task.py 48.8 KB
Newer Older
1
import abc
2
from dataclasses import dataclass, field, asdict
3

4
import os
5
import re
6
import ast
lintangsutawika's avatar
lintangsutawika committed
7
import yaml
lintangsutawika's avatar
lintangsutawika committed
8
import logging
9
10
11
import evaluate
import random
import itertools
12
import functools
13
from tqdm import tqdm
14
15
16
17

import datasets
import numpy as np

baberabb's avatar
baberabb committed
18
from typing import Union, List, Any, Tuple, Literal
19
from collections.abc import Callable
20

21
from lm_eval import utils
22
from lm_eval.api import samplers
haileyschoelkopf's avatar
haileyschoelkopf committed
23
from lm_eval.api.instance import Instance
lintangsutawika's avatar
lintangsutawika committed
24
from lm_eval.api.filter import FilterEnsemble
25
26
27

from lm_eval.prompts import get_prompt
from lm_eval.filters import build_filter_ensemble
lintangsutawika's avatar
lintangsutawika committed
28
29
30
31
32
33
from lm_eval.api.metrics import (
    mean,
    weighted_perplexity,
    bits_per_byte,
)
from lm_eval.api.registry import (
haileyschoelkopf's avatar
haileyschoelkopf committed
34
    get_metric,
35
36
37
    get_evaluate,
    get_aggregation,
    METRIC_REGISTRY,
38
    DEFAULT_METRIC_REGISTRY,
lintangsutawika's avatar
lintangsutawika committed
39
)
40

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

lintangsutawika's avatar
lintangsutawika committed
48

49
eval_logger = logging.getLogger("lm-eval")
50

lintangsutawika's avatar
lintangsutawika committed
51

52
53
@dataclass
class TaskConfig(dict):
54
    # task naming/registry
55
    task: str = None
lintangsutawika's avatar
lintangsutawika committed
56
    task_alias: str = None
57
    group: Union[str, list] = None
lintangsutawika's avatar
lintangsutawika committed
58
    group_alias: Union[str, list] = None
59
60
61
    # HF dataset options.
    # which dataset to use,
    # and what splits for what purpose
62
63
    dataset_path: str = None
    dataset_name: str = None
64
    dataset_kwargs: dict = None
65
66
67
    training_split: str = None
    validation_split: str = None
    test_split: str = None
lintangsutawika's avatar
lintangsutawika committed
68
    fewshot_split: str = None  # TODO: assert that this not None if num_fewshot > 0. (?) assert if this is same split as one evaling (?)
69
70
    # formatting / prompting options.
    # see docs/advanced_task_guide.md for more info
71
    process_docs: Callable = None
72
73
    doc_to_text: Union[Callable, str] = None
    doc_to_target: Union[Callable, str] = None
lintangsutawika's avatar
lintangsutawika committed
74
    doc_to_choice: Union[Callable, str, dict, list] = None
lintangsutawika's avatar
lintangsutawika committed
75
    process_results: Union[Callable, str] = None
76
    use_prompt: str = None
77
    description: str = ""
78
79
    target_delimiter: str = " "
    fewshot_delimiter: str = "\n\n"
haileyschoelkopf's avatar
haileyschoelkopf committed
80
    fewshot_config: dict = None
81
    # runtime configuration options
82
    num_fewshot: int = None
83
    # scoring options
84
    metric_list: list = None
85
    output_type: str = "generate_until"
86
    generation_kwargs: dict = None
87
    repeats: int = 1
lintangsutawika's avatar
lintangsutawika committed
88
    filter_list: Union[str, list] = None
89
90
    should_decontaminate: bool = False
    doc_to_decontamination_query: str = None
91

lintangsutawika's avatar
lintangsutawika committed
92
93
94
    metadata: Union[
        str, list
    ] = None  # by default, not used in the code. allows for users to pass arbitrary info to tasks
95

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

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

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

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

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

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

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

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

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

haileyschoelkopf's avatar
haileyschoelkopf committed
140
141
142
143
144
145
146
147
148
149
        :return: dict
            A printable dictionary version of the TaskConfig object.

        # TODO: should any default value in the TaskConfig not be printed?
        """
        cfg_dict = asdict(self)
        # remove values that are `None`
        for k, v in list(cfg_dict.items()):
            if v is None:
                cfg_dict.pop(k)
haileyschoelkopf's avatar
haileyschoelkopf committed
150
151
152
            elif isinstance(v, Callable):
                # TODO: this should handle Promptsource template objects as a separate case?
                cfg_dict[k] = str(v)
haileyschoelkopf's avatar
haileyschoelkopf committed
153
        return cfg_dict
154

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

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

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

    VERSION = None
167

168
169
170
171
172
173
174
175
    # The name of the `Task` benchmark as denoted in the HuggingFace datasets Hub
    # or a path to a custom `datasets` loading script.
    DATASET_PATH: str = None

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

    OUTPUT_TYPE: str = None
lintangsutawika's avatar
lintangsutawika committed
176

177
178
179
180
181
182
    def __init__(
        self,
        data_dir=None,
        cache_dir=None,
        download_mode=None,
        config=None,
Ethan Smith's avatar
Ethan Smith committed
183
    ) -> None:
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
        """
        :param data_dir: str
            Stores the path to a local folder containing the `Task`'s data files.
            Use this to specify the path to manually downloaded data (usually when
            the dataset is not publicly accessible).
        :param cache_dir: str
            The directory to read/write the `Task` dataset. This follows the
            HuggingFace `datasets` API with the default cache directory located at:
                `~/.cache/huggingface/datasets`
            NOTE: You can change the cache location globally for a given process
            to another directory:
                `export HF_DATASETS_CACHE="/path/to/another/directory"`
        :param download_mode: datasets.DownloadMode
            How to treat pre-existing `Task` downloads and data.
            - `datasets.DownloadMode.REUSE_DATASET_IF_EXISTS`
                Reuse download and reuse dataset.
            - `datasets.DownloadMode.REUSE_CACHE_IF_EXISTS`
                Reuse download with fresh dataset.
            - `datasets.DownloadMode.FORCE_REDOWNLOAD`
                Fresh download and fresh dataset.
        """
        self.download(data_dir, cache_dir, download_mode)
        self._training_docs = None
        self._fewshot_docs = None
        self._instances = None

lintangsutawika's avatar
lintangsutawika committed
210
        self._config = TaskConfig({**config}) if config else TaskConfig()
211

lintangsutawika's avatar
lintangsutawika committed
212
        self._filters = [build_filter_ensemble("none", [["take_first", None]])]
213

Ethan Smith's avatar
Ethan Smith committed
214
    def download(self, data_dir=None, cache_dir=None, download_mode=None) -> None:
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
        """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.
        """
239
240
241
242
243
244
245
        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,
        )
246

247
248
249
250
251
    @property
    def config(self):
        """Returns the TaskConfig associated with this class."""
        return self._config

252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
    @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 []

288
289
290
291
292
293
294
295
296
297
    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
298
            eval_logger.warning(
299
                "has_training_docs and has_validation_docs are False"
300
                ", using test_docs as fewshot_docs but this is not recommended."
lintangsutawika's avatar
lintangsutawika committed
301
            )
302
303
            return self.test_docs()

304
305
306
307
308
309
310
311
312
313
    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
314

315
316
317
318
319
320
321
322
323
324
325
326
327
    @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
328
    def doc_to_decontamination_query(self, doc) -> None:
329
330
331
332
333
334
335
336
337
338
339
340
341
        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
342
    def build_all_requests(self, limit=None, rank=None, world_size=None) -> None:
343
344
345
346
347
348
349
350
351
352
        """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!"

353
        eval_logger.info(f"Building contexts for task on rank {rank}...")
354

355
        instances = []
356
357
        for doc_id, doc in utils.create_iterator(
            enumerate(docs), rank, world_size, limit
lintangsutawika's avatar
lintangsutawika committed
358
        ):
359
            # sample fewshot context #TODO: need to offset doc_id by rank now!
360
            fewshot_ctx = self.fewshot_context(
361
                doc,
362
                0 if self.config.num_fewshot is None else self.config.num_fewshot,
363
            )
364

365
            # TODO: we should override self.config.repeats if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
366
367
368
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
369
                metadata=(self.config["task"], doc_id, self.config.repeats),
lintangsutawika's avatar
lintangsutawika committed
370
            )
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395

            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
396
            The number of times each instance in a dataset is inferred on. Defaults to 1,
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
            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
415
    def aggregation(self):
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
        """
        :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
432
433
434
435
436
437
438
439
440
441
    @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))

442
    @utils.positional_deprecated
lintangsutawika's avatar
lintangsutawika committed
443
    def fewshot_context(
444
445
446
447
448
        self,
        doc,
        num_fewshot,
        rnd=random.Random(1234),
        description=None,
lintangsutawika's avatar
lintangsutawika committed
449
    ):
450
451
452
453
454
455
456
        """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
457
458
459
460
461
        :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
        self._metric_fn_list = {}
        self._metric_fn_kwargs = {}
558
        self._aggregation_list = {}
559
        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:
566
                metric = get_metric(metric_name)
567
                self._metric_fn_list[metric_name] = metric["function"]
lintangsutawika's avatar
lintangsutawika committed
568
                self._metric_fn_kwargs[metric_name] = {}
569
570
                self._aggregation_list = metric["aggregation"]
                self._higher_is_better[metric_name] = metric["is_higher_better"]
571
        else:
572
            for metric_config in self.config.metric_list:
573
                assert "metric" in metric_config
574
                from_registry = False
575
576
577
578
                metric_name = metric_config["metric"]
                kwargs = {
                    key: metric_config[key]
                    for key in metric_config
Chris's avatar
Chris committed
579
                    if key
lintangsutawika's avatar
lintangsutawika committed
580
                    not in ["metric", "aggregation", "higher_is_better", "use_hf_evaluate"]
581
                }
582
                use_hf_evaluate = (
lintangsutawika's avatar
lintangsutawika committed
583
584
                    "use_hf_evaluate" in metric_config
                    and metric_config["use_hf_evaluate"] is True
Chris's avatar
Chris committed
585
                )
586

lintangsutawika's avatar
lintangsutawika committed
587
588
589
590
                if self.config.process_results is not None:
                    metric_fn = None
                    kwargs = {}
                elif callable(metric_name):
591
592
593
                    metric_fn = metric_name.__call__
                    metric_name = metric_name.__name__
                else:
594
                    assert type(metric_name) == str
lintangsutawika's avatar
lintangsutawika committed
595
                    use_metric_for_agg = True
596
597
598
599
600
                    if use_hf_evaluate:
                        metric_fn = get_evaluate(metric_name, **kwargs)
                    elif metric_name in METRIC_REGISTRY:
                        metric = get_metric(metric_name, **kwargs)
                        metric_fn = metric["function"]
601
602

                self._metric_fn_kwargs[metric_name] = kwargs
603
                self._metric_fn_list[metric_name] = metric_fn
lintangsutawika's avatar
lintangsutawika committed
604

605
606
607
608
609
610
611
                if "aggregation" in metric_config:
                    agg_name = metric_config["aggregation"]
                    if isinstance(agg_name, str):
                        self._aggregation_list[metric_name] = get_aggregation(agg_name)
                    elif callable(agg_name):  # noqa: E721
                        self._aggregation_list[metric_name] = agg_name
                else:
lintangsutawika's avatar
lintangsutawika committed
612
613
614
615
                    if use_hf_evaluate:
                        self._aggregation_list[metric_name] = metric_fn
                    elif (metric_name in METRIC_REGISTRY) and ("aggregation" in metric):
                        self._aggregation_list[metric_name] = metric["aggregation"]
616
617
618
619
620
621

                if "higher_is_better" in metric_config:
                    self._higher_is_better[metric_name] = metric_config[
                        "higher_is_better"
                    ]
                else:
lintangsutawika's avatar
lintangsutawika committed
622
                    self._higher_is_better[metric_name] = metric["higher_is_better"]
623

624
        self.download(self.config.dataset_kwargs)
625
626
627
        self._training_docs = None
        self._fewshot_docs = None

628
        if self.config.filter_list is not None:
lintangsutawika's avatar
lintangsutawika committed
629
            self._filters = []
630
            for filter_config in self.config.filter_list:
lintangsutawika's avatar
lintangsutawika committed
631
632
633
634
635
636
637
                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
638
639
640
                        }
                        components.append([function["function"], kwargs])
                    filter_pipeline = build_filter_ensemble(filter_name, components)
lintangsutawika's avatar
lintangsutawika committed
641
                self._filters.append(filter_pipeline)
lintangsutawika's avatar
lintangsutawika committed
642
        else:
643
            self._filters = [build_filter_ensemble("none", [["take_first", None]])]
644

645
646
        if self.config.use_prompt is not None:
            eval_logger.info(f"loading prompt {self.config.use_prompt}")
647
            self.prompt = get_prompt(
648
                self.config.use_prompt, self.DATASET_PATH, self.DATASET_NAME
lintangsutawika's avatar
lintangsutawika committed
649
            )
650
651
652
        else:
            self.prompt = None

lintangsutawika's avatar
lintangsutawika committed
653
        if self.fewshot_docs() is not None:
haileyschoelkopf's avatar
haileyschoelkopf committed
654
            self.sampler = samplers.get_sampler(
haileyschoelkopf's avatar
haileyschoelkopf committed
655
656
657
                self.config.fewshot_config.get("sampler", "default")
                if self.config.fewshot_config
                else "default"
haileyschoelkopf's avatar
haileyschoelkopf committed
658
            )(list(self.fewshot_docs()), self, rnd=random.Random(1234))
659

660
        if self.has_test_docs():
661
            self.task_docs = self.test_docs()
662
        elif self.has_validation_docs():
663
            self.task_docs = self.validation_docs()
664
665
666
667
668
        else:
            assert (
                False
            ), f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"

669
        # Test One Doc
670
        self.features = list(self.task_docs.features.keys())
671
672
        self.multiple_input = 0
        self.multiple_target = 0
673
        test_doc = self.task_docs[0]
674
        test_text = self.doc_to_text(test_doc)
675
        test_target = self.doc_to_target(test_doc)
lintangsutawika's avatar
lintangsutawika committed
676

677
        if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
678
679
680
            test_choice = self.doc_to_choice(test_doc)
            if type(test_choice) is not list:
                eval_logger.error("doc_to_choice must return list")
681
682
            else:
                num_choice = len(test_choice)
683

684
685
            if type(test_text) is int:
                self.multiple_input = num_choice
686
687
        else:
            test_choice = None
688

689
        if type(test_target) is list:
690
            self.multiple_target = len(test_target)
691
        else:
lintangsutawika's avatar
lintangsutawika committed
692
            if (type(test_target) is int) and (test_choice is not None):
lintangsutawika's avatar
lintangsutawika committed
693
                test_target = test_choice[test_target]
694
            else:
lintangsutawika's avatar
lintangsutawika committed
695
                test_target = str(test_target)
696

697
698
699
        if test_choice is not None:
            check_choices = test_choice
        else:
lintangsutawika's avatar
lintangsutawika committed
700
            check_choices = [test_target]
701
702
703
704
        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 = (
705
706
                    True
                    if self.config.target_delimiter.rstrip()
707
                    != self.config.target_delimiter
708
                    else False
709
                )
710

711
712
713
714
715
716
                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(
717
                        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'
718
719
                    )

Ethan Smith's avatar
Ethan Smith committed
720
    def download(self, dataset_kwargs=None) -> None:
721
722
723
724
725
726
        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
727
    def has_training_docs(self) -> bool:
728
        if self.config.training_split is not None:
729
730
731
732
            return True
        else:
            return False

baberabb's avatar
baberabb committed
733
    def has_validation_docs(self) -> bool:
734
        if self.config.validation_split is not None:
735
736
737
738
            return True
        else:
            return False

baberabb's avatar
baberabb committed
739
    def has_test_docs(self) -> bool:
740
        if self.config.test_split is not None:
741
742
743
744
            return True
        else:
            return False

baberabb's avatar
baberabb committed
745
    def training_docs(self) -> datasets.Dataset:
746
        if self.has_training_docs():
747
748
749
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.training_split]
750
                )
751
            return self.dataset[self.config.training_split]
752

baberabb's avatar
baberabb committed
753
    def validation_docs(self) -> datasets.Dataset:
754
        if self.has_validation_docs():
755
756
757
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.validation_split]
758
                )
759
            return self.dataset[self.config.validation_split]
760

baberabb's avatar
baberabb committed
761
    def test_docs(self) -> datasets.Dataset:
762
        if self.has_test_docs():
763
764
765
            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]
766

767
    def fewshot_docs(self):
768
769
        if self.config.fewshot_split is not None:
            return self.dataset[self.config.fewshot_split]
770
        else:
771
            if (self.config.num_fewshot is not None) and (self.config.num_fewshot > 0):
772
                eval_logger.warning(
773
                    f"Task '{self.config.task}': "
774
775
776
777
                    "num_fewshot > 0 but fewshot_split is None. "
                    "using preconfigured rule."
                )
            return super().fewshot_docs()
778

lintangsutawika's avatar
lintangsutawika committed
779
780
781
782
783
784
785
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
    @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)

812
813
814
815
816
817
818
819
    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

820
    def should_decontaminate(self):
821
        return self.config.should_decontaminate
822
823

    def doc_to_decontamination_query(self, doc):
824
        if self.config.should_decontaminate:
825
826
            if self.config.doc_to_decontamination_query is None:
                return self.doc_to_text(doc)
827
            else:
828
829
830
831
832
833
834
835
836
837
838
                doc_to_decontamination_query = self.config.doc_to_decontamination_query
                if doc_to_decontamination_query in self.features:
                    return doc[doc_to_decontamination_query]
                elif callable(doc_to_decontamination_query):
                    return doc_to_decontamination_query(doc)
                else:
                    return ast.literal_eval(
                        utils.apply_template(
                            self.config.doc_to_decontamination_query, doc
                        )
                    )
839

840
841
842
843
844
845
846
847
848
849
850
851
    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):
852
853
        if self.prompt is not None:
            doc_to_text = self.prompt
854
        else:
855
            doc_to_text = self.config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
856

857
858
859
        if type(doc_to_text) == int:
            return doc_to_text
        elif type(doc_to_text) == str:
860
            if doc_to_text in self.features:
861
                # if self.config.doc_to_choice is not None:
862
863
                #     return self.doc_to_choice(doc)[doc[doc_to_text]]
                # else:
864
865
                return doc[doc_to_text]
            else:
lintangsutawika's avatar
lintangsutawika committed
866
                text_string = utils.apply_template(doc_to_text, doc)
lintangsutawika's avatar
lintangsutawika committed
867
                if text_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
868
869
870
                    return ast.literal_eval(text_string)
                else:
                    return text_string
871
        elif callable(doc_to_text):
872
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
873
        # Used when applying a Promptsource template
874
        elif hasattr(doc_to_text, "apply"):
875
876
877
878
879
            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")
880
                return self.config.fewshot_delimiter
881
        else:
882
            print(type(doc_to_text))
883
            raise TypeError
884

885
    def doc_to_target(self, doc: dict) -> Union[int, str, list]:
886
887
        if self.prompt is not None:
            doc_to_target = self.prompt
888
        else:
889
            doc_to_target = self.config.doc_to_target
890

891
892
893
        if type(doc_to_target) == int:
            return doc_to_target
        elif type(doc_to_target) == str:
894
            if doc_to_target in self.features:
895
                # if self.config.doc_to_choice is not None:
896
897
898
                #     return self.doc_to_choice(doc)[doc[doc_to_target]]
                # else:
                return doc[doc_to_target]
899
            else:
lintangsutawika's avatar
lintangsutawika committed
900
                target_string = utils.apply_template(doc_to_target, doc)
lintangsutawika's avatar
lintangsutawika committed
901
                if target_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
902
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
903
904
905
906
907
                elif (
                    len(target_string) >= 2
                    and (target_string[0] == "[")
                    and (target_string[-1] == "]")
                ):
baberabb's avatar
baberabb committed
908
909
910
911
                    try:
                        return ast.literal_eval(target_string)
                    except (SyntaxError, ValueError):
                        return target_string
lintangsutawika's avatar
lintangsutawika committed
912
913
                else:
                    return target_string
914
915
        elif type(doc_to_target) == list:
            return doc_to_target
916
        elif callable(doc_to_target):
917
            return doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
918
        # Used when applying a Promptsource template
919
        elif hasattr(doc_to_target, "apply"):
920
            applied_prompt = doc_to_target.apply(doc)
921
922
923
924
            if len(applied_prompt) == 2:
                return applied_prompt[1]
            else:
                eval_logger.warning("Applied prompt returns empty string")
925
                return self.config.fewshot_delimiter
926
927
        else:
            raise TypeError
928

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

        if type(doc_to_choice) == str:
938
939
940
941
            if doc_to_choice in self.features:
                return doc[doc_to_choice]
            else:
                return ast.literal_eval(utils.apply_template(doc_to_choice, doc))
942
943
944
945
946
947
948
949
950
951
        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
952

baberabb's avatar
baberabb committed
953
954
955
    def construct_requests(
        self, doc: dict, ctx: str, **kwargs
    ) -> Union[List[Instance], Instance]:
956
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
957
            arguments = (ctx, self.doc_to_target(doc))
958
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
959
            arguments = (self.doc_to_target(doc),)
960
        elif self.OUTPUT_TYPE == "multiple_choice":
961
            choices = self.doc_to_choice(doc)
962
            target_delimiter = self.config.target_delimiter
963
964
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
965
                cont = self.doc_to_target(doc)
966
                arguments = [(ctx, f"{target_delimiter}{cont}") for ctx in choices]
967
            else:
968
                # Otherwise they are placed in the continuation
969
                arguments = [(ctx, f"{target_delimiter}{cont}") for cont in choices]
970

971
            request_list = [
972
973
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
974
                    doc=doc,
975
                    arguments=arg,
976
                    idx=i,
977
978
                    **kwargs,
                )
979
                for i, arg in enumerate(arguments)
980
            ]
981
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
982
            if "acc_mutual_info" in self._metric_fn_list.keys():
983
984
985
                # if we are calculating multiple choice accuracy
                # using mutual information instead of raw loglikelihood as metric, need unconditional lls.

lintangsutawika's avatar
lintangsutawika committed
986
                # here mutual info refers to calculating
987
988
989
990
991
992
                # 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
993
                            doc=doc,
994
                            arguments=("", "{}".format(choice)),
995
996
997
                            idx=i,
                            **kwargs,
                        )
lintangsutawika's avatar
lintangsutawika committed
998
                        for i, choice in enumerate(choices)
999
1000
1001
                    ]
                )
            return request_list
lintangsutawika's avatar
lintangsutawika committed
1002

1003
        elif self.OUTPUT_TYPE == "generate_until":
1004
            arguments = (ctx, self.config.generation_kwargs)
lintangsutawika's avatar
lintangsutawika committed
1005
1006

        return Instance(
lintangsutawika's avatar
lintangsutawika committed
1007
1008
            request_type=self.OUTPUT_TYPE, doc=doc, arguments=arguments, idx=0, **kwargs
        )
1009
1010

    def process_results(self, doc, results):
1011
1012
1013
1014

        # Process results returns 1 of X things per doc/results
        # 1. A score
        # 2. Components to be processed later to obtained a score. such as gold and prediction
1015
1016
        if callable(self.config.process_results):
            return self.config.process_results(doc, results)
lintangsutawika's avatar
lintangsutawika committed
1017

1018
        result_dict = {}
1019
        use_metric = list(self._metric_fn_list.keys())
1020
1021
1022
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
1023
1024
1025
1026
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
1027
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
1028
            (loglikelihood,) = results
1029
1030
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
            return {
                **(
                    {"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 {}
                ),
            }
1048
        elif self.OUTPUT_TYPE == "multiple_choice":
1049
            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
1050

1051
            # retrieve choices in List[str] form, to compute choice lengths, etc.
1052
            choices = self.doc_to_choice(doc)
1053
1054
            completion_len = np.array([float(len(i)) for i in choices])

1055
1056
            if (
                2 * len(choices) == len(lls)
1057
                and "acc_mutual_info" in self._metric_fn_list.keys()
1058
1059
1060
1061
1062
1063
1064
            ):
                # 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]
1065

1066
1067
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
1068

1069
1070
            if self.multiple_input:
                gold = self.doc_to_text(doc)
1071
            else:
1072
                gold = self.doc_to_target(doc)
1073
1074

            gold_index_error = False
1075
            if isinstance(gold, list):
Lintang Sutawika's avatar
Lintang Sutawika committed
1076
1077
                gold = [i if i < len(choices) else -100 for i in gold]
                if -100 in gold:
1078
1079
                    gold_index_error = True
            else:
1080
                if isinstance(gold, int):
Lintang Sutawika's avatar
Lintang Sutawika committed
1081
                    gold = gold if gold < len(choices) else -100
1082
                elif isinstance(gold, str):
Lintang Sutawika's avatar
Lintang Sutawika committed
1083
                    gold = choices.index(gold) if gold in choices else -100
lintangsutawika's avatar
lintangsutawika committed
1084

Lintang Sutawika's avatar
Lintang Sutawika committed
1085
                if gold == -100:
1086
1087
1088
1089
                    gold_index_error = True

            if gold_index_error:
                eval_logger.warning(
lintangsutawika's avatar
lintangsutawika committed
1090
                    f"Label index was not in within range of available choices,"
1091
1092
                    f"Sample:\n\n{doc}\n\n"
                )
lintangsutawika's avatar
lintangsutawika committed
1093

1094
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
1095
1096
                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
1097
                exact_match = int(any([is_greedy[i] if i != -100 else 0 for i in gold]))
lintangsutawika's avatar
lintangsutawika committed
1098
1099
1100
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
1101
                # 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
1102
                exact_match = int(is_greedy[gold]) if gold != -100 else 0
1103

1104
            # gold, lls, is_greedy, completion_len
1105
            result_dict = {
1106
1107
                **({"acc": acc} if "acc" in use_metric else {}),
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
1108
                **({"exact_match": exact_match} if "exact_match" in use_metric else {}),
1109
1110
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
1111
1112
            }

1113
            if "acc_mutual_info" in use_metric:
lintangsutawika's avatar
lintangsutawika committed
1114
1115
1116
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
1117
1118
1119
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

1120
        elif self.OUTPUT_TYPE == "generate_until":
1121
            gold = self.doc_to_target(doc)
Chris's avatar
Chris committed
1122
            result = results[0]
1123
            if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1124
                # If you set doc_to_choice,
lintangsutawika's avatar
lintangsutawika committed
1125
                # it assumes that doc_to_target returns a number.
1126
1127
                choices = self.doc_to_choice(doc)
                gold = choices[gold]
1128
1129
            # we expect multiple_targets to be a list.
            elif self.multiple_target:
baberabb's avatar
baberabb committed
1130
                gold = list(gold)
Chris's avatar
Chris committed
1131
1132
1133
            elif type(gold) != type(result):
                # cast gold to the same type as result
                gold = type(result)(gold)
1134

lintangsutawika's avatar
lintangsutawika committed
1135
            for metric in self._metric_fn_list.keys():
1136
1137
                result_dict[metric] = (gold, result)
                continue
haileyschoelkopf's avatar
haileyschoelkopf committed
1138
1139
1140
1141
1142
                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
1143
1144
1145
1146
                    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
1147
                    for gold_option in gold:
1148
                        try:
1149
                            result_score = self._metric_fn_list[metric](
1150
1151
                                references=[gold_option],
                                predictions=[result],
1152
                                **self._metric_fn_kwargs[metric],
1153
                            )
baberabb's avatar
baberabb committed
1154
1155
1156
                        except (
                            TypeError
                        ):  # TODO: this is hacky and I don't want to do it
1157
                            result_score = self._metric_fn_list[metric](
haileyschoelkopf's avatar
haileyschoelkopf committed
1158
1159
1160
                                [gold_option, result]
                            )
                        if isinstance(result_score, dict):
haileyschoelkopf's avatar
haileyschoelkopf committed
1161
                            # TODO: this handles the case where HF evaluate returns a dict.
1162
                            result_score = result_score[metric]
haileyschoelkopf's avatar
haileyschoelkopf committed
1163
                        scores.append(result_score)
haileyschoelkopf's avatar
haileyschoelkopf committed
1164
                    if any(scores):
1165
                        result_score = 1.0
haileyschoelkopf's avatar
haileyschoelkopf committed
1166
                    else:
1167
                        result_score = 0.0
haileyschoelkopf's avatar
haileyschoelkopf committed
1168
                else:
1169
                    try:
1170
                        result_score = self._metric_fn_list[metric](
1171
1172
1173
                            references=[gold],
                            predictions=[result],
                        )
1174
                    except TypeError:  # needed for now in order to use a different interface between our own metrics and HF Evaluate metrics
1175
                        result_score = self._metric_fn_list[metric]([gold, result])
1176
1177
1178
1179
                    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
1180
        else:
lintangsutawika's avatar
lintangsutawika committed
1181
1182
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
1183
                "'loglikelihood', 'loglikelihood_rolling', 'generate_until' or 'multiple_choice'",
1184
            )
1185
1186
1187

        return result_dict

1188
1189
    def aggregation(self):
        return self._aggregation_list
1190
1191

    def higher_is_better(self):
haileyschoelkopf's avatar
haileyschoelkopf committed
1192
        return self._higher_is_better
1193
1194
1195
1196
1197


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

baberabb's avatar
baberabb committed
1198
    def doc_to_target(self, doc: dict) -> str:
1199
1200
        return " " + doc["choices"][doc["gold"]]

baberabb's avatar
baberabb committed
1201
    def construct_requests(self, doc: dict, ctx: str, **kwargs) -> List[Instance]:
1202
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1203
1204
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1205
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1206
                doc=doc,
1207
                arguments=(ctx, " {}".format(choice)),
1208
                idx=i,
1209
1210
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1211
1212
            for i, choice in enumerate(doc["choices"])
        ]
1213

baberabb's avatar
baberabb committed
1214
    def process_results(self, doc: dict, results: List[Tuple[float, bool]]) -> dict:
lintangsutawika's avatar
lintangsutawika committed
1215
1216
1217
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
        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
1229
    def higher_is_better(self) -> dict:
1230
1231
1232
1233
1234
        return {
            "acc": True,
            "acc_norm": True,
        }

1235
    def aggregation(self) -> dict:
1236
1237
1238
1239
1240
1241
        return {
            "acc": mean,
            "acc_norm": mean,
        }


lintangsutawika's avatar
lintangsutawika committed
1242
class PerplexityTask(Task):
1243
1244
    OUTPUT_TYPE = "loglikelihood_rolling"

baberabb's avatar
baberabb committed
1245
    def has_training_docs(self) -> bool:
1246
1247
        return False

baberabb's avatar
baberabb committed
1248
    def fewshot_examples(self, k: int, rnd) -> List:
1249
1250
1251
        assert k == 0
        return []

baberabb's avatar
baberabb committed
1252
    def fewshot_context(self, doc: dict, num_fewshot: int) -> Literal[""]:
1253
1254
1255
1256
1257
1258
        assert (
            num_fewshot == 0
        ), "The number of fewshot examples must be 0 for perplexity tasks."

        return ""

baberabb's avatar
baberabb committed
1259
    def higher_is_better(self) -> dict:
1260
1261
1262
1263
1264
1265
1266
1267
1268
        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
1269
    def doc_to_text(self, doc) -> str:
1270
1271
1272
1273
1274
        return ""

    def doc_to_target(self, doc):
        return doc

baberabb's avatar
baberabb committed
1275
    def construct_requests(self, doc: dict, ctx: Union[str, None], **kwargs):
1276
1277
        assert not ctx

lintangsutawika's avatar
lintangsutawika committed
1278
1279
1280
1281
1282
1283
1284
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1285

baberabb's avatar
baberabb committed
1286
    def process_results(self, doc: dict, results: float) -> dict:
1287
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1288
1289
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1290
1291
1292
1293
1294
1295
        return {
            "word_perplexity": (loglikelihood, words),
            "byte_perplexity": (loglikelihood, bytes_),
            "bits_per_byte": (loglikelihood, bytes_),
        }

1296
    def aggregation(self) -> dict:
1297
1298
1299
1300
1301
1302
1303
        return {
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
            "bits_per_byte": bits_per_byte,
        }

    @classmethod
baberabb's avatar
baberabb committed
1304
    def count_bytes(cls, doc) -> int:
1305
1306
1307
        return len(doc.encode("utf-8"))

    @classmethod
baberabb's avatar
baberabb committed
1308
    def count_words(cls, doc) -> int:
1309
1310
        """Downstream tasks with custom word boundaries should override this!"""
        return len(re.split(r"\s+", doc))