task.py 49.4 KB
Newer Older
1
import abc
2
import ast
lintangsutawika's avatar
lintangsutawika committed
3
import logging
4
import random
5
6
7
8
import re
from collections.abc import Callable
from dataclasses import asdict, dataclass
from typing import Any, List, Literal, Tuple, Union
9
10
11
12
13

import datasets
import numpy as np

from lm_eval import utils
14
from lm_eval.api import samplers
haileyschoelkopf's avatar
haileyschoelkopf committed
15
from lm_eval.api.instance import Instance
lintangsutawika's avatar
lintangsutawika committed
16
from lm_eval.api.metrics import (
17
    bits_per_byte,
lintangsutawika's avatar
lintangsutawika committed
18
19
20
21
    mean,
    weighted_perplexity,
)
from lm_eval.api.registry import (
22
23
    AGGREGATION_REGISTRY,
    DEFAULT_METRIC_REGISTRY,
haileyschoelkopf's avatar
haileyschoelkopf committed
24
    get_aggregation,
25
    get_metric,
26
    get_metric_aggregation,
haileyschoelkopf's avatar
haileyschoelkopf committed
27
    is_higher_better,
lintangsutawika's avatar
lintangsutawika committed
28
)
29
30
31
from lm_eval.filters import build_filter_ensemble
from lm_eval.prompts import get_prompt

32

33
34
35
36
ALL_OUTPUT_TYPES = [
    "loglikelihood",
    "multiple_choice",
    "loglikelihood_rolling",
37
    "generate_until",
38
39
]

lintangsutawika's avatar
lintangsutawika committed
40

41
eval_logger = logging.getLogger("lm-eval")
42

lintangsutawika's avatar
lintangsutawika committed
43

44
45
@dataclass
class TaskConfig(dict):
46
    # task naming/registry
47
    task: str = None
lintangsutawika's avatar
lintangsutawika committed
48
    task_alias: str = None
49
    group: Union[str, list] = None
lintangsutawika's avatar
lintangsutawika committed
50
    group_alias: Union[str, list] = None
51
52
53
    # HF dataset options.
    # which dataset to use,
    # and what splits for what purpose
54
55
    dataset_path: str = None
    dataset_name: str = None
56
    dataset_kwargs: dict = None
57
58
59
    training_split: str = None
    validation_split: str = None
    test_split: str = None
lintangsutawika's avatar
lintangsutawika committed
60
    fewshot_split: str = None  # TODO: assert that this not None if num_fewshot > 0. (?) assert if this is same split as one evaling (?)
61
62
    # formatting / prompting options.
    # see docs/advanced_task_guide.md for more info
63
    process_docs: Callable = None
64
65
    doc_to_text: Union[Callable, str] = None
    doc_to_target: Union[Callable, str] = None
lintangsutawika's avatar
lintangsutawika committed
66
    doc_to_choice: Union[Callable, str, dict, list] = None
lintangsutawika's avatar
lintangsutawika committed
67
    process_results: Union[Callable, str] = None
68
    use_prompt: str = None
69
    description: str = ""
70
71
    target_delimiter: str = " "
    fewshot_delimiter: str = "\n\n"
haileyschoelkopf's avatar
haileyschoelkopf committed
72
    fewshot_config: dict = None
73
    # runtime configuration options
74
    num_fewshot: int = None
75
    # scoring options
76
    metric_list: list = None
77
    output_type: str = "generate_until"
78
    generation_kwargs: dict = None
79
    repeats: int = 1
lintangsutawika's avatar
lintangsutawika committed
80
    filter_list: Union[str, list] = None
81
82
    should_decontaminate: bool = False
    doc_to_decontamination_query: str = None
83

lintangsutawika's avatar
lintangsutawika committed
84
85
86
    metadata: Union[
        str, list
    ] = None  # by default, not used in the code. allows for users to pass arbitrary info to tasks
87

Ethan Smith's avatar
Ethan Smith committed
88
    def __post_init__(self) -> None:
Lintang Sutawika's avatar
Lintang Sutawika committed
89
        if self.generation_kwargs is not None:
90
            if self.output_type != "generate_until":
Lintang Sutawika's avatar
Lintang Sutawika committed
91
                eval_logger.warning(
92
                    f"[{self.task}] passed `generation_kwargs`, but not using `output_type: generate_until`!"
Lintang Sutawika's avatar
Lintang Sutawika committed
93
                )
94
                assert self.output_type != "generate_until"
Lintang Sutawika's avatar
Lintang Sutawika committed
95
96
97
98
99
100
101

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

            if "until" not in self.generation_kwargs:
102
                self.generation_kwargs["until"] = [self.fewshot_delimiter]
Lintang Sutawika's avatar
Lintang Sutawika committed
103
        else:
104
            if self.output_type == "generate_until":
Lintang Sutawika's avatar
Lintang Sutawika committed
105
106
                # ensure that we greedily generate in absence of explicit arguments otherwise
                self.generation_kwargs = {
Lintang Sutawika's avatar
Lintang Sutawika committed
107
                    "until": None
108
109
                    if self.fewshot_delimiter is None
                    else [self.fewshot_delimiter],
Lintang Sutawika's avatar
Lintang Sutawika committed
110
111
                    "do_sample": False,
                }
112

haileyschoelkopf's avatar
haileyschoelkopf committed
113
114
        # TODO: how to make TaskConfigs be de- and re-serializable, even when using the !function constructor?

115
116
117
    def __getitem__(self, item):
        return getattr(self, item)

118
119
120
    def __setitem__(self, item, value):
        return setattr(self, item, value)

121
    def to_dict(self):
122
123
        """dumps the current config as a dictionary object, as a printable format.
        null fields will not be printed.
haileyschoelkopf's avatar
haileyschoelkopf committed
124
        Used for dumping results alongside full task configuration
125

haileyschoelkopf's avatar
haileyschoelkopf committed
126
127
128
129
130
131
132
133
134
135
        :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
136
137
138
            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
139
        return cfg_dict
140

141
142
143
144
145
146
147
148
149
150
151
152

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
153

154
155
156
157
158
159
160
161
    # 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
162

163
164
165
166
167
168
    def __init__(
        self,
        data_dir=None,
        cache_dir=None,
        download_mode=None,
        config=None,
Ethan Smith's avatar
Ethan Smith committed
169
    ) -> None:
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
        """
        :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
196
        self._config = TaskConfig({**config}) if config else TaskConfig()
197

lintangsutawika's avatar
lintangsutawika committed
198
        self._filters = [build_filter_ensemble("none", [["take_first", None]])]
199

Ethan Smith's avatar
Ethan Smith committed
200
    def download(self, data_dir=None, cache_dir=None, download_mode=None) -> None:
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
        """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.
        """
225
226
227
228
229
230
231
        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,
        )
232

233
234
235
236
237
    @property
    def config(self):
        """Returns the TaskConfig associated with this class."""
        return self._config

238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
    @abc.abstractmethod
    def has_training_docs(self):
        """Whether the task has a training set"""
        pass

    @abc.abstractmethod
    def has_validation_docs(self):
        """Whether the task has a validation set"""
        pass

    @abc.abstractmethod
    def has_test_docs(self):
        """Whether the task has a test set"""
        pass

    def training_docs(self):
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

    def validation_docs(self):
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

    def test_docs(self):
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

274
275
276
277
278
279
280
281
282
283
    def fewshot_docs(self):
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        if self.has_training_docs():
            return self.training_docs()
        elif self.has_validation_docs():
            return self.validation_docs()
        else:
lintangsutawika's avatar
lintangsutawika committed
284
            eval_logger.warning(
285
                "has_training_docs and has_validation_docs are False"
286
                ", using test_docs as fewshot_docs but this is not recommended."
lintangsutawika's avatar
lintangsutawika committed
287
            )
288
289
            return self.test_docs()

290
291
292
293
294
295
296
297
298
299
    def _process_doc(self, doc):
        """
        Override this to process (detokenize, strip, replace, etc.) individual
        documents. This can be used in a map over documents of a data split.
        E.g. `map(self._process_doc, self.dataset["validation"])`

        :return: dict
            The processed version of the specified `doc`.
        """
        return doc
lintangsutawika's avatar
lintangsutawika committed
300

301
302
303
304
305
306
307
308
309
310
311
312
313
    @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
314
    def doc_to_decontamination_query(self, doc) -> None:
315
316
317
318
319
320
321
322
323
324
325
326
327
        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
328
    def build_all_requests(self, limit=None, rank=None, world_size=None) -> None:
329
330
331
332
333
334
        """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:
335
            assert False, f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"
336

337
        eval_logger.info(f"Building contexts for task on rank {rank}...")
338

339
        instances = []
340
341
        for doc_id, doc in utils.create_iterator(
            enumerate(docs), rank, world_size, limit
lintangsutawika's avatar
lintangsutawika committed
342
        ):
343
            # sample fewshot context #TODO: need to offset doc_id by rank now!
344
            fewshot_ctx = self.fewshot_context(
345
                doc,
346
                0 if self.config.num_fewshot is None else self.config.num_fewshot,
347
            )
348

349
            # TODO: we should override self.config.repeats if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
350
351
352
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
353
                metadata=(self.config["task"], doc_id, self.config.repeats),
lintangsutawika's avatar
lintangsutawika committed
354
            )
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379

            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
380
            The number of times each instance in a dataset is inferred on. Defaults to 1,
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
            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
416
417
418
419
420
421
422
423
424
425
    @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))

426
    @utils.positional_deprecated
lintangsutawika's avatar
lintangsutawika committed
427
    def fewshot_context(
428
429
430
431
432
        self,
        doc,
        num_fewshot,
        rnd=random.Random(1234),
        description=None,
lintangsutawika's avatar
lintangsutawika committed
433
    ):
434
435
436
437
438
439
440
        """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
441
442
443
444
445
        :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.
446
447
448
        :returns: str
            The fewshot context.
        """
lintangsutawika's avatar
lintangsutawika committed
449
450
451
452
        assert (
            rnd is not None
        ), "A `random.Random` generator argument must be provided to `rnd`"

453
        description = description if description else ""
454
455

        if num_fewshot == 0:
lintangsutawika's avatar
lintangsutawika committed
456
            labeled_examples = ""
457
        else:
lintangsutawika's avatar
lintangsutawika committed
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
            # 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
482
            )
483
484

        example = self.doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
485
        return description + labeled_examples + example
486
487

    def apply_filters(self):
lintangsutawika's avatar
lintangsutawika committed
488
489
        if hasattr(self, "_filters"):
            for f in self._filters:
lintangsutawika's avatar
lintangsutawika committed
490
                f.apply(self._instances, None)
lintangsutawika's avatar
lintangsutawika committed
491
492
493
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances
494

baberabb's avatar
baberabb committed
495
    def dump_config(self) -> dict:
496
        """Returns a dictionary representing the task's config.
497
498
499
500
501

        :returns: str
            The fewshot context.
        """
        # TODO: this should only return the overrides applied to a non-YAML task's configuration.
502
        # (num_fewshot)
503
        return self.config.to_dict()
504

505
506

class ConfigurableTask(Task):
507
    VERSION = "Yaml"
508
    OUTPUT_TYPE = None
509
    CONFIG = None
510
511
512

    def __init__(
        self, data_dir=None, cache_dir=None, download_mode=None, config: dict = None
Ethan Smith's avatar
Ethan Smith committed
513
    ) -> None:  # TODO no super() call here
514
        # Get pre-configured attributes
515
        self._config = self.CONFIG
516

517
        # Use new configurations if there was no preconfiguration
518
        if self.config is None:
519
            self._config = TaskConfig(**config)
520
521
        # Overwrite configs
        else:
lintangsutawika's avatar
lintangsutawika committed
522
            if config is not None:
523
                self._config.__dict__.update(config)
524

525
        if self.config is None:
lintangsutawika's avatar
lintangsutawika committed
526
527
528
            raise ValueError(
                "Must pass a config to ConfigurableTask, either in cls.CONFIG or `config` kwarg"
            )
529

530
531
532
        if self.config.output_type is not None:
            assert self.config.output_type in ALL_OUTPUT_TYPES
            self.OUTPUT_TYPE = self.config.output_type
533

534
535
        if self.config.dataset_path is not None:
            self.DATASET_PATH = self.config.dataset_path
536

537
538
        if self.config.dataset_name is not None:
            self.DATASET_NAME = self.config.dataset_name
539

540
541
542
543
        self._metric_fn_list = {}
        self._metric_fn_kwargs = {}
        self._aggregation_list = {}
        self._higher_is_better = {}
544

545
        if self.config.metric_list is None:
546
            # TODO: handle this in TaskConfig.__post_init__ ?
547
548
            _metric_list = DEFAULT_METRIC_REGISTRY[self.config.output_type]

549
            for metric_name in _metric_list:
haileyschoelkopf's avatar
haileyschoelkopf committed
550
                self._metric_fn_list[metric_name] = get_metric(metric_name)
lintangsutawika's avatar
lintangsutawika committed
551
                self._metric_fn_kwargs[metric_name] = {}
552
553
554
                self._aggregation_list[metric_name] = get_metric_aggregation(
                    metric_name
                )
haileyschoelkopf's avatar
haileyschoelkopf committed
555
                self._higher_is_better[metric_name] = is_higher_better(metric_name)
556
        else:
557
            for metric_config in self.config.metric_list:
558
559
560
561
562
                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
563
564
                    if key
                    not in ["metric", "aggregation", "higher_is_better", "hf_evaluate"]
565
                }
Chris's avatar
Chris committed
566
567
568
569
                hf_evaluate_metric = (
                    "hf_evaluate" in metric_config
                    and metric_config["hf_evaluate"] is True
                )
570

571
                if self.config.process_results is not None:
572
573
                    self._metric_fn_list[metric_name] = None
                    self._metric_fn_kwargs[metric_name] = {}
574
575
576
577
578
579
                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
580
581
582
                    self._metric_fn_list[metric_name] = get_metric(
                        metric_name, hf_evaluate_metric
                    )
583
                    self._metric_fn_kwargs[metric_name] = kwargs
lintangsutawika's avatar
lintangsutawika committed
584

585
                if "aggregation" in metric_config:
586
                    agg_name = metric_config["aggregation"]
587
                    if isinstance(agg_name, str):
haileyschoelkopf's avatar
haileyschoelkopf committed
588
                        self._aggregation_list[metric_name] = get_aggregation(agg_name)
589
                    elif callable(agg_name):  # noqa: E721
590
591
592
                        self._aggregation_list[metric_name] = metric_config[
                            "aggregation"
                        ]
593
                else:
594
                    INV_AGG_REGISTRY = {v: k for k, v in AGGREGATION_REGISTRY.items()}
lintangsutawika's avatar
lintangsutawika committed
595
                    metric_agg = get_metric_aggregation(metric_name)
596
                    eval_logger.warning(
baberabb's avatar
baberabb committed
597
                        f"[Task: {self._config.task}] metric {metric_name} is defined, but aggregation is not. "
598
599
                        f"using default "
                        f"aggregation={INV_AGG_REGISTRY[metric_agg]}"
600
                    )
601
                    self._aggregation_list[metric_name] = metric_agg
lintangsutawika's avatar
lintangsutawika committed
602

603
604
605
606
607
608
                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
609
                        f"[Task: {self._config.task}] metric {metric_name} is defined, but higher_is_better is not. "
610
                        f"using default "
haileyschoelkopf's avatar
haileyschoelkopf committed
611
                        f"higher_is_better={is_higher_better(metric_name)}"
612
                    )
haileyschoelkopf's avatar
haileyschoelkopf committed
613
                    self._higher_is_better[metric_name] = is_higher_better(metric_name)
614

615
        self.download(self.config.dataset_kwargs)
616
617
618
        self._training_docs = None
        self._fewshot_docs = None

619
        if self.config.filter_list is not None:
lintangsutawika's avatar
lintangsutawika committed
620
            self._filters = []
621
            for filter_config in self.config.filter_list:
lintangsutawika's avatar
lintangsutawika committed
622
623
624
625
626
627
628
                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
629
630
631
                        }
                        components.append([function["function"], kwargs])
                    filter_pipeline = build_filter_ensemble(filter_name, components)
lintangsutawika's avatar
lintangsutawika committed
632
                self._filters.append(filter_pipeline)
lintangsutawika's avatar
lintangsutawika committed
633
        else:
634
            self._filters = [build_filter_ensemble("none", [["take_first", None]])]
635

636
637
        if self.config.use_prompt is not None:
            eval_logger.info(f"loading prompt {self.config.use_prompt}")
638
            self.prompt = get_prompt(
639
                self.config.use_prompt, self.DATASET_PATH, self.DATASET_NAME
lintangsutawika's avatar
lintangsutawika committed
640
            )
641
642
643
        else:
            self.prompt = None

lintangsutawika's avatar
lintangsutawika committed
644
        if self.fewshot_docs() is not None:
haileyschoelkopf's avatar
haileyschoelkopf committed
645
            self.sampler = samplers.get_sampler(
haileyschoelkopf's avatar
haileyschoelkopf committed
646
647
648
                self.config.fewshot_config.get("sampler", "default")
                if self.config.fewshot_config
                else "default"
haileyschoelkopf's avatar
haileyschoelkopf committed
649
            )(list(self.fewshot_docs()), self, rnd=random.Random(1234))
650

651
        if self.has_test_docs():
652
            self.task_docs = self.test_docs()
653
        elif self.has_validation_docs():
654
            self.task_docs = self.validation_docs()
655
        else:
656
            assert False, f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"
657

658
        # Test One Doc
659
        self.features = list(self.task_docs.features.keys())
660
661
        self.multiple_input = 0
        self.multiple_target = 0
662
        test_doc = self.task_docs[0]
663
        test_text = self.doc_to_text(test_doc)
664
        test_target = self.doc_to_target(test_doc)
lintangsutawika's avatar
lintangsutawika committed
665

666
        if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
667
            test_choice = self.doc_to_choice(test_doc)
668
            if not isinstance(test_choice, list):
lintangsutawika's avatar
lintangsutawika committed
669
                eval_logger.error("doc_to_choice must return list")
670
671
            else:
                num_choice = len(test_choice)
672

673
            if isinstance(test_text, int):
674
                self.multiple_input = num_choice
675
676
        else:
            test_choice = None
677

678
        if isinstance(test_target, list):
679
            self.multiple_target = len(test_target)
680
        else:
681
            if (isinstance(test_target, int)) and (test_choice is not None):
lintangsutawika's avatar
lintangsutawika committed
682
                test_target = test_choice[test_target]
683
            else:
lintangsutawika's avatar
lintangsutawika committed
684
                test_target = str(test_target)
685

686
687
688
        if test_choice is not None:
            check_choices = test_choice
        else:
lintangsutawika's avatar
lintangsutawika committed
689
            check_choices = [test_target]
690
691
692
693
        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 = (
694
695
                    True
                    if self.config.target_delimiter.rstrip()
696
                    != self.config.target_delimiter
697
                    else False
698
                )
699

700
                if delimiter_has_whitespace and choice_has_whitespace:
701
702
                    eval_logger.debug(
                        f'Both target_delimiter "{self.config.target_delimiter}" and target choice: "{choice}" have whitespace'
703
704
                    )
                elif (not delimiter_has_whitespace) and (not choice_has_whitespace):
705
                    eval_logger.debug(
706
                        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'
707
708
                    )

Ethan Smith's avatar
Ethan Smith committed
709
    def download(self, dataset_kwargs=None) -> None:
710
711
712
713
714
715
        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
716
    def has_training_docs(self) -> bool:
717
        if self.config.training_split is not None:
718
719
720
721
            return True
        else:
            return False

baberabb's avatar
baberabb committed
722
    def has_validation_docs(self) -> bool:
723
        if self.config.validation_split is not None:
724
725
726
727
            return True
        else:
            return False

baberabb's avatar
baberabb committed
728
    def has_test_docs(self) -> bool:
729
        if self.config.test_split is not None:
730
731
732
733
            return True
        else:
            return False

baberabb's avatar
baberabb committed
734
    def training_docs(self) -> datasets.Dataset:
735
        if self.has_training_docs():
736
737
738
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.training_split]
739
                )
740
            return self.dataset[self.config.training_split]
741

baberabb's avatar
baberabb committed
742
    def validation_docs(self) -> datasets.Dataset:
743
        if self.has_validation_docs():
744
745
746
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.validation_split]
747
                )
748
            return self.dataset[self.config.validation_split]
749

baberabb's avatar
baberabb committed
750
    def test_docs(self) -> datasets.Dataset:
751
        if self.has_test_docs():
752
753
754
            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]
755

756
    def fewshot_docs(self):
757
758
        if self.config.fewshot_split is not None:
            return self.dataset[self.config.fewshot_split]
759
        else:
760
            if (self.config.num_fewshot is not None) and (self.config.num_fewshot > 0):
761
                eval_logger.warning(
762
                    f"Task '{self.config.task}': "
763
764
765
766
                    "num_fewshot > 0 but fewshot_split is None. "
                    "using preconfigured rule."
                )
            return super().fewshot_docs()
767

lintangsutawika's avatar
lintangsutawika committed
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
    @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)
790
791
792
793
794
795
796
797
798
799
800
801
802
        if self.multiple_input:
            return labeled_examples
        else:
            if isinstance(example, str):
                return labeled_examples + example
            elif isinstance(example, list):
                return [labeled_examples + ex for ex in example]
            elif isinstance(example, int):
                if self.config.doc_to_choice is not None:
                    choices = self.doc_to_choice(doc)
                    return labeled_examples + choices[example]
                else:
                    return labeled_examples + str(example)
lintangsutawika's avatar
lintangsutawika committed
803

804
805
806
807
808
809
810
811
    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

812
    def should_decontaminate(self):
813
        return self.config.should_decontaminate
814
815

    def doc_to_decontamination_query(self, doc):
816
        if self.config.should_decontaminate:
817
818
            if self.config.doc_to_decontamination_query is None:
                return self.doc_to_text(doc)
819
            else:
820
821
822
823
824
825
826
827
828
829
830
                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
                        )
                    )
831

832
833
834
835
836
837
838
839
840
841
842
843
    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):
844
845
        if self.prompt is not None:
            doc_to_text = self.prompt
846
        else:
847
            doc_to_text = self.config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
848

849
        if isinstance(doc_to_text, int):
850
            return doc_to_text
851
        elif isinstance(doc_to_text, str):
852
            if doc_to_text in self.features:
853
                # if self.config.doc_to_choice is not None:
854
855
                #     return self.doc_to_choice(doc)[doc[doc_to_text]]
                # else:
856
857
                return doc[doc_to_text]
            else:
lintangsutawika's avatar
lintangsutawika committed
858
                text_string = utils.apply_template(doc_to_text, doc)
lintangsutawika's avatar
lintangsutawika committed
859
                if text_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
860
861
862
                    return ast.literal_eval(text_string)
                else:
                    return text_string
863
        elif callable(doc_to_text):
864
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
865
        # Used when applying a Promptsource template
866
        elif hasattr(doc_to_text, "apply"):
867
868
869
870
871
            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")
872
                return self.config.fewshot_delimiter
873
        else:
874
            print(type(doc_to_text))
875
            raise TypeError
876

877
    def doc_to_target(self, doc: dict) -> Union[int, str, list]:
878
879
        if self.prompt is not None:
            doc_to_target = self.prompt
880
        else:
881
            doc_to_target = self.config.doc_to_target
882

883
        if isinstance(doc_to_target, int):
884
            return doc_to_target
885
        elif isinstance(doc_to_target, str):
886
            if doc_to_target in self.features:
887
                # if self.config.doc_to_choice is not None:
888
889
890
                #     return self.doc_to_choice(doc)[doc[doc_to_target]]
                # else:
                return doc[doc_to_target]
891
            else:
lintangsutawika's avatar
lintangsutawika committed
892
                target_string = utils.apply_template(doc_to_target, doc)
lintangsutawika's avatar
lintangsutawika committed
893
                if target_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
894
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
895
896
897
898
899
                elif (
                    len(target_string) >= 2
                    and (target_string[0] == "[")
                    and (target_string[-1] == "]")
                ):
baberabb's avatar
baberabb committed
900
901
902
903
                    try:
                        return ast.literal_eval(target_string)
                    except (SyntaxError, ValueError):
                        return target_string
lintangsutawika's avatar
lintangsutawika committed
904
905
                else:
                    return target_string
906
        elif isinstance(doc_to_target, list):
907
            return doc_to_target
908
        elif callable(doc_to_target):
909
            return doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
910
        # Used when applying a Promptsource template
911
        elif hasattr(doc_to_target, "apply"):
912
            applied_prompt = doc_to_target.apply(doc)
913
914
915
916
            if len(applied_prompt) == 2:
                return applied_prompt[1]
            else:
                eval_logger.warning("Applied prompt returns empty string")
917
                return self.config.fewshot_delimiter
918
919
        else:
            raise TypeError
920

baberabb's avatar
baberabb committed
921
    def doc_to_choice(self, doc: Any) -> List[str]:
922
923
        if self.prompt is not None:
            doc_to_choice = self.prompt
924
        elif self.config.doc_to_choice is None:
925
926
            eval_logger.error("doc_to_choice was called but not set in config")
        else:
927
            doc_to_choice = self.config.doc_to_choice
928

929
        if isinstance(doc_to_choice, str):
930
931
932
933
            if doc_to_choice in self.features:
                return doc[doc_to_choice]
            else:
                return ast.literal_eval(utils.apply_template(doc_to_choice, doc))
934
        elif isinstance(doc_to_choice, list):
935
            return doc_to_choice
936
        elif isinstance(doc_to_choice, dict):
937
938
939
940
941
942
943
            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
944

baberabb's avatar
baberabb committed
945
946
947
    def construct_requests(
        self, doc: dict, ctx: str, **kwargs
    ) -> Union[List[Instance], Instance]:
948
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
949
            arguments = (ctx, self.doc_to_target(doc))
950
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
951
            arguments = (self.doc_to_target(doc),)
952
        elif self.OUTPUT_TYPE == "multiple_choice":
953
            choices = self.doc_to_choice(doc)
954
            target_delimiter = self.config.target_delimiter
955
956
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
957
                cont = self.doc_to_target(doc)
958
959
960
                arguments = [
                    (ctx + choice, f"{target_delimiter}{cont}") for choice in choices
                ]
961
            else:
962
                # Otherwise they are placed in the continuation
963
                arguments = [(ctx, f"{target_delimiter}{cont}") for cont in choices]
964

965
            request_list = [
966
967
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
968
                    doc=doc,
969
                    arguments=arg,
970
                    idx=i,
971
972
                    **kwargs,
                )
973
                for i, arg in enumerate(arguments)
974
            ]
975
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
976
            if "acc_mutual_info" in self._metric_fn_list.keys():
977
978
979
                # if we are calculating multiple choice accuracy
                # using mutual information instead of raw loglikelihood as metric, need unconditional lls.

lintangsutawika's avatar
lintangsutawika committed
980
                # here mutual info refers to calculating
981
982
983
984
985
986
                # 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
987
                            doc=doc,
988
                            arguments=("", "{}".format(choice)),
989
990
991
                            idx=i,
                            **kwargs,
                        )
lintangsutawika's avatar
lintangsutawika committed
992
                        for i, choice in enumerate(choices)
993
994
995
                    ]
                )
            return request_list
lintangsutawika's avatar
lintangsutawika committed
996

997
        elif self.OUTPUT_TYPE == "generate_until":
998
            arguments = (ctx, self.config.generation_kwargs)
lintangsutawika's avatar
lintangsutawika committed
999
1000

        return Instance(
lintangsutawika's avatar
lintangsutawika committed
1001
1002
            request_type=self.OUTPUT_TYPE, doc=doc, arguments=arguments, idx=0, **kwargs
        )
1003
1004

    def process_results(self, doc, results):
1005
1006
        if callable(self.config.process_results):
            return self.config.process_results(doc, results)
lintangsutawika's avatar
lintangsutawika committed
1007

1008
        result_dict = {}
1009
        use_metric = list(self._metric_fn_list.keys())
1010
1011
1012
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
1013
1014
1015
1016
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
1017
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
1018
            (loglikelihood,) = results
1019
1020
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
haileyschoelkopf's avatar
haileyschoelkopf committed
1021
            return {
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
                **(
                    {"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
1037
            }
1038
        elif self.OUTPUT_TYPE == "multiple_choice":
1039
            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
1040

1041
            # retrieve choices in List[str] form, to compute choice lengths, etc.
1042
            choices = self.doc_to_choice(doc)
1043
1044
            completion_len = np.array([float(len(i)) for i in choices])

1045
1046
            if (
                2 * len(choices) == len(lls)
1047
                and "acc_mutual_info" in self._metric_fn_list.keys()
1048
1049
1050
1051
1052
1053
1054
            ):
                # 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]
1055

1056
1057
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
1058

1059
1060
            if self.multiple_input:
                gold = self.doc_to_text(doc)
1061
            else:
1062
                gold = self.doc_to_target(doc)
1063
1064

            gold_index_error = False
1065
            if isinstance(gold, list):
Lintang Sutawika's avatar
Lintang Sutawika committed
1066
1067
                gold = [i if i < len(choices) else -100 for i in gold]
                if -100 in gold:
1068
1069
                    gold_index_error = True
            else:
1070
                if isinstance(gold, int):
Lintang Sutawika's avatar
Lintang Sutawika committed
1071
                    gold = gold if gold < len(choices) else -100
1072
                elif isinstance(gold, str):
Lintang Sutawika's avatar
Lintang Sutawika committed
1073
                    gold = choices.index(gold) if gold in choices else -100
lintangsutawika's avatar
lintangsutawika committed
1074

Lintang Sutawika's avatar
Lintang Sutawika committed
1075
                if gold == -100:
1076
1077
1078
1079
                    gold_index_error = True

            if gold_index_error:
                eval_logger.warning(
lintangsutawika's avatar
lintangsutawika committed
1080
                    f"Label index was not in within range of available choices,"
1081
1082
                    f"Sample:\n\n{doc}\n\n"
                )
lintangsutawika's avatar
lintangsutawika committed
1083

1084
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
1085
1086
                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
1087
                exact_match = int(any([is_greedy[i] if i != -100 else 0 for i in gold]))
lintangsutawika's avatar
lintangsutawika committed
1088
1089
1090
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
1091
                # 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
1092
                exact_match = int(is_greedy[gold]) if gold != -100 else 0
1093
1094

            result_dict = {
1095
                **({"acc": acc} if "acc" in use_metric else {}),
1096
1097
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
1098
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
1099
                **({"exact_match": exact_match} if "exact_match" in use_metric else {}),
1100
1101
            }

1102
            if "acc_mutual_info" in use_metric:
lintangsutawika's avatar
lintangsutawika committed
1103
1104
1105
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
1106
1107
1108
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

1109
        elif self.OUTPUT_TYPE == "generate_until":
1110
            gold = self.doc_to_target(doc)
Chris's avatar
Chris committed
1111
            result = results[0]
1112
            if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1113
                # If you set doc_to_choice,
lintangsutawika's avatar
lintangsutawika committed
1114
                # it assumes that doc_to_target returns a number.
1115
1116
                choices = self.doc_to_choice(doc)
                gold = choices[gold]
1117
1118
            # we expect multiple_targets to be a list.
            elif self.multiple_target:
baberabb's avatar
baberabb committed
1119
                gold = list(gold)
Chris's avatar
Chris committed
1120
1121
1122
            elif type(gold) != type(result):
                # cast gold to the same type as result
                gold = type(result)(gold)
1123

lintangsutawika's avatar
lintangsutawika committed
1124
            for metric in self._metric_fn_list.keys():
haileyschoelkopf's avatar
haileyschoelkopf committed
1125
1126
1127
1128
1129
                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
1130
1131
1132
1133
                    if not isinstance(gold, list):
                        # sometimes, a multiple_target dataset has exceptions where one doc has only one string answer
                        # print(gold)
                        gold = [gold]
1134
1135
1136
1137
1138
1139
1140
1141
                    if metric == "exact_match":
                        result = [result for _ in range(len(gold))]
                        scores = self._metric_fn_list[metric](
                            references=gold,
                            predictions=result,
                            **self._metric_fn_kwargs[metric],
                        )[metric]
                        result_score = 1.0 if scores > 0.0 else 0.0
haileyschoelkopf's avatar
haileyschoelkopf committed
1142
                    else:
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
                        for gold_option in gold:
                            try:
                                result_score = self._metric_fn_list[metric](
                                    references=[gold_option],
                                    predictions=[result],
                                    **self._metric_fn_kwargs[metric],
                                )
                            except (
                                TypeError
                            ):  # TODO: this is hacky and I don't want to do it
                                result_score = self._metric_fn_list[metric](
                                    [gold_option, result]
                                )
                            if isinstance(result_score, dict):
                                # TODO: this handles the case where HF evaluate returns a dict.
                                result_score = result_score[metric]
                            scores.append(result_score)
                        if any(scores):
                            result_score = 1.0
                        else:
                            result_score = 0.0
haileyschoelkopf's avatar
haileyschoelkopf committed
1164
                else:
1165
                    try:
1166
                        result_score = self._metric_fn_list[metric](
1167
1168
                            references=[gold],
                            predictions=[result],
1169
                            **self._metric_fn_kwargs[metric],
1170
                        )
1171
                    except TypeError:  # needed for now in order to use a different interface between our own metrics and HF Evaluate metrics
1172
                        result_score = self._metric_fn_list[metric]([gold, result])
1173
1174
1175
1176
                    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
1177
        else:
lintangsutawika's avatar
lintangsutawika committed
1178
1179
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
1180
                "'loglikelihood', 'loglikelihood_rolling', 'generate_until' or 'multiple_choice'",
1181
            )
1182
1183
1184
1185
1186
1187
1188

        return result_dict

    def aggregation(self):
        return self._aggregation_list

    def higher_is_better(self):
haileyschoelkopf's avatar
haileyschoelkopf committed
1189
        return self._higher_is_better
1190
1191
1192
1193
1194


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

baberabb's avatar
baberabb committed
1195
    def doc_to_target(self, doc: dict) -> str:
1196
1197
        return " " + doc["choices"][doc["gold"]]

baberabb's avatar
baberabb committed
1198
    def construct_requests(self, doc: dict, ctx: str, **kwargs) -> List[Instance]:
1199
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1200
1201
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1202
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1203
                doc=doc,
1204
                arguments=(ctx, " {}".format(choice)),
1205
                idx=i,
1206
1207
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1208
1209
            for i, choice in enumerate(doc["choices"])
        ]
1210

baberabb's avatar
baberabb committed
1211
    def process_results(self, doc: dict, results: List[Tuple[float, bool]]) -> dict:
lintangsutawika's avatar
lintangsutawika committed
1212
1213
1214
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
        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
1226
    def higher_is_better(self) -> dict:
1227
1228
1229
1230
1231
        return {
            "acc": True,
            "acc_norm": True,
        }

baberabb's avatar
baberabb committed
1232
    def aggregation(self) -> dict:
1233
1234
1235
1236
1237
1238
        return {
            "acc": mean,
            "acc_norm": mean,
        }


lintangsutawika's avatar
lintangsutawika committed
1239
class PerplexityTask(Task):
1240
1241
    OUTPUT_TYPE = "loglikelihood_rolling"

baberabb's avatar
baberabb committed
1242
    def has_training_docs(self) -> bool:
1243
1244
        return False

baberabb's avatar
baberabb committed
1245
    def fewshot_examples(self, k: int, rnd) -> List:
1246
1247
1248
        assert k == 0
        return []

baberabb's avatar
baberabb committed
1249
    def fewshot_context(self, doc: dict, num_fewshot: int) -> Literal[""]:
1250
1251
1252
1253
1254
1255
        assert (
            num_fewshot == 0
        ), "The number of fewshot examples must be 0 for perplexity tasks."

        return ""

baberabb's avatar
baberabb committed
1256
    def higher_is_better(self) -> dict:
1257
1258
1259
1260
1261
1262
1263
1264
1265
        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
1266
    def doc_to_text(self, doc) -> str:
1267
1268
1269
1270
1271
        return ""

    def doc_to_target(self, doc):
        return doc

baberabb's avatar
baberabb committed
1272
    def construct_requests(self, doc: dict, ctx: Union[str, None], **kwargs):
1273
1274
        assert not ctx

lintangsutawika's avatar
lintangsutawika committed
1275
1276
1277
1278
1279
1280
1281
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1282

baberabb's avatar
baberabb committed
1283
    def process_results(self, doc: dict, results: float) -> dict:
1284
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1285
1286
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1287
1288
1289
1290
1291
1292
        return {
            "word_perplexity": (loglikelihood, words),
            "byte_perplexity": (loglikelihood, bytes_),
            "bits_per_byte": (loglikelihood, bytes_),
        }

baberabb's avatar
baberabb committed
1293
    def aggregation(self) -> dict:
1294
1295
1296
1297
1298
1299
1300
        return {
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
            "bits_per_byte": bits_per_byte,
        }

    @classmethod
baberabb's avatar
baberabb committed
1301
    def count_bytes(cls, doc) -> int:
1302
1303
1304
        return len(doc.encode("utf-8"))

    @classmethod
baberabb's avatar
baberabb committed
1305
    def count_words(cls, doc) -> int:
1306
1307
        """Downstream tasks with custom word boundaries should override this!"""
        return len(re.split(r"\s+", doc))