task.py 45.8 KB
Newer Older
1
import abc
2
from dataclasses import dataclass, field, asdict
3
4

import re
5
import ast
lintangsutawika's avatar
lintangsutawika committed
6
import yaml
7
8
9
import evaluate
import random
import itertools
10
import functools
11
from tqdm import tqdm
12
13
14
15

import datasets
import numpy as np

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

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

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

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

50
51
52

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

lintangsutawika's avatar
lintangsutawika committed
89
    metadata: str = None  # by default, not used in the code. allows for users to pass arbitrary info to tasks
90

91

Ethan Smith's avatar
Ethan Smith committed
92
    def __post_init__(self) -> None:
lintangsutawika's avatar
lintangsutawika committed
93
94
95
        if "." in self.dataset_path:
            import inspect
            from importlib import import_module
lintangsutawika's avatar
format  
lintangsutawika committed
96

lintangsutawika's avatar
lintangsutawika committed
97
            self.dataset_path = inspect.getfile(import_module(self.dataset_path))
98

Lintang Sutawika's avatar
Lintang Sutawika committed
99
100
101
        if self.generation_kwargs is not None:
            if self.output_type != "greedy_until":
                eval_logger.warning(
102
                    "passed `generation_kwargs`, but not using `output_type: greedy_until`!"
Lintang Sutawika's avatar
Lintang Sutawika committed
103
                )
104
                assert self.output_type != "greedy_until"
Lintang Sutawika's avatar
Lintang Sutawika committed
105
106
107
108
109
110
111

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

            if "until" not in self.generation_kwargs:
112
                self.generation_kwargs["until"] = [self.fewshot_delimiter]
Lintang Sutawika's avatar
Lintang Sutawika committed
113
114
115
116
        else:
            if self.output_type == "greedy_until":
                # ensure that we greedily generate in absence of explicit arguments otherwise
                self.generation_kwargs = {
Lintang Sutawika's avatar
Lintang Sutawika committed
117
                    "until": None
118
119
                    if self.fewshot_delimiter is None
                    else [self.fewshot_delimiter],
Lintang Sutawika's avatar
Lintang Sutawika committed
120
121
122
                    "do_sample": False,
                    "temperature": 0.0,
                }
123

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

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

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

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

haileyschoelkopf's avatar
haileyschoelkopf committed
137
138
139
140
141
142
143
144
145
146
        :return: dict
            A printable dictionary version of the TaskConfig object.

        # TODO: should any default value in the TaskConfig not be printed?
        """
        cfg_dict = asdict(self)
        # remove values that are `None`
        for k, v in list(cfg_dict.items()):
            if v is None:
                cfg_dict.pop(k)
haileyschoelkopf's avatar
haileyschoelkopf committed
147
148
149
            elif isinstance(v, Callable):
                # TODO: this should handle Promptsource template objects as a separate case?
                cfg_dict[k] = str(v)
haileyschoelkopf's avatar
haileyschoelkopf committed
150
        return cfg_dict
151

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

class Task(abc.ABC):
    """A task represents an entire benchmark including its dataset, problems,
    answers, and evaluation methods. See BoolQ for a simple example implementation

    A `doc` can be any python object which represents one instance of evaluation.
    This is usually a dictionary e.g.
        {"question": ..., "answer": ...} or
        {"question": ..., question, answer)
    """

    VERSION = None
164

165
166
167
168
169
170
171
172
    # The name of the `Task` benchmark as denoted in the HuggingFace datasets Hub
    # or a path to a custom `datasets` loading script.
    DATASET_PATH: str = None

    # The name of a subset within `DATASET_PATH`.
    DATASET_NAME: str = None

    OUTPUT_TYPE: str = None
lintangsutawika's avatar
lintangsutawika committed
173

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

haileyschoelkopf's avatar
haileyschoelkopf committed
207
        self._config = TaskConfig(**config) if config else TaskConfig()
208
209
210

        if not hasattr(self, "_filters"):
            self._filters = []
lintangsutawika's avatar
lintangsutawika committed
211
            for name, components in self._config.get(
212
                "filters", [["none", [["take_first", None]]]]
lintangsutawika's avatar
lintangsutawika committed
213
            ):
214
215
216
                filter_pipeline = build_filter_ensemble(name, components)
                self._filters.append(filter_pipeline)

lintangsutawika's avatar
lintangsutawika committed
217
        self.sampler = samplers.Sampler(
218
219
            list(self.fewshot_docs()), self, rnd=random.Random(1234)
        )
220

Ethan Smith's avatar
Ethan Smith committed
221
    def download(self, data_dir=None, cache_dir=None, download_mode=None) -> None:
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
        """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.
        """
246
247
248
249
250
251
252
        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,
        )
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289

    @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 []

290
291
292
293
294
295
296
297
298
299
    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
300
            eval_logger.warning(
301
                "has_training_docs and has_validation_docs are False"
302
                ", using test_docs as fewshot_docs but this is not recommended."
lintangsutawika's avatar
lintangsutawika committed
303
            )
304
305
            return self.test_docs()

306
307
308
309
310
311
312
313
314
315
    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
316

317
318
319
320
321
322
323
324
325
326
327
328
329
    @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
330
    def doc_to_decontamination_query(self, doc) -> None:
331
332
333
334
335
336
337
338
339
340
341
342
343
        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
344
    def build_all_requests(self, limit=None, rank=None, world_size=None) -> None:
345
346
347
348
349
350
351
352
353
354
        """Build a set of Instances for a task, and store them in task.instances"""
        if self.has_test_docs():
            docs = self.test_docs()
        elif self.has_validation_docs():
            docs = self.validation_docs()
        else:
            assert (
                False
            ), f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"

355
356
357
358
        eval_logger.info(
            f"Building contexts for task '{self._config.task}' on rank {rank}..."
        )

359
        instances = []
360
361
        for doc_id, doc in utils.create_iterator(
            enumerate(docs), rank, world_size, limit
lintangsutawika's avatar
lintangsutawika committed
362
        ):
363
            # sample fewshot context #TODO: need to offset doc_id by rank now!
364
            fewshot_ctx = self.fewshot_context(
365
366
                doc,
                self._config.num_fewshot,
367
            )
368

haileyschoelkopf's avatar
haileyschoelkopf committed
369
            # TODO: we should override self._config.repeats if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
370
371
372
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
373
                metadata=(self._config["task"], doc_id, self._config.repeats),
lintangsutawika's avatar
lintangsutawika committed
374
            )
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399

            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
400
            The number of times each instance in a dataset is inferred on. Defaults to 1,
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
            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
436
437
438
439
440
441
442
443
444
445
    @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))

446
    @utils.positional_deprecated
447
    def fewshot_context(self, doc, num_fewshot):
448
449
450
451
452
453
454
455
456
457
458
459
        """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:
460
461
            # always prepend the (possibly empty) task description
            labeled_examples = self._config.description
462
        else:
lintangsutawika's avatar
lintangsutawika committed
463
464
465
            labeled_examples = self._config.description + self.sampler.get_context(
                doc, num_fewshot
            )
466
467

        example = self.doc_to_text(doc)
468
469
470
471
        if type(example) == str:
            return labeled_examples + example
        elif type(example) == list:
            return [labeled_examples + ex for ex in example]
472
        elif type(example) == int:
lintangsutawika's avatar
lintangsutawika committed
473
474
475
476
477
            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)
478
479

    def apply_filters(self):
lintangsutawika's avatar
lintangsutawika committed
480
481
482
483
484
485
        if hasattr(self, "_filters"):
            for f in self._filters:
                f.apply(self._instances)
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances
486

baberabb's avatar
baberabb committed
487
    def dump_config(self) -> dict:
488
        """Returns a dictionary representing the task's config.
489
490
491
492
493

        :returns: str
            The fewshot context.
        """
        # TODO: this should only return the overrides applied to a non-YAML task's configuration.
494
        # (num_fewshot)
495
496
        return self._config.to_dict()

497
498

class ConfigurableTask(Task):
499
    VERSION = "Yaml"
500
    OUTPUT_TYPE = None
501
    CONFIG = None
502
503
504

    def __init__(
        self, data_dir=None, cache_dir=None, download_mode=None, config: dict = None
Ethan Smith's avatar
Ethan Smith committed
505
    ) -> None:  # TODO no super() call here
506
        # Get pre-configured attributes
507
        self._config = self.CONFIG
508

509
510
        # Use new configurations if there was no preconfiguration
        if self._config is None:
511
            self._config = TaskConfig(**config)
512
513
        # Overwrite configs
        else:
lintangsutawika's avatar
lintangsutawika committed
514
            if config is not None:
515
                self._config.__dict__.update(config)
516

517
        if self._config is None:
lintangsutawika's avatar
lintangsutawika committed
518
519
520
            raise ValueError(
                "Must pass a config to ConfigurableTask, either in cls.CONFIG or `config` kwarg"
            )
521
522

        if self._config.output_type is not None:
523
            assert self._config.output_type in ALL_OUTPUT_TYPES
524
525
            self.OUTPUT_TYPE = self._config.output_type

526
527
528
529
530
531
        if self._config.dataset_path is not None:
            self.DATASET_PATH = self._config.dataset_path

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

532
533
534
535
        self._metric_fn_list = {}
        self._metric_fn_kwargs = {}
        self._aggregation_list = {}
        self._higher_is_better = {}
536

537
        _metric_list = DEFAULT_METRIC_REGISTRY[self._config.output_type]
538
        if self._config.metric_list is None:
539
            # TODO: handle this in TaskConfig.__post_init__ ?
540
            for metric_name in _metric_list:
haileyschoelkopf's avatar
haileyschoelkopf committed
541
542
                self._metric_fn_list[metric_name] = get_metric(metric_name)
                self._aggregation_list[metric_name] = get_default_aggregation(
543
                    metric_name
haileyschoelkopf's avatar
haileyschoelkopf committed
544
545
                )
                self._higher_is_better[metric_name] = is_higher_better(metric_name)
546
547
548
549
550
551
552
553
554
        else:
            for metric_config in self._config.metric_list:
                assert "metric" in metric_config
                metric_name = metric_config["metric"]
                kwargs = {
                    key: metric_config[key]
                    for key in metric_config
                    if key not in ["metric", "aggregation", "higher_is_better"]
                }
555

556
                if self._config.process_results is not None:
557
558
                    self._metric_fn_list[metric_name] = None
                    self._metric_fn_kwargs[metric_name] = {}
559
560
561
562
563
564
565
566
                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:
                    self._metric_fn_list[metric_name] = get_metric(metric_name)
                    self._metric_fn_kwargs[metric_name] = kwargs
lintangsutawika's avatar
lintangsutawika committed
567

568
                if "aggregation" in metric_config:
569
                    agg_name = metric_config["aggregation"]
570
                    if type(agg_name) == str:
haileyschoelkopf's avatar
haileyschoelkopf committed
571
                        self._aggregation_list[metric_name] = get_aggregation(agg_name)
572
573
574
575
                    elif callable(agg_name):
                        self._aggregation_list[metric_name] = metric_config[
                            "aggregation"
                        ]
576
                else:
577
                    INV_AGG_REGISTRY = {v: k for k, v in AGGREGATION_REGISTRY.items()}
haileyschoelkopf's avatar
haileyschoelkopf committed
578
                    metric_agg = get_default_aggregation(metric_name)
579
                    eval_logger.warning(
580
581
582
                        f"metric {metric_name} is defined, but aggregation is not. "
                        f"using default "
                        f"aggregation={INV_AGG_REGISTRY[metric_agg]}"
583
                    )
584
                    self._aggregation_list[metric_name] = metric_agg
lintangsutawika's avatar
lintangsutawika committed
585

586
587
588
589
590
591
                if "higher_is_better" in metric_config:
                    self._higher_is_better[metric_name] = metric_config[
                        "higher_is_better"
                    ]
                else:
                    eval_logger.warning(
592
593
                        f"metric {metric_name} is defined, but higher_is_better is not. "
                        f"using default "
haileyschoelkopf's avatar
haileyschoelkopf committed
594
                        f"higher_is_better={is_higher_better(metric_name)}"
595
                    )
haileyschoelkopf's avatar
haileyschoelkopf committed
596
                    self._higher_is_better[metric_name] = is_higher_better(metric_name)
597

598
        self.download(self._config.dataset_kwargs)
599
600
601
        self._training_docs = None
        self._fewshot_docs = None

lintangsutawika's avatar
lintangsutawika committed
602
        if self._config.filter_list is not None:
lintangsutawika's avatar
lintangsutawika committed
603
            self._filters = []
lintangsutawika's avatar
lintangsutawika committed
604
605
606
607
608
609
610
611
            for filter_config in self._config.filter_list:
                for filter_pipeline in filter_config:
                    filter_name = filter_config["name"]
                    filter_functions = filter_config["filter"]
                    components = []
                    for function in filter_functions:
                        kwargs = {
                            key: function[key] for key in function if key != "function"
lintangsutawika's avatar
lintangsutawika committed
612
613
614
                        }
                        components.append([function["function"], kwargs])
                    filter_pipeline = build_filter_ensemble(filter_name, components)
lintangsutawika's avatar
lintangsutawika committed
615
                self._filters.append(filter_pipeline)
lintangsutawika's avatar
lintangsutawika committed
616
        else:
617
            self._filters = [build_filter_ensemble("none", [["take_first", None]])]
618
619

        if self._config.use_prompt is not None:
lintangsutawika's avatar
lintangsutawika committed
620
            eval_logger.info(f"loading prompt {self._config.use_prompt}")
621
            self.prompt = get_prompt(
lintangsutawika's avatar
lintangsutawika committed
622
623
                self._config.use_prompt, self.DATASET_PATH, self.DATASET_NAME
            )
624
625
626
        else:
            self.prompt = None

lintangsutawika's avatar
lintangsutawika committed
627
628
        if self.fewshot_docs() is not None:
            self.sampler = samplers.Sampler(
629
                list(self.fewshot_docs()), self, rnd=random.Random(1234)
630
            )
631

632
        if self.has_test_docs():
633
            self.task_docs = self.test_docs()
634
        elif self.has_validation_docs():
635
            self.task_docs = self.validation_docs()
636
637
638
639
640
        else:
            assert (
                False
            ), f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"

641
        # Test One Doc
642
        self.features = list(self.task_docs.features.keys())
643
644
        self.multiple_input = 0
        self.multiple_target = 0
645
        test_doc = self.task_docs[0]
646
        test_text = self.doc_to_text(test_doc)
647
        test_target = self.doc_to_target(test_doc)
lintangsutawika's avatar
lintangsutawika committed
648
649
650
651
652

        if self._config.doc_to_choice is not None:
            test_choice = self.doc_to_choice(test_doc)
            if type(test_choice) is not list:
                eval_logger.error("doc_to_choice must return list")
653
654
            else:
                num_choice = len(test_choice)
655

656
657
            if type(test_text) is int:
                self.multiple_input = num_choice
lintangsutawika's avatar
lintangsutawika committed
658
659
        else:
            test_choice = None
660

661
        if type(test_target) is list:
662
            self.multiple_target = len(test_target)
lintangsutawika's avatar
lintangsutawika committed
663
664
665
666
667
        else:
            if (type(test_target) is int) and (test_choice is not None):
                test_target = test_choice[test_target]
            else:
                test_target = str(test_target)
668

lintangsutawika's avatar
lintangsutawika committed
669
670
671
        if test_choice is not None:
            check_choices = test_choice
        else:
lintangsutawika's avatar
lintangsutawika committed
672
            check_choices = [test_target]
lintangsutawika's avatar
lintangsutawika committed
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687

        for choice in check_choices:
            choice_has_whitespace = True if " " in choice else False
            delimiter_has_whitespace = (
                True if " " in self._config.target_delimiter else False
            )

            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(
                    f'Both target_delimiter and target choice: "{choice}" does not have whitespace, ignore if the language you are evaluating on does not require/use whitespace'
                )
688

Ethan Smith's avatar
Ethan Smith committed
689
    def download(self, dataset_kwargs=None) -> None:
690
691
692
693
694
695
        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
696
    def has_training_docs(self) -> bool:
697
698
699
700
701
        if self._config.training_split is not None:
            return True
        else:
            return False

baberabb's avatar
baberabb committed
702
    def has_validation_docs(self) -> bool:
703
704
705
706
707
        if self._config.validation_split is not None:
            return True
        else:
            return False

baberabb's avatar
baberabb committed
708
    def has_test_docs(self) -> bool:
709
710
711
712
713
        if self._config.test_split is not None:
            return True
        else:
            return False

baberabb's avatar
baberabb committed
714
    def training_docs(self) -> datasets.Dataset:
715
        if self.has_training_docs():
716
            if self._config.process_docs is not None:
717
718
719
                return self._config.process_docs(
                    self.dataset[self._config.training_split]
                )
720
721
            return self.dataset[self._config.training_split]

baberabb's avatar
baberabb committed
722
    def validation_docs(self) -> datasets.Dataset:
723
        if self.has_validation_docs():
724
            if self._config.process_docs is not None:
725
726
727
                return self._config.process_docs(
                    self.dataset[self._config.validation_split]
                )
728
729
            return self.dataset[self._config.validation_split]

baberabb's avatar
baberabb committed
730
    def test_docs(self) -> datasets.Dataset:
731
        if self.has_test_docs():
732
            if self._config.process_docs is not None:
733
                return self._config.process_docs(self.dataset[self._config.test_split])
734
735
            return self.dataset[self._config.test_split]

736
    def fewshot_docs(self):
737
        if self._config.fewshot_split is not None:
738
            return self.dataset[self._config.fewshot_split]
739
740
741
        else:
            if self._config.num_fewshot > 0:
                eval_logger.warning(
haileyschoelkopf's avatar
haileyschoelkopf committed
742
                    f"Task '{self._config.task}': "
743
744
745
746
                    "num_fewshot > 0 but fewshot_split is None. "
                    "using preconfigured rule."
                )
            return super().fewshot_docs()
747

748
749
750
751
752
753
754
755
756
    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

757
758
759
760
761
    def should_decontaminate(self):
        return self._config.should_decontaminate

    def doc_to_decontamination_query(self, doc):
        if self._config.should_decontaminate:
762
763
764
765
766
767
            if self._config.doc_to_decontamination_query in self.features:
                return doc[self._config.doc_to_decontamination_query]
            else:
                return ast.literal_eval(
                    utils.apply_template(self._config.doc_to_decontamination_query, doc)
                )
768

769
770
771
772
773
774
775
776
777
778
779
780
    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):
781
782
        if self.prompt is not None:
            doc_to_text = self.prompt
783
784
        else:
            doc_to_text = self._config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
785

786
787
788
        if type(doc_to_text) == int:
            return doc_to_text
        elif type(doc_to_text) == str:
789
            if doc_to_text in self.features:
790
791
792
                # if self._config.doc_to_choice is not None:
                #     return self.doc_to_choice(doc)[doc[doc_to_text]]
                # else:
793
794
                return doc[doc_to_text]
            else:
lintangsutawika's avatar
lintangsutawika committed
795
                text_string = utils.apply_template(doc_to_text, doc)
lintangsutawika's avatar
lintangsutawika committed
796
                if text_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
797
798
799
                    return ast.literal_eval(text_string)
                else:
                    return text_string
800
        elif callable(doc_to_text):
801
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
802
        # Used when applying a Promptsource template
803
        elif hasattr(doc_to_text, "apply"):
804
805
806
807
808
            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")
809
                return self._config.fewshot_delimiter
810
        else:
811
            print(type(doc_to_text))
812
            raise TypeError
813

814
    def doc_to_target(self, doc: dict) -> Union[int, str, list]:
815
816
        if self.prompt is not None:
            doc_to_target = self.prompt
817
818
819
        else:
            doc_to_target = self._config.doc_to_target

820
821
822
        if type(doc_to_target) == int:
            return doc_to_target
        elif type(doc_to_target) == str:
823
            if doc_to_target in self.features:
824
825
826
827
                # if self._config.doc_to_choice is not None:
                #     return self.doc_to_choice(doc)[doc[doc_to_target]]
                # else:
                return doc[doc_to_target]
828
            else:
lintangsutawika's avatar
lintangsutawika committed
829
                target_string = utils.apply_template(doc_to_target, doc)
lintangsutawika's avatar
lintangsutawika committed
830
                if target_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
831
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
832
833
834
835
836
                elif (
                    len(target_string) >= 2
                    and (target_string[0] == "[")
                    and (target_string[-1] == "]")
                ):
837
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
838
839
                else:
                    return target_string
840
841
        elif type(doc_to_target) == list:
            return doc_to_target
842
        elif callable(doc_to_target):
843
            return doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
844
        # Used when applying a Promptsource template
845
        elif hasattr(doc_to_target, "apply"):
846
            applied_prompt = doc_to_target.apply(doc)
847
848
849
850
            if len(applied_prompt) == 2:
                return applied_prompt[1]
            else:
                eval_logger.warning("Applied prompt returns empty string")
851
                return self._config.fewshot_delimiter
852
853
        else:
            raise TypeError
854

baberabb's avatar
baberabb committed
855
    def doc_to_choice(self, doc: Any) -> List[str]:
856
857
        if self.prompt is not None:
            doc_to_choice = self.prompt
lintangsutawika's avatar
lintangsutawika committed
858
        elif self._config.doc_to_choice is None:
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
            eval_logger.error("doc_to_choice was called but not set in config")
        else:
            doc_to_choice = self._config.doc_to_choice

        if type(doc_to_choice) == str:
            return ast.literal_eval(utils.apply_template(doc_to_choice, doc))
        elif type(doc_to_choice) == list:
            return doc_to_choice
        elif type(doc_to_choice) == dict:
            return list(doc_to_choice.values())
        elif callable(doc_to_choice):
            return doc_to_choice(doc)
        elif hasattr(doc_to_choice, "get_answer_choices_list"):
            return doc_to_choice.get_answer_choices_list(doc)
        else:
            raise TypeError
875

876
    def gold_alias(self, doc):
877
878
879
880
881
        # returns a version of the gold target answer to a document,
        # which should be passed into metric for scoring as the ground truth.

        # in multiple_choice tasks, this should be castable to an int corresponding to the index
        # within the answer choices, while doc_to_target is the string version of {{answer_choices[gold]}}.
lintangsutawika's avatar
lintangsutawika committed
882
        if self._config.gold_alias is not None:
883
884
            doc_to_target = self._config.gold_alias
        else:
lintangsutawika's avatar
lintangsutawika committed
885
            return self.doc_to_target(doc)
886
887
888
889
890
891
892
893
894
895

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

baberabb's avatar
baberabb committed
896
897
898
    def construct_requests(
        self, doc: dict, ctx: str, **kwargs
    ) -> Union[List[Instance], Instance]:
899
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
900
            arguments = (ctx, self.doc_to_target(doc))
901
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
902
            arguments = (self.doc_to_target(doc),)
903
        elif self.OUTPUT_TYPE == "multiple_choice":
904
            choices = self.doc_to_choice(doc)
905
            target_delimiter = self._config.target_delimiter
906
907
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
908
                cont = self.doc_to_target(doc)
909
                arguments = [(ctx, f"{target_delimiter}{cont}") for ctx in choices]
910
            else:
911
                # Otherwise they are placed in the continuation
912
                arguments = [(ctx, f"{target_delimiter}{cont}") for cont in choices]
913

914
            request_list = [
915
916
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
917
                    doc=doc,
918
                    arguments=arg,
919
                    idx=i,
920
921
                    **kwargs,
                )
922
                for i, arg in enumerate(arguments)
923
            ]
924
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
925
            if "acc_mutual_info" in self._metric_fn_list.keys():
926
927
928
                # if we are calculating multiple choice accuracy
                # using mutual information instead of raw loglikelihood as metric, need unconditional lls.

lintangsutawika's avatar
lintangsutawika committed
929
                # here mutual info refers to calculating
930
931
932
933
934
935
                # 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
936
                            doc=doc,
937
                            arguments=("", "{}".format(choice)),
938
939
940
                            idx=i,
                            **kwargs,
                        )
lintangsutawika's avatar
lintangsutawika committed
941
                        for i, choice in enumerate(choices)
942
943
944
                    ]
                )
            return request_list
lintangsutawika's avatar
lintangsutawika committed
945

946
        elif self.OUTPUT_TYPE == "greedy_until":
947
            arguments = (ctx, self._config.generation_kwargs)
lintangsutawika's avatar
lintangsutawika committed
948
949

        return Instance(
lintangsutawika's avatar
lintangsutawika committed
950
951
            request_type=self.OUTPUT_TYPE, doc=doc, arguments=arguments, idx=0, **kwargs
        )
952
953

    def process_results(self, doc, results):
lintangsutawika's avatar
lintangsutawika committed
954
955
        if callable(self._config.process_results):
            return self._config.process_results(doc, results)
lintangsutawika's avatar
lintangsutawika committed
956

957
        result_dict = {}
958
        use_metric = list(self._metric_fn_list.keys())
959
960
961
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
962
963
964
965
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
966
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
967
            (loglikelihood,) = results
968
969
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
haileyschoelkopf's avatar
haileyschoelkopf committed
970
            return {
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
                **(
                    {"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
986
            }
987
        elif self.OUTPUT_TYPE == "multiple_choice":
988
            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
989

990
            # retrieve choices in List[str] form, to compute choice lengths, etc.
991
            choices = self.doc_to_choice(doc)
992
993
            completion_len = np.array([float(len(i)) for i in choices])

994
995
            if (
                2 * len(choices) == len(lls)
996
                and "acc_mutual_info" in self._metric_fn_list.keys()
997
998
999
1000
1001
1002
1003
            ):
                # 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]
1004

1005
1006
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
1007

1008
1009
            if self.multiple_input:
                gold = self.doc_to_text(doc)
1010
            else:
1011
                gold = self.doc_to_target(doc)
1012
1013
1014

            gold_index_error = False
            if type(gold) is list:
Lintang Sutawika's avatar
Lintang Sutawika committed
1015
1016
                gold = [i if i < len(choices) else -100 for i in gold]
                if -100 in gold:
1017
1018
1019
                    gold_index_error = True
            else:
                if type(gold) is int:
Lintang Sutawika's avatar
Lintang Sutawika committed
1020
                    gold = gold if gold < len(choices) else -100
1021
                elif type(gold) is str:
Lintang Sutawika's avatar
Lintang Sutawika committed
1022
                    gold = choices.index(gold) if gold in choices else -100
lintangsutawika's avatar
lintangsutawika committed
1023

Lintang Sutawika's avatar
Lintang Sutawika committed
1024
                if gold == -100:
1025
1026
1027
1028
                    gold_index_error = True

            if gold_index_error:
                eval_logger.warning(
lintangsutawika's avatar
lintangsutawika committed
1029
                    f"Label index was not in within range of available choices,"
1030
1031
                    f"Sample:\n\n{doc}\n\n"
                )
lintangsutawika's avatar
lintangsutawika committed
1032

1033
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
1034
1035
                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
1036
                exact_match = int(any([is_greedy[i] if i != -100 else 0 for i in gold]))
lintangsutawika's avatar
lintangsutawika committed
1037
1038
1039
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
1040
                # 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
1041
                exact_match = int(is_greedy[gold]) if gold != -100 else 0
1042
1043

            result_dict = {
1044
                **({"acc": acc} if "acc" in use_metric else {}),
1045
1046
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
1047
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
1048
                **({"exact_match": exact_match} if "exact_match" in use_metric else {}),
1049
1050
            }

1051
            if "acc_mutual_info" in use_metric:
lintangsutawika's avatar
lintangsutawika committed
1052
1053
1054
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
1055
1056
1057
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

1058
        elif self.OUTPUT_TYPE == "greedy_until":
1059
            gold = self.doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
1060
            if self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1061
                # If you set doc_to_choice,
lintangsutawika's avatar
lintangsutawika committed
1062
                # it assumes that doc_to_target returns a number.
1063
1064
                choices = self.doc_to_choice(doc)
                gold = choices[gold]
lintangsutawika's avatar
lintangsutawika committed
1065
1066
            else:
                gold = str(gold)
1067

lintangsutawika's avatar
lintangsutawika committed
1068
            result = results[0]
lintangsutawika's avatar
lintangsutawika committed
1069
            for metric in self._metric_fn_list.keys():
haileyschoelkopf's avatar
haileyschoelkopf committed
1070
1071
1072
1073
1074
1075
                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 = []
                    for gold_option in gold:
1076
                        try:
1077
                            result_score = self._metric_fn_list[metric](
1078
1079
                                references=[gold_option],
                                predictions=[result],
1080
                                **self._metric_fn_kwargs[metric],
1081
1082
                            )
                        except TypeError:  # TODO: this is hacky and I don't want to do it
1083
                            result_score = self._metric_fn_list[metric](
haileyschoelkopf's avatar
haileyschoelkopf committed
1084
1085
1086
                                [gold_option, result]
                            )
                        if isinstance(result_score, dict):
haileyschoelkopf's avatar
haileyschoelkopf committed
1087
                            # TODO: this handles the case where HF evaluate returns a dict.
1088
                            result_score = result_score[metric]
haileyschoelkopf's avatar
haileyschoelkopf committed
1089
                        scores.append(result_score)
haileyschoelkopf's avatar
haileyschoelkopf committed
1090
                    if any(scores):
haileyschoelkopf's avatar
haileyschoelkopf committed
1091
                        result_score = 1.0
haileyschoelkopf's avatar
haileyschoelkopf committed
1092
                    else:
haileyschoelkopf's avatar
haileyschoelkopf committed
1093
                        result_score = 0.0
haileyschoelkopf's avatar
haileyschoelkopf committed
1094
                else:
1095
                    try:
1096
                        result_score = self._metric_fn_list[metric](
1097
1098
                            references=[gold],
                            predictions=[result],
1099
                            **self._metric_fn_kwargs[metric],
1100
                        )
1101
1102
                    except TypeError:  # needed for now in order to use a different interface between our own metrics and HF Evaluate metrics
                        result_score = self._metric_fn_list[metric]([gold, result])
1103
1104
1105
1106
                    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
1107
        else:
lintangsutawika's avatar
lintangsutawika committed
1108
1109
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
1110
                "'loglikelihood', 'loglikelihood_rolling', 'greedy_until' or 'multiple_choice'",
1111
            )
1112
1113
1114
1115
1116
1117
1118

        return result_dict

    def aggregation(self):
        return self._aggregation_list

    def higher_is_better(self):
haileyschoelkopf's avatar
haileyschoelkopf committed
1119
        return self._higher_is_better
1120
1121
1122
1123
1124


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

baberabb's avatar
baberabb committed
1125
    def doc_to_target(self, doc: dict) -> str:
1126
1127
        return " " + doc["choices"][doc["gold"]]

baberabb's avatar
baberabb committed
1128
    def construct_requests(self, doc: dict, ctx: str, **kwargs) -> List[Instance]:
1129
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1130
1131
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1132
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1133
                doc=doc,
1134
                arguments=(ctx, " {}".format(choice)),
1135
                idx=i,
1136
1137
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1138
1139
            for i, choice in enumerate(doc["choices"])
        ]
1140

baberabb's avatar
baberabb committed
1141
    def process_results(self, doc: dict, results: List[Tuple[float, bool]]) -> dict:
lintangsutawika's avatar
lintangsutawika committed
1142
1143
1144
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
        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
1156
    def higher_is_better(self) -> dict:
1157
1158
1159
1160
1161
        return {
            "acc": True,
            "acc_norm": True,
        }

baberabb's avatar
baberabb committed
1162
    def aggregation(self) -> dict:
1163
1164
1165
1166
1167
1168
        return {
            "acc": mean,
            "acc_norm": mean,
        }


lintangsutawika's avatar
lintangsutawika committed
1169
class PerplexityTask(Task):
1170
1171
    OUTPUT_TYPE = "loglikelihood_rolling"

baberabb's avatar
baberabb committed
1172
    def has_training_docs(self) -> bool:
1173
1174
        return False

baberabb's avatar
baberabb committed
1175
    def fewshot_examples(self, k: int, rnd) -> List:
1176
1177
1178
        assert k == 0
        return []

baberabb's avatar
baberabb committed
1179
    def fewshot_context(self, doc: dict, num_fewshot: int) -> Literal[""]:
1180
1181
1182
1183
1184
1185
        assert (
            num_fewshot == 0
        ), "The number of fewshot examples must be 0 for perplexity tasks."

        return ""

baberabb's avatar
baberabb committed
1186
    def higher_is_better(self) -> dict:
1187
1188
1189
1190
1191
1192
1193
1194
1195
        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
1196
    def doc_to_text(self, doc) -> str:
1197
1198
1199
1200
1201
        return ""

    def doc_to_target(self, doc):
        return doc

baberabb's avatar
baberabb committed
1202
    def construct_requests(self, doc: dict, ctx: Union[str, None], **kwargs):
1203
1204
        assert not ctx

lintangsutawika's avatar
lintangsutawika committed
1205
1206
1207
1208
1209
1210
1211
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1212

baberabb's avatar
baberabb committed
1213
    def process_results(self, doc: dict, results: float) -> dict:
1214
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1215
1216
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1217
1218
1219
1220
1221
1222
        return {
            "word_perplexity": (loglikelihood, words),
            "byte_perplexity": (loglikelihood, bytes_),
            "bits_per_byte": (loglikelihood, bytes_),
        }

baberabb's avatar
baberabb committed
1223
    def aggregation(self) -> dict:
1224
1225
1226
1227
1228
1229
1230
        return {
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
            "bits_per_byte": bits_per_byte,
        }

    @classmethod
baberabb's avatar
baberabb committed
1231
    def count_bytes(cls, doc) -> int:
1232
1233
1234
        return len(doc.encode("utf-8"))

    @classmethod
baberabb's avatar
baberabb committed
1235
    def count_words(cls, doc) -> int:
1236
1237
        """Downstream tasks with custom word boundaries should override this!"""
        return len(re.split(r"\s+", doc))