot context string that is made up of a prepended description
        (if provided), the `num_fewshot` number of examples, and an appended prompt example.

        :param doc: str
            The document as returned from training_docs, validation_docs, or test_docs.
        :param num_fewshot: int
            The number of fewshot examples to provide in the returned context string.
lintangsutawika's avatar
lintangsutawika committed
455
456
457
458
459
460
461
        :param provide_description: bool
            Not implemented, and this option is deprecated and will be removed in a future version in favor of a different description providing method
        :param rnd: random.Random
            The pseudo-random number generator used to randomly sample examples.
            WARNING: This is currently a required arg although it's optionalized with a default `None`.
        :param description: str
            The task's description that will be prepended to the fewshot examples.
462
463
464
        :returns: str
            The fewshot context.
        """
lintangsutawika's avatar
lintangsutawika committed
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
        assert (
            rnd is not None
        ), "A `random.Random` generator argument must be provided to `rnd`"
        assert not provide_description, (
            "The `provide_description` arg will be removed in future versions. To prepend "
            "a custom description to the context, supply the corresponding string via the "
            "`description` arg."
        )
        if provide_description is not None:
            # nudge people to not specify it at all
            print(
                "WARNING: provide_description is deprecated and will be removed in a future version in favor of description_dict"
            )

        description = description + "\n\n" if description else ""
480
481

        if num_fewshot == 0:
lintangsutawika's avatar
lintangsutawika committed
482
            labeled_examples = ""
483
        else:
lintangsutawika's avatar
lintangsutawika committed
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
            # for sets with no training docs, draw from other set *but ensure no overlap with current doc*
            if self.has_training_docs():
                fewshotex = self.fewshot_examples(k=num_fewshot, rnd=rnd)
            else:
                if self._fewshot_docs is None:
                    self._fewshot_docs = list(
                        self.validation_docs()
                        if self.has_validation_docs()
                        else self.test_docs()
                    )

                fewshotex = rnd.sample(self._fewshot_docs, num_fewshot + 1)

                # get rid of the doc that's the one we're evaluating, if it's in the fewshot
                fewshotex = [x for x in fewshotex if x != doc][:num_fewshot]

            labeled_examples = (
                "\n\n".join(
                    [
                        self.doc_to_text(doc) + self.doc_to_target(doc)
                        for doc in fewshotex
                    ]
                )
                + "\n\n"
lintangsutawika's avatar
lintangsutawika committed
508
            )
509
510

        example = self.doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
511
        return description + labeled_examples + example
512
513

    def apply_filters(self):
lintangsutawika's avatar
lintangsutawika committed
514
515
        if hasattr(self, "_filters"):
            for f in self._filters:
lintangsutawika's avatar
lintangsutawika committed
516
                f.apply(self._instances, None)
lintangsutawika's avatar
lintangsutawika committed
517
518
519
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances
520

baberabb's avatar
baberabb committed
521
    def dump_config(self) -> dict:
522
        """Returns a dictionary representing the task's config.
523
524
525
526
527

        :returns: str
            The fewshot context.
        """
        # TODO: this should only return the overrides applied to a non-YAML task's configuration.
528
        # (num_fewshot)
529
        return self.config.to_dict()
530

531
532

class ConfigurableTask(Task):
533
    VERSION = "Yaml"
534
    OUTPUT_TYPE = None
535
    CONFIG = None
536
537
538

    def __init__(
        self, data_dir=None, cache_dir=None, download_mode=None, config: dict = None
Ethan Smith's avatar
Ethan Smith committed
539
    ) -> None:  # TODO no super() call here
540
        # Get pre-configured attributes
541
        self._config = self.CONFIG
542

543
        # Use new configurations if there was no preconfiguration
544
        if self.config is None:
545
            self._config = TaskConfig(**config)
546
547
        # Overwrite configs
        else:
lintangsutawika's avatar
lintangsutawika committed
548
            if config is not None:
549
                self._config.__dict__.update(config)
550

551
        if self.config is None:
lintangsutawika's avatar
lintangsutawika committed
552
553
554
            raise ValueError(
                "Must pass a config to ConfigurableTask, either in cls.CONFIG or `config` kwarg"
            )
555

556
557
558
        if self.config.output_type is not None:
            assert self.config.output_type in ALL_OUTPUT_TYPES
            self.OUTPUT_TYPE = self.config.output_type
559

560
561
        if self.config.dataset_path is not None:
            self.DATASET_PATH = self.config.dataset_path
562

563
564
        if self.config.dataset_name is not None:
            self.DATASET_NAME = self.config.dataset_name
565

566
567
568
569
        self._metric_fn_list = {}
        self._metric_fn_kwargs = {}
        self._aggregation_list = {}
        self._higher_is_better = {}
570

571
        if self.config.metric_list is None:
572
            # TODO: handle this in TaskConfig.__post_init__ ?
573
574
            _metric_list = DEFAULT_METRIC_REGISTRY[self.config.output_type]

575
            for metric_name in _metric_list:
haileyschoelkopf's avatar
haileyschoelkopf committed
576
                self._metric_fn_list[metric_name] = get_metric(metric_name)
lintangsutawika's avatar
lintangsutawika committed
577
                self._metric_fn_kwargs[metric_name] = {}
578
579
580
                self._aggregation_list[metric_name] = get_metric_aggregation(
                    metric_name
                )
haileyschoelkopf's avatar
haileyschoelkopf committed
581
                self._higher_is_better[metric_name] = is_higher_better(metric_name)
582
        else:
583
            for metric_config in self.config.metric_list:
584
585
586
587
588
                assert "metric" in metric_config
                metric_name = metric_config["metric"]
                kwargs = {
                    key: metric_config[key]
                    for key in metric_config
Chris's avatar
Chris committed
589
590
                    if key
                    not in ["metric", "aggregation", "higher_is_better", "hf_evaluate"]
591
                }
Chris's avatar
Chris committed
592
593
594
595
                hf_evaluate_metric = (
                    "hf_evaluate" in metric_config
                    and metric_config["hf_evaluate"] is True
                )
596

597
                if self.config.process_results is not None:
598
599
                    self._metric_fn_list[metric_name] = None
                    self._metric_fn_kwargs[metric_name] = {}
600
601
602
603
604
605
                elif callable(metric_name):
                    metric_fn = metric_name.__call__
                    metric_name = metric_name.__name__
                    self._metric_fn_list[metric_name] = metric_fn
                    self._metric_fn_kwargs[metric_name] = kwargs
                else:
Chris's avatar
Chris committed
606
607
608
                    self._metric_fn_list[metric_name] = get_metric(
                        metric_name, hf_evaluate_metric
                    )
609
                    self._metric_fn_kwargs[metric_name] = kwargs
lintangsutawika's avatar
lintangsutawika committed
610

611
                if "aggregation" in metric_config:
612
                    agg_name = metric_config["aggregation"]
613
                    if type(agg_name) == str:
haileyschoelkopf's avatar
haileyschoelkopf committed
614
                        self._aggregation_list[metric_name] = get_aggregation(agg_name)
615
616
617
618
                    elif callable(agg_name):
                        self._aggregation_list[metric_name] = metric_config[
                            "aggregation"
                        ]
619
                else:
620
                    INV_AGG_REGISTRY = {v: k for k, v in AGGREGATION_REGISTRY.items()}
lintangsutawika's avatar
lintangsutawika committed
621
                    metric_agg = get_metric_aggregation(metric_name)
622
                    eval_logger.warning(
baberabb's avatar
baberabb committed
623
                        f"[Task: {self._config.task}] metric {metric_name} is defined, but aggregation is not. "
624
625
                        f"using default "
                        f"aggregation={INV_AGG_REGISTRY[metric_agg]}"
626
                    )
627
                    self._aggregation_list[metric_name] = metric_agg
lintangsutawika's avatar
lintangsutawika committed
628

629
630
631
632
633
634
                if "higher_is_better" in metric_config:
                    self._higher_is_better[metric_name] = metric_config[
                        "higher_is_better"
                    ]
                else:
                    eval_logger.warning(
baberabb's avatar
baberabb committed
635
                        f"[Task: {self._config.task}] metric {metric_name} is defined, but higher_is_better is not. "
636
                        f"using default "
haileyschoelkopf's avatar
haileyschoelkopf committed
637
                        f"higher_is_better={is_higher_better(metric_name)}"
638
                    )
haileyschoelkopf's avatar
haileyschoelkopf committed
639
                    self._higher_is_better[metric_name] = is_higher_better(metric_name)
640

641
        self.download(self.config.dataset_kwargs)
642
643
644
        self._training_docs = None
        self._fewshot_docs = None

645
        if self.config.filter_list is not None:
lintangsutawika's avatar
lintangsutawika committed
646
            self._filters = []
647
            for filter_config in self.config.filter_list:
lintangsutawika's avatar
lintangsutawika committed
648
649
650
651
652
653
654
                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
655
656
657
                        }
                        components.append([function["function"], kwargs])
                    filter_pipeline = build_filter_ensemble(filter_name, components)
lintangsutawika's avatar
lintangsutawika committed
658
                self._filters.append(filter_pipeline)
lintangsutawika's avatar
lintangsutawika committed
659
        else:
660
            self._filters = [build_filter_ensemble("none", [["take_first", None]])]
661

662
663
        if self.config.use_prompt is not None:
            eval_logger.info(f"loading prompt {self.config.use_prompt}")
664
            self.prompt = get_prompt(
665
                self.config.use_prompt, self.DATASET_PATH, self.DATASET_NAME
lintangsutawika's avatar
lintangsutawika committed
666
            )
667
668
669
        else:
            self.prompt = None

lintangsutawika's avatar
lintangsutawika committed
670
        if self.fewshot_docs() is not None:
haileyschoelkopf's avatar
haileyschoelkopf committed
671
            self.sampler = samplers.get_sampler(
haileyschoelkopf's avatar
haileyschoelkopf committed
672
673
674
                self.config.fewshot_config.get("sampler", "default")
                if self.config.fewshot_config
                else "default"
haileyschoelkopf's avatar
haileyschoelkopf committed
675
            )(list(self.fewshot_docs()), self, rnd=random.Random(1234))
676

677
        if self.has_test_docs():
678
            self.task_docs = self.test_docs()
679
        elif self.has_validation_docs():
680
            self.task_docs = self.validation_docs()
681
682
683
684
685
        else:
            assert (
                False
            ), f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"

686
        # Test One Doc
687
        self.features = list(self.task_docs.features.keys())
688
689
        self.multiple_input = 0
        self.multiple_target = 0
690
        test_doc = self.task_docs[0]
691
        test_text = self.doc_to_text(test_doc)
692
        test_target = self.doc_to_target(test_doc)
lintangsutawika's avatar
lintangsutawika committed
693

694
        if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
695
696
697
            test_choice = self.doc_to_choice(test_doc)
            if type(test_choice) is not list:
                eval_logger.error("doc_to_choice must return list")
698
699
            else:
                num_choice = len(test_choice)
700

701
702
            if type(test_text) is int:
                self.multiple_input = num_choice
703
704
        else:
            test_choice = None
705

706
        if type(test_target) is list:
707
            self.multiple_target = len(test_target)
708
        else:
lintangsutawika's avatar
lintangsutawika committed
709
            if (type(test_target) is int) and (test_choice is not None):
lintangsutawika's avatar
lintangsutawika committed
710
                test_target = test_choice[test_target]
711
            else:
lintangsutawika's avatar
lintangsutawika committed
712
                test_target = str(test_target)
713

714
715
716
        if test_choice is not None:
            check_choices = test_choice
        else:
lintangsutawika's avatar
lintangsutawika committed
717
            check_choices = [test_target]
718
719
720
721
        if self.config.doc_to_choice is not None:
            for choice in check_choices:
                choice_has_whitespace = True if choice[0].isspace() else False
                delimiter_has_whitespace = (
722
723
                    True
                    if self.config.target_delimiter.rstrip()
724
                    != self.config.target_delimiter
725
                    else False
726
                )
727

728
729
730
731
732
733
                if delimiter_has_whitespace and choice_has_whitespace:
                    eval_logger.warning(
                        f'Both target_delimiter and target choice: "{choice}" have whitespace'
                    )
                elif (not delimiter_has_whitespace) and (not choice_has_whitespace):
                    eval_logger.warning(
734
                        f'Both target_delimiter "{self.config.target_delimiter}" and target choice: "{choice}" do not have whitespace, ignore if the language you are evaluating on does not require/use whitespace'
735
736
                    )

Ethan Smith's avatar
Ethan Smith committed
737
    def download(self, dataset_kwargs=None) -> None:
738
739
740
741
742
743
        self.dataset = datasets.load_dataset(
            path=self.DATASET_PATH,
            name=self.DATASET_NAME,
            **dataset_kwargs if dataset_kwargs is not None else {},
        )

baberabb's avatar
baberabb committed
744
    def has_training_docs(self) -> bool:
745
        if self.config.training_split is not None:
746
747
748
749
            return True
        else:
            return False

baberabb's avatar
baberabb committed
750
    def has_validation_docs(self) -> bool:
751
        if self.config.validation_split is not None:
752
753
754
755
            return True
        else:
            return False

baberabb's avatar
baberabb committed
756
    def has_test_docs(self) -> bool:
757
        if self.config.test_split is not None:
758
759
760
761
            return True
        else:
            return False

baberabb's avatar
baberabb committed
762
    def training_docs(self) -> datasets.Dataset:
763
        if self.has_training_docs():
764
765
766
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.training_split]
767
                )
768
            return self.dataset[self.config.training_split]
769

baberabb's avatar
baberabb committed
770
    def validation_docs(self) -> datasets.Dataset:
771
        if self.has_validation_docs():
772
773
774
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.validation_split]
775
                )
776
            return self.dataset[self.config.validation_split]
777

baberabb's avatar
baberabb committed
778
    def test_docs(self) -> datasets.Dataset:
779
        if self.has_test_docs():
780
781
782
            if self.config.process_docs is not None:
                return self.config.process_docs(self.dataset[self.config.test_split])
            return self.dataset[self.config.test_split]
783

784
    def fewshot_docs(self):
785
786
        if self.config.fewshot_split is not None:
            return self.dataset[self.config.fewshot_split]
787
        else:
788
            if self.config.num_fewshot > 0:
789
                eval_logger.warning(
790
                    f"Task '{self.config.task}': "
791
792
793
794
                    "num_fewshot > 0 but fewshot_split is None. "
                    "using preconfigured rule."
                )
            return super().fewshot_docs()
795

lintangsutawika's avatar
lintangsutawika committed
796
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
    @utils.positional_deprecated
    def fewshot_context(self, doc, num_fewshot):
        """Returns a fewshot context string that is made up of a prepended description
        (if provided), the `num_fewshot` number of examples, and an appended prompt example.

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

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

        example = self.doc_to_text(doc)
        if type(example) == str:
            return labeled_examples + example
        elif type(example) == list:
            return [labeled_examples + ex for ex in example]
        elif type(example) == int:
            if self.config.doc_to_choice is not None:
                choices = self.doc_to_choice(doc)
                return labeled_examples + choices[example]
            else:
                return labeled_examples + str(example)

829
830
831
832
833
834
835
836
    def apply_filters(self):
        if hasattr(self, "_filters"):
            for f in self._filters:
                f.apply(self._instances, self.task_docs)
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances

837
    def should_decontaminate(self):
838
        return self.config.should_decontaminate
839
840

    def doc_to_decontamination_query(self, doc):
841
842
843
        if self.config.should_decontaminate:
            if self.config.doc_to_decontamination_query in self.features:
                return doc[self.config.doc_to_decontamination_query]
844
845
            else:
                return ast.literal_eval(
846
                    utils.apply_template(self.config.doc_to_decontamination_query, doc)
847
                )
848

849
850
851
852
853
854
855
856
857
858
859
860
    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):
861
862
        if self.prompt is not None:
            doc_to_text = self.prompt
863
        else:
864
            doc_to_text = self.config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
865

866
867
868
        if type(doc_to_text) == int:
            return doc_to_text
        elif type(doc_to_text) == str:
869
            if doc_to_text in self.features:
870
                # if self.config.doc_to_choice is not None:
871
872
                #     return self.doc_to_choice(doc)[doc[doc_to_text]]
                # else:
873
874
                return doc[doc_to_text]
            else:
lintangsutawika's avatar
lintangsutawika committed
875
                text_string = utils.apply_template(doc_to_text, doc)
lintangsutawika's avatar
lintangsutawika committed
876
                if text_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
877
878
879
                    return ast.literal_eval(text_string)
                else:
                    return text_string
880
        elif callable(doc_to_text):
881
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
882
        # Used when applying a Promptsource template
883
        elif hasattr(doc_to_text, "apply"):
884
885
886
887
888
            applied_prompt = doc_to_text.apply(doc)
            if len(applied_prompt) == 2:
                return applied_prompt[0]
            else:
                eval_logger.warning("Applied prompt returns empty string")
889
                return self.config.fewshot_delimiter
890
        else:
891
            print(type(doc_to_text))
892
            raise TypeError
893

894
    def doc_to_target(self, doc: dict) -> Union[int, str, list]:
895
896
        if self.prompt is not None:
            doc_to_target = self.prompt
897
        else:
898
            doc_to_target = self.config.doc_to_target
899

900
901
902
        if type(doc_to_target) == int:
            return doc_to_target
        elif type(doc_to_target) == str:
903
            if doc_to_target in self.features:
904
                # if self.config.doc_to_choice is not None:
905
906
907
                #     return self.doc_to_choice(doc)[doc[doc_to_target]]
                # else:
                return doc[doc_to_target]
908
            else:
lintangsutawika's avatar
lintangsutawika committed
909
                target_string = utils.apply_template(doc_to_target, doc)
lintangsutawika's avatar
lintangsutawika committed
910
                if target_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
911
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
912
913
914
915
916
                elif (
                    len(target_string) >= 2
                    and (target_string[0] == "[")
                    and (target_string[-1] == "]")
                ):
baberabb's avatar
baberabb committed
917
918
919
920
                    try:
                        return ast.literal_eval(target_string)
                    except (SyntaxError, ValueError):
                        return target_string
lintangsutawika's avatar
lintangsutawika committed
921
922
                else:
                    return target_string
923
924
        elif type(doc_to_target) == list:
            return doc_to_target
925
        elif callable(doc_to_target):
926
            return doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
927
        # Used when applying a Promptsource template
928
        elif hasattr(doc_to_target, "apply"):
929
            applied_prompt = doc_to_target.apply(doc)
930
931
932
933
            if len(applied_prompt) == 2:
                return applied_prompt[1]
            else:
                eval_logger.warning("Applied prompt returns empty string")
934
                return self.config.fewshot_delimiter
935
936
        else:
            raise TypeError
937

baberabb's avatar
baberabb committed
938
    def doc_to_choice(self, doc: Any) -> List[str]:
939
940
        if self.prompt is not None:
            doc_to_choice = self.prompt
941
        elif self.config.doc_to_choice is None:
942
943
            eval_logger.error("doc_to_choice was called but not set in config")
        else:
944
            doc_to_choice = self.config.doc_to_choice
945
946
947
948
949
950
951
952
953
954
955
956
957

        if type(doc_to_choice) == str:
            return ast.literal_eval(utils.apply_template(doc_to_choice, doc))
        elif type(doc_to_choice) == list:
            return doc_to_choice
        elif type(doc_to_choice) == dict:
            return list(doc_to_choice.values())
        elif callable(doc_to_choice):
            return doc_to_choice(doc)
        elif hasattr(doc_to_choice, "get_answer_choices_list"):
            return doc_to_choice.get_answer_choices_list(doc)
        else:
            raise TypeError
958

baberabb's avatar
baberabb committed
959
960
961
    def construct_requests(
        self, doc: dict, ctx: str, **kwargs
    ) -> Union[List[Instance], Instance]:
962
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
963
            arguments = (ctx, self.doc_to_target(doc))
964
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
965
            arguments = (self.doc_to_target(doc),)
966
        elif self.OUTPUT_TYPE == "multiple_choice":
967
            choices = self.doc_to_choice(doc)
968
            target_delimiter = self.config.target_delimiter
969
970
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
971
                cont = self.doc_to_target(doc)
972
                arguments = [(ctx, f"{target_delimiter}{cont}") for ctx in choices]
973
            else:
974
                # Otherwise they are placed in the continuation
975
                arguments = [(ctx, f"{target_delimiter}{cont}") for cont in choices]
976

977
            request_list = [
978
979
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
980
                    doc=doc,
981
                    arguments=arg,
982
                    idx=i,
983
984
                    **kwargs,
                )
985
                for i, arg in enumerate(arguments)
986
            ]
987
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
988
            if "acc_mutual_info" in self._metric_fn_list.keys():
989
990
991
                # if we are calculating multiple choice accuracy
                # using mutual information instead of raw loglikelihood as metric, need unconditional lls.

lintangsutawika's avatar
lintangsutawika committed
992
                # here mutual info refers to calculating
993
994
995
996
997
998
                # 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
999
                            doc=doc,
1000
                            arguments=("", "{}".format(choice)),
1001
1002
1003
                            idx=i,
                            **kwargs,
                        )
lintangsutawika's avatar
lintangsutawika committed
1004
                        for i, choice in enumerate(choices)
1005
1006
1007
                    ]
                )
            return request_list
lintangsutawika's avatar
lintangsutawika committed
1008

1009
        elif self.OUTPUT_TYPE == "generate_until":
1010
            arguments = (ctx, self.config.generation_kwargs)
lintangsutawika's avatar
lintangsutawika committed
1011
1012

        return Instance(
lintangsutawika's avatar
lintangsutawika committed
1013
1014
            request_type=self.OUTPUT_TYPE, doc=doc, arguments=arguments, idx=0, **kwargs
        )
1015
1016

    def process_results(self, doc, results):
1017
1018
        if callable(self.config.process_results):
            return self.config.process_results(doc, results)
lintangsutawika's avatar
lintangsutawika committed
1019

1020
        result_dict = {}
1021
        use_metric = list(self._metric_fn_list.keys())
1022
1023
1024
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
1025
1026
1027
1028
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
1029
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
1030
            (loglikelihood,) = results
1031
1032
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
haileyschoelkopf's avatar
haileyschoelkopf committed
1033
            return {
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
                **(
                    {"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
1049
            }
1050
        elif self.OUTPUT_TYPE == "multiple_choice":
1051
            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
1052

1053
            # retrieve choices in List[str] form, to compute choice lengths, etc.
1054
            choices = self.doc_to_choice(doc)
1055
1056
            completion_len = np.array([float(len(i)) for i in choices])

1057
1058
            if (
                2 * len(choices) == len(lls)
1059
                and "acc_mutual_info" in self._metric_fn_list.keys()
1060
1061
1062
1063
1064
1065
1066
            ):
                # 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]
1067

1068
1069
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
1070

1071
1072
            if self.multiple_input:
                gold = self.doc_to_text(doc)
1073
            else:
1074
                gold = self.doc_to_target(doc)
1075
1076
1077

            gold_index_error = False
            if type(gold) is list:
Lintang Sutawika's avatar
Lintang Sutawika committed
1078
1079
                gold = [i if i < len(choices) else -100 for i in gold]
                if -100 in gold:
1080
1081
1082
                    gold_index_error = True
            else:
                if type(gold) is int:
Lintang Sutawika's avatar
Lintang Sutawika committed
1083
                    gold = gold if gold < len(choices) else -100
1084
                elif type(gold) is str:
Lintang Sutawika's avatar
Lintang Sutawika committed
1085
                    gold = choices.index(gold) if gold in choices else -100
lintangsutawika's avatar
lintangsutawika committed
1086

Lintang Sutawika's avatar
Lintang Sutawika committed
1087
                if gold == -100:
1088
1089
1090
1091
                    gold_index_error = True

            if gold_index_error:
                eval_logger.warning(
lintangsutawika's avatar
lintangsutawika committed
1092
                    f"Label index was not in within range of available choices,"
1093
1094
                    f"Sample:\n\n{doc}\n\n"
                )
lintangsutawika's avatar
lintangsutawika committed
1095

1096
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
1097
1098
                acc = 1.0 if pred in gold else 0.0
                acc_norm = 1.0 if pred_norm in gold else 0.0
Lintang Sutawika's avatar
Lintang Sutawika committed
1099
                exact_match = int(any([is_greedy[i] if i != -100 else 0 for i in gold]))
lintangsutawika's avatar
lintangsutawika committed
1100
1101
1102
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
1103
                # TODO: this gets score of 0 on arc_challenge for pythia-70m. need to test that this works properly
Lintang Sutawika's avatar
Lintang Sutawika committed
1104
                exact_match = int(is_greedy[gold]) if gold != -100 else 0
1105
1106

            result_dict = {
1107
                **({"acc": acc} if "acc" in use_metric else {}),
1108
1109
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
1110
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
1111
                **({"exact_match": exact_match} if "exact_match" in use_metric else {}),
1112
1113
            }

1114
            if "acc_mutual_info" in use_metric:
lintangsutawika's avatar
lintangsutawika committed
1115
1116
1117
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
1118
1119
1120
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

1121
        elif self.OUTPUT_TYPE == "generate_until":
1122
            gold = self.doc_to_target(doc)
Chris's avatar
Chris committed
1123
            result = results[0]
1124
            if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1125
                # If you set doc_to_choice,
lintangsutawika's avatar
lintangsutawika committed
1126
                # it assumes that doc_to_target returns a number.
1127
1128
                choices = self.doc_to_choice(doc)
                gold = choices[gold]
1129
1130
            # we expect multiple_targets to be a list.
            elif self.multiple_target:
baberabb's avatar
baberabb committed
1131
                gold = list(gold)
Chris's avatar
Chris committed
1132
1133
1134
            elif type(gold) != type(result):
                # cast gold to the same type as result
                gold = type(result)(gold)
1135

lintangsutawika's avatar
lintangsutawika committed
1136
            for metric in self._metric_fn_list.keys():
haileyschoelkopf's avatar
haileyschoelkopf committed
1137
1138
1139
1140
1141
                if self.multiple_target:
                    # in the case where we have multiple targets,
                    # return true if any are true
                    # TODO: this may break for multipLe_target, non zero-or-1 metrics
                    scores = []
haileyschoelkopf's avatar
haileyschoelkopf committed
1142
1143
1144
1145
                    if not isinstance(gold, list):
                        # sometimes, a multiple_target dataset has exceptions where one doc has only one string answer
                        # print(gold)
                        gold = [gold]
haileyschoelkopf's avatar
haileyschoelkopf committed
1146
                    for gold_option in gold:
1147
                        try:
1148
                            result_score = self._metric_fn_list[metric](
1149
1150
                                references=[gold_option],
                                predictions=[result],
1151
                                **self._metric_fn_kwargs[metric],
1152
                            )
baberabb's avatar
baberabb committed
1153
1154
1155
                        except (
                            TypeError
                        ):  # TODO: this is hacky and I don't want to do it
1156
                            result_score = self._metric_fn_list[metric](
haileyschoelkopf's avatar
haileyschoelkopf committed
1157
1158
1159
                                [gold_option, result]
                            )
                        if isinstance(result_score, dict):
haileyschoelkopf's avatar
haileyschoelkopf committed
1160
                            # TODO: this handles the case where HF evaluate returns a dict.
1161
                            result_score = result_score[metric]
haileyschoelkopf's avatar
haileyschoelkopf committed
1162
                        scores.append(result_score)
haileyschoelkopf's avatar
haileyschoelkopf committed
1163
                    if any(scores):
1164
                        result_score = 1.0
haileyschoelkopf's avatar
haileyschoelkopf committed
1165
                    else:
1166
                        result_score = 0.0
haileyschoelkopf's avatar
haileyschoelkopf committed
1167
                else:
1168
                    try:
1169
                        result_score = self._metric_fn_list[metric](
1170
1171
                            references=[gold],
                            predictions=[result],
1172
                            **self._metric_fn_kwargs[metric],
1173
                        )
baberabb's avatar
baberabb committed
1174
1175
1176
                    except (
                        TypeError
                    ):  # needed for now in order to use a different interface between our own metrics and HF Evaluate metrics
1177
                        result_score = self._metric_fn_list[metric]([gold, result])
1178
1179
1180
1181
                    if isinstance(result_score, dict):
                        # TODO: this handles the case where HF evaluate returns a dict.
                        result_score = result_score[metric]
                result_dict[metric] = result_score
1182
        else:
lintangsutawika's avatar
lintangsutawika committed
1183
1184
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
1185
                "'loglikelihood', 'loglikelihood_rolling', 'generate_until' or 'multiple_choice'",
1186
            )
1187
1188
1189
1190
1191
1192
1193

        return result_dict

    def aggregation(self):
        return self._aggregation_list

    def higher_is_better(self):
haileyschoelkopf's avatar
haileyschoelkopf committed
1194
        return self._higher_is_better
1195
1196
1197
1198
1199


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

baberabb's avatar
baberabb committed
1200
    def doc_to_target(self, doc: dict) -> str:
1201
1202
        return " " + doc["choices"][doc["gold"]]

baberabb's avatar
baberabb committed
1203
    def construct_requests(self, doc: dict, ctx: str, **kwargs) -> List[Instance]:
1204
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1205
1206
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1207
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1208
                doc=doc,
1209
                arguments=(ctx, " {}".format(choice)),
1210
                idx=i,
1211
1212
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1213
1214
            for i, choice in enumerate(doc["choices"])
        ]
1215

baberabb's avatar
baberabb committed
1216
    def process_results(self, doc: dict, results: List[Tuple[float, bool]]) -> dict:
lintangsutawika's avatar
lintangsutawika committed
1217
1218
1219
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
        gold = doc["gold"]

        acc = 1.0 if np.argmax(results) == gold else 0.0
        completion_len = np.array([float(len(i)) for i in doc["choices"]])
        acc_norm = 1.0 if np.argmax(results / completion_len) == gold else 0.0

        return {
            "acc": acc,
            "acc_norm": acc_norm,
        }

baberabb's avatar
baberabb committed
1231
    def higher_is_better(self) -> dict:
1232
1233
1234
1235
1236
        return {
            "acc": True,
            "acc_norm": True,
        }

baberabb's avatar
baberabb committed
1237
    def aggregation(self) -> dict:
1238
1239
1240
1241
1242
1243
        return {
            "acc": mean,
            "acc_norm": mean,
        }


lintangsutawika's avatar
lintangsutawika committed
1244
class PerplexityTask(Task):
1245
1246
    OUTPUT_TYPE = "loglikelihood_rolling"

baberabb's avatar
baberabb committed
1247
    def has_training_docs(self) -> bool:
1248
1249
        return False

baberabb's avatar
baberabb committed
1250
    def fewshot_examples(self, k: int, rnd) -> List:
1251
1252
1253
        assert k == 0
        return []

baberabb's avatar
baberabb committed
1254
    def fewshot_context(self, doc: dict, num_fewshot: int) -> Literal[""]:
1255
1256
1257
1258
1259
1260
        assert (
            num_fewshot == 0
        ), "The number of fewshot examples must be 0 for perplexity tasks."

        return ""

baberabb's avatar
baberabb committed
1261
    def higher_is_better(self) -> dict:
1262
1263
1264
1265
1266
1267
1268
1269
1270
        return {
            "word_perplexity": False,
            "byte_perplexity": False,
            "bits_per_byte": False,
        }

    def doc_to_decontamination_query(self, doc):
        return doc

Ethan Smith's avatar
Ethan Smith committed
1271
    def doc_to_text(self, doc) -> str:
1272
1273
1274
1275
1276
        return ""

    def doc_to_target(self, doc):
        return doc

baberabb's avatar
baberabb committed
1277
    def construct_requests(self, doc: dict, ctx: Union[str, None], **kwargs):
1278
1279
        assert not ctx

lintangsutawika's avatar
lintangsutawika committed
1280
1281
1282
1283
1284
1285
1286
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1287

baberabb's avatar
baberabb committed
1288
    def process_results(self, doc: dict, results: float) -> dict:
1289
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1290
1291
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1292
1293
1294
1295
1296
1297
        return {
            "word_perplexity": (loglikelihood, words),
            "byte_perplexity": (loglikelihood, bytes_),
            "bits_per_byte": (loglikelihood, bytes_),
        }

baberabb's avatar
baberabb committed
1298
    def aggregation(self) -> dict:
1299
1300
1301
1302
1303
1304
1305
        return {
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
            "bits_per_byte": bits_per_byte,
        }

    @classmethod
baberabb's avatar
baberabb committed
1306
    def count_bytes(cls, doc) -> int:
1307
1308
1309
        return len(doc.encode("utf-8"))

    @classmethod
baberabb's avatar
baberabb committed
1310
    def count_words(cls, doc) -> int:
1311
1312
        """Downstream tasks with custom word boundaries should override this!"""
        return len(re.split(r"\s+", doc))