task.py 47.1 KB
Newer Older
1
import abc
2
from dataclasses import dataclass, field, asdict
3
4

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

import datasets
import numpy as np

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

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

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

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

50
51
52

@dataclass
class TaskConfig(dict):
53
    # task naming/registry
54
    task: str = None
55
    group: Union[str, list] = None
56
57
58
    # HF dataset options.
    # which dataset to use,
    # and what splits for what purpose
59
60
    dataset_path: str = None
    dataset_name: str = None
61
    dataset_kwargs: dict = None
62
63
64
    training_split: str = None
    validation_split: str = None
    test_split: str = None
lintangsutawika's avatar
lintangsutawika committed
65
    fewshot_split: str = None  # TODO: assert that this not None if num_fewshot > 0. (?) assert if this is same split as one evaling (?)
66
67
    # formatting / prompting options.
    # see docs/advanced_task_guide.md for more info
68
    process_docs: Callable = None
69
70
    doc_to_text: Union[Callable, str] = None
    doc_to_target: Union[Callable, str] = None
lintangsutawika's avatar
lintangsutawika committed
71
    doc_to_choice: Union[Callable, str, dict, list] = None
72
    gold_alias: Union[Callable, str] = None
lintangsutawika's avatar
lintangsutawika committed
73
    process_results: Union[Callable, str] = None
74
    use_prompt: str = None
75
    description: str = ""
76
77
    target_delimiter: str = " "
    fewshot_delimiter: str = "\n\n"
haileyschoelkopf's avatar
haileyschoelkopf committed
78
    fewshot_config: dict = None
79
    # runtime configuration options
80
    num_fewshot: int = 0
81
    # scoring options
82
    metric_list: list = None
83
    output_type: str = "greedy_until"
84
    generation_kwargs: dict = None
85
    repeats: int = 1
lintangsutawika's avatar
lintangsutawika committed
86
    filter_list: Union[str, list] = None
87
88
    should_decontaminate: bool = False
    doc_to_decontamination_query: str = None
89

lintangsutawika's avatar
lintangsutawika committed
90
    metadata: str = None  # by default, not used in the code. allows for users to pass arbitrary info to tasks
91

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

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

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

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

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

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

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

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

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

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

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

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

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

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

    VERSION = None
164

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

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

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

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

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

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

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

Ethan Smith's avatar
Ethan Smith committed
221
    def download(self, data_dir=None, cache_dir=None, download_mode=None) -> None:
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
        """Downloads and returns the task dataset.
        Override this method to download the dataset from a custom API.

        :param data_dir: str
            Stores the path to a local folder containing the `Task`'s data files.
            Use this to specify the path to manually downloaded data (usually when
            the dataset is not publicly accessible).
        :param cache_dir: str
            The directory to read/write the `Task` dataset. This follows the
            HuggingFace `datasets` API with the default cache directory located at:
                `~/.cache/huggingface/datasets`
            NOTE: You can change the cache location globally for a given process
            by setting the shell environment variable, `HF_DATASETS_CACHE`,
            to another directory:
                `export HF_DATASETS_CACHE="/path/to/another/directory"`
        :param download_mode: datasets.DownloadMode
            How to treat pre-existing `Task` downloads and data.
            - `datasets.DownloadMode.REUSE_DATASET_IF_EXISTS`
                Reuse download and reuse dataset.
            - `datasets.DownloadMode.REUSE_CACHE_IF_EXISTS`
                Reuse download with fresh dataset.
            - `datasets.DownloadMode.FORCE_REDOWNLOAD`
                Fresh download and fresh dataset.
        """
246
247
248
249
250
251
252
        self.dataset = datasets.load_dataset(
            path=self.DATASET_PATH,
            name=self.DATASET_NAME,
            data_dir=data_dir,
            cache_dir=cache_dir,
            download_mode=download_mode,
        )
253

254
255
256
257
258
    @property
    def config(self):
        """Returns the TaskConfig associated with this class."""
        return self._config

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
292
293
294
    @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 []

295
296
297
298
299
300
301
302
303
304
    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
305
            eval_logger.warning(
306
                "has_training_docs and has_validation_docs are False"
307
                ", using test_docs as fewshot_docs but this is not recommended."
lintangsutawika's avatar
lintangsutawika committed
308
            )
309
310
            return self.test_docs()

311
312
313
314
315
316
317
318
319
320
    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
321

322
323
324
325
326
327
328
329
330
331
332
333
334
    @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
335
    def doc_to_decontamination_query(self, doc) -> None:
336
337
338
339
340
341
342
343
344
345
346
347
348
        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
349
    def build_all_requests(self, limit=None, rank=None, world_size=None) -> None:
350
351
352
353
354
355
356
357
358
359
        """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!"

360
        eval_logger.info(
361
            f"Building contexts for task '{self.config.task}' on rank {rank}..."
362
363
        )

364
        instances = []
365
366
        for doc_id, doc in utils.create_iterator(
            enumerate(docs), rank, world_size, limit
lintangsutawika's avatar
lintangsutawika committed
367
        ):
368
            # sample fewshot context #TODO: need to offset doc_id by rank now!
369
            fewshot_ctx = self.fewshot_context(
370
                doc,
371
                self.config.num_fewshot,
372
            )
373

374
            # TODO: we should override self.config.repeats if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
375
376
377
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
378
                metadata=(self.config["task"], doc_id, self.config.repeats),
lintangsutawika's avatar
lintangsutawika committed
379
            )
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404

            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
405
            The number of times each instance in a dataset is inferred on. Defaults to 1,
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
438
439
440
            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
441
442
443
444
445
446
447
448
449
450
    @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))

451
    @utils.positional_deprecated
452
    def fewshot_context(self, doc, num_fewshot):
453
454
455
456
457
458
459
460
461
462
463
464
        """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:
465
            # always prepend the (possibly empty) task description
466
            labeled_examples = self.config.description
467
        else:
468
            labeled_examples = self.config.description + self.sampler.get_context(
lintangsutawika's avatar
lintangsutawika committed
469
470
                doc, num_fewshot
            )
471
472

        example = self.doc_to_text(doc)
473
474
475
476
        if type(example) == str:
            return labeled_examples + example
        elif type(example) == list:
            return [labeled_examples + ex for ex in example]
477
        elif type(example) == int:
478
            if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
479
480
481
482
                choices = self.doc_to_choice(doc)
                return labeled_examples + choices[example]
            else:
                return labeled_examples + str(example)
483
484

    def apply_filters(self):
lintangsutawika's avatar
lintangsutawika committed
485
486
487
488
489
490
        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
491

baberabb's avatar
baberabb committed
492
    def dump_config(self) -> dict:
493
        """Returns a dictionary representing the task's config.
494
495
496
497
498

        :returns: str
            The fewshot context.
        """
        # TODO: this should only return the overrides applied to a non-YAML task's configuration.
499
        # (num_fewshot)
500
        return self.config.to_dict()
501

502
503

class ConfigurableTask(Task):
504
    VERSION = "Yaml"
505
    OUTPUT_TYPE = None
506
    CONFIG = None
507
508
509

    def __init__(
        self, data_dir=None, cache_dir=None, download_mode=None, config: dict = None
Ethan Smith's avatar
Ethan Smith committed
510
    ) -> None:  # TODO no super() call here
511
        # Get pre-configured attributes
512
        self._config = self.CONFIG
513

514
        # Use new configurations if there was no preconfiguration
515
        if self.config is None:
516
            self._config = TaskConfig(**config)
517
518
        # Overwrite configs
        else:
lintangsutawika's avatar
lintangsutawika committed
519
            if config is not None:
520
                self._config.__dict__.update(config)
521

522
        if self.config is None:
lintangsutawika's avatar
lintangsutawika committed
523
524
525
            raise ValueError(
                "Must pass a config to ConfigurableTask, either in cls.CONFIG or `config` kwarg"
            )
526

527
528
529
        if self.config.output_type is not None:
            assert self.config.output_type in ALL_OUTPUT_TYPES
            self.OUTPUT_TYPE = self.config.output_type
530

531
532
        if self.config.dataset_path is not None:
            self.DATASET_PATH = self.config.dataset_path
533

534
535
        if self.config.dataset_name is not None:
            self.DATASET_NAME = self.config.dataset_name
536

537
538
539
540
        self._metric_fn_list = {}
        self._metric_fn_kwargs = {}
        self._aggregation_list = {}
        self._higher_is_better = {}
541

542
543
        _metric_list = DEFAULT_METRIC_REGISTRY[self.config.output_type]
        if self.config.metric_list is None:
544
            # TODO: handle this in TaskConfig.__post_init__ ?
545
            for metric_name in _metric_list:
haileyschoelkopf's avatar
haileyschoelkopf committed
546
547
                self._metric_fn_list[metric_name] = get_metric(metric_name)
                self._aggregation_list[metric_name] = get_default_aggregation(
548
                    metric_name
haileyschoelkopf's avatar
haileyschoelkopf committed
549
550
                )
                self._higher_is_better[metric_name] = is_higher_better(metric_name)
551
        else:
552
            for metric_config in self.config.metric_list:
553
554
555
556
557
                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
558
559
                    if key
                    not in ["metric", "aggregation", "higher_is_better", "hf_evaluate"]
560
                }
Chris's avatar
Chris committed
561
562
563
564
                hf_evaluate_metric = (
                    "hf_evaluate" in metric_config
                    and metric_config["hf_evaluate"] is True
                )
565

566
                if self.config.process_results is not None:
567
568
                    self._metric_fn_list[metric_name] = None
                    self._metric_fn_kwargs[metric_name] = {}
569
570
571
572
573
574
                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
575
576
577
                    self._metric_fn_list[metric_name] = get_metric(
                        metric_name, hf_evaluate_metric
                    )
578
                    self._metric_fn_kwargs[metric_name] = kwargs
lintangsutawika's avatar
lintangsutawika committed
579

580
                if "aggregation" in metric_config:
581
                    agg_name = metric_config["aggregation"]
582
                    if type(agg_name) == str:
haileyschoelkopf's avatar
haileyschoelkopf committed
583
                        self._aggregation_list[metric_name] = get_aggregation(agg_name)
584
585
586
587
                    elif callable(agg_name):
                        self._aggregation_list[metric_name] = metric_config[
                            "aggregation"
                        ]
588
                else:
589
                    INV_AGG_REGISTRY = {v: k for k, v in AGGREGATION_REGISTRY.items()}
haileyschoelkopf's avatar
haileyschoelkopf committed
590
                    metric_agg = get_default_aggregation(metric_name)
591
                    eval_logger.warning(
baberabb's avatar
baberabb committed
592
                        f"[Task: {self._config.task}] metric {metric_name} is defined, but aggregation is not. "
593
594
                        f"using default "
                        f"aggregation={INV_AGG_REGISTRY[metric_agg]}"
595
                    )
596
                    self._aggregation_list[metric_name] = metric_agg
lintangsutawika's avatar
lintangsutawika committed
597

598
599
600
601
602
603
                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
604
                        f"[Task: {self._config.task}] metric {metric_name} is defined, but higher_is_better is not. "
605
                        f"using default "
haileyschoelkopf's avatar
haileyschoelkopf committed
606
                        f"higher_is_better={is_higher_better(metric_name)}"
607
                    )
haileyschoelkopf's avatar
haileyschoelkopf committed
608
                    self._higher_is_better[metric_name] = is_higher_better(metric_name)
609

610
        self.download(self.config.dataset_kwargs)
611
612
613
        self._training_docs = None
        self._fewshot_docs = None

614
        if self.config.filter_list is not None:
lintangsutawika's avatar
lintangsutawika committed
615
            self._filters = []
616
            for filter_config in self.config.filter_list:
lintangsutawika's avatar
lintangsutawika committed
617
618
619
620
621
622
623
                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
624
625
626
                        }
                        components.append([function["function"], kwargs])
                    filter_pipeline = build_filter_ensemble(filter_name, components)
lintangsutawika's avatar
lintangsutawika committed
627
                self._filters.append(filter_pipeline)
lintangsutawika's avatar
lintangsutawika committed
628
        else:
629
            self._filters = [build_filter_ensemble("none", [["take_first", None]])]
630

631
632
        if self.config.use_prompt is not None:
            eval_logger.info(f"loading prompt {self.config.use_prompt}")
633
            self.prompt = get_prompt(
634
                self.config.use_prompt, self.DATASET_PATH, self.DATASET_NAME
lintangsutawika's avatar
lintangsutawika committed
635
            )
636
637
638
        else:
            self.prompt = None

lintangsutawika's avatar
lintangsutawika committed
639
        if self.fewshot_docs() is not None:
haileyschoelkopf's avatar
haileyschoelkopf committed
640
            self.sampler = samplers.get_sampler(
haileyschoelkopf's avatar
haileyschoelkopf committed
641
642
643
                self.config.fewshot_config.get("sampler", "default")
                if self.config.fewshot_config
                else "default"
haileyschoelkopf's avatar
haileyschoelkopf committed
644
            )(list(self.fewshot_docs()), self, rnd=random.Random(1234))
645

646
        if self.has_test_docs():
647
            self.task_docs = self.test_docs()
648
        elif self.has_validation_docs():
649
            self.task_docs = self.validation_docs()
650
651
652
653
654
        else:
            assert (
                False
            ), f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"

655
        # Test One Doc
656
        self.features = list(self.task_docs.features.keys())
657
658
        self.multiple_input = 0
        self.multiple_target = 0
659
        test_doc = self.task_docs[0]
660
        test_text = self.doc_to_text(test_doc)
661
        test_target = self.doc_to_target(test_doc)
lintangsutawika's avatar
lintangsutawika committed
662

663
        if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
664
665
666
            test_choice = self.doc_to_choice(test_doc)
            if type(test_choice) is not list:
                eval_logger.error("doc_to_choice must return list")
667
668
            else:
                num_choice = len(test_choice)
669

670
671
            if type(test_text) is int:
                self.multiple_input = num_choice
672
673
        else:
            test_choice = None
674

675
        if type(test_target) is list:
676
            self.multiple_target = len(test_target)
677
        else:
lintangsutawika's avatar
lintangsutawika committed
678
            if (type(test_target) is int) and (test_choice is not None):
lintangsutawika's avatar
lintangsutawika committed
679
                test_target = test_choice[test_target]
680
            else:
lintangsutawika's avatar
lintangsutawika committed
681
                test_target = str(test_target)
682

683
684
685
        if test_choice is not None:
            check_choices = test_choice
        else:
lintangsutawika's avatar
lintangsutawika committed
686
            check_choices = [test_target]
687
688
689
690
691
        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 = (
                    True if self.config.target_delimiter[-1].isspace() else False
692
                )
693

694
695
696
697
698
699
700
701
702
                if delimiter_has_whitespace and choice_has_whitespace:
                    eval_logger.warning(
                        f'Both target_delimiter and target choice: "{choice}" have whitespace'
                    )
                elif (not delimiter_has_whitespace) and (not choice_has_whitespace):
                    eval_logger.warning(
                        f'Both target_delimiter and target choice: "{choice}" does not have whitespace, ignore if the language you are evaluating on does not require/use whitespace'
                    )

Ethan Smith's avatar
Ethan Smith committed
703
    def download(self, dataset_kwargs=None) -> None:
704
705
706
707
708
709
        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
710
    def has_training_docs(self) -> bool:
711
        if self.config.training_split is not None:
712
713
714
715
            return True
        else:
            return False

baberabb's avatar
baberabb committed
716
    def has_validation_docs(self) -> bool:
717
        if self.config.validation_split is not None:
718
719
720
721
            return True
        else:
            return False

baberabb's avatar
baberabb committed
722
    def has_test_docs(self) -> bool:
723
        if self.config.test_split is not None:
724
725
726
727
            return True
        else:
            return False

baberabb's avatar
baberabb committed
728
    def training_docs(self) -> datasets.Dataset:
729
        if self.has_training_docs():
730
731
732
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.training_split]
733
                )
734
            return self.dataset[self.config.training_split]
735

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

baberabb's avatar
baberabb committed
744
    def test_docs(self) -> datasets.Dataset:
745
        if self.has_test_docs():
746
747
748
            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]
749

750
    def fewshot_docs(self):
751
752
        if self.config.fewshot_split is not None:
            return self.dataset[self.config.fewshot_split]
753
        else:
754
            if self.config.num_fewshot > 0:
755
                eval_logger.warning(
756
                    f"Task '{self.config.task}': "
757
758
759
760
                    "num_fewshot > 0 but fewshot_split is None. "
                    "using preconfigured rule."
                )
            return super().fewshot_docs()
761

762
763
764
765
766
767
768
769
770
    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

771
    def should_decontaminate(self):
772
        return self.config.should_decontaminate
773
774

    def doc_to_decontamination_query(self, doc):
775
776
777
        if self.config.should_decontaminate:
            if self.config.doc_to_decontamination_query in self.features:
                return doc[self.config.doc_to_decontamination_query]
778
779
            else:
                return ast.literal_eval(
780
                    utils.apply_template(self.config.doc_to_decontamination_query, doc)
781
                )
782

783
784
785
786
787
788
789
790
791
792
793
794
    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):
795
796
        if self.prompt is not None:
            doc_to_text = self.prompt
797
        else:
798
            doc_to_text = self.config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
799

800
801
802
        if type(doc_to_text) == int:
            return doc_to_text
        elif type(doc_to_text) == str:
803
            if doc_to_text in self.features:
804
                # if self.config.doc_to_choice is not None:
805
806
                #     return self.doc_to_choice(doc)[doc[doc_to_text]]
                # else:
807
808
                return doc[doc_to_text]
            else:
lintangsutawika's avatar
lintangsutawika committed
809
                text_string = utils.apply_template(doc_to_text, doc)
lintangsutawika's avatar
lintangsutawika committed
810
                if text_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
811
812
813
                    return ast.literal_eval(text_string)
                else:
                    return text_string
814
        elif callable(doc_to_text):
815
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
816
        # Used when applying a Promptsource template
817
        elif hasattr(doc_to_text, "apply"):
818
819
820
821
822
            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")
823
                return self.config.fewshot_delimiter
824
        else:
825
            print(type(doc_to_text))
826
            raise TypeError
827

828
    def doc_to_target(self, doc: dict) -> Union[int, str, list]:
829
830
        if self.prompt is not None:
            doc_to_target = self.prompt
831
        else:
832
            doc_to_target = self.config.doc_to_target
833

834
835
836
        if type(doc_to_target) == int:
            return doc_to_target
        elif type(doc_to_target) == str:
837
            if doc_to_target in self.features:
838
                # if self.config.doc_to_choice is not None:
839
840
841
                #     return self.doc_to_choice(doc)[doc[doc_to_target]]
                # else:
                return doc[doc_to_target]
842
            else:
lintangsutawika's avatar
lintangsutawika committed
843
                target_string = utils.apply_template(doc_to_target, doc)
lintangsutawika's avatar
lintangsutawika committed
844
                if target_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
845
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
846
847
848
849
850
                elif (
                    len(target_string) >= 2
                    and (target_string[0] == "[")
                    and (target_string[-1] == "]")
                ):
baberabb's avatar
baberabb committed
851
852
853
854
                    try:
                        return ast.literal_eval(target_string)
                    except (SyntaxError, ValueError):
                        return target_string
lintangsutawika's avatar
lintangsutawika committed
855
856
                else:
                    return target_string
857
858
        elif type(doc_to_target) == list:
            return doc_to_target
859
        elif callable(doc_to_target):
860
            return doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
861
        # Used when applying a Promptsource template
862
        elif hasattr(doc_to_target, "apply"):
863
            applied_prompt = doc_to_target.apply(doc)
864
865
866
867
            if len(applied_prompt) == 2:
                return applied_prompt[1]
            else:
                eval_logger.warning("Applied prompt returns empty string")
868
                return self.config.fewshot_delimiter
869
870
        else:
            raise TypeError
871

baberabb's avatar
baberabb committed
872
    def doc_to_choice(self, doc: Any) -> List[str]:
873
874
        if self.prompt is not None:
            doc_to_choice = self.prompt
875
        elif self.config.doc_to_choice is None:
876
877
            eval_logger.error("doc_to_choice was called but not set in config")
        else:
878
            doc_to_choice = self.config.doc_to_choice
879
880
881
882
883
884
885
886
887
888
889
890
891

        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
892

893
    def gold_alias(self, doc):
894
895
896
897
898
        # 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]}}.
899
900
        if self.config.gold_alias is not None:
            doc_to_target = self.config.gold_alias
901
        else:
lintangsutawika's avatar
lintangsutawika committed
902
            return self.doc_to_target(doc)
903
904
905
906
907
908
909
910
911
912

        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
913
914
915
    def construct_requests(
        self, doc: dict, ctx: str, **kwargs
    ) -> Union[List[Instance], Instance]:
916
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
917
            arguments = (ctx, self.doc_to_target(doc))
918
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
919
            arguments = (self.doc_to_target(doc),)
920
        elif self.OUTPUT_TYPE == "multiple_choice":
921
            choices = self.doc_to_choice(doc)
922
            target_delimiter = self.config.target_delimiter
923
924
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
925
                cont = self.doc_to_target(doc)
926
                arguments = [(ctx, f"{target_delimiter}{cont}") for ctx in choices]
927
            else:
928
                # Otherwise they are placed in the continuation
929
                arguments = [(ctx, f"{target_delimiter}{cont}") for cont in choices]
930

931
            request_list = [
932
933
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
934
                    doc=doc,
935
                    arguments=arg,
936
                    idx=i,
937
938
                    **kwargs,
                )
939
                for i, arg in enumerate(arguments)
940
            ]
941
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
942
            if "acc_mutual_info" in self._metric_fn_list.keys():
943
944
945
                # if we are calculating multiple choice accuracy
                # using mutual information instead of raw loglikelihood as metric, need unconditional lls.

lintangsutawika's avatar
lintangsutawika committed
946
                # here mutual info refers to calculating
947
948
949
950
951
952
                # 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
953
                            doc=doc,
954
                            arguments=("", "{}".format(choice)),
955
956
957
                            idx=i,
                            **kwargs,
                        )
lintangsutawika's avatar
lintangsutawika committed
958
                        for i, choice in enumerate(choices)
959
960
961
                    ]
                )
            return request_list
lintangsutawika's avatar
lintangsutawika committed
962

963
        elif self.OUTPUT_TYPE == "greedy_until":
964
            arguments = (ctx, self.config.generation_kwargs)
lintangsutawika's avatar
lintangsutawika committed
965
966

        return Instance(
lintangsutawika's avatar
lintangsutawika committed
967
968
            request_type=self.OUTPUT_TYPE, doc=doc, arguments=arguments, idx=0, **kwargs
        )
969
970
971

    def process_results(self, doc, results):

972
973
        if callable(self.config.process_results):
            return self.config.process_results(doc, results)
lintangsutawika's avatar
lintangsutawika committed
974

975
        result_dict = {}
976
        use_metric = list(self._metric_fn_list.keys())
977
978
979
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
980
981
982
983
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
984
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
985
            (loglikelihood,) = results
986
987
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
haileyschoelkopf's avatar
haileyschoelkopf committed
988
            return {
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
                **(
                    {"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
1004
            }
1005
        elif self.OUTPUT_TYPE == "multiple_choice":
1006
            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
1007

1008
            # retrieve choices in List[str] form, to compute choice lengths, etc.
1009
            choices = self.doc_to_choice(doc)
1010
1011
            completion_len = np.array([float(len(i)) for i in choices])

1012
1013
            if (
                2 * len(choices) == len(lls)
1014
                and "acc_mutual_info" in self._metric_fn_list.keys()
1015
1016
1017
1018
1019
1020
1021
            ):
                # 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]
1022

1023
1024
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
1025

1026
1027
            if self.multiple_input:
                gold = self.doc_to_text(doc)
1028
            else:
1029
                gold = self.doc_to_target(doc)
1030
1031
1032

            gold_index_error = False
            if type(gold) is list:
Lintang Sutawika's avatar
Lintang Sutawika committed
1033
1034
                gold = [i if i < len(choices) else -100 for i in gold]
                if -100 in gold:
1035
1036
1037
                    gold_index_error = True
            else:
                if type(gold) is int:
Lintang Sutawika's avatar
Lintang Sutawika committed
1038
                    gold = gold if gold < len(choices) else -100
1039
                elif type(gold) is str:
Lintang Sutawika's avatar
Lintang Sutawika committed
1040
                    gold = choices.index(gold) if gold in choices else -100
lintangsutawika's avatar
lintangsutawika committed
1041

Lintang Sutawika's avatar
Lintang Sutawika committed
1042
                if gold == -100:
1043
1044
1045
1046
                    gold_index_error = True

            if gold_index_error:
                eval_logger.warning(
lintangsutawika's avatar
lintangsutawika committed
1047
                    f"Label index was not in within range of available choices,"
1048
1049
                    f"Sample:\n\n{doc}\n\n"
                )
lintangsutawika's avatar
lintangsutawika committed
1050

1051
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
1052
1053
                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
1054
                exact_match = int(any([is_greedy[i] if i != -100 else 0 for i in gold]))
lintangsutawika's avatar
lintangsutawika committed
1055
1056
1057
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
1058
                # 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
1059
                exact_match = int(is_greedy[gold]) if gold != -100 else 0
1060
1061

            result_dict = {
1062
                **({"acc": acc} if "acc" in use_metric else {}),
1063
1064
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
1065
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
1066
                **({"exact_match": exact_match} if "exact_match" 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)
Chris's avatar
Chris committed
1078
            result = results[0]
1079
            if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1080
                # If you set doc_to_choice,
lintangsutawika's avatar
lintangsutawika committed
1081
                # it assumes that doc_to_target returns a number.
1082
1083
                choices = self.doc_to_choice(doc)
                gold = choices[gold]
1084
1085
            # we expect multiple_targets to be a list.
            elif self.multiple_target:
baberabb's avatar
baberabb committed
1086
                gold = list(gold)
Chris's avatar
Chris committed
1087
1088
1089
            elif type(gold) != type(result):
                # cast gold to the same type as result
                gold = type(result)(gold)
1090

lintangsutawika's avatar
lintangsutawika committed
1091
            for metric in self._metric_fn_list.keys():
haileyschoelkopf's avatar
haileyschoelkopf committed
1092
1093
1094
1095
1096
                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
1097
1098
1099
1100
                    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
1101
                    for gold_option in gold:
1102
                        try:
1103
                            result_score = self._metric_fn_list[metric](
1104
1105
                                references=[gold_option],
                                predictions=[result],
1106
                                **self._metric_fn_kwargs[metric],
1107
1108
                            )
                        except TypeError:  # TODO: this is hacky and I don't want to do it
1109
                            result_score = self._metric_fn_list[metric](
haileyschoelkopf's avatar
haileyschoelkopf committed
1110
1111
1112
                                [gold_option, result]
                            )
                        if isinstance(result_score, dict):
haileyschoelkopf's avatar
haileyschoelkopf committed
1113
                            # TODO: this handles the case where HF evaluate returns a dict.
1114
                            result_score = result_score[metric]
haileyschoelkopf's avatar
haileyschoelkopf committed
1115
                        scores.append(result_score)
haileyschoelkopf's avatar
haileyschoelkopf committed
1116
                    if any(scores):
1117
                        result_score = 1.0
haileyschoelkopf's avatar
haileyschoelkopf committed
1118
                    else:
1119
                        result_score = 0.0
haileyschoelkopf's avatar
haileyschoelkopf committed
1120
                else:
1121
                    try:
1122
                        result_score = self._metric_fn_list[metric](
1123
1124
                            references=[gold],
                            predictions=[result],
1125
                            **self._metric_fn_kwargs[metric],
1126
                        )
1127
1128
                    except TypeError:  # needed for now in order to use a different interface between our own metrics and HF Evaluate metrics
                        result_score = self._metric_fn_list[metric]([gold, result])
1129
1130
1131
1132
                    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
1133
        else:
lintangsutawika's avatar
lintangsutawika committed
1134
1135
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
1136
                "'loglikelihood', 'loglikelihood_rolling', 'greedy_until' or 'multiple_choice'",
1137
            )
1138
1139
1140
1141
1142
1143
1144

        return result_dict

    def aggregation(self):
        return self._aggregation_list

    def higher_is_better(self):
haileyschoelkopf's avatar
haileyschoelkopf committed
1145
        return self._higher_is_better
1146
1147
1148
1149
1150


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

baberabb's avatar
baberabb committed
1151
    def doc_to_target(self, doc: dict) -> str:
1152
1153
        return " " + doc["choices"][doc["gold"]]

baberabb's avatar
baberabb committed
1154
    def construct_requests(self, doc: dict, ctx: str, **kwargs) -> List[Instance]:
1155
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1156
1157
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1158
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1159
                doc=doc,
1160
                arguments=(ctx, " {}".format(choice)),
1161
                idx=i,
1162
1163
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1164
1165
            for i, choice in enumerate(doc["choices"])
        ]
1166

baberabb's avatar
baberabb committed
1167
    def process_results(self, doc: dict, results: List[Tuple[float, bool]]) -> dict:
lintangsutawika's avatar
lintangsutawika committed
1168
1169
1170
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
        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
1182
    def higher_is_better(self) -> dict:
1183
1184
1185
1186
1187
        return {
            "acc": True,
            "acc_norm": True,
        }

baberabb's avatar
baberabb committed
1188
    def aggregation(self) -> dict:
1189
1190
1191
1192
1193
1194
        return {
            "acc": mean,
            "acc_norm": mean,
        }


lintangsutawika's avatar
lintangsutawika committed
1195
class PerplexityTask(Task):
1196
1197
    OUTPUT_TYPE = "loglikelihood_rolling"

baberabb's avatar
baberabb committed
1198
    def has_training_docs(self) -> bool:
1199
1200
        return False

baberabb's avatar
baberabb committed
1201
    def fewshot_examples(self, k: int, rnd) -> List:
1202
1203
1204
        assert k == 0
        return []

baberabb's avatar
baberabb committed
1205
    def fewshot_context(self, doc: dict, num_fewshot: int) -> Literal[""]:
1206
1207
1208
1209
1210
1211
        assert (
            num_fewshot == 0
        ), "The number of fewshot examples must be 0 for perplexity tasks."

        return ""

baberabb's avatar
baberabb committed
1212
    def higher_is_better(self) -> dict:
1213
1214
1215
1216
1217
1218
1219
1220
1221
        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
1222
    def doc_to_text(self, doc) -> str:
1223
1224
1225
1226
1227
        return ""

    def doc_to_target(self, doc):
        return doc

baberabb's avatar
baberabb committed
1228
    def construct_requests(self, doc: dict, ctx: Union[str, None], **kwargs):
1229
1230
        assert not ctx

lintangsutawika's avatar
lintangsutawika committed
1231
1232
1233
1234
1235
1236
1237
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1238

baberabb's avatar
baberabb committed
1239
    def process_results(self, doc: dict, results: float) -> dict:
1240
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1241
1242
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1243
1244
1245
1246
1247
1248
        return {
            "word_perplexity": (loglikelihood, words),
            "byte_perplexity": (loglikelihood, bytes_),
            "bits_per_byte": (loglikelihood, bytes_),
        }

baberabb's avatar
baberabb committed
1249
    def aggregation(self) -> dict:
1250
1251
1252
1253
1254
1255
1256
        return {
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
            "bits_per_byte": bits_per_byte,
        }

    @classmethod
baberabb's avatar
baberabb committed
1257
    def count_bytes(cls, doc) -> int:
1258
1259
1260
        return len(doc.encode("utf-8"))

    @classmethod
baberabb's avatar
baberabb committed
1261
    def count_words(cls, doc) -> int:
1262
1263
        """Downstream tasks with custom word boundaries should override this!"""
        return len(re.split(r"\s+", doc))