task.py 46.8 KB
Newer Older
1
import abc
2
import ast
3
import functools
Herbie Bradley's avatar
Herbie Bradley committed
4
5
6
7
8
9
import itertools
import random
import re
from collections.abc import Callable
from dataclasses import asdict, dataclass, field
from typing import Any, List, Literal, Tuple, Union
10
11

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

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

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

48
49
50

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

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

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

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

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

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

            if "until" not in self.generation_kwargs:
109
                self.generation_kwargs["until"] = [self.fewshot_delimiter]
Lintang Sutawika's avatar
Lintang Sutawika committed
110
111
112
113
        else:
            if self.output_type == "greedy_until":
                # ensure that we greedily generate in absence of explicit arguments otherwise
                self.generation_kwargs = {
Lintang Sutawika's avatar
Lintang Sutawika committed
114
                    "until": None
115
116
                    if self.fewshot_delimiter is None
                    else [self.fewshot_delimiter],
Lintang Sutawika's avatar
Lintang Sutawika committed
117
118
119
                    "do_sample": False,
                    "temperature": 0.0,
                }
120

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

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

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

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

haileyschoelkopf's avatar
haileyschoelkopf committed
134
135
136
137
138
139
140
141
142
143
        :return: dict
            A printable dictionary version of the TaskConfig object.

        # TODO: should any default value in the TaskConfig not be printed?
        """
        cfg_dict = asdict(self)
        # remove values that are `None`
        for k, v in list(cfg_dict.items()):
            if v is None:
                cfg_dict.pop(k)
haileyschoelkopf's avatar
haileyschoelkopf committed
144
145
146
            elif isinstance(v, Callable):
                # TODO: this should handle Promptsource template objects as a separate case?
                cfg_dict[k] = str(v)
haileyschoelkopf's avatar
haileyschoelkopf committed
147
        return cfg_dict
148

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

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

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

    VERSION = None
161

162
163
164
165
166
167
168
169
    # The name of the `Task` benchmark as denoted in the HuggingFace datasets Hub
    # or a path to a custom `datasets` loading script.
    DATASET_PATH: str = None

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

    OUTPUT_TYPE: str = None
lintangsutawika's avatar
lintangsutawika committed
170

171
172
173
174
175
176
    def __init__(
        self,
        data_dir=None,
        cache_dir=None,
        download_mode=None,
        config=None,
Ethan Smith's avatar
Ethan Smith committed
177
    ) -> None:
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
        """
        :param data_dir: str
            Stores the path to a local folder containing the `Task`'s data files.
            Use this to specify the path to manually downloaded data (usually when
            the dataset is not publicly accessible).
        :param cache_dir: str
            The directory to read/write the `Task` dataset. This follows the
            HuggingFace `datasets` API with the default cache directory located at:
                `~/.cache/huggingface/datasets`
            NOTE: You can change the cache location globally for a given process
            to another directory:
                `export HF_DATASETS_CACHE="/path/to/another/directory"`
        :param download_mode: datasets.DownloadMode
            How to treat pre-existing `Task` downloads and data.
            - `datasets.DownloadMode.REUSE_DATASET_IF_EXISTS`
                Reuse download and reuse dataset.
            - `datasets.DownloadMode.REUSE_CACHE_IF_EXISTS`
                Reuse download with fresh dataset.
            - `datasets.DownloadMode.FORCE_REDOWNLOAD`
                Fresh download and fresh dataset.
        """
        self.download(data_dir, cache_dir, download_mode)
        self._training_docs = None
        self._fewshot_docs = None
        self._instances = None

haileyschoelkopf's avatar
haileyschoelkopf committed
204
        self._config = TaskConfig(**config) if config else TaskConfig()
205
206
207

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

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

Ethan Smith's avatar
Ethan Smith committed
218
    def download(self, data_dir=None, cache_dir=None, download_mode=None) -> None:
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
        """Downloads and returns the task dataset.
        Override this method to download the dataset from a custom API.

        :param data_dir: str
            Stores the path to a local folder containing the `Task`'s data files.
            Use this to specify the path to manually downloaded data (usually when
            the dataset is not publicly accessible).
        :param cache_dir: str
            The directory to read/write the `Task` dataset. This follows the
            HuggingFace `datasets` API with the default cache directory located at:
                `~/.cache/huggingface/datasets`
            NOTE: You can change the cache location globally for a given process
            by setting the shell environment variable, `HF_DATASETS_CACHE`,
            to another directory:
                `export HF_DATASETS_CACHE="/path/to/another/directory"`
        :param download_mode: datasets.DownloadMode
            How to treat pre-existing `Task` downloads and data.
            - `datasets.DownloadMode.REUSE_DATASET_IF_EXISTS`
                Reuse download and reuse dataset.
            - `datasets.DownloadMode.REUSE_CACHE_IF_EXISTS`
                Reuse download with fresh dataset.
            - `datasets.DownloadMode.FORCE_REDOWNLOAD`
                Fresh download and fresh dataset.
        """
243
244
245
246
247
248
249
        self.dataset = datasets.load_dataset(
            path=self.DATASET_PATH,
            name=self.DATASET_NAME,
            data_dir=data_dir,
            cache_dir=cache_dir,
            download_mode=download_mode,
        )
250

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

256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
    @abc.abstractmethod
    def has_training_docs(self):
        """Whether the task has a training set"""
        pass

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

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

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

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

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

292
293
294
295
296
297
298
299
300
301
    def fewshot_docs(self):
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        if self.has_training_docs():
            return self.training_docs()
        elif self.has_validation_docs():
            return self.validation_docs()
        else:
lintangsutawika's avatar
lintangsutawika committed
302
            eval_logger.warning(
303
                "has_training_docs and has_validation_docs are False"
304
                ", using test_docs as fewshot_docs but this is not recommended."
lintangsutawika's avatar
lintangsutawika committed
305
            )
306
307
            return self.test_docs()

308
309
310
311
312
313
314
315
316
317
    def _process_doc(self, doc):
        """
        Override this to process (detokenize, strip, replace, etc.) individual
        documents. This can be used in a map over documents of a data split.
        E.g. `map(self._process_doc, self.dataset["validation"])`

        :return: dict
            The processed version of the specified `doc`.
        """
        return doc
lintangsutawika's avatar
lintangsutawika committed
318

319
320
321
322
323
324
325
326
327
328
329
330
331
    @property
    def instances(self):
        """After calling `task.build_all_requests()`, tasks
        maintain a list of the dataset instances which will be evaluated.
        """
        return self._instances

    def fewshot_examples(self, k, rnd):
        if self._training_docs is None:
            self._training_docs = list(self.training_docs())

        return rnd.sample(self._training_docs, k)

Ethan Smith's avatar
Ethan Smith committed
332
    def doc_to_decontamination_query(self, doc) -> None:
333
334
335
336
337
338
339
340
341
342
343
344
345
        print(
            "Override doc_to_decontamination_query with document specific decontamination query."
        )
        assert False

    @abc.abstractmethod
    def doc_to_text(self, doc):
        pass

    @abc.abstractmethod
    def doc_to_target(self, doc):
        pass

Ethan Smith's avatar
Ethan Smith committed
346
    def build_all_requests(self, limit=None, rank=None, world_size=None) -> None:
347
348
349
350
351
352
353
354
355
356
        """Build a set of Instances for a task, and store them in task.instances"""
        if self.has_test_docs():
            docs = self.test_docs()
        elif self.has_validation_docs():
            docs = self.validation_docs()
        else:
            assert (
                False
            ), f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"

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

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

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

            if not isinstance(inst, list):
                inst = [inst]

            instances.extend(inst)

        self._instances = instances
        assert len(self._instances) != 0, "task.build_requests() did not find any docs!"

    @abc.abstractmethod
    def construct_requests(self, doc, ctx, **kwargs):
        """Uses RequestFactory to construct Requests and returns an iterable of
        Requests which will be sent to the LM.

        :param doc:
            The document as returned from training_docs, validation_docs, or test_docs.
        :param ctx: str
            The context string, generated by fewshot_context. This includes the natural
            language description, as well as the few shot examples, and the question
            part of the document for `doc`.
        :param doc_idx: int
            The index of a document within `self.test_docs()` or `self.validation_docs()`,
            whichever is the main split used.
        :param repeats: int
        TODO: update this docstring
lintangsutawika's avatar
lintangsutawika committed
402
            The number of times each instance in a dataset is inferred on. Defaults to 1,
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
            can be increased for techniques like majority voting.
        """
        pass

    @abc.abstractmethod
    def process_results(self, doc, results):
        """Take a single document and the LM results and evaluates, returning a
        dict where keys are the names of submetrics and values are the values of
        the metric for that one document

        :param doc:
            The document as returned from training_docs, validation_docs, or test_docs.
        :param results:
            The results of the requests created in construct_requests.
        """
        pass

    @abc.abstractmethod
    def aggregation(self):
        """
        :returns: {str: [metric_score] -> float}
            A dictionary where keys are the names of submetrics and values are
            functions that aggregate a list of metric scores
        """
        pass

    @abc.abstractmethod
    def higher_is_better(self):
        """
        :returns: {str: bool}
            A dictionary where keys are the names of submetrics and values are
            whether a higher value of the submetric is better
        """
        pass

haileyschoelkopf's avatar
haileyschoelkopf committed
438
439
440
441
442
443
444
445
446
447
    @classmethod
    def count_bytes(cls, doc):
        """Used for byte-level perplexity metrics in rolling loglikelihood"""
        return len(doc.encode("utf-8"))

    @classmethod
    def count_words(cls, doc):
        """Downstream loglikelihood_rolling perplexity tasks with custom word boundaries should override this!"""
        return len(re.split(r"\s+", doc))

448
    @utils.positional_deprecated
449
    def fewshot_context(self, doc, num_fewshot):
450
451
452
453
454
455
456
457
458
459
460
461
        """Returns a fewshot context string that is made up of a prepended description
        (if provided), the `num_fewshot` number of examples, and an appended prompt example.

        :param doc: str
            The document as returned from training_docs, validation_docs, or test_docs.
        :param num_fewshot: int
            The number of fewshot examples to provide in the returned context string.
        :returns: str
            The fewshot context.
        """

        if num_fewshot == 0:
462
            # always prepend the (possibly empty) task description
463
            labeled_examples = self.config.description
464
        else:
465
            labeled_examples = self.config.description + self.sampler.get_context(
lintangsutawika's avatar
lintangsutawika committed
466
467
                doc, num_fewshot
            )
468
469

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

    def apply_filters(self):
lintangsutawika's avatar
lintangsutawika committed
482
483
484
485
486
487
        if hasattr(self, "_filters"):
            for f in self._filters:
                f.apply(self._instances)
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances
488

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

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

499
500

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

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

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

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

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

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

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

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

539
540
        _metric_list = DEFAULT_METRIC_REGISTRY[self.config.output_type]
        if self.config.metric_list is None:
541
            # TODO: handle this in TaskConfig.__post_init__ ?
542
            for metric_name in _metric_list:
haileyschoelkopf's avatar
haileyschoelkopf committed
543
544
                self._metric_fn_list[metric_name] = get_metric(metric_name)
                self._aggregation_list[metric_name] = get_default_aggregation(
545
                    metric_name
haileyschoelkopf's avatar
haileyschoelkopf committed
546
547
                )
                self._higher_is_better[metric_name] = is_higher_better(metric_name)
548
        else:
549
            for metric_config in self.config.metric_list:
550
551
552
553
554
555
556
                assert "metric" in metric_config
                metric_name = metric_config["metric"]
                kwargs = {
                    key: metric_config[key]
                    for key in metric_config
                    if key not in ["metric", "aggregation", "higher_is_better"]
                }
557

558
                if self.config.process_results is not None:
559
560
                    self._metric_fn_list[metric_name] = None
                    self._metric_fn_kwargs[metric_name] = {}
561
562
563
564
565
566
567
568
                elif callable(metric_name):
                    metric_fn = metric_name.__call__
                    metric_name = metric_name.__name__
                    self._metric_fn_list[metric_name] = metric_fn
                    self._metric_fn_kwargs[metric_name] = kwargs
                else:
                    self._metric_fn_list[metric_name] = get_metric(metric_name)
                    self._metric_fn_kwargs[metric_name] = kwargs
lintangsutawika's avatar
lintangsutawika committed
569

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

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

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

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

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

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

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

643
        # Test One Doc
Herbie Bradley's avatar
Herbie Bradley committed
644
645
        # self.features = ["text", "meta"]
        # return None
646
        self.features = list(self.task_docs.features.keys())
647
648
        self.multiple_input = 0
        self.multiple_target = 0
649
        test_doc = self.task_docs[0]
650
        test_text = self.doc_to_text(test_doc)
651
        test_target = self.doc_to_target(test_doc)
lintangsutawika's avatar
lintangsutawika committed
652

653
        if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
654
655
656
            test_choice = self.doc_to_choice(test_doc)
            if type(test_choice) is not list:
                eval_logger.error("doc_to_choice must return list")
657
658
            else:
                num_choice = len(test_choice)
659

660
661
            if type(test_text) is int:
                self.multiple_input = num_choice
662
663
        else:
            test_choice = None
664

665
        if type(test_target) is list:
666
            self.multiple_target = len(test_target)
667
        else:
lintangsutawika's avatar
lintangsutawika committed
668
            if (type(test_target) is int) and (test_choice is not None):
lintangsutawika's avatar
lintangsutawika committed
669
                test_target = test_choice[test_target]
670
            else:
lintangsutawika's avatar
lintangsutawika committed
671
                test_target = str(test_target)
672

673
674
675
        if test_choice is not None:
            check_choices = test_choice
        else:
lintangsutawika's avatar
lintangsutawika committed
676
            check_choices = [test_target]
677
678
679
680

        for choice in check_choices:
            choice_has_whitespace = True if " " in choice else False
            delimiter_has_whitespace = (
681
                True if " " in self.config.target_delimiter else False
682
683
684
685
686
687
688
            )

            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):
689
                eval_logger.warning(
690
                    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'
691
                )
692

Ethan Smith's avatar
Ethan Smith committed
693
    def download(self, dataset_kwargs=None) -> None:
694
695
696
697
698
699
        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
700
    def has_training_docs(self) -> bool:
701
        if self.config.training_split is not None:
702
703
704
705
            return True
        else:
            return False

baberabb's avatar
baberabb committed
706
    def has_validation_docs(self) -> bool:
707
        if self.config.validation_split is not None:
708
709
710
711
            return True
        else:
            return False

baberabb's avatar
baberabb committed
712
    def has_test_docs(self) -> bool:
713
        if self.config.test_split is not None:
714
715
716
717
            return True
        else:
            return False

baberabb's avatar
baberabb committed
718
    def training_docs(self) -> datasets.Dataset:
719
        if self.has_training_docs():
720
721
722
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.training_split]
723
                )
724
            return self.dataset[self.config.training_split]
725

baberabb's avatar
baberabb committed
726
    def validation_docs(self) -> datasets.Dataset:
727
        if self.has_validation_docs():
728
729
730
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.validation_split]
731
                )
732
            return self.dataset[self.config.validation_split]
733

baberabb's avatar
baberabb committed
734
    def test_docs(self) -> datasets.Dataset:
735
        if self.has_test_docs():
736
737
738
            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]
739

740
    def fewshot_docs(self):
741
742
        if self.config.fewshot_split is not None:
            return self.dataset[self.config.fewshot_split]
743
        else:
744
            if self.config.num_fewshot > 0:
745
                eval_logger.warning(
746
                    f"Task '{self.config.task}': "
747
748
749
                    "num_fewshot > 0 but fewshot_split is None. "
                    "using preconfigured rule."
                )
Herbie Bradley's avatar
Herbie Bradley committed
750
751
752
                return super().fewshot_docs()
            else:
                return None
753

754
755
756
757
758
759
760
761
    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

762
    def should_decontaminate(self):
763
        return self.config.should_decontaminate
764
765

    def doc_to_decontamination_query(self, doc):
766
767
768
        if self.config.should_decontaminate:
            if self.config.doc_to_decontamination_query in self.features:
                return doc[self.config.doc_to_decontamination_query]
769
770
            else:
                return ast.literal_eval(
771
                    utils.apply_template(self.config.doc_to_decontamination_query, doc)
772
                )
773

774
775
776
777
778
779
780
781
782
783
784
785
    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):
786
787
        if self.prompt is not None:
            doc_to_text = self.prompt
788
        else:
789
            doc_to_text = self.config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
790

791
792
793
        if type(doc_to_text) == int:
            return doc_to_text
        elif type(doc_to_text) == str:
794
            if doc_to_text in self.features:
795
                # if self.config.doc_to_choice is not None:
796
797
                #     return self.doc_to_choice(doc)[doc[doc_to_text]]
                # else:
798
799
                return doc[doc_to_text]
            else:
lintangsutawika's avatar
lintangsutawika committed
800
                text_string = utils.apply_template(doc_to_text, doc)
lintangsutawika's avatar
lintangsutawika committed
801
                if text_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
802
803
804
                    return ast.literal_eval(text_string)
                else:
                    return text_string
805
        elif callable(doc_to_text):
806
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
807
        # Used when applying a Promptsource template
808
        elif hasattr(doc_to_text, "apply"):
809
810
811
812
813
            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")
814
                return self.config.fewshot_delimiter
815
        else:
816
            print(type(doc_to_text))
817
            raise TypeError
818

819
    def doc_to_target(self, doc: dict) -> Union[int, str, list]:
820
821
        if self.prompt is not None:
            doc_to_target = self.prompt
822
        else:
823
            doc_to_target = self.config.doc_to_target
824

825
826
827
        if type(doc_to_target) == int:
            return doc_to_target
        elif type(doc_to_target) == str:
828
            if doc_to_target in self.features:
829
                # if self.config.doc_to_choice is not None:
830
831
832
                #     return self.doc_to_choice(doc)[doc[doc_to_target]]
                # else:
                return doc[doc_to_target]
833
            else:
lintangsutawika's avatar
lintangsutawika committed
834
                target_string = utils.apply_template(doc_to_target, doc)
Herbie Bradley's avatar
Herbie Bradley committed
835
                # return target_string
lintangsutawika's avatar
lintangsutawika committed
836
                if target_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
837
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
838
839
840
841
842
                elif (
                    len(target_string) >= 2
                    and (target_string[0] == "[")
                    and (target_string[-1] == "]")
                ):
843
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
844
845
                else:
                    return target_string
846
847
        elif type(doc_to_target) == list:
            return doc_to_target
848
        elif callable(doc_to_target):
849
            return doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
850
        # Used when applying a Promptsource template
851
        elif hasattr(doc_to_target, "apply"):
852
            applied_prompt = doc_to_target.apply(doc)
853
854
855
856
            if len(applied_prompt) == 2:
                return applied_prompt[1]
            else:
                eval_logger.warning("Applied prompt returns empty string")
857
                return self.config.fewshot_delimiter
858
859
        else:
            raise TypeError
860

baberabb's avatar
baberabb committed
861
    def doc_to_choice(self, doc: Any) -> List[str]:
862
863
        if self.prompt is not None:
            doc_to_choice = self.prompt
864
        elif self.config.doc_to_choice is None:
865
866
            eval_logger.error("doc_to_choice was called but not set in config")
        else:
867
            doc_to_choice = self.config.doc_to_choice
868
869
870
871
872
873
874
875
876
877
878
879
880

        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
881

882
    def gold_alias(self, doc):
883
884
885
886
887
        # 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]}}.
888
889
        if self.config.gold_alias is not None:
            doc_to_target = self.config.gold_alias
890
        else:
lintangsutawika's avatar
lintangsutawika committed
891
            return self.doc_to_target(doc)
892
893
894
895
896
897
898
899
900
901

        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
902
903
904
    def construct_requests(
        self, doc: dict, ctx: str, **kwargs
    ) -> Union[List[Instance], Instance]:
905
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
906
            arguments = (ctx, self.doc_to_target(doc))
907
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
908
            arguments = (self.doc_to_target(doc),)
909
        elif self.OUTPUT_TYPE == "multiple_choice":
910
            choices = self.doc_to_choice(doc)
911
            target_delimiter = self.config.target_delimiter
912
913
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
914
                cont = self.doc_to_target(doc)
915
                arguments = [(ctx, f"{target_delimiter}{cont}") for ctx in choices]
916
            else:
917
                # Otherwise they are placed in the continuation
918
                arguments = [(ctx, f"{target_delimiter}{cont}") for cont in choices]
919

920
            request_list = [
921
922
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
923
                    doc=doc,
924
                    arguments=arg,
925
                    idx=i,
926
927
                    **kwargs,
                )
928
                for i, arg in enumerate(arguments)
929
            ]
930
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
931
            if "acc_mutual_info" in self._metric_fn_list.keys():
932
933
934
                # if we are calculating multiple choice accuracy
                # using mutual information instead of raw loglikelihood as metric, need unconditional lls.

lintangsutawika's avatar
lintangsutawika committed
935
                # here mutual info refers to calculating
936
937
938
939
940
941
                # 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
942
                            doc=doc,
943
                            arguments=("", "{}".format(choice)),
944
945
946
                            idx=i,
                            **kwargs,
                        )
lintangsutawika's avatar
lintangsutawika committed
947
                        for i, choice in enumerate(choices)
948
949
950
                    ]
                )
            return request_list
lintangsutawika's avatar
lintangsutawika committed
951

952
        elif self.OUTPUT_TYPE == "greedy_until":
953
            arguments = (ctx, self.config.generation_kwargs)
lintangsutawika's avatar
lintangsutawika committed
954
955

        return Instance(
lintangsutawika's avatar
lintangsutawika committed
956
957
            request_type=self.OUTPUT_TYPE, doc=doc, arguments=arguments, idx=0, **kwargs
        )
958
959

    def process_results(self, doc, results):
960
961
        if callable(self.config.process_results):
            return self.config.process_results(doc, results)
lintangsutawika's avatar
lintangsutawika committed
962

963
        result_dict = {}
964
        use_metric = list(self._metric_fn_list.keys())
965
966
967
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
968
969
970
971
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
972
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
973
            (loglikelihood,) = results
974
975
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
haileyschoelkopf's avatar
haileyschoelkopf committed
976
            return {
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
                **(
                    {"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
992
            }
993
        elif self.OUTPUT_TYPE == "multiple_choice":
994
            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
995

996
            # retrieve choices in List[str] form, to compute choice lengths, etc.
997
            choices = self.doc_to_choice(doc)
998
999
            completion_len = np.array([float(len(i)) for i in choices])

1000
            if 2 * len(choices) == len(lls) and "acc_mutual_info" in use_metric:
1001
1002
1003
1004
1005
1006
                # 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]
1007

1008
1009
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
1010

1011
1012
            if self.multiple_input:
                gold = self.doc_to_text(doc)
1013
            else:
1014
                gold = self.doc_to_target(doc)
1015
1016
1017

            gold_index_error = False
            if type(gold) is list:
Lintang Sutawika's avatar
Lintang Sutawika committed
1018
1019
                gold = [i if i < len(choices) else -100 for i in gold]
                if -100 in gold:
1020
1021
1022
                    gold_index_error = True
            else:
                if type(gold) is int:
Lintang Sutawika's avatar
Lintang Sutawika committed
1023
                    gold = gold if gold < len(choices) else -100
1024
                elif type(gold) is str:
Lintang Sutawika's avatar
Lintang Sutawika committed
1025
                    gold = choices.index(gold) if gold in choices else -100
lintangsutawika's avatar
lintangsutawika committed
1026

Lintang Sutawika's avatar
Lintang Sutawika committed
1027
                if gold == -100:
1028
1029
1030
1031
                    gold_index_error = True

            if gold_index_error:
                eval_logger.warning(
lintangsutawika's avatar
lintangsutawika committed
1032
                    f"Label index was not in within range of available choices,"
1033
1034
                    f"Sample:\n\n{doc}\n\n"
                )
1035
1036
1037

            if "ece" in use_metric:
                # Convert lls from log-probabilities to normalized probabilities
1038
1039
                norm_probs: list[float] = np.exp(lls - sp.logsumexp(lls)).tolist()
                calib_scores: list[float] = [0.0] * len(choices)
1040
1041
1042
1043
1044
                if isinstance(gold, list):
                    for g in gold:
                        calib_scores[g] = 1.0
                else:
                    calib_scores[gold] = 1.0
1045
                calibration_probs: dict[str, list[float]] = {
1046
1047
1048
1049
                    "probs": norm_probs,
                    "scores": calib_scores,
                }

1050
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
1051
1052
                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
1053
                exact_match = int(any([is_greedy[i] if i != -100 else 0 for i in gold]))
lintangsutawika's avatar
lintangsutawika committed
1054
1055
1056
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
1057
                # 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
1058
                exact_match = int(is_greedy[gold]) if gold != -100 else 0
1059
1060

            result_dict = {
1061
                **({"acc": acc} if "acc" in use_metric else {}),
1062
1063
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
1064
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
1065
                **({"exact_match": exact_match} if "exact_match" in use_metric else {}),
1066
                **({"ece": calibration_probs} if "ece" in use_metric else {}),
1067
1068
            }

1069
            if "acc_mutual_info" in use_metric:
lintangsutawika's avatar
lintangsutawika committed
1070
1071
1072
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
1073
1074
1075
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

1076
        elif self.OUTPUT_TYPE == "greedy_until":
1077
            gold = self.doc_to_target(doc)
1078
            if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1079
                # If you set doc_to_choice,
lintangsutawika's avatar
lintangsutawika committed
1080
                # it assumes that doc_to_target returns a number.
1081
1082
                choices = self.doc_to_choice(doc)
                gold = choices[gold]
lintangsutawika's avatar
lintangsutawika committed
1083
1084
            else:
                gold = str(gold)
1085

lintangsutawika's avatar
lintangsutawika committed
1086
            result = results[0]
lintangsutawika's avatar
lintangsutawika committed
1087
            for metric in self._metric_fn_list.keys():
haileyschoelkopf's avatar
haileyschoelkopf committed
1088
1089
1090
1091
1092
1093
                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:
1094
                        try:
1095
                            result_score = self._metric_fn_list[metric](
1096
1097
                                references=[gold_option],
                                predictions=[result],
1098
                                **self._metric_fn_kwargs[metric],
1099
                            )
Herbie Bradley's avatar
Herbie Bradley committed
1100
1101
1102
                        except (
                            TypeError
                        ):  # TODO: this is hacky and I don't want to do it
1103
                            result_score = self._metric_fn_list[metric](
haileyschoelkopf's avatar
haileyschoelkopf committed
1104
1105
1106
                                [gold_option, result]
                            )
                        if isinstance(result_score, dict):
haileyschoelkopf's avatar
haileyschoelkopf committed
1107
                            # TODO: this handles the case where HF evaluate returns a dict.
1108
                            result_score = result_score[metric]
haileyschoelkopf's avatar
haileyschoelkopf committed
1109
                        scores.append(result_score)
haileyschoelkopf's avatar
haileyschoelkopf committed
1110
                    if any(scores):
1111
                        result_score = 1.0
haileyschoelkopf's avatar
haileyschoelkopf committed
1112
                    else:
1113
                        result_score = 0.0
haileyschoelkopf's avatar
haileyschoelkopf committed
1114
                else:
1115
                    try:
1116
                        result_score = self._metric_fn_list[metric](
1117
1118
                            references=[gold],
                            predictions=[result],
1119
                            **self._metric_fn_kwargs[metric],
1120
                        )
Herbie Bradley's avatar
Herbie Bradley committed
1121
1122
1123
                    except (
                        TypeError
                    ):  # needed for now in order to use a different interface between our own metrics and HF Evaluate metrics
1124
                        result_score = self._metric_fn_list[metric]([gold, result])
1125
1126
1127
1128
                    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
1129
        else:
lintangsutawika's avatar
lintangsutawika committed
1130
1131
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
1132
                "'loglikelihood', 'loglikelihood_rolling', 'greedy_until' or 'multiple_choice'",
1133
            )
1134
1135
1136
1137
1138
1139
1140

        return result_dict

    def aggregation(self):
        return self._aggregation_list

    def higher_is_better(self):
haileyschoelkopf's avatar
haileyschoelkopf committed
1141
        return self._higher_is_better
1142
1143
1144
1145
1146


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

baberabb's avatar
baberabb committed
1147
    def doc_to_target(self, doc: dict) -> str:
1148
1149
        return " " + doc["choices"][doc["gold"]]

baberabb's avatar
baberabb committed
1150
    def construct_requests(self, doc: dict, ctx: str, **kwargs) -> List[Instance]:
1151
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1152
1153
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1154
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1155
                doc=doc,
1156
                arguments=(ctx, " {}".format(choice)),
1157
                idx=i,
1158
1159
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1160
1161
            for i, choice in enumerate(doc["choices"])
        ]
1162

baberabb's avatar
baberabb committed
1163
    def process_results(self, doc: dict, results: List[Tuple[float, bool]]) -> dict:
lintangsutawika's avatar
lintangsutawika committed
1164
1165
1166
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
        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
1178
    def higher_is_better(self) -> dict:
1179
1180
1181
1182
1183
        return {
            "acc": True,
            "acc_norm": True,
        }

baberabb's avatar
baberabb committed
1184
    def aggregation(self) -> dict:
1185
1186
1187
1188
1189
1190
        return {
            "acc": mean,
            "acc_norm": mean,
        }


lintangsutawika's avatar
lintangsutawika committed
1191
class PerplexityTask(Task):
1192
1193
    OUTPUT_TYPE = "loglikelihood_rolling"

baberabb's avatar
baberabb committed
1194
    def has_training_docs(self) -> bool:
1195
1196
        return False

baberabb's avatar
baberabb committed
1197
    def fewshot_examples(self, k: int, rnd) -> List:
1198
1199
1200
        assert k == 0
        return []

baberabb's avatar
baberabb committed
1201
    def fewshot_context(self, doc: dict, num_fewshot: int) -> Literal[""]:
1202
1203
1204
1205
1206
1207
        assert (
            num_fewshot == 0
        ), "The number of fewshot examples must be 0 for perplexity tasks."

        return ""

baberabb's avatar
baberabb committed
1208
    def higher_is_better(self) -> dict:
1209
1210
1211
1212
1213
1214
1215
1216
1217
        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
1218
    def doc_to_text(self, doc) -> str:
1219
1220
1221
1222
1223
        return ""

    def doc_to_target(self, doc):
        return doc

baberabb's avatar
baberabb committed
1224
    def construct_requests(self, doc: dict, ctx: Union[str, None], **kwargs):
1225
1226
        assert not ctx

lintangsutawika's avatar
lintangsutawika committed
1227
1228
1229
1230
1231
1232
1233
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1234

baberabb's avatar
baberabb committed
1235
    def process_results(self, doc: dict, results: float) -> dict:
1236
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1237
1238
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1239
1240
1241
1242
1243
1244
        return {
            "word_perplexity": (loglikelihood, words),
            "byte_perplexity": (loglikelihood, bytes_),
            "bits_per_byte": (loglikelihood, bytes_),
        }

baberabb's avatar
baberabb committed
1245
    def aggregation(self) -> dict:
1246
1247
1248
1249
1250
1251
1252
        return {
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
            "bits_per_byte": bits_per_byte,
        }

    @classmethod
baberabb's avatar
baberabb committed
1253
    def count_bytes(cls, doc) -> int:
1254
1255
1256
        return len(doc.encode("utf-8"))

    @classmethod
baberabb's avatar
baberabb committed
1257
    def count_words(cls, doc) -> int:
1258
1259
        """Downstream tasks with custom word boundaries should override this!"""
        return len(re.split(r"\s+", doc))