task.py 48.7 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
    get_metric,
    get_aggregation,
36
    get_metric_aggregation,
haileyschoelkopf's avatar
haileyschoelkopf committed
37
    is_higher_better,
38
39
    DEFAULT_METRIC_REGISTRY,
    OUTPUT_TYPE_REGISTRY,
lintangsutawika's avatar
lintangsutawika committed
40
41
    AGGREGATION_REGISTRY,
)
42

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

50
51
52

@dataclass
class TaskConfig(dict):
53
    # task naming/registry
54
    task: str = None
lintangsutawika's avatar
lintangsutawika committed
55
    task_alias: str = None
56
    group: Union[str, list] = None
lintangsutawika's avatar
lintangsutawika committed
57
    group_alias: Union[str, list] = None
58
59
60
    # HF dataset options.
    # which dataset to use,
    # and what splits for what purpose
61
62
    dataset_path: str = None
    dataset_name: str = None
63
    dataset_kwargs: dict = None
64
65
66
    training_split: str = None
    validation_split: str = None
    test_split: str = None
lintangsutawika's avatar
lintangsutawika committed
67
    fewshot_split: str = None  # TODO: assert that this not None if num_fewshot > 0. (?) assert if this is same split as one evaling (?)
68
69
    # formatting / prompting options.
    # see docs/advanced_task_guide.md for more info
70
    process_docs: Callable = None
71
72
    doc_to_text: Union[Callable, str] = None
    doc_to_target: Union[Callable, str] = None
lintangsutawika's avatar
lintangsutawika committed
73
    doc_to_choice: Union[Callable, str, dict, list] = None
lintangsutawika's avatar
lintangsutawika committed
74
    process_results: Union[Callable, str] = None
75
    use_prompt: str = None
76
    description: str = ""
77
78
    target_delimiter: str = " "
    fewshot_delimiter: str = "\n\n"
haileyschoelkopf's avatar
haileyschoelkopf committed
79
    fewshot_config: dict = None
80
    # runtime configuration options
81
    num_fewshot: int = 0
82
    # scoring options
83
    metric_list: list = None
84
    output_type: str = "generate_until"
85
    generation_kwargs: dict = None
86
    repeats: int = 1
lintangsutawika's avatar
lintangsutawika committed
87
    filter_list: Union[str, list] = None
88
89
    should_decontaminate: bool = False
    doc_to_decontamination_query: str = None
90

lintangsutawika's avatar
lintangsutawika committed
91
    metadata: str = None  # by default, not used in the code. allows for users to pass arbitrary info to tasks
92

Ethan Smith's avatar
Ethan Smith committed
93
    def __post_init__(self) -> None:
lintangsutawika's avatar
lintangsutawika committed
94
        if self.dataset_path and ("." in self.dataset_path):
lintangsutawika's avatar
lintangsutawika committed
95
96
            import inspect
            from importlib import import_module
lintangsutawika's avatar
format  
lintangsutawika committed
97

lintangsutawika's avatar
lintangsutawika committed
98
            self.dataset_path = inspect.getfile(import_module(self.dataset_path))
99

Lintang Sutawika's avatar
Lintang Sutawika committed
100
        if self.generation_kwargs is not None:
101
            if self.output_type != "generate_until":
Lintang Sutawika's avatar
Lintang Sutawika committed
102
                eval_logger.warning(
103
                    f"[{self.task}] passed `generation_kwargs`, but not using `output_type: generate_until`!"
Lintang Sutawika's avatar
Lintang Sutawika committed
104
                )
105
                assert self.output_type != "generate_until"
Lintang Sutawika's avatar
Lintang Sutawika committed
106
107
108
109
110
111
112

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

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

lintangsutawika's avatar
lintangsutawika committed
207
        self._config = TaskConfig({**config}) if config else TaskConfig()
208

lintangsutawika's avatar
lintangsutawika committed
209
        self._filters = [build_filter_ensemble("none", [["take_first", None]])]
210

Ethan Smith's avatar
Ethan Smith committed
211
    def download(self, data_dir=None, cache_dir=None, download_mode=None) -> None:
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
        """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.
        """
236
237
238
239
240
241
242
        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,
        )
243

244
245
246
247
248
    @property
    def config(self):
        """Returns the TaskConfig associated with this class."""
        return self._config

249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
    @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 []

285
286
287
288
289
290
291
292
293
294
    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
295
            eval_logger.warning(
296
                "has_training_docs and has_validation_docs are False"
297
                ", using test_docs as fewshot_docs but this is not recommended."
lintangsutawika's avatar
lintangsutawika committed
298
            )
299
300
            return self.test_docs()

301
302
303
304
305
306
307
308
309
310
    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
311

312
313
314
315
316
317
318
319
320
321
322
323
324
    @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
325
    def doc_to_decontamination_query(self, doc) -> None:
326
327
328
329
330
331
332
333
334
335
336
337
338
        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
339
    def build_all_requests(self, limit=None, rank=None, world_size=None) -> None:
340
341
342
343
344
345
346
347
348
349
        """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!"

350
        eval_logger.info(f"Building contexts for task on rank {rank}...")
351

352
        instances = []
353
354
        for doc_id, doc in utils.create_iterator(
            enumerate(docs), rank, world_size, limit
lintangsutawika's avatar
lintangsutawika committed
355
        ):
356
            # sample fewshot context #TODO: need to offset doc_id by rank now!
357
            fewshot_ctx = self.fewshot_context(
358
                doc,
359
                self.config.num_fewshot,
360
            )
361

362
            # TODO: we should override self.config.repeats if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
363
364
365
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
366
                metadata=(self.config["task"], doc_id, self.config.repeats),
lintangsutawika's avatar
lintangsutawika committed
367
            )
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392

            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
393
            The number of times each instance in a dataset is inferred on. Defaults to 1,
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
            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
429
430
431
432
433
434
435
436
437
438
    @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))

439
    @utils.positional_deprecated
lintangsutawika's avatar
lintangsutawika committed
440
    def fewshot_context(
441
442
443
444
445
446
        self,
        doc,
        num_fewshot,
        provide_description=None,
        rnd=random.Random(1234),
        description=None,
lintangsutawika's avatar
lintangsutawika committed
447
    ):
448
449
450
451
452
453
454
        """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.
lintangsutawika's avatar
lintangsutawika committed
455
456
457
458
459
460
461
        :param provide_description: bool
            Not implemented, and this option is deprecated and will be removed in a future version in favor of a different description providing method
        :param rnd: random.Random
            The pseudo-random number generator used to randomly sample examples.
            WARNING: This is currently a required arg although it's optionalized with a default `None`.
        :param description: str
            The task's description that will be prepended to the fewshot examples.
462
463
464
        :returns: str
            The fewshot context.
        """
lintangsutawika's avatar
lintangsutawika committed
465
466
467
468
        assert (
            rnd is not None
        ), "A `random.Random` generator argument must be provided to `rnd`"

469
        description = description if description else ""
470
471

        if num_fewshot == 0:
lintangsutawika's avatar
lintangsutawika committed
472
            labeled_examples = ""
473
        else:
lintangsutawika's avatar
lintangsutawika committed
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
            # for sets with no training docs, draw from other set *but ensure no overlap with current doc*
            if self.has_training_docs():
                fewshotex = self.fewshot_examples(k=num_fewshot, rnd=rnd)
            else:
                if self._fewshot_docs is None:
                    self._fewshot_docs = list(
                        self.validation_docs()
                        if self.has_validation_docs()
                        else self.test_docs()
                    )

                fewshotex = rnd.sample(self._fewshot_docs, num_fewshot + 1)

                # get rid of the doc that's the one we're evaluating, if it's in the fewshot
                fewshotex = [x for x in fewshotex if x != doc][:num_fewshot]

            labeled_examples = (
                "\n\n".join(
                    [
                        self.doc_to_text(doc) + self.doc_to_target(doc)
                        for doc in fewshotex
                    ]
                )
                + "\n\n"
lintangsutawika's avatar
lintangsutawika committed
498
            )
499
500

        example = self.doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
501
        return description + labeled_examples + example
502
503

    def apply_filters(self):
lintangsutawika's avatar
lintangsutawika committed
504
505
        if hasattr(self, "_filters"):
            for f in self._filters:
lintangsutawika's avatar
lintangsutawika committed
506
                f.apply(self._instances, None)
lintangsutawika's avatar
lintangsutawika committed
507
508
509
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances
510

baberabb's avatar
baberabb committed
511
    def dump_config(self) -> dict:
512
        """Returns a dictionary representing the task's config.
513
514
515
516
517

        :returns: str
            The fewshot context.
        """
        # TODO: this should only return the overrides applied to a non-YAML task's configuration.
518
        # (num_fewshot)
519
        return self.config.to_dict()
520

521
522

class ConfigurableTask(Task):
523
    VERSION = "Yaml"
524
    OUTPUT_TYPE = None
525
    CONFIG = None
526
527
528

    def __init__(
        self, data_dir=None, cache_dir=None, download_mode=None, config: dict = None
Ethan Smith's avatar
Ethan Smith committed
529
    ) -> None:  # TODO no super() call here
530
        # Get pre-configured attributes
531
        self._config = self.CONFIG
532

533
        # Use new configurations if there was no preconfiguration
534
        if self.config is None:
535
            self._config = TaskConfig(**config)
536
537
        # Overwrite configs
        else:
lintangsutawika's avatar
lintangsutawika committed
538
            if config is not None:
539
                self._config.__dict__.update(config)
540

541
        if self.config is None:
lintangsutawika's avatar
lintangsutawika committed
542
543
544
            raise ValueError(
                "Must pass a config to ConfigurableTask, either in cls.CONFIG or `config` kwarg"
            )
545

546
547
548
        if self.config.output_type is not None:
            assert self.config.output_type in ALL_OUTPUT_TYPES
            self.OUTPUT_TYPE = self.config.output_type
549

550
551
        if self.config.dataset_path is not None:
            self.DATASET_PATH = self.config.dataset_path
552

553
554
        if self.config.dataset_name is not None:
            self.DATASET_NAME = self.config.dataset_name
555

556
557
558
559
        self._metric_fn_list = {}
        self._metric_fn_kwargs = {}
        self._aggregation_list = {}
        self._higher_is_better = {}
560

561
        if self.config.metric_list is None:
562
            # TODO: handle this in TaskConfig.__post_init__ ?
563
564
            _metric_list = DEFAULT_METRIC_REGISTRY[self.config.output_type]

565
            for metric_name in _metric_list:
haileyschoelkopf's avatar
haileyschoelkopf committed
566
                self._metric_fn_list[metric_name] = get_metric(metric_name)
lintangsutawika's avatar
lintangsutawika committed
567
                self._metric_fn_kwargs[metric_name] = {}
568
569
570
                self._aggregation_list[metric_name] = get_metric_aggregation(
                    metric_name
                )
haileyschoelkopf's avatar
haileyschoelkopf committed
571
                self._higher_is_better[metric_name] = is_higher_better(metric_name)
572
        else:
573
            for metric_config in self.config.metric_list:
574
575
576
577
578
                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
579
580
                    if key
                    not in ["metric", "aggregation", "higher_is_better", "hf_evaluate"]
581
                }
Chris's avatar
Chris committed
582
583
584
585
                hf_evaluate_metric = (
                    "hf_evaluate" in metric_config
                    and metric_config["hf_evaluate"] is True
                )
586

587
                if self.config.process_results is not None:
588
589
                    self._metric_fn_list[metric_name] = None
                    self._metric_fn_kwargs[metric_name] = {}
590
591
592
593
594
595
                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
596
597
598
                    self._metric_fn_list[metric_name] = get_metric(
                        metric_name, hf_evaluate_metric
                    )
599
                    self._metric_fn_kwargs[metric_name] = kwargs
lintangsutawika's avatar
lintangsutawika committed
600

601
                if "aggregation" in metric_config:
602
                    agg_name = metric_config["aggregation"]
603
                    if type(agg_name) == str:
haileyschoelkopf's avatar
haileyschoelkopf committed
604
                        self._aggregation_list[metric_name] = get_aggregation(agg_name)
605
606
607
608
                    elif callable(agg_name):
                        self._aggregation_list[metric_name] = metric_config[
                            "aggregation"
                        ]
609
                else:
610
                    INV_AGG_REGISTRY = {v: k for k, v in AGGREGATION_REGISTRY.items()}
lintangsutawika's avatar
lintangsutawika committed
611
                    metric_agg = get_metric_aggregation(metric_name)
612
                    eval_logger.warning(
baberabb's avatar
baberabb committed
613
                        f"[Task: {self._config.task}] metric {metric_name} is defined, but aggregation is not. "
614
615
                        f"using default "
                        f"aggregation={INV_AGG_REGISTRY[metric_agg]}"
616
                    )
617
                    self._aggregation_list[metric_name] = metric_agg
lintangsutawika's avatar
lintangsutawika committed
618

619
620
621
622
623
624
                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
625
                        f"[Task: {self._config.task}] metric {metric_name} is defined, but higher_is_better is not. "
626
                        f"using default "
haileyschoelkopf's avatar
haileyschoelkopf committed
627
                        f"higher_is_better={is_higher_better(metric_name)}"
628
                    )
haileyschoelkopf's avatar
haileyschoelkopf committed
629
                    self._higher_is_better[metric_name] = is_higher_better(metric_name)
630

631
        self.download(self.config.dataset_kwargs)
632
633
634
        self._training_docs = None
        self._fewshot_docs = None

635
        if self.config.filter_list is not None:
lintangsutawika's avatar
lintangsutawika committed
636
            self._filters = []
637
            for filter_config in self.config.filter_list:
lintangsutawika's avatar
lintangsutawika committed
638
639
640
641
642
643
644
                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
645
646
647
                        }
                        components.append([function["function"], kwargs])
                    filter_pipeline = build_filter_ensemble(filter_name, components)
lintangsutawika's avatar
lintangsutawika committed
648
                self._filters.append(filter_pipeline)
lintangsutawika's avatar
lintangsutawika committed
649
        else:
650
            self._filters = [build_filter_ensemble("none", [["take_first", None]])]
651

652
653
        if self.config.use_prompt is not None:
            eval_logger.info(f"loading prompt {self.config.use_prompt}")
654
            self.prompt = get_prompt(
655
                self.config.use_prompt, self.DATASET_PATH, self.DATASET_NAME
lintangsutawika's avatar
lintangsutawika committed
656
            )
657
658
659
        else:
            self.prompt = None

lintangsutawika's avatar
lintangsutawika committed
660
        if self.fewshot_docs() is not None:
haileyschoelkopf's avatar
haileyschoelkopf committed
661
            self.sampler = samplers.get_sampler(
haileyschoelkopf's avatar
haileyschoelkopf committed
662
663
664
                self.config.fewshot_config.get("sampler", "default")
                if self.config.fewshot_config
                else "default"
haileyschoelkopf's avatar
haileyschoelkopf committed
665
            )(list(self.fewshot_docs()), self, rnd=random.Random(1234))
666

667
        if self.has_test_docs():
668
            self.task_docs = self.test_docs()
669
        elif self.has_validation_docs():
670
            self.task_docs = self.validation_docs()
671
672
673
674
675
        else:
            assert (
                False
            ), f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"

676
        # Test One Doc
677
        self.features = list(self.task_docs.features.keys())
678
679
        self.multiple_input = 0
        self.multiple_target = 0
680
        test_doc = self.task_docs[0]
681
        test_text = self.doc_to_text(test_doc)
682
        test_target = self.doc_to_target(test_doc)
lintangsutawika's avatar
lintangsutawika committed
683

684
        if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
685
686
687
            test_choice = self.doc_to_choice(test_doc)
            if type(test_choice) is not list:
                eval_logger.error("doc_to_choice must return list")
688
689
            else:
                num_choice = len(test_choice)
690

691
692
            if type(test_text) is int:
                self.multiple_input = num_choice
693
694
        else:
            test_choice = None
695

696
        if type(test_target) is list:
697
            self.multiple_target = len(test_target)
698
        else:
lintangsutawika's avatar
lintangsutawika committed
699
            if (type(test_target) is int) and (test_choice is not None):
lintangsutawika's avatar
lintangsutawika committed
700
                test_target = test_choice[test_target]
701
            else:
lintangsutawika's avatar
lintangsutawika committed
702
                test_target = str(test_target)
703

704
705
706
        if test_choice is not None:
            check_choices = test_choice
        else:
lintangsutawika's avatar
lintangsutawika committed
707
            check_choices = [test_target]
708
709
710
711
        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 = (
712
713
                    True
                    if self.config.target_delimiter.rstrip()
714
                    != self.config.target_delimiter
715
                    else False
716
                )
717

718
719
720
721
722
723
                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(
724
                        f'Both target_delimiter "{self.config.target_delimiter}" and target choice: "{choice}" do not have whitespace, ignore if the language you are evaluating on does not require/use whitespace'
725
726
                    )

Ethan Smith's avatar
Ethan Smith committed
727
    def download(self, dataset_kwargs=None) -> None:
728
729
730
731
732
733
        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
734
    def has_training_docs(self) -> bool:
735
        if self.config.training_split is not None:
736
737
738
739
            return True
        else:
            return False

baberabb's avatar
baberabb committed
740
    def has_validation_docs(self) -> bool:
741
        if self.config.validation_split is not None:
742
743
744
745
            return True
        else:
            return False

baberabb's avatar
baberabb committed
746
    def has_test_docs(self) -> bool:
747
        if self.config.test_split is not None:
748
749
750
751
            return True
        else:
            return False

baberabb's avatar
baberabb committed
752
    def training_docs(self) -> datasets.Dataset:
753
        if self.has_training_docs():
754
755
756
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.training_split]
757
                )
758
            return self.dataset[self.config.training_split]
759

baberabb's avatar
baberabb committed
760
    def validation_docs(self) -> datasets.Dataset:
761
        if self.has_validation_docs():
762
763
764
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.validation_split]
765
                )
766
            return self.dataset[self.config.validation_split]
767

baberabb's avatar
baberabb committed
768
    def test_docs(self) -> datasets.Dataset:
769
        if self.has_test_docs():
770
771
772
            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]
773

774
    def fewshot_docs(self):
775
776
        if self.config.fewshot_split is not None:
            return self.dataset[self.config.fewshot_split]
777
        else:
778
            if self.config.num_fewshot > 0:
779
                eval_logger.warning(
780
                    f"Task '{self.config.task}': "
781
782
783
784
                    "num_fewshot > 0 but fewshot_split is None. "
                    "using preconfigured rule."
                )
            return super().fewshot_docs()
785

lintangsutawika's avatar
lintangsutawika committed
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
    @utils.positional_deprecated
    def fewshot_context(self, doc, num_fewshot):
        """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:
            # always prepend the (possibly empty) task description
            labeled_examples = self.config.description
        else:
            labeled_examples = self.config.description + self.sampler.get_context(
                doc, num_fewshot
            )

        example = self.doc_to_text(doc)
        if type(example) == str:
            return labeled_examples + example
        elif type(example) == list:
            return [labeled_examples + ex for ex in example]
        elif type(example) == int:
            if self.config.doc_to_choice is not None:
                choices = self.doc_to_choice(doc)
                return labeled_examples + choices[example]
            else:
                return labeled_examples + str(example)

819
820
821
822
823
824
825
826
    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

827
    def should_decontaminate(self):
828
        return self.config.should_decontaminate
829
830

    def doc_to_decontamination_query(self, doc):
831
832
833
        if self.config.should_decontaminate:
            if self.config.doc_to_decontamination_query in self.features:
                return doc[self.config.doc_to_decontamination_query]
834
835
            else:
                return ast.literal_eval(
836
                    utils.apply_template(self.config.doc_to_decontamination_query, doc)
837
                )
838

839
840
841
842
843
844
845
846
847
848
849
850
    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):
851
852
        if self.prompt is not None:
            doc_to_text = self.prompt
853
        else:
854
            doc_to_text = self.config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
855

856
857
858
        if type(doc_to_text) == int:
            return doc_to_text
        elif type(doc_to_text) == str:
859
            if doc_to_text in self.features:
860
                # if self.config.doc_to_choice is not None:
861
862
                #     return self.doc_to_choice(doc)[doc[doc_to_text]]
                # else:
863
864
                return doc[doc_to_text]
            else:
lintangsutawika's avatar
lintangsutawika committed
865
                text_string = utils.apply_template(doc_to_text, doc)
lintangsutawika's avatar
lintangsutawika committed
866
                if text_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
867
868
869
                    return ast.literal_eval(text_string)
                else:
                    return text_string
870
        elif callable(doc_to_text):
871
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
872
        # Used when applying a Promptsource template
873
        elif hasattr(doc_to_text, "apply"):
874
875
876
877
878
            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")
879
                return self.config.fewshot_delimiter
880
        else:
881
            print(type(doc_to_text))
882
            raise TypeError
883

884
    def doc_to_target(self, doc: dict) -> Union[int, str, list]:
885
886
        if self.prompt is not None:
            doc_to_target = self.prompt
887
        else:
888
            doc_to_target = self.config.doc_to_target
889

890
891
892
        if type(doc_to_target) == int:
            return doc_to_target
        elif type(doc_to_target) == str:
893
            if doc_to_target in self.features:
894
                # if self.config.doc_to_choice is not None:
895
896
897
                #     return self.doc_to_choice(doc)[doc[doc_to_target]]
                # else:
                return doc[doc_to_target]
898
            else:
lintangsutawika's avatar
lintangsutawika committed
899
                target_string = utils.apply_template(doc_to_target, doc)
lintangsutawika's avatar
lintangsutawika committed
900
                if target_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
901
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
902
903
904
905
906
                elif (
                    len(target_string) >= 2
                    and (target_string[0] == "[")
                    and (target_string[-1] == "]")
                ):
baberabb's avatar
baberabb committed
907
908
909
910
                    try:
                        return ast.literal_eval(target_string)
                    except (SyntaxError, ValueError):
                        return target_string
lintangsutawika's avatar
lintangsutawika committed
911
912
                else:
                    return target_string
913
914
        elif type(doc_to_target) == list:
            return doc_to_target
915
        elif callable(doc_to_target):
916
            return doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
917
        # Used when applying a Promptsource template
918
        elif hasattr(doc_to_target, "apply"):
919
            applied_prompt = doc_to_target.apply(doc)
920
921
922
923
            if len(applied_prompt) == 2:
                return applied_prompt[1]
            else:
                eval_logger.warning("Applied prompt returns empty string")
924
                return self.config.fewshot_delimiter
925
926
        else:
            raise TypeError
927

baberabb's avatar
baberabb committed
928
    def doc_to_choice(self, doc: Any) -> List[str]:
929
930
        if self.prompt is not None:
            doc_to_choice = self.prompt
931
        elif self.config.doc_to_choice is None:
932
933
            eval_logger.error("doc_to_choice was called but not set in config")
        else:
934
            doc_to_choice = self.config.doc_to_choice
935
936
937
938
939
940
941
942
943
944
945
946
947

        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
948

baberabb's avatar
baberabb committed
949
950
951
    def construct_requests(
        self, doc: dict, ctx: str, **kwargs
    ) -> Union[List[Instance], Instance]:
952
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
953
            arguments = (ctx, self.doc_to_target(doc))
954
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
955
            arguments = (self.doc_to_target(doc),)
956
        elif self.OUTPUT_TYPE == "multiple_choice":
957
            choices = self.doc_to_choice(doc)
958
            target_delimiter = self.config.target_delimiter
959
960
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
961
                cont = self.doc_to_target(doc)
962
                arguments = [(ctx, f"{target_delimiter}{cont}") for ctx in choices]
963
            else:
964
                # Otherwise they are placed in the continuation
965
                arguments = [(ctx, f"{target_delimiter}{cont}") for cont in choices]
966

967
            request_list = [
968
969
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
970
                    doc=doc,
971
                    arguments=arg,
972
                    idx=i,
973
974
                    **kwargs,
                )
975
                for i, arg in enumerate(arguments)
976
            ]
977
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
978
            if "acc_mutual_info" in self._metric_fn_list.keys():
979
980
981
                # if we are calculating multiple choice accuracy
                # using mutual information instead of raw loglikelihood as metric, need unconditional lls.

lintangsutawika's avatar
lintangsutawika committed
982
                # here mutual info refers to calculating
983
984
985
986
987
988
                # 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
989
                            doc=doc,
990
                            arguments=("", "{}".format(choice)),
991
992
993
                            idx=i,
                            **kwargs,
                        )
lintangsutawika's avatar
lintangsutawika committed
994
                        for i, choice in enumerate(choices)
995
996
997
                    ]
                )
            return request_list
lintangsutawika's avatar
lintangsutawika committed
998

999
        elif self.OUTPUT_TYPE == "generate_until":
1000
            arguments = (ctx, self.config.generation_kwargs)
lintangsutawika's avatar
lintangsutawika committed
1001
1002

        return Instance(
lintangsutawika's avatar
lintangsutawika committed
1003
1004
            request_type=self.OUTPUT_TYPE, doc=doc, arguments=arguments, idx=0, **kwargs
        )
1005
1006

    def process_results(self, doc, results):
1007
1008
        if callable(self.config.process_results):
            return self.config.process_results(doc, results)
lintangsutawika's avatar
lintangsutawika committed
1009

1010
        result_dict = {}
1011
        use_metric = list(self._metric_fn_list.keys())
1012
1013
1014
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
1015
1016
1017
1018
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
1019
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
1020
            (loglikelihood,) = results
1021
1022
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
haileyschoelkopf's avatar
haileyschoelkopf committed
1023
            return {
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
                **(
                    {"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
1039
            }
1040
        elif self.OUTPUT_TYPE == "multiple_choice":
1041
            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
1042

1043
            # retrieve choices in List[str] form, to compute choice lengths, etc.
1044
            choices = self.doc_to_choice(doc)
1045
1046
            completion_len = np.array([float(len(i)) for i in choices])

1047
1048
            if (
                2 * len(choices) == len(lls)
1049
                and "acc_mutual_info" in self._metric_fn_list.keys()
1050
1051
1052
1053
1054
1055
1056
            ):
                # 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]
1057

1058
1059
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
1060

1061
1062
            if self.multiple_input:
                gold = self.doc_to_text(doc)
1063
            else:
1064
                gold = self.doc_to_target(doc)
1065
1066
1067

            gold_index_error = False
            if type(gold) is list:
Lintang Sutawika's avatar
Lintang Sutawika committed
1068
1069
                gold = [i if i < len(choices) else -100 for i in gold]
                if -100 in gold:
1070
1071
1072
                    gold_index_error = True
            else:
                if type(gold) is int:
Lintang Sutawika's avatar
Lintang Sutawika committed
1073
                    gold = gold if gold < len(choices) else -100
1074
                elif type(gold) is str:
Lintang Sutawika's avatar
Lintang Sutawika committed
1075
                    gold = choices.index(gold) if gold in choices else -100
lintangsutawika's avatar
lintangsutawika committed
1076

Lintang Sutawika's avatar
Lintang Sutawika committed
1077
                if gold == -100:
1078
1079
1080
1081
                    gold_index_error = True

            if gold_index_error:
                eval_logger.warning(
lintangsutawika's avatar
lintangsutawika committed
1082
                    f"Label index was not in within range of available choices,"
1083
1084
                    f"Sample:\n\n{doc}\n\n"
                )
lintangsutawika's avatar
lintangsutawika committed
1085

1086
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
1087
1088
                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
1089
                exact_match = int(any([is_greedy[i] if i != -100 else 0 for i in gold]))
lintangsutawika's avatar
lintangsutawika committed
1090
1091
1092
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
1093
                # 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
1094
                exact_match = int(is_greedy[gold]) if gold != -100 else 0
1095
1096

            result_dict = {
1097
                **({"acc": acc} if "acc" in use_metric else {}),
1098
1099
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
1100
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
1101
                **({"exact_match": exact_match} if "exact_match" in use_metric else {}),
1102
1103
            }

1104
            if "acc_mutual_info" in use_metric:
lintangsutawika's avatar
lintangsutawika committed
1105
1106
1107
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
1108
1109
1110
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

1111
        elif self.OUTPUT_TYPE == "generate_until":
1112
            gold = self.doc_to_target(doc)
Chris's avatar
Chris committed
1113
            result = results[0]
1114
            if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1115
                # If you set doc_to_choice,
lintangsutawika's avatar
lintangsutawika committed
1116
                # it assumes that doc_to_target returns a number.
1117
1118
                choices = self.doc_to_choice(doc)
                gold = choices[gold]
1119
1120
            # we expect multiple_targets to be a list.
            elif self.multiple_target:
baberabb's avatar
baberabb committed
1121
                gold = list(gold)
Chris's avatar
Chris committed
1122
1123
1124
            elif type(gold) != type(result):
                # cast gold to the same type as result
                gold = type(result)(gold)
1125

lintangsutawika's avatar
lintangsutawika committed
1126
            for metric in self._metric_fn_list.keys():
haileyschoelkopf's avatar
haileyschoelkopf committed
1127
1128
1129
1130
1131
                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
1132
1133
1134
1135
                    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
1136
                    for gold_option in gold:
1137
                        try:
1138
                            result_score = self._metric_fn_list[metric](
1139
1140
                                references=[gold_option],
                                predictions=[result],
1141
                                **self._metric_fn_kwargs[metric],
1142
                            )
baberabb's avatar
baberabb committed
1143
1144
1145
                        except (
                            TypeError
                        ):  # TODO: this is hacky and I don't want to do it
1146
                            result_score = self._metric_fn_list[metric](
haileyschoelkopf's avatar
haileyschoelkopf committed
1147
1148
1149
                                [gold_option, result]
                            )
                        if isinstance(result_score, dict):
haileyschoelkopf's avatar
haileyschoelkopf committed
1150
                            # TODO: this handles the case where HF evaluate returns a dict.
1151
                            result_score = result_score[metric]
haileyschoelkopf's avatar
haileyschoelkopf committed
1152
                        scores.append(result_score)
haileyschoelkopf's avatar
haileyschoelkopf committed
1153
                    if any(scores):
1154
                        result_score = 1.0
haileyschoelkopf's avatar
haileyschoelkopf committed
1155
                    else:
1156
                        result_score = 0.0
haileyschoelkopf's avatar
haileyschoelkopf committed
1157
                else:
1158
                    try:
1159
                        result_score = self._metric_fn_list[metric](
1160
1161
                            references=[gold],
                            predictions=[result],
1162
                            **self._metric_fn_kwargs[metric],
1163
                        )
baberabb's avatar
baberabb committed
1164
1165
1166
                    except (
                        TypeError
                    ):  # needed for now in order to use a different interface between our own metrics and HF Evaluate metrics
1167
                        result_score = self._metric_fn_list[metric]([gold, result])
1168
1169
1170
1171
                    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
1172
        else:
lintangsutawika's avatar
lintangsutawika committed
1173
1174
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
1175
                "'loglikelihood', 'loglikelihood_rolling', 'generate_until' or 'multiple_choice'",
1176
            )
1177
1178
1179
1180
1181
1182
1183

        return result_dict

    def aggregation(self):
        return self._aggregation_list

    def higher_is_better(self):
haileyschoelkopf's avatar
haileyschoelkopf committed
1184
        return self._higher_is_better
1185
1186
1187
1188
1189


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

baberabb's avatar
baberabb committed
1190
    def doc_to_target(self, doc: dict) -> str:
1191
1192
        return " " + doc["choices"][doc["gold"]]

baberabb's avatar
baberabb committed
1193
    def construct_requests(self, doc: dict, ctx: str, **kwargs) -> List[Instance]:
1194
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1195
1196
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1197
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1198
                doc=doc,
1199
                arguments=(ctx, " {}".format(choice)),
1200
                idx=i,
1201
1202
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1203
1204
            for i, choice in enumerate(doc["choices"])
        ]
1205

baberabb's avatar
baberabb committed
1206
    def process_results(self, doc: dict, results: List[Tuple[float, bool]]) -> dict:
lintangsutawika's avatar
lintangsutawika committed
1207
1208
1209
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
        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
1221
    def higher_is_better(self) -> dict:
1222
1223
1224
1225
1226
        return {
            "acc": True,
            "acc_norm": True,
        }

baberabb's avatar
baberabb committed
1227
    def aggregation(self) -> dict:
1228
1229
1230
1231
1232
1233
        return {
            "acc": mean,
            "acc_norm": mean,
        }


lintangsutawika's avatar
lintangsutawika committed
1234
class PerplexityTask(Task):
1235
1236
    OUTPUT_TYPE = "loglikelihood_rolling"

baberabb's avatar
baberabb committed
1237
    def has_training_docs(self) -> bool:
1238
1239
        return False

baberabb's avatar
baberabb committed
1240
    def fewshot_examples(self, k: int, rnd) -> List:
1241
1242
1243
        assert k == 0
        return []

baberabb's avatar
baberabb committed
1244
    def fewshot_context(self, doc: dict, num_fewshot: int) -> Literal[""]:
1245
1246
1247
1248
1249
1250
        assert (
            num_fewshot == 0
        ), "The number of fewshot examples must be 0 for perplexity tasks."

        return ""

baberabb's avatar
baberabb committed
1251
    def higher_is_better(self) -> dict:
1252
1253
1254
1255
1256
1257
1258
1259
1260
        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
1261
    def doc_to_text(self, doc) -> str:
1262
1263
1264
1265
1266
        return ""

    def doc_to_target(self, doc):
        return doc

baberabb's avatar
baberabb committed
1267
    def construct_requests(self, doc: dict, ctx: Union[str, None], **kwargs):
1268
1269
        assert not ctx

lintangsutawika's avatar
lintangsutawika committed
1270
1271
1272
1273
1274
1275
1276
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1277

baberabb's avatar
baberabb committed
1278
    def process_results(self, doc: dict, results: float) -> dict:
1279
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1280
1281
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1282
1283
1284
1285
1286
1287
        return {
            "word_perplexity": (loglikelihood, words),
            "byte_perplexity": (loglikelihood, bytes_),
            "bits_per_byte": (loglikelihood, bytes_),
        }

baberabb's avatar
baberabb committed
1288
    def aggregation(self) -> dict:
1289
1290
1291
1292
1293
1294
1295
        return {
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
            "bits_per_byte": bits_per_byte,
        }

    @classmethod
baberabb's avatar
baberabb committed
1296
    def count_bytes(cls, doc) -> int:
1297
1298
1299
        return len(doc.encode("utf-8"))

    @classmethod
baberabb's avatar
baberabb committed
1300
    def count_words(cls, doc) -> int:
1301
1302
        """Downstream tasks with custom word boundaries should override this!"""
        return len(re.split(r"\s+", doc))