task.py 49.2 KB
Newer Older
1
import abc
2
import ast
lintangsutawika's avatar
lintangsutawika committed
3
import logging
4
import os
5
import random
6
7
8
9
import re
from collections.abc import Callable
from dataclasses import asdict, dataclass
from typing import Any, List, Literal, Tuple, Union
10
11
12
13
14

import datasets
import numpy as np

from lm_eval import utils
15
from lm_eval.api import samplers
haileyschoelkopf's avatar
haileyschoelkopf committed
16
from lm_eval.api.instance import Instance
lintangsutawika's avatar
lintangsutawika committed
17
from lm_eval.api.metrics import (
18
    bits_per_byte,
lintangsutawika's avatar
lintangsutawika committed
19
20
21
22
23
    mean,
    weighted_perplexity,
    bits_per_byte,
)
from lm_eval.api.registry import (
haileyschoelkopf's avatar
haileyschoelkopf committed
24
    get_metric,
25
26
27
    get_evaluate,
    get_aggregation,
    METRIC_REGISTRY,
28
    DEFAULT_METRIC_REGISTRY,
lintangsutawika's avatar
lintangsutawika committed
29
)
30
31
32
from lm_eval.filters import build_filter_ensemble
from lm_eval.prompts import get_prompt

33

34
35
36
37
ALL_OUTPUT_TYPES = [
    "loglikelihood",
    "multiple_choice",
    "loglikelihood_rolling",
38
    "generate_until",
39
40
]

lintangsutawika's avatar
lintangsutawika committed
41

42
eval_logger = logging.getLogger("lm-eval")
43

lintangsutawika's avatar
lintangsutawika committed
44

45
46
@dataclass
class TaskConfig(dict):
47
    # task naming/registry
48
    task: str = None
lintangsutawika's avatar
lintangsutawika committed
49
    task_alias: str = None
50
    group: Union[str, list] = None
lintangsutawika's avatar
lintangsutawika committed
51
    group_alias: Union[str, list] = None
52
53
54
    # HF dataset options.
    # which dataset to use,
    # and what splits for what purpose
55
56
    dataset_path: str = None
    dataset_name: str = None
57
    dataset_kwargs: dict = None
58
59
60
    training_split: str = None
    validation_split: str = None
    test_split: str = None
lintangsutawika's avatar
lintangsutawika committed
61
    fewshot_split: str = None  # TODO: assert that this not None if num_fewshot > 0. (?) assert if this is same split as one evaling (?)
62
63
    # formatting / prompting options.
    # see docs/advanced_task_guide.md for more info
64
    process_docs: Callable = None
65
66
    doc_to_text: Union[Callable, str] = None
    doc_to_target: Union[Callable, str] = None
lintangsutawika's avatar
lintangsutawika committed
67
    doc_to_choice: Union[Callable, str, dict, list] = None
lintangsutawika's avatar
lintangsutawika committed
68
    process_results: Union[Callable, str] = None
69
    use_prompt: str = None
70
    description: str = ""
71
72
    target_delimiter: str = " "
    fewshot_delimiter: str = "\n\n"
haileyschoelkopf's avatar
haileyschoelkopf committed
73
    fewshot_config: dict = None
74
    # runtime configuration options
75
    num_fewshot: int = None
76
    # scoring options
77
    metric_list: list = None
78
    output_type: str = "generate_until"
79
    generation_kwargs: dict = None
80
    repeats: int = 1
lintangsutawika's avatar
lintangsutawika committed
81
    filter_list: Union[str, list] = None
82
83
    should_decontaminate: bool = False
    doc_to_decontamination_query: str = None
84

lintangsutawika's avatar
lintangsutawika committed
85
86
87
    metadata: Union[
        str, list
    ] = None  # by default, not used in the code. allows for users to pass arbitrary info to tasks
88

Ethan Smith's avatar
Ethan Smith committed
89
    def __post_init__(self) -> None:
90
        if self.dataset_path and os.path.exists(os.path.dirname(self.dataset_path)):
lintangsutawika's avatar
lintangsutawika committed
91
92
            import inspect
            from importlib import import_module
lintangsutawika's avatar
format  
lintangsutawika committed
93

lintangsutawika's avatar
lintangsutawika committed
94
            self.dataset_path = inspect.getfile(import_module(self.dataset_path))
95

Lintang Sutawika's avatar
Lintang Sutawika committed
96
        if self.generation_kwargs is not None:
97
            if self.output_type != "generate_until":
Lintang Sutawika's avatar
Lintang Sutawika committed
98
                eval_logger.warning(
99
                    f"[{self.task}] passed `generation_kwargs`, but not using `output_type: generate_until`!"
Lintang Sutawika's avatar
Lintang Sutawika committed
100
                )
101
                assert self.output_type != "generate_until"
Lintang Sutawika's avatar
Lintang Sutawika committed
102
103
104
105
106
107
108

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

            if "until" not in self.generation_kwargs:
109
                self.generation_kwargs["until"] = [self.fewshot_delimiter]
Lintang Sutawika's avatar
Lintang Sutawika committed
110
        else:
111
            if self.output_type == "generate_until":
Lintang Sutawika's avatar
Lintang Sutawika committed
112
113
                # ensure that we greedily generate in absence of explicit arguments otherwise
                self.generation_kwargs = {
Lintang Sutawika's avatar
Lintang Sutawika committed
114
                    "until": None
115
116
                    if self.fewshot_delimiter is None
                    else [self.fewshot_delimiter],
Lintang Sutawika's avatar
Lintang Sutawika committed
117
118
                    "do_sample": False,
                }
119

haileyschoelkopf's avatar
haileyschoelkopf committed
120
121
        # TODO: how to make TaskConfigs be de- and re-serializable, even when using the !function constructor?

122
123
124
    def __getitem__(self, item):
        return getattr(self, item)

125
126
127
    def __setitem__(self, item, value):
        return setattr(self, item, value)

128
    def to_dict(self):
129
130
        """dumps the current config as a dictionary object, as a printable format.
        null fields will not be printed.
haileyschoelkopf's avatar
haileyschoelkopf committed
131
        Used for dumping results alongside full task configuration
132

haileyschoelkopf's avatar
haileyschoelkopf committed
133
134
135
136
137
138
139
140
141
142
        :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
143
144
145
            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
146
        return cfg_dict
147

148
149
150
151
152
153
154
155
156
157
158
159

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
160

161
162
163
164
165
166
167
168
    # 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
169

170
171
172
173
174
175
    def __init__(
        self,
        data_dir=None,
        cache_dir=None,
        download_mode=None,
        config=None,
Ethan Smith's avatar
Ethan Smith committed
176
    ) -> None:
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
        """
        :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
203
        self._config = TaskConfig({**config}) if config else TaskConfig()
204

lintangsutawika's avatar
lintangsutawika committed
205
        self._filters = [build_filter_ensemble("none", [["take_first", None]])]
206

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

240
241
242
243
244
    @property
    def config(self):
        """Returns the TaskConfig associated with this class."""
        return self._config

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
274
275
276
277
278
279
280
    @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 []

281
282
283
284
285
286
287
288
289
290
    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
291
            eval_logger.warning(
292
                "has_training_docs and has_validation_docs are False"
293
                ", using test_docs as fewshot_docs but this is not recommended."
lintangsutawika's avatar
lintangsutawika committed
294
            )
295
296
            return self.test_docs()

297
298
299
300
301
302
303
304
305
306
    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
307

308
309
310
311
312
313
314
315
316
317
318
319
320
    @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
321
    def doc_to_decontamination_query(self, doc) -> None:
322
323
324
325
326
327
328
329
330
331
332
333
334
        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
335
    def build_all_requests(self, limit=None, rank=None, world_size=None) -> None:
336
337
338
339
340
341
        """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:
342
            assert False, f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"
343

344
        eval_logger.info(f"Building contexts for task on rank {rank}...")
345

346
        instances = []
347
348
        for doc_id, doc in utils.create_iterator(
            enumerate(docs), rank, world_size, limit
lintangsutawika's avatar
lintangsutawika committed
349
        ):
350
            # sample fewshot context #TODO: need to offset doc_id by rank now!
351
            fewshot_ctx = self.fewshot_context(
352
                doc,
353
                0 if self.config.num_fewshot is None else self.config.num_fewshot,
354
            )
355

356
            # TODO: we should override self.config.repeats if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
357
358
359
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
360
                metadata=(self.config["task"], doc_id, self.config.repeats),
lintangsutawika's avatar
lintangsutawika committed
361
            )
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

            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
387
            The number of times each instance in a dataset is inferred on. Defaults to 1,
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
            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
406
    def aggregation(self):
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
        """
        :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
423
424
425
426
427
428
429
430
431
432
    @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))

433
    @utils.positional_deprecated
lintangsutawika's avatar
lintangsutawika committed
434
    def fewshot_context(
435
436
437
438
439
        self,
        doc,
        num_fewshot,
        rnd=random.Random(1234),
        description=None,
lintangsutawika's avatar
lintangsutawika committed
440
    ):
441
442
443
444
445
446
447
        """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
448
449
450
451
452
        :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.
453
454
455
        :returns: str
            The fewshot context.
        """
lintangsutawika's avatar
lintangsutawika committed
456
457
458
459
        assert (
            rnd is not None
        ), "A `random.Random` generator argument must be provided to `rnd`"

460
        description = description if description else ""
461
462

        if num_fewshot == 0:
lintangsutawika's avatar
lintangsutawika committed
463
            labeled_examples = ""
464
        else:
lintangsutawika's avatar
lintangsutawika committed
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
            # 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
489
            )
490
491

        example = self.doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
492
        return description + labeled_examples + example
493
494

    def apply_filters(self):
lintangsutawika's avatar
lintangsutawika committed
495
496
        if hasattr(self, "_filters"):
            for f in self._filters:
lintangsutawika's avatar
lintangsutawika committed
497
                f.apply(self._instances, None)
lintangsutawika's avatar
lintangsutawika committed
498
499
500
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances
501

baberabb's avatar
baberabb committed
502
    def dump_config(self) -> dict:
503
        """Returns a dictionary representing the task's config.
504
505
506
507
508

        :returns: str
            The fewshot context.
        """
        # TODO: this should only return the overrides applied to a non-YAML task's configuration.
509
        # (num_fewshot)
510
        return self.config.to_dict()
511

512
513

class ConfigurableTask(Task):
514
    VERSION = "Yaml"
515
    OUTPUT_TYPE = None
516
    CONFIG = None
517
518
519

    def __init__(
        self, data_dir=None, cache_dir=None, download_mode=None, config: dict = None
Ethan Smith's avatar
Ethan Smith committed
520
    ) -> None:  # TODO no super() call here
521
        # Get pre-configured attributes
522
        self._config = self.CONFIG
523

524
        # Use new configurations if there was no preconfiguration
525
        if self.config is None:
526
            self._config = TaskConfig(**config)
527
528
        # Overwrite configs
        else:
lintangsutawika's avatar
lintangsutawika committed
529
            if config is not None:
530
                self._config.__dict__.update(config)
531

532
        if self.config is None:
lintangsutawika's avatar
lintangsutawika committed
533
534
535
            raise ValueError(
                "Must pass a config to ConfigurableTask, either in cls.CONFIG or `config` kwarg"
            )
536

537
538
539
        if self.config.output_type is not None:
            assert self.config.output_type in ALL_OUTPUT_TYPES
            self.OUTPUT_TYPE = self.config.output_type
540

541
542
        if self.config.dataset_path is not None:
            self.DATASET_PATH = self.config.dataset_path
543

544
545
        if self.config.dataset_name is not None:
            self.DATASET_NAME = self.config.dataset_name
546

547
548
        self._metric_fn_list = {}
        self._metric_fn_kwargs = {}
549
        self._aggregation_list = {}
550
        self._higher_is_better = {}
551

552
        if self.config.metric_list is None:
553
            # TODO: handle this in TaskConfig.__post_init__ ?
554
555
            _metric_list = DEFAULT_METRIC_REGISTRY[self.config.output_type]

556
            for metric_name in _metric_list:
557
                metric = get_metric(metric_name)
558
                self._metric_fn_list[metric_name] = metric["function"]
lintangsutawika's avatar
lintangsutawika committed
559
                self._metric_fn_kwargs[metric_name] = {}
560
561
                self._aggregation_list = metric["aggregation"]
                self._higher_is_better[metric_name] = metric["is_higher_better"]
562
        else:
563
            for metric_config in self.config.metric_list:
564
565
566
567
568
                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
569
                    if key
570
571
572
573
574
575
                    not in [
                        "metric",
                        "aggregation",
                        "higher_is_better",
                        "use_hf_evaluate",
                    ]
576
                }
577
                use_hf_evaluate = (
lintangsutawika's avatar
lintangsutawika committed
578
579
                    "use_hf_evaluate" in metric_config
                    and metric_config["use_hf_evaluate"] is True
Chris's avatar
Chris committed
580
                )
581

lintangsutawika's avatar
lintangsutawika committed
582
583
584
585
                if self.config.process_results is not None:
                    metric_fn = None
                    kwargs = {}
                elif callable(metric_name):
586
587
588
                    metric_fn = metric_name.__call__
                    metric_name = metric_name.__name__
                else:
589
590
591
592
593
594
                    assert type(metric_name) == str
                    if use_hf_evaluate:
                        metric_fn = get_evaluate(metric_name, **kwargs)
                    elif metric_name in METRIC_REGISTRY:
                        metric = get_metric(metric_name, **kwargs)
                        metric_fn = metric["function"]
595
596

                self._metric_fn_kwargs[metric_name] = kwargs
597
                self._metric_fn_list[metric_name] = metric_fn
lintangsutawika's avatar
lintangsutawika committed
598

599
600
601
602
603
                # Ignores aggregation if the metric set
                # is a registered metric
                # for backward compatibility
                if metric_name in METRIC_REGISTRY and ("aggregation" not in metric):
                    self._aggregation_list[metric_name] = metric_fn
604
                else:
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
                    if "aggregation" in metric_config:

                        agg_name = metric_config["aggregation"]
                        if isinstance(agg_name, str):
                            self._aggregation_list[metric_name] = get_aggregation(
                                agg_name
                            )
                        elif callable(agg_name):  # noqa: E721
                            self._aggregation_list[metric_name] = agg_name
                    else:
                        if use_hf_evaluate:
                            self._aggregation_list[metric_name] = metric_fn
                        elif (metric_name in METRIC_REGISTRY) and (
                            "aggregation" in metric
                        ):
                            self._aggregation_list[metric_name] = metric["aggregation"]
621
622
623
624
625
626

                if "higher_is_better" in metric_config:
                    self._higher_is_better[metric_name] = metric_config[
                        "higher_is_better"
                    ]
                else:
lintangsutawika's avatar
lintangsutawika committed
627
                    self._higher_is_better[metric_name] = metric["higher_is_better"]
628

629
        self.download(self.config.dataset_kwargs)
630
631
632
        self._training_docs = None
        self._fewshot_docs = None

633
        if self.config.filter_list is not None:
lintangsutawika's avatar
lintangsutawika committed
634
            self._filters = []
635
            for filter_config in self.config.filter_list:
lintangsutawika's avatar
lintangsutawika committed
636
637
638
639
640
641
642
                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
643
644
645
                        }
                        components.append([function["function"], kwargs])
                    filter_pipeline = build_filter_ensemble(filter_name, components)
lintangsutawika's avatar
lintangsutawika committed
646
                self._filters.append(filter_pipeline)
lintangsutawika's avatar
lintangsutawika committed
647
        else:
648
            self._filters = [build_filter_ensemble("none", [["take_first", None]])]
649

650
651
        if self.config.use_prompt is not None:
            eval_logger.info(f"loading prompt {self.config.use_prompt}")
652
            self.prompt = get_prompt(
653
                self.config.use_prompt, self.DATASET_PATH, self.DATASET_NAME
lintangsutawika's avatar
lintangsutawika committed
654
            )
655
656
657
        else:
            self.prompt = None

lintangsutawika's avatar
lintangsutawika committed
658
        if self.fewshot_docs() is not None:
haileyschoelkopf's avatar
haileyschoelkopf committed
659
            self.sampler = samplers.get_sampler(
haileyschoelkopf's avatar
haileyschoelkopf committed
660
661
662
                self.config.fewshot_config.get("sampler", "default")
                if self.config.fewshot_config
                else "default"
haileyschoelkopf's avatar
haileyschoelkopf committed
663
            )(list(self.fewshot_docs()), self, rnd=random.Random(1234))
664

665
        if self.has_test_docs():
666
            self.task_docs = self.test_docs()
667
        elif self.has_validation_docs():
668
            self.task_docs = self.validation_docs()
669
        else:
670
            assert False, f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"
671

672
        # Test One Doc
673
        self.features = list(self.task_docs.features.keys())
674
675
        self.multiple_input = 0
        self.multiple_target = 0
676
        test_doc = self.task_docs[0]
677
        test_text = self.doc_to_text(test_doc)
678
        test_target = self.doc_to_target(test_doc)
lintangsutawika's avatar
lintangsutawika committed
679

680
        if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
681
            test_choice = self.doc_to_choice(test_doc)
682
            if not isinstance(test_choice, list):
lintangsutawika's avatar
lintangsutawika committed
683
                eval_logger.error("doc_to_choice must return list")
684
685
            else:
                num_choice = len(test_choice)
686

687
            if isinstance(test_text, int):
688
                self.multiple_input = num_choice
689
690
        else:
            test_choice = None
691

692
        if isinstance(test_target, list):
693
            self.multiple_target = len(test_target)
694
        else:
695
            if (isinstance(test_target, int)) and (test_choice is not None):
lintangsutawika's avatar
lintangsutawika committed
696
                test_target = test_choice[test_target]
697
            else:
lintangsutawika's avatar
lintangsutawika committed
698
                test_target = str(test_target)
699

700
701
702
        if test_choice is not None:
            check_choices = test_choice
        else:
lintangsutawika's avatar
lintangsutawika committed
703
            check_choices = [test_target]
704
705
706
707
        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 = (
708
709
                    True
                    if self.config.target_delimiter.rstrip()
710
                    != self.config.target_delimiter
711
                    else False
712
                )
713

714
715
716
717
718
719
                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(
720
                        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'
721
722
                    )

Ethan Smith's avatar
Ethan Smith committed
723
    def download(self, dataset_kwargs=None) -> None:
724
725
726
727
728
729
        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
730
    def has_training_docs(self) -> bool:
731
        if self.config.training_split is not None:
732
733
734
735
            return True
        else:
            return False

baberabb's avatar
baberabb committed
736
    def has_validation_docs(self) -> bool:
737
        if self.config.validation_split is not None:
738
739
740
741
            return True
        else:
            return False

baberabb's avatar
baberabb committed
742
    def has_test_docs(self) -> bool:
743
        if self.config.test_split is not None:
744
745
746
747
            return True
        else:
            return False

baberabb's avatar
baberabb committed
748
    def training_docs(self) -> datasets.Dataset:
749
        if self.has_training_docs():
750
751
752
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.training_split]
753
                )
754
            return self.dataset[self.config.training_split]
755

baberabb's avatar
baberabb committed
756
    def validation_docs(self) -> datasets.Dataset:
757
        if self.has_validation_docs():
758
759
760
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.validation_split]
761
                )
762
            return self.dataset[self.config.validation_split]
763

baberabb's avatar
baberabb committed
764
    def test_docs(self) -> datasets.Dataset:
765
        if self.has_test_docs():
766
767
768
            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]
769

770
    def fewshot_docs(self):
771
772
        if self.config.fewshot_split is not None:
            return self.dataset[self.config.fewshot_split]
773
        else:
774
            if (self.config.num_fewshot is not None) and (self.config.num_fewshot > 0):
775
                eval_logger.warning(
776
                    f"Task '{self.config.task}': "
777
778
779
780
                    "num_fewshot > 0 but fewshot_split is None. "
                    "using preconfigured rule."
                )
            return super().fewshot_docs()
781

lintangsutawika's avatar
lintangsutawika committed
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
    @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)
804
        if isinstance(example, str):
lintangsutawika's avatar
lintangsutawika committed
805
            return labeled_examples + example
806
        elif isinstance(example, list):
lintangsutawika's avatar
lintangsutawika committed
807
            return [labeled_examples + ex for ex in example]
808
        elif isinstance(example, int):
lintangsutawika's avatar
lintangsutawika committed
809
810
811
812
813
814
            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)

815
816
817
818
819
820
821
822
    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

823
    def should_decontaminate(self):
824
        return self.config.should_decontaminate
825
826

    def doc_to_decontamination_query(self, doc):
827
        if self.config.should_decontaminate:
828
829
            if self.config.doc_to_decontamination_query is None:
                return self.doc_to_text(doc)
830
            else:
831
832
833
834
835
836
837
838
839
840
841
                doc_to_decontamination_query = self.config.doc_to_decontamination_query
                if doc_to_decontamination_query in self.features:
                    return doc[doc_to_decontamination_query]
                elif callable(doc_to_decontamination_query):
                    return doc_to_decontamination_query(doc)
                else:
                    return ast.literal_eval(
                        utils.apply_template(
                            self.config.doc_to_decontamination_query, doc
                        )
                    )
842

843
844
845
846
847
848
849
850
851
852
853
854
    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):
855
856
        if self.prompt is not None:
            doc_to_text = self.prompt
857
        else:
858
            doc_to_text = self.config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
859

860
        if isinstance(doc_to_text, int):
861
            return doc_to_text
862
        elif isinstance(doc_to_text, str):
863
            if doc_to_text in self.features:
864
                # if self.config.doc_to_choice is not None:
865
866
                #     return self.doc_to_choice(doc)[doc[doc_to_text]]
                # else:
867
868
                return doc[doc_to_text]
            else:
lintangsutawika's avatar
lintangsutawika committed
869
                text_string = utils.apply_template(doc_to_text, doc)
lintangsutawika's avatar
lintangsutawika committed
870
                if text_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
871
872
873
                    return ast.literal_eval(text_string)
                else:
                    return text_string
874
        elif callable(doc_to_text):
875
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
876
        # Used when applying a Promptsource template
877
        elif hasattr(doc_to_text, "apply"):
878
879
880
881
882
            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")
883
                return self.config.fewshot_delimiter
884
        else:
885
            print(type(doc_to_text))
886
            raise TypeError
887

888
    def doc_to_target(self, doc: dict) -> Union[int, str, list]:
889
890
        if self.prompt is not None:
            doc_to_target = self.prompt
891
        else:
892
            doc_to_target = self.config.doc_to_target
893

894
        if isinstance(doc_to_target, int):
895
            return doc_to_target
896
        elif isinstance(doc_to_target, str):
897
            if doc_to_target in self.features:
898
                # if self.config.doc_to_choice is not None:
899
900
901
                #     return self.doc_to_choice(doc)[doc[doc_to_target]]
                # else:
                return doc[doc_to_target]
902
            else:
lintangsutawika's avatar
lintangsutawika committed
903
                target_string = utils.apply_template(doc_to_target, doc)
lintangsutawika's avatar
lintangsutawika committed
904
                if target_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
905
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
906
907
908
909
910
                elif (
                    len(target_string) >= 2
                    and (target_string[0] == "[")
                    and (target_string[-1] == "]")
                ):
baberabb's avatar
baberabb committed
911
912
913
914
                    try:
                        return ast.literal_eval(target_string)
                    except (SyntaxError, ValueError):
                        return target_string
lintangsutawika's avatar
lintangsutawika committed
915
916
                else:
                    return target_string
917
        elif isinstance(doc_to_target, list):
918
            return doc_to_target
919
        elif callable(doc_to_target):
920
            return doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
921
        # Used when applying a Promptsource template
922
        elif hasattr(doc_to_target, "apply"):
923
            applied_prompt = doc_to_target.apply(doc)
924
925
926
927
            if len(applied_prompt) == 2:
                return applied_prompt[1]
            else:
                eval_logger.warning("Applied prompt returns empty string")
928
                return self.config.fewshot_delimiter
929
930
        else:
            raise TypeError
931

baberabb's avatar
baberabb committed
932
    def doc_to_choice(self, doc: Any) -> List[str]:
933
934
        if self.prompt is not None:
            doc_to_choice = self.prompt
935
        elif self.config.doc_to_choice is None:
936
937
            eval_logger.error("doc_to_choice was called but not set in config")
        else:
938
            doc_to_choice = self.config.doc_to_choice
939

940
        if isinstance(doc_to_choice, str):
941
942
943
944
            if doc_to_choice in self.features:
                return doc[doc_to_choice]
            else:
                return ast.literal_eval(utils.apply_template(doc_to_choice, doc))
945
        elif isinstance(doc_to_choice, list):
946
            return doc_to_choice
947
        elif isinstance(doc_to_choice, dict):
948
949
950
951
952
953
954
            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
955

baberabb's avatar
baberabb committed
956
957
958
    def construct_requests(
        self, doc: dict, ctx: str, **kwargs
    ) -> Union[List[Instance], Instance]:
959
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
960
            arguments = (ctx, self.doc_to_target(doc))
961
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
962
            arguments = (self.doc_to_target(doc),)
963
        elif self.OUTPUT_TYPE == "multiple_choice":
964
            choices = self.doc_to_choice(doc)
965
            target_delimiter = self.config.target_delimiter
966
967
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
968
                cont = self.doc_to_target(doc)
969
                arguments = [(ctx, f"{target_delimiter}{cont}") for ctx in choices]
970
            else:
971
                # Otherwise they are placed in the continuation
972
                arguments = [(ctx, f"{target_delimiter}{cont}") for cont in choices]
973

974
            request_list = [
975
976
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
977
                    doc=doc,
978
                    arguments=arg,
979
                    idx=i,
980
981
                    **kwargs,
                )
982
                for i, arg in enumerate(arguments)
983
            ]
984
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
985
            if "acc_mutual_info" in self._metric_fn_list.keys():
986
987
988
                # if we are calculating multiple choice accuracy
                # using mutual information instead of raw loglikelihood as metric, need unconditional lls.

lintangsutawika's avatar
lintangsutawika committed
989
                # here mutual info refers to calculating
990
991
992
993
994
995
                # 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
996
                            doc=doc,
997
                            arguments=("", "{}".format(choice)),
998
999
1000
                            idx=i,
                            **kwargs,
                        )
lintangsutawika's avatar
lintangsutawika committed
1001
                        for i, choice in enumerate(choices)
1002
1003
1004
                    ]
                )
            return request_list
lintangsutawika's avatar
lintangsutawika committed
1005

1006
        elif self.OUTPUT_TYPE == "generate_until":
1007
            arguments = (ctx, self.config.generation_kwargs)
lintangsutawika's avatar
lintangsutawika committed
1008
1009

        return Instance(
lintangsutawika's avatar
lintangsutawika committed
1010
1011
            request_type=self.OUTPUT_TYPE, doc=doc, arguments=arguments, idx=0, **kwargs
        )
1012
1013

    def process_results(self, doc, results):
1014
1015
1016
1017

        # Process results returns 1 of X things per doc/results
        # 1. A score
        # 2. Components to be processed later to obtained a score. such as gold and prediction
1018
1019
        if callable(self.config.process_results):
            return self.config.process_results(doc, results)
lintangsutawika's avatar
lintangsutawika committed
1020

1021
        result_dict = {}
1022
        use_metric = list(self._metric_fn_list.keys())
1023
1024
1025
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
1026
1027
1028
1029
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
1030
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
1031
            (loglikelihood,) = results
1032
1033
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
            return {
                **(
                    {"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 {}
                ),
            }
1051
        elif self.OUTPUT_TYPE == "multiple_choice":
1052
            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
1053

1054
            # retrieve choices in List[str] form, to compute choice lengths, etc.
1055
            choices = self.doc_to_choice(doc)
1056
1057
            completion_len = np.array([float(len(i)) for i in choices])

1058
1059
            if (
                2 * len(choices) == len(lls)
1060
                and "acc_mutual_info" in self._metric_fn_list.keys()
1061
1062
1063
1064
1065
1066
1067
            ):
                # 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]
1068

1069
1070
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
1071

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

            gold_index_error = False
1078
            if isinstance(gold, list):
Lintang Sutawika's avatar
Lintang Sutawika committed
1079
1080
                gold = [i if i < len(choices) else -100 for i in gold]
                if -100 in gold:
1081
1082
                    gold_index_error = True
            else:
1083
                if isinstance(gold, int):
Lintang Sutawika's avatar
Lintang Sutawika committed
1084
                    gold = gold if gold < len(choices) else -100
1085
                elif isinstance(gold, str):
Lintang Sutawika's avatar
Lintang Sutawika committed
1086
                    gold = choices.index(gold) if gold in choices else -100
lintangsutawika's avatar
lintangsutawika committed
1087

Lintang Sutawika's avatar
Lintang Sutawika committed
1088
                if gold == -100:
1089
1090
1091
1092
                    gold_index_error = True

            if gold_index_error:
                eval_logger.warning(
lintangsutawika's avatar
lintangsutawika committed
1093
                    f"Label index was not in within range of available choices,"
1094
1095
                    f"Sample:\n\n{doc}\n\n"
                )
lintangsutawika's avatar
lintangsutawika committed
1096

1097
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
1098
1099
                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
1100
                exact_match = int(any([is_greedy[i] if i != -100 else 0 for i in gold]))
lintangsutawika's avatar
lintangsutawika committed
1101
1102
1103
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
1104
                # 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
1105
                exact_match = int(is_greedy[gold]) if gold != -100 else 0
1106

1107
            # gold, lls, is_greedy, completion_len
1108
            result_dict = {
1109
1110
                **({"acc": acc} if "acc" in use_metric else {}),
                **({"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
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
1114
1115
            }

1116
            if "acc_mutual_info" in use_metric:
lintangsutawika's avatar
lintangsutawika committed
1117
1118
1119
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
1120
1121
1122
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

1123
        elif self.OUTPUT_TYPE == "generate_until":
1124
            gold = self.doc_to_target(doc)
Chris's avatar
Chris committed
1125
            result = results[0]
1126
            if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1127
                # If you set doc_to_choice,
lintangsutawika's avatar
lintangsutawika committed
1128
                # it assumes that doc_to_target returns a number.
1129
1130
                choices = self.doc_to_choice(doc)
                gold = choices[gold]
1131
1132
            # we expect multiple_targets to be a list.
            elif self.multiple_target:
baberabb's avatar
baberabb committed
1133
                gold = list(gold)
Chris's avatar
Chris committed
1134
1135
1136
            elif type(gold) != type(result):
                # cast gold to the same type as result
                gold = type(result)(gold)
1137

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

        return result_dict

1192
1193
    def aggregation(self):
        return self._aggregation_list
1194
1195

    def higher_is_better(self):
haileyschoelkopf's avatar
haileyschoelkopf committed
1196
        return self._higher_is_better
1197
1198
1199
1200
1201


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

baberabb's avatar
baberabb committed
1202
    def doc_to_target(self, doc: dict) -> str:
1203
1204
        return " " + doc["choices"][doc["gold"]]

baberabb's avatar
baberabb committed
1205
    def construct_requests(self, doc: dict, ctx: str, **kwargs) -> List[Instance]:
1206
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1207
1208
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1209
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1210
                doc=doc,
1211
                arguments=(ctx, " {}".format(choice)),
1212
                idx=i,
1213
1214
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1215
1216
            for i, choice in enumerate(doc["choices"])
        ]
1217

baberabb's avatar
baberabb committed
1218
    def process_results(self, doc: dict, results: List[Tuple[float, bool]]) -> dict:
lintangsutawika's avatar
lintangsutawika committed
1219
1220
1221
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
        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
1233
    def higher_is_better(self) -> dict:
1234
1235
1236
1237
1238
        return {
            "acc": True,
            "acc_norm": True,
        }

1239
    def aggregation(self) -> dict:
1240
1241
1242
1243
1244
1245
        return {
            "acc": mean,
            "acc_norm": mean,
        }


lintangsutawika's avatar
lintangsutawika committed
1246
class PerplexityTask(Task):
1247
1248
    OUTPUT_TYPE = "loglikelihood_rolling"

baberabb's avatar
baberabb committed
1249
    def has_training_docs(self) -> bool:
1250
1251
        return False

baberabb's avatar
baberabb committed
1252
    def fewshot_examples(self, k: int, rnd) -> List:
1253
1254
1255
        assert k == 0
        return []

baberabb's avatar
baberabb committed
1256
    def fewshot_context(self, doc: dict, num_fewshot: int) -> Literal[""]:
1257
1258
1259
1260
1261
1262
        assert (
            num_fewshot == 0
        ), "The number of fewshot examples must be 0 for perplexity tasks."

        return ""

baberabb's avatar
baberabb committed
1263
    def higher_is_better(self) -> dict:
1264
1265
1266
1267
1268
1269
1270
1271
1272
        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
1273
    def doc_to_text(self, doc) -> str:
1274
1275
1276
1277
1278
        return ""

    def doc_to_target(self, doc):
        return doc

baberabb's avatar
baberabb committed
1279
    def construct_requests(self, doc: dict, ctx: Union[str, None], **kwargs):
1280
1281
        assert not ctx

lintangsutawika's avatar
lintangsutawika committed
1282
1283
1284
1285
1286
1287
1288
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1289

baberabb's avatar
baberabb committed
1290
    def process_results(self, doc: dict, results: float) -> dict:
1291
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1292
1293
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1294
1295
1296
1297
1298
1299
        return {
            "word_perplexity": (loglikelihood, words),
            "byte_perplexity": (loglikelihood, bytes_),
            "bits_per_byte": (loglikelihood, bytes_),
        }

1300
    def aggregation(self) -> dict:
1301
1302
1303
1304
1305
1306
1307
        return {
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
            "bits_per_byte": bits_per_byte,
        }

    @classmethod
baberabb's avatar
baberabb committed
1308
    def count_bytes(cls, doc) -> int:
1309
1310
1311
        return len(doc.encode("utf-8"))

    @classmethod
baberabb's avatar
baberabb committed
1312
    def count_words(cls, doc) -> int:
1313
1314
        """Downstream tasks with custom word boundaries should override this!"""
        return len(re.split(r"\s+", doc))