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

Ethan Smith's avatar
Ethan Smith committed
91
    def __post_init__(self) -> None:
lintangsutawika's avatar
lintangsutawika committed
92
93
94
        if "." in self.dataset_path:
            import inspect
            from importlib import import_module
lintangsutawika's avatar
format  
lintangsutawika committed
95

lintangsutawika's avatar
lintangsutawika committed
96
            self.dataset_path = inspect.getfile(import_module(self.dataset_path))
97

Lintang Sutawika's avatar
Lintang Sutawika committed
98
99
100
        if self.generation_kwargs is not None:
            if self.output_type != "greedy_until":
                eval_logger.warning(
101
                    "passed `generation_kwargs`, but not using `output_type: greedy_until`!"
Lintang Sutawika's avatar
Lintang Sutawika committed
102
                )
103
                assert self.output_type != "greedy_until"
Lintang Sutawika's avatar
Lintang Sutawika committed
104
105
106
107
108
109
110

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

            if "until" not in self.generation_kwargs:
111
                self.generation_kwargs["until"] = [self.fewshot_delimiter]
Lintang Sutawika's avatar
Lintang Sutawika committed
112
113
114
115
        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
116
                    "until": None
117
118
                    if self.fewshot_delimiter is None
                    else [self.fewshot_delimiter],
Lintang Sutawika's avatar
Lintang Sutawika committed
119
120
121
                    "do_sample": False,
                    "temperature": 0.0,
                }
122

haileyschoelkopf's avatar
haileyschoelkopf committed
123
124
        # TODO: how to make TaskConfigs be de- and re-serializable, even when using the !function constructor?

125
126
127
    def __getitem__(self, item):
        return getattr(self, item)

128
129
130
    def __setitem__(self, item, value):
        return setattr(self, item, value)

131
    def to_dict(self):
132
133
        """dumps the current config as a dictionary object, as a printable format.
        null fields will not be printed.
haileyschoelkopf's avatar
haileyschoelkopf committed
134
        Used for dumping results alongside full task configuration
135

haileyschoelkopf's avatar
haileyschoelkopf committed
136
137
138
139
140
141
142
143
144
145
        :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
146
147
148
            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
149
        return cfg_dict
150

151
152
153
154
155
156
157
158
159
160
161
162

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
163

164
165
166
167
168
169
170
171
    # 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
172

173
174
175
176
177
178
    def __init__(
        self,
        data_dir=None,
        cache_dir=None,
        download_mode=None,
        config=None,
Ethan Smith's avatar
Ethan Smith committed
179
    ) -> None:
180
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
        """
        :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
206
        self._config = TaskConfig(**config) if config else TaskConfig()
207
208
209

        if not hasattr(self, "_filters"):
            self._filters = []
lintangsutawika's avatar
lintangsutawika committed
210
            for name, components in self._config.get(
211
                "filters", [["none", [["take_first", None]]]]
lintangsutawika's avatar
lintangsutawika committed
212
            ):
213
214
215
                filter_pipeline = build_filter_ensemble(name, components)
                self._filters.append(filter_pipeline)

lintangsutawika's avatar
lintangsutawika committed
216
        self.sampler = samplers.Sampler(
217
218
            list(self.fewshot_docs()), self, rnd=random.Random(1234)
        )
219

Ethan Smith's avatar
Ethan Smith committed
220
    def download(self, data_dir=None, cache_dir=None, download_mode=None) -> None:
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
        """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.
        """
245
246
247
248
249
250
251
        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,
        )
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288

    @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 []

289
290
291
292
293
294
295
296
297
298
    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
299
            eval_logger.warning(
300
                "has_training_docs and has_validation_docs are False"
301
                ", using test_docs as fewshot_docs but this is not recommended."
lintangsutawika's avatar
lintangsutawika committed
302
            )
303
304
            return self.test_docs()

305
306
307
308
309
310
311
312
313
314
    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
315

316
317
318
319
320
321
322
323
324
325
326
327
328
    @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
329
    def doc_to_decontamination_query(self, doc) -> None:
330
331
332
333
334
335
336
337
338
339
340
341
342
        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
343
    def build_all_requests(self, limit=None, rank=None, world_size=None) -> None:
344
345
346
347
348
349
350
351
352
353
        """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!"

354
355
356
357
        eval_logger.info(
            f"Building contexts for task '{self._config.task}' on rank {rank}..."
        )

358
        instances = []
359
360
        for doc_id, doc in utils.create_iterator(
            enumerate(docs), rank, world_size, limit
lintangsutawika's avatar
lintangsutawika committed
361
        ):
362
            # sample fewshot context #TODO: need to offset doc_id by rank now!
363
            fewshot_ctx = self.fewshot_context(
364
365
                doc,
                self._config.num_fewshot,
366
            )
367

haileyschoelkopf's avatar
haileyschoelkopf committed
368
            # TODO: we should override self._config.repeats if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
369
370
371
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
372
                metadata=(self._config["task"], doc_id, self._config.repeats),
lintangsutawika's avatar
lintangsutawika committed
373
            )
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

            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
399
            The number of times each instance in a dataset is inferred on. Defaults to 1,
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
            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
435
436
437
438
439
440
441
442
443
444
    @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))

445
    @utils.positional_deprecated
446
    def fewshot_context(self, doc, num_fewshot):
447
448
449
450
451
452
453
454
455
456
457
458
        """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:
459
460
            # always prepend the (possibly empty) task description
            labeled_examples = self._config.description
461
        else:
lintangsutawika's avatar
lintangsutawika committed
462
463
464
            labeled_examples = self._config.description + self.sampler.get_context(
                doc, num_fewshot
            )
465
466

        example = self.doc_to_text(doc)
467
468
469
470
        if type(example) == str:
            return labeled_examples + example
        elif type(example) == list:
            return [labeled_examples + ex for ex in example]
471
        elif type(example) == int:
lintangsutawika's avatar
lintangsutawika committed
472
473
474
475
476
            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)
477
478

    def apply_filters(self):
lintangsutawika's avatar
lintangsutawika committed
479
480
481
482
483
484
        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
485

baberabb's avatar
baberabb committed
486
    def dump_config(self) -> dict:
487
        """Returns a dictionary representing the task's config.
488
489
490
491
492

        :returns: str
            The fewshot context.
        """
        # TODO: this should only return the overrides applied to a non-YAML task's configuration.
493
        # (num_fewshot)
494
495
        return self._config.to_dict()

496
497

class ConfigurableTask(Task):
498
    VERSION = "Yaml"
499
    OUTPUT_TYPE = None
500
    CONFIG = None
501
502
503

    def __init__(
        self, data_dir=None, cache_dir=None, download_mode=None, config: dict = None
Ethan Smith's avatar
Ethan Smith committed
504
    ) -> None:  # TODO no super() call here
505
        # Get pre-configured attributes
506
        self._config = self.CONFIG
507

508
509
        # Use new configurations if there was no preconfiguration
        if self._config is None:
510
            self._config = TaskConfig(**config)
511
512
        # Overwrite configs
        else:
lintangsutawika's avatar
lintangsutawika committed
513
            if config is not None:
514
                self._config.__dict__.update(config)
515

516
        if self._config is None:
lintangsutawika's avatar
lintangsutawika committed
517
518
519
            raise ValueError(
                "Must pass a config to ConfigurableTask, either in cls.CONFIG or `config` kwarg"
            )
520
521

        if self._config.output_type is not None:
522
            assert self._config.output_type in ALL_OUTPUT_TYPES
523
524
            self.OUTPUT_TYPE = self._config.output_type

525
526
527
528
529
530
        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

531
532
533
534
        self._metric_fn_list = {}
        self._metric_fn_kwargs = {}
        self._aggregation_list = {}
        self._higher_is_better = {}
535

536
        _metric_list = DEFAULT_METRIC_REGISTRY[self._config.output_type]
537
        if self._config.metric_list is None:
538
            # TODO: handle this in TaskConfig.__post_init__ ?
539
            for metric_name in _metric_list:
haileyschoelkopf's avatar
haileyschoelkopf committed
540
541
                self._metric_fn_list[metric_name] = get_metric(metric_name)
                self._aggregation_list[metric_name] = get_default_aggregation(
542
                    metric_name
haileyschoelkopf's avatar
haileyschoelkopf committed
543
544
                )
                self._higher_is_better[metric_name] = is_higher_better(metric_name)
545
546
547
548
549
550
551
552
553
        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"]
                }
554

555
                if self._config.process_results is not None:
556
557
                    self._metric_fn_list[metric_name] = None
                    self._metric_fn_kwargs[metric_name] = {}
558
559
560
561
562
563
564
565
                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
566

567
                if "aggregation" in metric_config:
568
                    agg_name = metric_config["aggregation"]
569
                    if type(agg_name) == str:
haileyschoelkopf's avatar
haileyschoelkopf committed
570
                        self._aggregation_list[metric_name] = get_aggregation(agg_name)
571
572
573
574
                    elif callable(agg_name):
                        self._aggregation_list[metric_name] = metric_config[
                            "aggregation"
                        ]
575
                else:
576
                    INV_AGG_REGISTRY = {v: k for k, v in AGGREGATION_REGISTRY.items()}
haileyschoelkopf's avatar
haileyschoelkopf committed
577
                    metric_agg = get_default_aggregation(metric_name)
578
                    eval_logger.warning(
579
580
581
                        f"metric {metric_name} is defined, but aggregation is not. "
                        f"using default "
                        f"aggregation={INV_AGG_REGISTRY[metric_agg]}"
582
                    )
583
                    self._aggregation_list[metric_name] = metric_agg
lintangsutawika's avatar
lintangsutawika committed
584

585
586
587
588
589
590
                if "higher_is_better" in metric_config:
                    self._higher_is_better[metric_name] = metric_config[
                        "higher_is_better"
                    ]
                else:
                    eval_logger.warning(
591
592
                        f"metric {metric_name} is defined, but higher_is_better is not. "
                        f"using default "
haileyschoelkopf's avatar
haileyschoelkopf committed
593
                        f"higher_is_better={is_higher_better(metric_name)}"
594
                    )
haileyschoelkopf's avatar
haileyschoelkopf committed
595
                    self._higher_is_better[metric_name] = is_higher_better(metric_name)
596

597
        self.download(self._config.dataset_kwargs)
598
599
600
        self._training_docs = None
        self._fewshot_docs = None

lintangsutawika's avatar
lintangsutawika committed
601
        if self._config.filter_list is not None:
lintangsutawika's avatar
lintangsutawika committed
602
            self._filters = []
lintangsutawika's avatar
lintangsutawika committed
603
604
605
606
607
608
609
610
            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
611
612
613
                        }
                        components.append([function["function"], kwargs])
                    filter_pipeline = build_filter_ensemble(filter_name, components)
lintangsutawika's avatar
lintangsutawika committed
614
                self._filters.append(filter_pipeline)
lintangsutawika's avatar
lintangsutawika committed
615
        else:
616
            self._filters = [build_filter_ensemble("none", [["take_first", None]])]
617
618

        if self._config.use_prompt is not None:
lintangsutawika's avatar
lintangsutawika committed
619
            eval_logger.info(f"loading prompt {self._config.use_prompt}")
620
            self.prompt = get_prompt(
lintangsutawika's avatar
lintangsutawika committed
621
622
                self._config.use_prompt, self.DATASET_PATH, self.DATASET_NAME
            )
623
624
625
        else:
            self.prompt = None

lintangsutawika's avatar
lintangsutawika committed
626
627
        if self.fewshot_docs() is not None:
            self.sampler = samplers.Sampler(
628
                list(self.fewshot_docs()), self, rnd=random.Random(1234)
629
            )
630

631
        if self.has_test_docs():
632
            self.task_docs = self.test_docs()
633
        elif self.has_validation_docs():
634
            self.task_docs = self.validation_docs()
635
636
637
638
639
        else:
            assert (
                False
            ), f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"

640
        # Test One Doc
641
        self.features = list(self.task_docs.features.keys())
642
643
        self.multiple_input = 0
        self.multiple_target = 0
644
        test_doc = self.task_docs[0]
645
        test_text = self.doc_to_text(test_doc)
646
        test_target = self.doc_to_target(test_doc)
lintangsutawika's avatar
lintangsutawika committed
647
648
649
650
651

        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")
652
653
            else:
                num_choice = len(test_choice)
654

655
656
            if type(test_text) is int:
                self.multiple_input = num_choice
lintangsutawika's avatar
lintangsutawika committed
657
658
        else:
            test_choice = None
659

660
        if type(test_target) is list:
661
            self.multiple_target = len(test_target)
lintangsutawika's avatar
lintangsutawika committed
662
663
664
665
666
        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)
667

lintangsutawika's avatar
lintangsutawika committed
668
669
670
        if test_choice is not None:
            check_choices = test_choice
        else:
lintangsutawika's avatar
lintangsutawika committed
671
            check_choices = [test_target]
lintangsutawika's avatar
lintangsutawika committed
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686

        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'
                )
687

Ethan Smith's avatar
Ethan Smith committed
688
    def download(self, dataset_kwargs=None) -> None:
689
690
691
692
693
694
        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
695
    def has_training_docs(self) -> bool:
696
697
698
699
700
        if self._config.training_split is not None:
            return True
        else:
            return False

baberabb's avatar
baberabb committed
701
    def has_validation_docs(self) -> bool:
702
703
704
705
706
        if self._config.validation_split is not None:
            return True
        else:
            return False

baberabb's avatar
baberabb committed
707
    def has_test_docs(self) -> bool:
708
709
710
711
712
        if self._config.test_split is not None:
            return True
        else:
            return False

baberabb's avatar
baberabb committed
713
    def training_docs(self) -> datasets.Dataset:
714
        if self.has_training_docs():
715
            if self._config.process_docs is not None:
716
717
718
                return self._config.process_docs(
                    self.dataset[self._config.training_split]
                )
719
720
            return self.dataset[self._config.training_split]

baberabb's avatar
baberabb committed
721
    def validation_docs(self) -> datasets.Dataset:
722
        if self.has_validation_docs():
723
            if self._config.process_docs is not None:
724
725
726
                return self._config.process_docs(
                    self.dataset[self._config.validation_split]
                )
727
728
            return self.dataset[self._config.validation_split]

baberabb's avatar
baberabb committed
729
    def test_docs(self) -> datasets.Dataset:
730
        if self.has_test_docs():
731
            if self._config.process_docs is not None:
732
                return self._config.process_docs(self.dataset[self._config.test_split])
733
734
            return self.dataset[self._config.test_split]

735
    def fewshot_docs(self):
736
        if self._config.fewshot_split is not None:
737
            return self.dataset[self._config.fewshot_split]
738
739
740
        else:
            if self._config.num_fewshot > 0:
                eval_logger.warning(
haileyschoelkopf's avatar
haileyschoelkopf committed
741
                    f"Task '{self._config.task}': "
742
743
744
745
                    "num_fewshot > 0 but fewshot_split is None. "
                    "using preconfigured rule."
                )
            return super().fewshot_docs()
746

747
748
749
750
751
752
753
754
755
    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

756
757
758
759
760
    def should_decontaminate(self):
        return self._config.should_decontaminate

    def doc_to_decontamination_query(self, doc):
        if self._config.should_decontaminate:
761
762
763
764
765
766
            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)
                )
767

768
769
770
771
772
773
774
775
776
777
778
779
    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):
780
781
        if self.prompt is not None:
            doc_to_text = self.prompt
782
783
        else:
            doc_to_text = self._config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
784

785
786
787
        if type(doc_to_text) == int:
            return doc_to_text
        elif type(doc_to_text) == str:
788
            if doc_to_text in self.features:
789
790
791
                # if self._config.doc_to_choice is not None:
                #     return self.doc_to_choice(doc)[doc[doc_to_text]]
                # else:
792
793
                return doc[doc_to_text]
            else:
lintangsutawika's avatar
lintangsutawika committed
794
                text_string = utils.apply_template(doc_to_text, doc)
lintangsutawika's avatar
lintangsutawika committed
795
                if text_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
796
797
798
                    return ast.literal_eval(text_string)
                else:
                    return text_string
799
        elif callable(doc_to_text):
800
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
801
        # Used when applying a Promptsource template
802
        elif hasattr(doc_to_text, "apply"):
803
804
805
806
807
            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")
808
                return self._config.fewshot_delimiter
809
        else:
810
            print(type(doc_to_text))
811
            raise TypeError
812

813
    def doc_to_target(self, doc: dict) -> Union[int, str, list]:
814
815
        if self.prompt is not None:
            doc_to_target = self.prompt
816
817
818
        else:
            doc_to_target = self._config.doc_to_target

819
820
821
        if type(doc_to_target) == int:
            return doc_to_target
        elif type(doc_to_target) == str:
822
            if doc_to_target in self.features:
823
824
825
826
                # 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]
827
            else:
lintangsutawika's avatar
lintangsutawika committed
828
                target_string = utils.apply_template(doc_to_target, doc)
lintangsutawika's avatar
lintangsutawika committed
829
                if target_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
830
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
831
832
833
834
835
                elif (
                    len(target_string) >= 2
                    and (target_string[0] == "[")
                    and (target_string[-1] == "]")
                ):
836
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
837
838
                else:
                    return target_string
839
840
        elif type(doc_to_target) == list:
            return doc_to_target
841
        elif callable(doc_to_target):
842
            return doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
843
        # Used when applying a Promptsource template
844
        elif hasattr(doc_to_target, "apply"):
845
            applied_prompt = doc_to_target.apply(doc)
846
847
848
849
            if len(applied_prompt) == 2:
                return applied_prompt[1]
            else:
                eval_logger.warning("Applied prompt returns empty string")
850
                return self._config.fewshot_delimiter
851
852
        else:
            raise TypeError
853

baberabb's avatar
baberabb committed
854
    def doc_to_choice(self, doc: Any) -> List[str]:
855
856
        if self.prompt is not None:
            doc_to_choice = self.prompt
lintangsutawika's avatar
lintangsutawika committed
857
        elif self._config.doc_to_choice is None:
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
            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
874

875
    def gold_alias(self, doc):
876
877
878
879
880
        # 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
881
        if self._config.gold_alias is not None:
882
883
            doc_to_target = self._config.gold_alias
        else:
lintangsutawika's avatar
lintangsutawika committed
884
            return self.doc_to_target(doc)
885
886
887
888
889
890
891
892
893
894

        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
895
896
897
    def construct_requests(
        self, doc: dict, ctx: str, **kwargs
    ) -> Union[List[Instance], Instance]:
898
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
899
            arguments = (ctx, self.doc_to_target(doc))
900
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
901
            arguments = (self.doc_to_target(doc),)
902
        elif self.OUTPUT_TYPE == "multiple_choice":
903
            choices = self.doc_to_choice(doc)
904
            target_delimiter = self._config.target_delimiter
905
906
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
907
                cont = self.doc_to_target(doc)
908
                arguments = [(ctx, f"{target_delimiter}{cont}") for ctx in choices]
909
            else:
910
                # Otherwise they are placed in the continuation
911
                arguments = [(ctx, f"{target_delimiter}{cont}") for cont in choices]
912

913
            request_list = [
914
915
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
916
                    doc=doc,
917
                    arguments=arg,
918
                    idx=i,
919
920
                    **kwargs,
                )
921
                for i, arg in enumerate(arguments)
922
            ]
923
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
924
            if "acc_mutual_info" in self._metric_fn_list.keys():
925
926
927
                # if we are calculating multiple choice accuracy
                # using mutual information instead of raw loglikelihood as metric, need unconditional lls.

lintangsutawika's avatar
lintangsutawika committed
928
                # here mutual info refers to calculating
929
930
931
932
933
934
                # 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
935
                            doc=doc,
936
                            arguments=("", "{}".format(choice)),
937
938
939
                            idx=i,
                            **kwargs,
                        )
lintangsutawika's avatar
lintangsutawika committed
940
                        for i, choice in enumerate(choices)
941
942
943
                    ]
                )
            return request_list
lintangsutawika's avatar
lintangsutawika committed
944

945
        elif self.OUTPUT_TYPE == "greedy_until":
946
            arguments = (ctx, self._config.generation_kwargs)
lintangsutawika's avatar
lintangsutawika committed
947
948

        return Instance(
lintangsutawika's avatar
lintangsutawika committed
949
950
            request_type=self.OUTPUT_TYPE, doc=doc, arguments=arguments, idx=0, **kwargs
        )
951
952

    def process_results(self, doc, results):
lintangsutawika's avatar
lintangsutawika committed
953
954
        if callable(self._config.process_results):
            return self._config.process_results(doc, results)
lintangsutawika's avatar
lintangsutawika committed
955

956
        result_dict = {}
957
        use_metric = list(self._metric_fn_list.keys())
958
959
960
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
961
962
963
964
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
965
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
966
            (loglikelihood,) = results
967
968
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
haileyschoelkopf's avatar
haileyschoelkopf committed
969
            return {
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
                **(
                    {"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
985
            }
986
        elif self.OUTPUT_TYPE == "multiple_choice":
987
            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
988

989
            # retrieve choices in List[str] form, to compute choice lengths, etc.
990
            choices = self.doc_to_choice(doc)
991
992
            completion_len = np.array([float(len(i)) for i in choices])

993
994
            if (
                2 * len(choices) == len(lls)
995
                and "acc_mutual_info" in self._metric_fn_list.keys()
996
997
998
999
1000
1001
1002
            ):
                # 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]
1003

1004
1005
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
1006

1007
1008
            if self.multiple_input:
                gold = self.doc_to_text(doc)
1009
            else:
1010
                gold = self.doc_to_target(doc)
1011
1012
1013

            gold_index_error = False
            if type(gold) is list:
Lintang Sutawika's avatar
Lintang Sutawika committed
1014
1015
                gold = [i if i < len(choices) else -100 for i in gold]
                if -100 in gold:
1016
1017
1018
                    gold_index_error = True
            else:
                if type(gold) is int:
Lintang Sutawika's avatar
Lintang Sutawika committed
1019
                    gold = gold if gold < len(choices) else -100
1020
                elif type(gold) is str:
Lintang Sutawika's avatar
Lintang Sutawika committed
1021
                    gold = choices.index(gold) if gold in choices else -100
lintangsutawika's avatar
lintangsutawika committed
1022

Lintang Sutawika's avatar
Lintang Sutawika committed
1023
                if gold == -100:
1024
1025
1026
1027
                    gold_index_error = True

            if gold_index_error:
                eval_logger.warning(
lintangsutawika's avatar
lintangsutawika committed
1028
                    f"Label index was not in within range of available choices,"
1029
1030
                    f"Sample:\n\n{doc}\n\n"
                )
lintangsutawika's avatar
lintangsutawika committed
1031

1032
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
1033
1034
                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
1035
                exact_match = int(any([is_greedy[i] if i != -100 else 0 for i in gold]))
lintangsutawika's avatar
lintangsutawika committed
1036
1037
1038
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
1039
                # 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
1040
                exact_match = int(is_greedy[gold]) if gold != -100 else 0
1041
1042

            result_dict = {
1043
                **({"acc": acc} if "acc" in use_metric else {}),
1044
1045
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
1046
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
1047
                **({"exact_match": exact_match} if "exact_match" in use_metric else {}),
1048
1049
            }

1050
            if "acc_mutual_info" in use_metric:
lintangsutawika's avatar
lintangsutawika committed
1051
1052
1053
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
1054
1055
1056
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

1057
        elif self.OUTPUT_TYPE == "greedy_until":
1058
            gold = self.doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
1059
            if self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1060
                # If you set doc_to_choice,
lintangsutawika's avatar
lintangsutawika committed
1061
                # it assumes that doc_to_target returns a number.
1062
1063
                choices = self.doc_to_choice(doc)
                gold = choices[gold]
lintangsutawika's avatar
lintangsutawika committed
1064
1065
            else:
                gold = str(gold)
1066

lintangsutawika's avatar
lintangsutawika committed
1067
            result = results[0]
lintangsutawika's avatar
lintangsutawika committed
1068
            for metric in self._metric_fn_list.keys():
haileyschoelkopf's avatar
haileyschoelkopf committed
1069
1070
1071
1072
1073
1074
                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:
1075
                        try:
1076
                            result_score = self._metric_fn_list[metric](
1077
1078
                                references=[gold_option],
                                predictions=[result],
1079
                                **self._metric_fn_kwargs[metric],
1080
1081
                            )
                        except TypeError:  # TODO: this is hacky and I don't want to do it
1082
                            result_score = self._metric_fn_list[metric](
haileyschoelkopf's avatar
haileyschoelkopf committed
1083
1084
1085
                                [gold_option, result]
                            )
                        if isinstance(result_score, dict):
haileyschoelkopf's avatar
haileyschoelkopf committed
1086
                            # TODO: this handles the case where HF evaluate returns a dict.
1087
                            result_score = result_score[metric]
haileyschoelkopf's avatar
haileyschoelkopf committed
1088
                        scores.append(result_score)
haileyschoelkopf's avatar
haileyschoelkopf committed
1089
                    if any(scores):
haileyschoelkopf's avatar
haileyschoelkopf committed
1090
                        result_score = 1.0
haileyschoelkopf's avatar
haileyschoelkopf committed
1091
                    else:
haileyschoelkopf's avatar
haileyschoelkopf committed
1092
                        result_score = 0.0
haileyschoelkopf's avatar
haileyschoelkopf committed
1093
                else:
1094
                    try:
1095
                        result_score = self._metric_fn_list[metric](
1096
1097
                            references=[gold],
                            predictions=[result],
1098
                            **self._metric_fn_kwargs[metric],
1099
                        )
1100
1101
                    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])
1102
1103
1104
1105
                    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
1106
        else:
lintangsutawika's avatar
lintangsutawika committed
1107
1108
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
1109
                "'loglikelihood', 'loglikelihood_rolling', 'greedy_until' or 'multiple_choice'",
1110
            )
1111
1112
1113
1114
1115
1116
1117

        return result_dict

    def aggregation(self):
        return self._aggregation_list

    def higher_is_better(self):
haileyschoelkopf's avatar
haileyschoelkopf committed
1118
        return self._higher_is_better
1119
1120
1121
1122
1123


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

baberabb's avatar
baberabb committed
1124
    def doc_to_target(self, doc: dict) -> str:
1125
1126
        return " " + doc["choices"][doc["gold"]]

baberabb's avatar
baberabb committed
1127
    def construct_requests(self, doc: dict, ctx: str, **kwargs) -> List[Instance]:
1128
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1129
1130
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1131
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1132
                doc=doc,
1133
                arguments=(ctx, " {}".format(choice)),
1134
                idx=i,
1135
1136
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1137
1138
            for i, choice in enumerate(doc["choices"])
        ]
1139

baberabb's avatar
baberabb committed
1140
    def process_results(self, doc: dict, results: List[Tuple[float, bool]]) -> dict:
lintangsutawika's avatar
lintangsutawika committed
1141
1142
1143
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
        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
1155
    def higher_is_better(self) -> dict:
1156
1157
1158
1159
1160
        return {
            "acc": True,
            "acc_norm": True,
        }

baberabb's avatar
baberabb committed
1161
    def aggregation(self) -> dict:
1162
1163
1164
1165
1166
1167
        return {
            "acc": mean,
            "acc_norm": mean,
        }


lintangsutawika's avatar
lintangsutawika committed
1168
class PerplexityTask(Task):
1169
1170
    OUTPUT_TYPE = "loglikelihood_rolling"

baberabb's avatar
baberabb committed
1171
    def has_training_docs(self) -> bool:
1172
1173
        return False

baberabb's avatar
baberabb committed
1174
    def fewshot_examples(self, k: int, rnd) -> List:
1175
1176
1177
        assert k == 0
        return []

baberabb's avatar
baberabb committed
1178
    def fewshot_context(self, doc: dict, num_fewshot: int) -> Literal[""]:
1179
1180
1181
1182
1183
1184
        assert (
            num_fewshot == 0
        ), "The number of fewshot examples must be 0 for perplexity tasks."

        return ""

baberabb's avatar
baberabb committed
1185
    def higher_is_better(self) -> dict:
1186
1187
1188
1189
1190
1191
1192
1193
1194
        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
1195
    def doc_to_text(self, doc) -> str:
1196
1197
1198
1199
1200
        return ""

    def doc_to_target(self, doc):
        return doc

baberabb's avatar
baberabb committed
1201
    def construct_requests(self, doc: dict, ctx: Union[str, None], **kwargs):
1202
1203
        assert not ctx

lintangsutawika's avatar
lintangsutawika committed
1204
1205
1206
1207
1208
1209
1210
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1211

baberabb's avatar
baberabb committed
1212
    def process_results(self, doc: dict, results: float) -> dict:
1213
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1214
1215
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1216
1217
1218
1219
1220
1221
        return {
            "word_perplexity": (loglikelihood, words),
            "byte_perplexity": (loglikelihood, bytes_),
            "bits_per_byte": (loglikelihood, bytes_),
        }

baberabb's avatar
baberabb committed
1222
    def aggregation(self) -> dict:
1223
1224
1225
1226
1227
1228
1229
        return {
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
            "bits_per_byte": bits_per_byte,
        }

    @classmethod
baberabb's avatar
baberabb committed
1230
    def count_bytes(cls, doc) -> int:
1231
1232
1233
        return len(doc.encode("utf-8"))

    @classmethod
baberabb's avatar
baberabb committed
1234
    def count_words(cls, doc) -> int:
1235
1236
        """Downstream tasks with custom word boundaries should override this!"""
        return len(re.split(r"\s+", doc))