task.py 46.5 KB
Newer Older
1
import abc
2
import ast
3
import functools
Herbie Bradley's avatar
Herbie Bradley committed
4
5
6
7
8
9
import itertools
import random
import re
from collections.abc import Callable
from dataclasses import asdict, dataclass, field
from typing import Any, List, Literal, Tuple, Union
10
11

import datasets
Herbie Bradley's avatar
Herbie Bradley committed
12
import evaluate
13
import numpy as np
Herbie Bradley's avatar
Herbie Bradley committed
14
15
16
import scipy.special as sp
import yaml
from tqdm import tqdm
17

18
from lm_eval import utils
19
from lm_eval.api import samplers
lintangsutawika's avatar
lintangsutawika committed
20
from lm_eval.api.filter import FilterEnsemble
Herbie Bradley's avatar
Herbie Bradley committed
21
from lm_eval.api.instance import Instance
lintangsutawika's avatar
lintangsutawika committed
22
23
from lm_eval.api.metrics import (
    bits_per_byte,
Herbie Bradley's avatar
Herbie Bradley committed
24
    mean,
lintangsutawika's avatar
lintangsutawika committed
25
    metric_max_over_ground_truths,
Herbie Bradley's avatar
Herbie Bradley committed
26
    weighted_perplexity,
lintangsutawika's avatar
lintangsutawika committed
27
28
)
from lm_eval.api.registry import (
Herbie Bradley's avatar
Herbie Bradley committed
29
30
31
    AGGREGATION_REGISTRY,
    DEFAULT_METRIC_REGISTRY,
    OUTPUT_TYPE_REGISTRY,
haileyschoelkopf's avatar
haileyschoelkopf committed
32
33
    get_aggregation,
    get_default_aggregation,
Herbie Bradley's avatar
Herbie Bradley committed
34
    get_metric,
haileyschoelkopf's avatar
haileyschoelkopf committed
35
    is_higher_better,
lintangsutawika's avatar
lintangsutawika committed
36
)
Herbie Bradley's avatar
Herbie Bradley committed
37
38
39
from lm_eval.filters import build_filter_ensemble
from lm_eval.logger import eval_logger
from lm_eval.prompts import get_prompt
40

41
42
43
44
45
46
47
ALL_OUTPUT_TYPES = [
    "loglikelihood",
    "multiple_choice",
    "loglikelihood_rolling",
    "greedy_until",
]

48
49
50

@dataclass
class TaskConfig(dict):
51
    # task naming/registry
52
    task: str = None
53
    group: Union[str, list] = None
54
55
56
    # HF dataset options.
    # which dataset to use,
    # and what splits for what purpose
57
58
    dataset_path: str = None
    dataset_name: str = None
59
    dataset_kwargs: dict = None
60
61
62
    training_split: str = None
    validation_split: str = None
    test_split: str = None
lintangsutawika's avatar
lintangsutawika committed
63
    fewshot_split: str = None  # TODO: assert that this not None if num_fewshot > 0. (?) assert if this is same split as one evaling (?)
64
65
    # formatting / prompting options.
    # see docs/advanced_task_guide.md for more info
66
    process_docs: Callable = None
67
68
    doc_to_text: Union[Callable, str] = None
    doc_to_target: Union[Callable, str] = None
lintangsutawika's avatar
lintangsutawika committed
69
    doc_to_choice: Union[Callable, str, dict, list] = None
70
    gold_alias: Union[Callable, str] = None
lintangsutawika's avatar
lintangsutawika committed
71
    process_results: Union[Callable, str] = None
72
    use_prompt: str = None
73
    description: str = ""
74
75
    target_delimiter: str = " "
    fewshot_delimiter: str = "\n\n"
76
    # runtime configuration options
77
    num_fewshot: int = 0
78
    # scoring options
79
    metric_list: list = None
80
    output_type: str = "greedy_until"
81
    generation_kwargs: dict = None
82
    repeats: int = 1
lintangsutawika's avatar
lintangsutawika committed
83
    filter_list: Union[str, list] = None
84
85
    should_decontaminate: bool = False
    doc_to_decontamination_query: str = None
86

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

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

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

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

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

            if "until" not in self.generation_kwargs:
109
                self.generation_kwargs["until"] = [self.fewshot_delimiter]
Lintang Sutawika's avatar
Lintang Sutawika committed
110
111
112
113
        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
114
                    "until": None
115
116
                    if self.fewshot_delimiter is None
                    else [self.fewshot_delimiter],
Lintang Sutawika's avatar
Lintang Sutawika committed
117
118
119
                    "do_sample": False,
                    "temperature": 0.0,
                }
120

haileyschoelkopf's avatar
haileyschoelkopf committed
121
122
        # TODO: how to make TaskConfigs be de- and re-serializable, even when using the !function constructor?

123
124
125
    def __getitem__(self, item):
        return getattr(self, item)

126
127
128
    def __setitem__(self, item, value):
        return setattr(self, item, value)

129
    def to_dict(self):
130
131
        """dumps the current config as a dictionary object, as a printable format.
        null fields will not be printed.
haileyschoelkopf's avatar
haileyschoelkopf committed
132
        Used for dumping results alongside full task configuration
133

haileyschoelkopf's avatar
haileyschoelkopf committed
134
135
136
137
138
139
140
141
142
143
        :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
144
145
146
            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
147
        return cfg_dict
148

149
150
151
152
153
154
155
156
157
158
159
160

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
161

162
163
164
165
166
167
168
169
    # 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
170

171
172
173
174
175
176
    def __init__(
        self,
        data_dir=None,
        cache_dir=None,
        download_mode=None,
        config=None,
Ethan Smith's avatar
Ethan Smith committed
177
    ) -> None:
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
        """
        :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
204
        self._config = TaskConfig(**config) if config else TaskConfig()
205
206
207

        if not hasattr(self, "_filters"):
            self._filters = []
lintangsutawika's avatar
lintangsutawika committed
208
            for name, components in self._config.get(
209
                "filters", [["none", [["take_first", None]]]]
lintangsutawika's avatar
lintangsutawika committed
210
            ):
211
212
213
                filter_pipeline = build_filter_ensemble(name, components)
                self._filters.append(filter_pipeline)

lintangsutawika's avatar
lintangsutawika committed
214
        self.sampler = samplers.Sampler(
215
216
            list(self.fewshot_docs()), self, rnd=random.Random(1234)
        )
217

Ethan Smith's avatar
Ethan Smith committed
218
    def download(self, data_dir=None, cache_dir=None, download_mode=None) -> None:
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
        """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.
        """
243
244
245
246
247
248
249
        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,
        )
250

251
252
253
254
255
    @property
    def config(self):
        """Returns the TaskConfig associated with this class."""
        return self._config

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
290
291
    @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 []

292
293
294
295
296
297
298
299
300
301
    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
302
            eval_logger.warning(
303
                "has_training_docs and has_validation_docs are False"
304
                ", using test_docs as fewshot_docs but this is not recommended."
lintangsutawika's avatar
lintangsutawika committed
305
            )
306
307
            return self.test_docs()

308
309
310
311
312
313
314
315
316
317
    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
318

319
320
321
322
323
324
325
326
327
328
329
330
331
    @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
332
    def doc_to_decontamination_query(self, doc) -> None:
333
334
335
336
337
338
339
340
341
342
343
344
345
        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
346
    def build_all_requests(self, limit=None, rank=None, world_size=None) -> None:
347
348
349
350
351
352
353
354
355
356
        """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!"

357
        eval_logger.info(
358
            f"Building contexts for task '{self.config.task}' on rank {rank}..."
359
360
        )

361
        instances = []
362
363
        for doc_id, doc in utils.create_iterator(
            enumerate(docs), rank, world_size, limit
lintangsutawika's avatar
lintangsutawika committed
364
        ):
365
            # sample fewshot context #TODO: need to offset doc_id by rank now!
366
            fewshot_ctx = self.fewshot_context(
367
                doc,
368
                self.config.num_fewshot,
369
            )
370

371
            # TODO: we should override self.config.repeats if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
372
373
374
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
375
                metadata=(self.config["task"], doc_id, self.config.repeats),
lintangsutawika's avatar
lintangsutawika committed
376
            )
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401

            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
402
            The number of times each instance in a dataset is inferred on. Defaults to 1,
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
436
437
            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
438
439
440
441
442
443
444
445
446
447
    @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))

448
    @utils.positional_deprecated
449
    def fewshot_context(self, doc, num_fewshot):
450
451
452
453
454
455
456
457
458
459
460
461
        """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:
462
            # always prepend the (possibly empty) task description
463
            labeled_examples = self.config.description
464
        else:
465
            labeled_examples = self.config.description + self.sampler.get_context(
lintangsutawika's avatar
lintangsutawika committed
466
467
                doc, num_fewshot
            )
468
469

        example = self.doc_to_text(doc)
470
471
472
473
        if type(example) == str:
            return labeled_examples + example
        elif type(example) == list:
            return [labeled_examples + ex for ex in example]
474
        elif type(example) == int:
475
            if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
476
477
478
479
                choices = self.doc_to_choice(doc)
                return labeled_examples + choices[example]
            else:
                return labeled_examples + str(example)
480
481

    def apply_filters(self):
lintangsutawika's avatar
lintangsutawika committed
482
483
484
485
486
487
        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
488

baberabb's avatar
baberabb committed
489
    def dump_config(self) -> dict:
490
        """Returns a dictionary representing the task's config.
491
492
493
494
495

        :returns: str
            The fewshot context.
        """
        # TODO: this should only return the overrides applied to a non-YAML task's configuration.
496
        # (num_fewshot)
497
        return self.config.to_dict()
498

499
500

class ConfigurableTask(Task):
501
    VERSION = "Yaml"
502
    OUTPUT_TYPE = None
503
    CONFIG = None
504
505
506

    def __init__(
        self, data_dir=None, cache_dir=None, download_mode=None, config: dict = None
Ethan Smith's avatar
Ethan Smith committed
507
    ) -> None:  # TODO no super() call here
508
        # Get pre-configured attributes
509
        self._config = self.CONFIG
510

511
        # Use new configurations if there was no preconfiguration
512
        if self.config is None:
513
            self._config = TaskConfig(**config)
514
515
        # Overwrite configs
        else:
lintangsutawika's avatar
lintangsutawika committed
516
            if config is not None:
517
                self._config.__dict__.update(config)
518

519
        if self.config is None:
lintangsutawika's avatar
lintangsutawika committed
520
521
522
            raise ValueError(
                "Must pass a config to ConfigurableTask, either in cls.CONFIG or `config` kwarg"
            )
523

524
525
526
        if self.config.output_type is not None:
            assert self.config.output_type in ALL_OUTPUT_TYPES
            self.OUTPUT_TYPE = self.config.output_type
527

528
529
        if self.config.dataset_path is not None:
            self.DATASET_PATH = self.config.dataset_path
530

531
532
        if self.config.dataset_name is not None:
            self.DATASET_NAME = self.config.dataset_name
533

534
535
536
537
        self._metric_fn_list = {}
        self._metric_fn_kwargs = {}
        self._aggregation_list = {}
        self._higher_is_better = {}
538

539
540
        _metric_list = DEFAULT_METRIC_REGISTRY[self.config.output_type]
        if self.config.metric_list is None:
541
            # TODO: handle this in TaskConfig.__post_init__ ?
542
            for metric_name in _metric_list:
haileyschoelkopf's avatar
haileyschoelkopf committed
543
544
                self._metric_fn_list[metric_name] = get_metric(metric_name)
                self._aggregation_list[metric_name] = get_default_aggregation(
545
                    metric_name
haileyschoelkopf's avatar
haileyschoelkopf committed
546
547
                )
                self._higher_is_better[metric_name] = is_higher_better(metric_name)
548
        else:
549
            for metric_config in self.config.metric_list:
550
551
552
553
554
555
556
                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"]
                }
557

558
                if self.config.process_results is not None:
559
560
                    self._metric_fn_list[metric_name] = None
                    self._metric_fn_kwargs[metric_name] = {}
561
562
563
564
565
566
567
568
                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
569

570
                if "aggregation" in metric_config:
571
                    agg_name = metric_config["aggregation"]
572
                    if type(agg_name) == str:
haileyschoelkopf's avatar
haileyschoelkopf committed
573
                        self._aggregation_list[metric_name] = get_aggregation(agg_name)
574
575
576
577
                    elif callable(agg_name):
                        self._aggregation_list[metric_name] = metric_config[
                            "aggregation"
                        ]
578
                else:
579
                    INV_AGG_REGISTRY = {v: k for k, v in AGGREGATION_REGISTRY.items()}
haileyschoelkopf's avatar
haileyschoelkopf committed
580
                    metric_agg = get_default_aggregation(metric_name)
581
                    eval_logger.warning(
582
583
584
                        f"metric {metric_name} is defined, but aggregation is not. "
                        f"using default "
                        f"aggregation={INV_AGG_REGISTRY[metric_agg]}"
585
                    )
586
                    self._aggregation_list[metric_name] = metric_agg
lintangsutawika's avatar
lintangsutawika committed
587

588
589
590
591
592
593
                if "higher_is_better" in metric_config:
                    self._higher_is_better[metric_name] = metric_config[
                        "higher_is_better"
                    ]
                else:
                    eval_logger.warning(
594
595
                        f"metric {metric_name} is defined, but higher_is_better is not. "
                        f"using default "
haileyschoelkopf's avatar
haileyschoelkopf committed
596
                        f"higher_is_better={is_higher_better(metric_name)}"
597
                    )
haileyschoelkopf's avatar
haileyschoelkopf committed
598
                    self._higher_is_better[metric_name] = is_higher_better(metric_name)
599

600
        self.download(self.config.dataset_kwargs)
601
602
603
        self._training_docs = None
        self._fewshot_docs = None

604
        if self.config.filter_list is not None:
lintangsutawika's avatar
lintangsutawika committed
605
            self._filters = []
606
            for filter_config in self.config.filter_list:
lintangsutawika's avatar
lintangsutawika committed
607
608
609
610
611
612
613
                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
614
615
616
                        }
                        components.append([function["function"], kwargs])
                    filter_pipeline = build_filter_ensemble(filter_name, components)
lintangsutawika's avatar
lintangsutawika committed
617
                self._filters.append(filter_pipeline)
lintangsutawika's avatar
lintangsutawika committed
618
        else:
619
            self._filters = [build_filter_ensemble("none", [["take_first", None]])]
620

621
622
        if self.config.use_prompt is not None:
            eval_logger.info(f"loading prompt {self.config.use_prompt}")
623
            self.prompt = get_prompt(
624
                self.config.use_prompt, self.DATASET_PATH, self.DATASET_NAME
lintangsutawika's avatar
lintangsutawika committed
625
            )
626
627
628
        else:
            self.prompt = None

lintangsutawika's avatar
lintangsutawika committed
629
630
        if self.fewshot_docs() is not None:
            self.sampler = samplers.Sampler(
631
                list(self.fewshot_docs()), self, rnd=random.Random(1234)
632
            )
633

634
        if self.has_test_docs():
635
            self.task_docs = self.test_docs()
636
        elif self.has_validation_docs():
637
            self.task_docs = self.validation_docs()
638
639
640
641
642
        else:
            assert (
                False
            ), f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"

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

651
        if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
652
653
654
            test_choice = self.doc_to_choice(test_doc)
            if type(test_choice) is not list:
                eval_logger.error("doc_to_choice must return list")
655
656
            else:
                num_choice = len(test_choice)
657

658
659
            if type(test_text) is int:
                self.multiple_input = num_choice
660
661
        else:
            test_choice = None
662

663
        if type(test_target) is list:
664
            self.multiple_target = len(test_target)
665
        else:
lintangsutawika's avatar
lintangsutawika committed
666
            if (type(test_target) is int) and (test_choice is not None):
lintangsutawika's avatar
lintangsutawika committed
667
                test_target = test_choice[test_target]
668
            else:
lintangsutawika's avatar
lintangsutawika committed
669
                test_target = str(test_target)
670

671
672
673
        if test_choice is not None:
            check_choices = test_choice
        else:
lintangsutawika's avatar
lintangsutawika committed
674
            check_choices = [test_target]
675
676
677
678

        for choice in check_choices:
            choice_has_whitespace = True if " " in choice else False
            delimiter_has_whitespace = (
679
                True if " " in self.config.target_delimiter else False
680
681
682
683
684
685
686
            )

            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):
687
                eval_logger.warning(
688
                    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'
689
                )
690

Ethan Smith's avatar
Ethan Smith committed
691
    def download(self, dataset_kwargs=None) -> None:
692
693
694
695
696
697
        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
698
    def has_training_docs(self) -> bool:
699
        if self.config.training_split is not None:
700
701
702
703
            return True
        else:
            return False

baberabb's avatar
baberabb committed
704
    def has_validation_docs(self) -> bool:
705
        if self.config.validation_split is not None:
706
707
708
709
            return True
        else:
            return False

baberabb's avatar
baberabb committed
710
    def has_test_docs(self) -> bool:
711
        if self.config.test_split is not None:
712
713
714
715
            return True
        else:
            return False

baberabb's avatar
baberabb committed
716
    def training_docs(self) -> datasets.Dataset:
717
        if self.has_training_docs():
718
719
720
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.training_split]
721
                )
722
            return self.dataset[self.config.training_split]
723

baberabb's avatar
baberabb committed
724
    def validation_docs(self) -> datasets.Dataset:
725
        if self.has_validation_docs():
726
727
728
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.validation_split]
729
                )
730
            return self.dataset[self.config.validation_split]
731

baberabb's avatar
baberabb committed
732
    def test_docs(self) -> datasets.Dataset:
733
        if self.has_test_docs():
734
735
736
            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]
737

738
    def fewshot_docs(self):
739
740
        if self.config.fewshot_split is not None:
            return self.dataset[self.config.fewshot_split]
741
        else:
742
            if self.config.num_fewshot > 0:
743
                eval_logger.warning(
744
                    f"Task '{self.config.task}': "
745
746
747
748
                    "num_fewshot > 0 but fewshot_split is None. "
                    "using preconfigured rule."
                )
            return super().fewshot_docs()
749

750
751
752
753
754
755
756
757
758
    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

759
    def should_decontaminate(self):
760
        return self.config.should_decontaminate
761
762

    def doc_to_decontamination_query(self, doc):
763
764
765
        if self.config.should_decontaminate:
            if self.config.doc_to_decontamination_query in self.features:
                return doc[self.config.doc_to_decontamination_query]
766
767
            else:
                return ast.literal_eval(
768
                    utils.apply_template(self.config.doc_to_decontamination_query, doc)
769
                )
770

771
772
773
774
775
776
777
778
779
780
781
782
    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):
783
784
        if self.prompt is not None:
            doc_to_text = self.prompt
785
        else:
786
            doc_to_text = self.config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
787

788
789
790
        if type(doc_to_text) == int:
            return doc_to_text
        elif type(doc_to_text) == str:
791
            if doc_to_text in self.features:
792
                # if self.config.doc_to_choice is not None:
793
794
                #     return self.doc_to_choice(doc)[doc[doc_to_text]]
                # else:
795
796
                return doc[doc_to_text]
            else:
lintangsutawika's avatar
lintangsutawika committed
797
                text_string = utils.apply_template(doc_to_text, doc)
lintangsutawika's avatar
lintangsutawika committed
798
                if text_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
799
800
801
                    return ast.literal_eval(text_string)
                else:
                    return text_string
802
        elif callable(doc_to_text):
803
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
804
        # Used when applying a Promptsource template
805
        elif hasattr(doc_to_text, "apply"):
806
807
808
809
810
            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")
811
                return self.config.fewshot_delimiter
812
        else:
813
            print(type(doc_to_text))
814
            raise TypeError
815

816
    def doc_to_target(self, doc: dict) -> Union[int, str, list]:
817
818
        if self.prompt is not None:
            doc_to_target = self.prompt
819
        else:
820
            doc_to_target = self.config.doc_to_target
821

822
823
824
        if type(doc_to_target) == int:
            return doc_to_target
        elif type(doc_to_target) == str:
825
            if doc_to_target in self.features:
826
                # if self.config.doc_to_choice is not None:
827
828
829
                #     return self.doc_to_choice(doc)[doc[doc_to_target]]
                # else:
                return doc[doc_to_target]
830
            else:
lintangsutawika's avatar
lintangsutawika committed
831
                target_string = utils.apply_template(doc_to_target, doc)
lintangsutawika's avatar
lintangsutawika committed
832
                if target_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
833
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
834
835
836
837
838
                elif (
                    len(target_string) >= 2
                    and (target_string[0] == "[")
                    and (target_string[-1] == "]")
                ):
839
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
840
841
                else:
                    return target_string
842
843
        elif type(doc_to_target) == list:
            return doc_to_target
844
        elif callable(doc_to_target):
845
            return doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
846
        # Used when applying a Promptsource template
847
        elif hasattr(doc_to_target, "apply"):
848
            applied_prompt = doc_to_target.apply(doc)
849
850
851
852
            if len(applied_prompt) == 2:
                return applied_prompt[1]
            else:
                eval_logger.warning("Applied prompt returns empty string")
853
                return self.config.fewshot_delimiter
854
855
        else:
            raise TypeError
856

baberabb's avatar
baberabb committed
857
    def doc_to_choice(self, doc: Any) -> List[str]:
858
859
        if self.prompt is not None:
            doc_to_choice = self.prompt
860
        elif self.config.doc_to_choice is None:
861
862
            eval_logger.error("doc_to_choice was called but not set in config")
        else:
863
            doc_to_choice = self.config.doc_to_choice
864
865
866
867
868
869
870
871
872
873
874
875
876

        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
877

878
    def gold_alias(self, doc):
879
880
881
882
883
        # 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]}}.
884
885
        if self.config.gold_alias is not None:
            doc_to_target = self.config.gold_alias
886
        else:
lintangsutawika's avatar
lintangsutawika committed
887
            return self.doc_to_target(doc)
888
889
890
891
892
893
894
895
896
897

        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
898
899
900
    def construct_requests(
        self, doc: dict, ctx: str, **kwargs
    ) -> Union[List[Instance], Instance]:
901
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
902
            arguments = (ctx, self.doc_to_target(doc))
903
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
904
            arguments = (self.doc_to_target(doc),)
905
        elif self.OUTPUT_TYPE == "multiple_choice":
906
            choices = self.doc_to_choice(doc)
907
            target_delimiter = self.config.target_delimiter
908
909
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
910
                cont = self.doc_to_target(doc)
911
                arguments = [(ctx, f"{target_delimiter}{cont}") for ctx in choices]
912
            else:
913
                # Otherwise they are placed in the continuation
914
                arguments = [(ctx, f"{target_delimiter}{cont}") for cont in choices]
915

916
            request_list = [
917
918
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
919
                    doc=doc,
920
                    arguments=arg,
921
                    idx=i,
922
923
                    **kwargs,
                )
924
                for i, arg in enumerate(arguments)
925
            ]
926
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
927
            if "acc_mutual_info" in self._metric_fn_list.keys():
928
929
930
                # if we are calculating multiple choice accuracy
                # using mutual information instead of raw loglikelihood as metric, need unconditional lls.

lintangsutawika's avatar
lintangsutawika committed
931
                # here mutual info refers to calculating
932
933
934
935
936
937
                # 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
938
                            doc=doc,
939
                            arguments=("", "{}".format(choice)),
940
941
942
                            idx=i,
                            **kwargs,
                        )
lintangsutawika's avatar
lintangsutawika committed
943
                        for i, choice in enumerate(choices)
944
945
946
                    ]
                )
            return request_list
lintangsutawika's avatar
lintangsutawika committed
947

948
        elif self.OUTPUT_TYPE == "greedy_until":
949
            arguments = (ctx, self.config.generation_kwargs)
lintangsutawika's avatar
lintangsutawika committed
950
951

        return Instance(
lintangsutawika's avatar
lintangsutawika committed
952
953
            request_type=self.OUTPUT_TYPE, doc=doc, arguments=arguments, idx=0, **kwargs
        )
954
955
956

    def process_results(self, doc, results):

957
958
        if callable(self.config.process_results):
            return self.config.process_results(doc, results)
lintangsutawika's avatar
lintangsutawika committed
959

960
        result_dict = {}
961
        use_metric = list(self._metric_fn_list.keys())
962
963
964
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
965
966
967
968
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
969
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
970
            (loglikelihood,) = results
971
972
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
haileyschoelkopf's avatar
haileyschoelkopf committed
973
            return {
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
                **(
                    {"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
989
            }
990
        elif self.OUTPUT_TYPE == "multiple_choice":
991
            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
992

993
            # retrieve choices in List[str] form, to compute choice lengths, etc.
994
            choices = self.doc_to_choice(doc)
995
996
            completion_len = np.array([float(len(i)) for i in choices])

997
            if 2 * len(choices) == len(lls) and "acc_mutual_info" in use_metric:
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"
                )
1032
1033
1034

            if "ece" in use_metric:
                # Convert lls from log-probabilities to normalized probabilities
1035
1036
                norm_probs: list[float] = np.exp(lls - sp.logsumexp(lls)).tolist()
                calib_scores: list[float] = [0.0] * len(choices)
1037
1038
1039
1040
1041
                if isinstance(gold, list):
                    for g in gold:
                        calib_scores[g] = 1.0
                else:
                    calib_scores[gold] = 1.0
1042
                calibration_probs: dict[str, list[float]] = {
1043
1044
1045
1046
                    "probs": norm_probs,
                    "scores": calib_scores,
                }

1047
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
1048
1049
                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
1050
                exact_match = int(any([is_greedy[i] if i != -100 else 0 for i in gold]))
lintangsutawika's avatar
lintangsutawika committed
1051
1052
1053
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
1054
                # 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
1055
                exact_match = int(is_greedy[gold]) if gold != -100 else 0
1056
1057

            result_dict = {
1058
                **({"acc": acc} if "acc" in use_metric else {}),
1059
1060
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
1061
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
1062
                **({"exact_match": exact_match} if "exact_match" in use_metric else {}),
1063
                **({"ece": calibration_probs} if "ece" in use_metric else {}),
1064
1065
            }

1066
            if "acc_mutual_info" in use_metric:
lintangsutawika's avatar
lintangsutawika committed
1067
1068
1069
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
1070
1071
1072
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

1073
        elif self.OUTPUT_TYPE == "greedy_until":
1074
            gold = self.doc_to_target(doc)
1075
            if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1076
                # If you set doc_to_choice,
lintangsutawika's avatar
lintangsutawika committed
1077
                # it assumes that doc_to_target returns a number.
1078
1079
                choices = self.doc_to_choice(doc)
                gold = choices[gold]
lintangsutawika's avatar
lintangsutawika committed
1080
1081
            else:
                gold = str(gold)
1082

lintangsutawika's avatar
lintangsutawika committed
1083
            result = results[0]
lintangsutawika's avatar
lintangsutawika committed
1084
            for metric in self._metric_fn_list.keys():
haileyschoelkopf's avatar
haileyschoelkopf committed
1085
1086
1087
1088
1089
1090
                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:
1091
                        try:
1092
                            result_score = self._metric_fn_list[metric](
1093
1094
                                references=[gold_option],
                                predictions=[result],
1095
                                **self._metric_fn_kwargs[metric],
1096
1097
                            )
                        except TypeError:  # TODO: this is hacky and I don't want to do it
1098
                            result_score = self._metric_fn_list[metric](
haileyschoelkopf's avatar
haileyschoelkopf committed
1099
1100
1101
                                [gold_option, result]
                            )
                        if isinstance(result_score, dict):
haileyschoelkopf's avatar
haileyschoelkopf committed
1102
                            # TODO: this handles the case where HF evaluate returns a dict.
1103
                            result_score = result_score[metric]
haileyschoelkopf's avatar
haileyschoelkopf committed
1104
                        scores.append(result_score)
haileyschoelkopf's avatar
haileyschoelkopf committed
1105
                    if any(scores):
1106
                        result_score = 1.0
haileyschoelkopf's avatar
haileyschoelkopf committed
1107
                    else:
1108
                        result_score = 0.0
haileyschoelkopf's avatar
haileyschoelkopf committed
1109
                else:
1110
                    try:
1111
                        result_score = self._metric_fn_list[metric](
1112
1113
                            references=[gold],
                            predictions=[result],
1114
                            **self._metric_fn_kwargs[metric],
1115
                        )
1116
1117
                    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])
1118
1119
1120
1121
                    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
1122
        else:
lintangsutawika's avatar
lintangsutawika committed
1123
1124
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
1125
                "'loglikelihood', 'loglikelihood_rolling', 'greedy_until' or 'multiple_choice'",
1126
            )
1127
1128
1129
1130
1131
1132
1133

        return result_dict

    def aggregation(self):
        return self._aggregation_list

    def higher_is_better(self):
haileyschoelkopf's avatar
haileyschoelkopf committed
1134
        return self._higher_is_better
1135
1136
1137
1138
1139


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

baberabb's avatar
baberabb committed
1140
    def doc_to_target(self, doc: dict) -> str:
1141
1142
        return " " + doc["choices"][doc["gold"]]

baberabb's avatar
baberabb committed
1143
    def construct_requests(self, doc: dict, ctx: str, **kwargs) -> List[Instance]:
1144
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1145
1146
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1147
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1148
                doc=doc,
1149
                arguments=(ctx, " {}".format(choice)),
1150
                idx=i,
1151
1152
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1153
1154
            for i, choice in enumerate(doc["choices"])
        ]
1155

baberabb's avatar
baberabb committed
1156
    def process_results(self, doc: dict, results: List[Tuple[float, bool]]) -> dict:
lintangsutawika's avatar
lintangsutawika committed
1157
1158
1159
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
        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
1171
    def higher_is_better(self) -> dict:
1172
1173
1174
1175
1176
        return {
            "acc": True,
            "acc_norm": True,
        }

baberabb's avatar
baberabb committed
1177
    def aggregation(self) -> dict:
1178
1179
1180
1181
1182
1183
        return {
            "acc": mean,
            "acc_norm": mean,
        }


lintangsutawika's avatar
lintangsutawika committed
1184
class PerplexityTask(Task):
1185
1186
    OUTPUT_TYPE = "loglikelihood_rolling"

baberabb's avatar
baberabb committed
1187
    def has_training_docs(self) -> bool:
1188
1189
        return False

baberabb's avatar
baberabb committed
1190
    def fewshot_examples(self, k: int, rnd) -> List:
1191
1192
1193
        assert k == 0
        return []

baberabb's avatar
baberabb committed
1194
    def fewshot_context(self, doc: dict, num_fewshot: int) -> Literal[""]:
1195
1196
1197
1198
1199
1200
        assert (
            num_fewshot == 0
        ), "The number of fewshot examples must be 0 for perplexity tasks."

        return ""

baberabb's avatar
baberabb committed
1201
    def higher_is_better(self) -> dict:
1202
1203
1204
1205
1206
1207
1208
1209
1210
        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
1211
    def doc_to_text(self, doc) -> str:
1212
1213
1214
1215
1216
        return ""

    def doc_to_target(self, doc):
        return doc

baberabb's avatar
baberabb committed
1217
    def construct_requests(self, doc: dict, ctx: Union[str, None], **kwargs):
1218
1219
        assert not ctx

lintangsutawika's avatar
lintangsutawika committed
1220
1221
1222
1223
1224
1225
1226
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1227

baberabb's avatar
baberabb committed
1228
    def process_results(self, doc: dict, results: float) -> dict:
1229
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1230
1231
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1232
1233
1234
1235
1236
1237
        return {
            "word_perplexity": (loglikelihood, words),
            "byte_perplexity": (loglikelihood, bytes_),
            "bits_per_byte": (loglikelihood, bytes_),
        }

baberabb's avatar
baberabb committed
1238
    def aggregation(self) -> dict:
1239
1240
1241
1242
1243
1244
1245
        return {
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
            "bits_per_byte": bits_per_byte,
        }

    @classmethod
baberabb's avatar
baberabb committed
1246
    def count_bytes(cls, doc) -> int:
1247
1248
1249
        return len(doc.encode("utf-8"))

    @classmethod
baberabb's avatar
baberabb committed
1250
    def count_words(cls, doc) -> int:
1251
1252
        """Downstream tasks with custom word boundaries should override this!"""
        return len(re.split(r"\s+", doc))