task.py 46.5 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
lintangsutawika's avatar
lintangsutawika committed
7
import logging
8
9
10
import evaluate
import random
import itertools
11
import functools
12
from tqdm import tqdm
13
14
15
16

import datasets
import numpy as np

baberabb's avatar
baberabb committed
17
from typing import Union, List, Any, Tuple, Literal
18
from collections.abc import Callable
19

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

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
    get_metric,
    get_aggregation,
36
    get_metric_aggregation,
haileyschoelkopf's avatar
haileyschoelkopf committed
37
    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
ALL_OUTPUT_TYPES = [
    "loglikelihood",
    "multiple_choice",
    "loglikelihood_rolling",
47
    "generate_until",
48
49
]

lintangsutawika's avatar
lintangsutawika committed
50

51
eval_logger = logging.getLogger("lm-eval")
52

lintangsutawika's avatar
lintangsutawika committed
53

54
55
@dataclass
class TaskConfig(dict):
56
    # task naming/registry
57
    task: str = None
lintangsutawika's avatar
lintangsutawika committed
58
    task_alias: str = None
59
    group: Union[str, list] = None
lintangsutawika's avatar
lintangsutawika committed
60
    group_alias: Union[str, list] = None
61
62
63
    # HF dataset options.
    # which dataset to use,
    # and what splits for what purpose
64
65
    dataset_path: str = None
    dataset_name: str = None
66
    dataset_kwargs: dict = None
67
68
69
    training_split: str = None
    validation_split: str = None
    test_split: str = None
lintangsutawika's avatar
lintangsutawika committed
70
    fewshot_split: str = None  # TODO: assert that this not None if num_fewshot > 0. (?) assert if this is same split as one evaling (?)
71
72
    # formatting / prompting options.
    # see docs/advanced_task_guide.md for more info
73
    process_docs: Callable = None
74
75
    doc_to_text: Union[Callable, str] = None
    doc_to_target: Union[Callable, str] = None
lintangsutawika's avatar
lintangsutawika committed
76
    doc_to_choice: Union[Callable, str, dict, list] = None
lintangsutawika's avatar
lintangsutawika committed
77
    process_results: Union[Callable, str] = None
78
    use_prompt: str = None
79
    description: str = ""
80
81
    target_delimiter: str = " "
    fewshot_delimiter: str = "\n\n"
haileyschoelkopf's avatar
haileyschoelkopf committed
82
    fewshot_config: dict = None
83
    # runtime configuration options
84
    num_fewshot: int = 0
85
    # scoring options
86
    metric_list: list = None
87
    output_type: str = "generate_until"
88
    generation_kwargs: dict = None
89
    repeats: int = 1
lintangsutawika's avatar
lintangsutawika committed
90
    filter_list: Union[str, list] = None
91
92
    should_decontaminate: bool = False
    doc_to_decontamination_query: str = None
93

lintangsutawika's avatar
lintangsutawika committed
94
    metadata: str = None  # by default, not used in the code. allows for users to pass arbitrary info to tasks
95

Ethan Smith's avatar
Ethan Smith committed
96
    def __post_init__(self) -> None:
lintangsutawika's avatar
lintangsutawika committed
97
98
99
        if "." in self.dataset_path:
            import inspect
            from importlib import import_module
lintangsutawika's avatar
format  
lintangsutawika committed
100

lintangsutawika's avatar
lintangsutawika committed
101
            self.dataset_path = inspect.getfile(import_module(self.dataset_path))
102

Lintang Sutawika's avatar
Lintang Sutawika committed
103
        if self.generation_kwargs is not None:
104
            if self.output_type != "generate_until":
Lintang Sutawika's avatar
Lintang Sutawika committed
105
                eval_logger.warning(
106
                    f"[{self.task}] passed `generation_kwargs`, but not using `output_type: generate_until`!"
Lintang Sutawika's avatar
Lintang Sutawika committed
107
                )
108
                assert self.output_type != "generate_until"
Lintang Sutawika's avatar
Lintang Sutawika committed
109
110
111
112
113
114
115

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

            if "until" not in self.generation_kwargs:
116
                self.generation_kwargs["until"] = [self.fewshot_delimiter]
Lintang Sutawika's avatar
Lintang Sutawika committed
117
        else:
118
            if self.output_type == "generate_until":
Lintang Sutawika's avatar
Lintang Sutawika committed
119
120
                # ensure that we greedily generate in absence of explicit arguments otherwise
                self.generation_kwargs = {
Lintang Sutawika's avatar
Lintang Sutawika committed
121
                    "until": None
122
123
                    if self.fewshot_delimiter is None
                    else [self.fewshot_delimiter],
Lintang Sutawika's avatar
Lintang Sutawika committed
124
125
                    "do_sample": False,
                }
126

haileyschoelkopf's avatar
haileyschoelkopf committed
127
128
        # TODO: how to make TaskConfigs be de- and re-serializable, even when using the !function constructor?

129
130
131
    def __getitem__(self, item):
        return getattr(self, item)

132
133
134
    def __setitem__(self, item, value):
        return setattr(self, item, value)

135
    def to_dict(self):
136
137
        """dumps the current config as a dictionary object, as a printable format.
        null fields will not be printed.
haileyschoelkopf's avatar
haileyschoelkopf committed
138
        Used for dumping results alongside full task configuration
139

haileyschoelkopf's avatar
haileyschoelkopf committed
140
141
142
143
144
145
146
147
148
149
        :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
150
151
152
            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
153
        return cfg_dict
154

155
156
157
158
159
160
161
162
163
164
165
166

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
167

168
169
170
171
172
173
174
175
    # 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
176

177
178
179
180
181
182
    def __init__(
        self,
        data_dir=None,
        cache_dir=None,
        download_mode=None,
        config=None,
Ethan Smith's avatar
Ethan Smith committed
183
    ) -> None:
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
        """
        :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
210
        self._config = TaskConfig(**config) if config else TaskConfig()
211
212
213

        if not hasattr(self, "_filters"):
            self._filters = []
lintangsutawika's avatar
lintangsutawika committed
214
            for name, components in self._config.get(
215
                "filters", [["none", [["take_first", None]]]]
lintangsutawika's avatar
lintangsutawika committed
216
            ):
217
218
219
                filter_pipeline = build_filter_ensemble(name, components)
                self._filters.append(filter_pipeline)

lintangsutawika's avatar
lintangsutawika committed
220
        self.sampler = samplers.Sampler(
221
222
            list(self.fewshot_docs()), self, rnd=random.Random(1234)
        )
223

Ethan Smith's avatar
Ethan Smith committed
224
    def download(self, data_dir=None, cache_dir=None, download_mode=None) -> None:
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
        """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.
        """
249
250
251
252
253
254
255
        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,
        )
256

257
258
259
260
261
    @property
    def config(self):
        """Returns the TaskConfig associated with this class."""
        return self._config

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
292
293
294
295
296
297
    @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 []

298
299
300
301
302
303
304
305
306
307
    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
308
            eval_logger.warning(
309
                "has_training_docs and has_validation_docs are False"
310
                ", using test_docs as fewshot_docs but this is not recommended."
lintangsutawika's avatar
lintangsutawika committed
311
            )
312
313
            return self.test_docs()

314
315
316
317
318
319
320
321
322
323
    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
324

325
326
327
328
329
330
331
332
333
334
335
336
337
    @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
338
    def doc_to_decontamination_query(self, doc) -> None:
339
340
341
342
343
344
345
346
347
348
349
350
351
        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
352
    def build_all_requests(self, limit=None, rank=None, world_size=None) -> None:
353
354
355
356
357
358
359
360
361
362
        """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!"

363
        eval_logger.info(
364
            f"Building contexts for task '{self.config.task}' on rank {rank}..."
365
366
        )

367
        instances = []
368
369
        for doc_id, doc in utils.create_iterator(
            enumerate(docs), rank, world_size, limit
lintangsutawika's avatar
lintangsutawika committed
370
        ):
371
            # sample fewshot context #TODO: need to offset doc_id by rank now!
372
            fewshot_ctx = self.fewshot_context(
373
                doc,
374
                self.config.num_fewshot,
375
            )
376

377
            # TODO: we should override self.config.repeats if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
378
379
380
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
381
                metadata=(self.config["task"], doc_id, self.config.repeats),
lintangsutawika's avatar
lintangsutawika committed
382
            )
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407

            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
408
            The number of times each instance in a dataset is inferred on. Defaults to 1,
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
438
439
440
441
442
443
            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
444
445
446
447
448
449
450
451
452
453
    @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))

454
    @utils.positional_deprecated
455
    def fewshot_context(self, doc, num_fewshot):
456
457
458
459
460
461
462
463
464
465
466
467
        """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:
468
            # always prepend the (possibly empty) task description
469
            labeled_examples = self.config.description
470
        else:
471
            labeled_examples = self.config.description + self.sampler.get_context(
lintangsutawika's avatar
lintangsutawika committed
472
473
                doc, num_fewshot
            )
474
475

        example = self.doc_to_text(doc)
476
477
478
479
        if type(example) == str:
            return labeled_examples + example
        elif type(example) == list:
            return [labeled_examples + ex for ex in example]
480
        elif type(example) == int:
481
            if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
482
483
484
485
                choices = self.doc_to_choice(doc)
                return labeled_examples + choices[example]
            else:
                return labeled_examples + str(example)
486
487

    def apply_filters(self):
lintangsutawika's avatar
lintangsutawika committed
488
489
490
491
492
493
        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
494

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

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

505
506

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

651
        if self.has_test_docs():
652
            self.task_docs = self.test_docs()
653
        elif self.has_validation_docs():
654
            self.task_docs = self.validation_docs()
655
656
657
658
659
        else:
            assert (
                False
            ), f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"

660
        # Test One Doc
661
        self.features = list(self.task_docs.features.keys())
662
663
        self.multiple_input = 0
        self.multiple_target = 0
664
        test_doc = self.task_docs[0]
665
        test_text = self.doc_to_text(test_doc)
666
        test_target = self.doc_to_target(test_doc)
lintangsutawika's avatar
lintangsutawika committed
667

668
        if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
669
670
671
            test_choice = self.doc_to_choice(test_doc)
            if type(test_choice) is not list:
                eval_logger.error("doc_to_choice must return list")
672
673
            else:
                num_choice = len(test_choice)
674

675
676
            if type(test_text) is int:
                self.multiple_input = num_choice
677
678
        else:
            test_choice = None
679

680
        if type(test_target) is list:
681
            self.multiple_target = len(test_target)
682
        else:
lintangsutawika's avatar
lintangsutawika committed
683
            if (type(test_target) is int) and (test_choice is not None):
lintangsutawika's avatar
lintangsutawika committed
684
                test_target = test_choice[test_target]
685
            else:
lintangsutawika's avatar
lintangsutawika committed
686
                test_target = str(test_target)
687

688
689
690
        if test_choice is not None:
            check_choices = test_choice
        else:
lintangsutawika's avatar
lintangsutawika committed
691
            check_choices = [test_target]
692
693
694
695
        if self.config.doc_to_choice is not None:
            for choice in check_choices:
                choice_has_whitespace = True if choice[0].isspace() else False
                delimiter_has_whitespace = (
696
697
                    True
                    if self.config.target_delimiter.rstrip()
698
                    != self.config.target_delimiter
699
                    else False
700
                )
701

702
703
704
705
706
707
                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(
708
                        f'Both target_delimiter "{self.config.target_delimiter}" and target choice: "{choice}" do not have whitespace, ignore if the language you are evaluating on does not require/use whitespace'
709
710
                    )

Ethan Smith's avatar
Ethan Smith committed
711
    def download(self, dataset_kwargs=None) -> None:
712
713
714
715
716
717
        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
718
    def has_training_docs(self) -> bool:
719
        if self.config.training_split is not None:
720
721
722
723
            return True
        else:
            return False

baberabb's avatar
baberabb committed
724
    def has_validation_docs(self) -> bool:
725
        if self.config.validation_split is not None:
726
727
728
729
            return True
        else:
            return False

baberabb's avatar
baberabb committed
730
    def has_test_docs(self) -> bool:
731
        if self.config.test_split is not None:
732
733
734
735
            return True
        else:
            return False

baberabb's avatar
baberabb committed
736
    def training_docs(self) -> datasets.Dataset:
737
        if self.has_training_docs():
738
739
740
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.training_split]
741
                )
742
            return self.dataset[self.config.training_split]
743

baberabb's avatar
baberabb committed
744
    def validation_docs(self) -> datasets.Dataset:
745
        if self.has_validation_docs():
746
747
748
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.validation_split]
749
                )
750
            return self.dataset[self.config.validation_split]
751

baberabb's avatar
baberabb committed
752
    def test_docs(self) -> datasets.Dataset:
753
        if self.has_test_docs():
754
755
756
            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]
757

758
    def fewshot_docs(self):
759
760
        if self.config.fewshot_split is not None:
            return self.dataset[self.config.fewshot_split]
761
        else:
762
            if self.config.num_fewshot > 0:
763
                eval_logger.warning(
764
                    f"Task '{self.config.task}': "
765
766
767
768
                    "num_fewshot > 0 but fewshot_split is None. "
                    "using preconfigured rule."
                )
            return super().fewshot_docs()
769

770
771
772
773
774
775
776
777
    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

778
    def should_decontaminate(self):
779
        return self.config.should_decontaminate
780
781

    def doc_to_decontamination_query(self, doc):
782
783
784
        if self.config.should_decontaminate:
            if self.config.doc_to_decontamination_query in self.features:
                return doc[self.config.doc_to_decontamination_query]
785
786
            else:
                return ast.literal_eval(
787
                    utils.apply_template(self.config.doc_to_decontamination_query, doc)
788
                )
789

790
791
792
793
794
795
796
797
798
799
800
801
    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):
802
803
        if self.prompt is not None:
            doc_to_text = self.prompt
804
        else:
805
            doc_to_text = self.config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
806

807
808
809
        if type(doc_to_text) == int:
            return doc_to_text
        elif type(doc_to_text) == str:
810
            if doc_to_text in self.features:
811
                # if self.config.doc_to_choice is not None:
812
813
                #     return self.doc_to_choice(doc)[doc[doc_to_text]]
                # else:
814
815
                return doc[doc_to_text]
            else:
lintangsutawika's avatar
lintangsutawika committed
816
                text_string = utils.apply_template(doc_to_text, doc)
lintangsutawika's avatar
lintangsutawika committed
817
                if text_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
818
819
820
                    return ast.literal_eval(text_string)
                else:
                    return text_string
821
        elif callable(doc_to_text):
822
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
823
        # Used when applying a Promptsource template
824
        elif hasattr(doc_to_text, "apply"):
825
826
827
828
829
            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")
830
                return self.config.fewshot_delimiter
831
        else:
832
            print(type(doc_to_text))
833
            raise TypeError
834

835
    def doc_to_target(self, doc: dict) -> Union[int, str, list]:
836
837
        if self.prompt is not None:
            doc_to_target = self.prompt
838
        else:
839
            doc_to_target = self.config.doc_to_target
840

841
842
843
        if type(doc_to_target) == int:
            return doc_to_target
        elif type(doc_to_target) == str:
844
            if doc_to_target in self.features:
845
                # if self.config.doc_to_choice is not None:
846
847
848
                #     return self.doc_to_choice(doc)[doc[doc_to_target]]
                # else:
                return doc[doc_to_target]
849
            else:
lintangsutawika's avatar
lintangsutawika committed
850
                target_string = utils.apply_template(doc_to_target, doc)
lintangsutawika's avatar
lintangsutawika committed
851
                if target_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
852
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
853
854
855
856
857
                elif (
                    len(target_string) >= 2
                    and (target_string[0] == "[")
                    and (target_string[-1] == "]")
                ):
baberabb's avatar
baberabb committed
858
859
860
861
                    try:
                        return ast.literal_eval(target_string)
                    except (SyntaxError, ValueError):
                        return target_string
lintangsutawika's avatar
lintangsutawika committed
862
863
                else:
                    return target_string
864
865
        elif type(doc_to_target) == list:
            return doc_to_target
866
        elif callable(doc_to_target):
867
            return doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
868
        # Used when applying a Promptsource template
869
        elif hasattr(doc_to_target, "apply"):
870
            applied_prompt = doc_to_target.apply(doc)
871
872
873
874
            if len(applied_prompt) == 2:
                return applied_prompt[1]
            else:
                eval_logger.warning("Applied prompt returns empty string")
875
                return self.config.fewshot_delimiter
876
877
        else:
            raise TypeError
878

baberabb's avatar
baberabb committed
879
    def doc_to_choice(self, doc: Any) -> List[str]:
880
881
        if self.prompt is not None:
            doc_to_choice = self.prompt
882
        elif self.config.doc_to_choice is None:
883
884
            eval_logger.error("doc_to_choice was called but not set in config")
        else:
885
            doc_to_choice = self.config.doc_to_choice
886
887
888
889
890
891
892
893
894
895
896
897
898

        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
899

baberabb's avatar
baberabb committed
900
901
902
    def construct_requests(
        self, doc: dict, ctx: str, **kwargs
    ) -> Union[List[Instance], Instance]:
903
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
904
            arguments = (ctx, self.doc_to_target(doc))
905
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
906
            arguments = (self.doc_to_target(doc),)
907
        elif self.OUTPUT_TYPE == "multiple_choice":
908
            choices = self.doc_to_choice(doc)
909
            target_delimiter = self.config.target_delimiter
910
911
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
912
                cont = self.doc_to_target(doc)
913
                arguments = [(ctx, f"{target_delimiter}{cont}") for ctx in choices]
914
            else:
915
                # Otherwise they are placed in the continuation
916
                arguments = [(ctx, f"{target_delimiter}{cont}") for cont in choices]
917

918
            request_list = [
919
920
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
921
                    doc=doc,
922
                    arguments=arg,
923
                    idx=i,
924
925
                    **kwargs,
                )
926
                for i, arg in enumerate(arguments)
927
            ]
928
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
929
            if "acc_mutual_info" in self._metric_fn_list.keys():
930
931
932
                # if we are calculating multiple choice accuracy
                # using mutual information instead of raw loglikelihood as metric, need unconditional lls.

lintangsutawika's avatar
lintangsutawika committed
933
                # here mutual info refers to calculating
934
935
936
937
938
939
                # 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
940
                            doc=doc,
941
                            arguments=("", "{}".format(choice)),
942
943
944
                            idx=i,
                            **kwargs,
                        )
lintangsutawika's avatar
lintangsutawika committed
945
                        for i, choice in enumerate(choices)
946
947
948
                    ]
                )
            return request_list
lintangsutawika's avatar
lintangsutawika committed
949

950
        elif self.OUTPUT_TYPE == "generate_until":
951
            arguments = (ctx, self.config.generation_kwargs)
lintangsutawika's avatar
lintangsutawika committed
952
953

        return Instance(
lintangsutawika's avatar
lintangsutawika committed
954
955
            request_type=self.OUTPUT_TYPE, doc=doc, arguments=arguments, idx=0, **kwargs
        )
956
957

    def process_results(self, doc, results):
958
959
        if callable(self.config.process_results):
            return self.config.process_results(doc, results)
lintangsutawika's avatar
lintangsutawika committed
960

961
        result_dict = {}
962
        use_metric = list(self._metric_fn_list.keys())
963
964
965
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
966
967
968
969
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
970
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
971
            (loglikelihood,) = results
972
973
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
haileyschoelkopf's avatar
haileyschoelkopf committed
974
            return {
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
                **(
                    {"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
990
            }
991
        elif self.OUTPUT_TYPE == "multiple_choice":
992
            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
993

994
            # retrieve choices in List[str] form, to compute choice lengths, etc.
995
            choices = self.doc_to_choice(doc)
996
997
            completion_len = np.array([float(len(i)) for i in choices])

998
999
            if (
                2 * len(choices) == len(lls)
1000
                and "acc_mutual_info" in self._metric_fn_list.keys()
1001
1002
1003
1004
1005
1006
1007
            ):
                # 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]
1008

1009
1010
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
1011

1012
1013
            if self.multiple_input:
                gold = self.doc_to_text(doc)
1014
            else:
1015
                gold = self.doc_to_target(doc)
1016
1017
1018

            gold_index_error = False
            if type(gold) is list:
Lintang Sutawika's avatar
Lintang Sutawika committed
1019
1020
                gold = [i if i < len(choices) else -100 for i in gold]
                if -100 in gold:
1021
1022
1023
                    gold_index_error = True
            else:
                if type(gold) is int:
Lintang Sutawika's avatar
Lintang Sutawika committed
1024
                    gold = gold if gold < len(choices) else -100
1025
                elif type(gold) is str:
Lintang Sutawika's avatar
Lintang Sutawika committed
1026
                    gold = choices.index(gold) if gold in choices else -100
lintangsutawika's avatar
lintangsutawika committed
1027

Lintang Sutawika's avatar
Lintang Sutawika committed
1028
                if gold == -100:
1029
1030
1031
1032
                    gold_index_error = True

            if gold_index_error:
                eval_logger.warning(
lintangsutawika's avatar
lintangsutawika committed
1033
                    f"Label index was not in within range of available choices,"
1034
1035
                    f"Sample:\n\n{doc}\n\n"
                )
lintangsutawika's avatar
lintangsutawika committed
1036

1037
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
1038
1039
                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
1040
                exact_match = int(any([is_greedy[i] if i != -100 else 0 for i in gold]))
lintangsutawika's avatar
lintangsutawika committed
1041
1042
1043
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
1044
                # 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
1045
                exact_match = int(is_greedy[gold]) if gold != -100 else 0
1046
1047

            result_dict = {
1048
                **({"acc": acc} if "acc" in use_metric else {}),
1049
1050
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
1051
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
1052
                **({"exact_match": exact_match} if "exact_match" in use_metric else {}),
1053
1054
            }

1055
            if "acc_mutual_info" in use_metric:
lintangsutawika's avatar
lintangsutawika committed
1056
1057
1058
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
1059
1060
1061
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

1062
        elif self.OUTPUT_TYPE == "generate_until":
1063
            gold = self.doc_to_target(doc)
Chris's avatar
Chris committed
1064
            result = results[0]
1065
            if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1066
                # If you set doc_to_choice,
lintangsutawika's avatar
lintangsutawika committed
1067
                # it assumes that doc_to_target returns a number.
1068
1069
                choices = self.doc_to_choice(doc)
                gold = choices[gold]
1070
1071
            # we expect multiple_targets to be a list.
            elif self.multiple_target:
baberabb's avatar
baberabb committed
1072
                gold = list(gold)
Chris's avatar
Chris committed
1073
1074
1075
            elif type(gold) != type(result):
                # cast gold to the same type as result
                gold = type(result)(gold)
1076

lintangsutawika's avatar
lintangsutawika committed
1077
            for metric in self._metric_fn_list.keys():
haileyschoelkopf's avatar
haileyschoelkopf committed
1078
1079
1080
1081
1082
                if self.multiple_target:
                    # in the case where we have multiple targets,
                    # return true if any are true
                    # TODO: this may break for multipLe_target, non zero-or-1 metrics
                    scores = []
haileyschoelkopf's avatar
haileyschoelkopf committed
1083
1084
1085
1086
                    if not isinstance(gold, list):
                        # sometimes, a multiple_target dataset has exceptions where one doc has only one string answer
                        # print(gold)
                        gold = [gold]
haileyschoelkopf's avatar
haileyschoelkopf committed
1087
                    for gold_option in gold:
1088
                        try:
1089
                            result_score = self._metric_fn_list[metric](
1090
1091
                                references=[gold_option],
                                predictions=[result],
1092
                                **self._metric_fn_kwargs[metric],
1093
                            )
baberabb's avatar
baberabb committed
1094
1095
1096
                        except (
                            TypeError
                        ):  # TODO: this is hacky and I don't want to do it
1097
                            result_score = self._metric_fn_list[metric](
haileyschoelkopf's avatar
haileyschoelkopf committed
1098
1099
1100
                                [gold_option, result]
                            )
                        if isinstance(result_score, dict):
haileyschoelkopf's avatar
haileyschoelkopf committed
1101
                            # TODO: this handles the case where HF evaluate returns a dict.
1102
                            result_score = result_score[metric]
haileyschoelkopf's avatar
haileyschoelkopf committed
1103
                        scores.append(result_score)
haileyschoelkopf's avatar
haileyschoelkopf committed
1104
                    if any(scores):
1105
                        result_score = 1.0
haileyschoelkopf's avatar
haileyschoelkopf committed
1106
                    else:
1107
                        result_score = 0.0
haileyschoelkopf's avatar
haileyschoelkopf committed
1108
                else:
1109
                    try:
1110
                        result_score = self._metric_fn_list[metric](
1111
1112
                            references=[gold],
                            predictions=[result],
1113
                            **self._metric_fn_kwargs[metric],
1114
                        )
baberabb's avatar
baberabb committed
1115
1116
1117
                    except (
                        TypeError
                    ):  # needed for now in order to use a different interface between our own metrics and HF Evaluate metrics
1118
                        result_score = self._metric_fn_list[metric]([gold, result])
1119
1120
1121
1122
                    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
1123
        else:
lintangsutawika's avatar
lintangsutawika committed
1124
1125
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
1126
                "'loglikelihood', 'loglikelihood_rolling', 'generate_until' or 'multiple_choice'",
1127
            )
1128
1129
1130
1131
1132
1133
1134

        return result_dict

    def aggregation(self):
        return self._aggregation_list

    def higher_is_better(self):
haileyschoelkopf's avatar
haileyschoelkopf committed
1135
        return self._higher_is_better
1136
1137
1138
1139
1140


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

baberabb's avatar
baberabb committed
1141
    def doc_to_target(self, doc: dict) -> str:
1142
1143
        return " " + doc["choices"][doc["gold"]]

baberabb's avatar
baberabb committed
1144
    def construct_requests(self, doc: dict, ctx: str, **kwargs) -> List[Instance]:
1145
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1146
1147
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1148
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1149
                doc=doc,
1150
                arguments=(ctx, " {}".format(choice)),
1151
                idx=i,
1152
1153
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1154
1155
            for i, choice in enumerate(doc["choices"])
        ]
1156

baberabb's avatar
baberabb committed
1157
    def process_results(self, doc: dict, results: List[Tuple[float, bool]]) -> dict:
lintangsutawika's avatar
lintangsutawika committed
1158
1159
1160
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
        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
1172
    def higher_is_better(self) -> dict:
1173
1174
1175
1176
1177
        return {
            "acc": True,
            "acc_norm": True,
        }

baberabb's avatar
baberabb committed
1178
    def aggregation(self) -> dict:
1179
1180
1181
1182
1183
1184
        return {
            "acc": mean,
            "acc_norm": mean,
        }


lintangsutawika's avatar
lintangsutawika committed
1185
class PerplexityTask(Task):
1186
1187
    OUTPUT_TYPE = "loglikelihood_rolling"

baberabb's avatar
baberabb committed
1188
    def has_training_docs(self) -> bool:
1189
1190
        return False

baberabb's avatar
baberabb committed
1191
    def fewshot_examples(self, k: int, rnd) -> List:
1192
1193
1194
        assert k == 0
        return []

baberabb's avatar
baberabb committed
1195
    def fewshot_context(self, doc: dict, num_fewshot: int) -> Literal[""]:
1196
1197
1198
1199
1200
1201
        assert (
            num_fewshot == 0
        ), "The number of fewshot examples must be 0 for perplexity tasks."

        return ""

baberabb's avatar
baberabb committed
1202
    def higher_is_better(self) -> dict:
1203
1204
1205
1206
1207
1208
1209
1210
1211
        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
1212
    def doc_to_text(self, doc) -> str:
1213
1214
1215
1216
1217
        return ""

    def doc_to_target(self, doc):
        return doc

baberabb's avatar
baberabb committed
1218
    def construct_requests(self, doc: dict, ctx: Union[str, None], **kwargs):
1219
1220
        assert not ctx

lintangsutawika's avatar
lintangsutawika committed
1221
1222
1223
1224
1225
1226
1227
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1228

baberabb's avatar
baberabb committed
1229
    def process_results(self, doc: dict, results: float) -> dict:
1230
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1231
1232
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1233
1234
1235
1236
1237
1238
        return {
            "word_perplexity": (loglikelihood, words),
            "byte_perplexity": (loglikelihood, bytes_),
            "bits_per_byte": (loglikelihood, bytes_),
        }

baberabb's avatar
baberabb committed
1239
    def aggregation(self) -> dict:
1240
1241
1242
1243
1244
1245
1246
        return {
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
            "bits_per_byte": bits_per_byte,
        }

    @classmethod
baberabb's avatar
baberabb committed
1247
    def count_bytes(cls, doc) -> int:
1248
1249
1250
        return len(doc.encode("utf-8"))

    @classmethod
baberabb's avatar
baberabb committed
1251
    def count_words(cls, doc) -> int:
1252
1253
        """Downstream tasks with custom word boundaries should override this!"""
        return len(re.split(r"\s+", doc))