task.py 60.7 KB
Newer Older
1
import abc
2
import ast
lintangsutawika's avatar
lintangsutawika committed
3
import logging
4
import random
5
6
import re
from collections.abc import Callable
7
from copy import deepcopy
8
from dataclasses import asdict, dataclass
9
from inspect import getsource
10
11
12
13
14
15
16
17
18
19
20
21
from typing import (
    Any,
    Dict,
    Iterable,
    Iterator,
    List,
    Literal,
    Mapping,
    Optional,
    Tuple,
    Union,
)
22
23
24

import datasets
import numpy as np
lintangsutawika's avatar
lintangsutawika committed
25
import shortuuid
26
from tqdm import tqdm
27
28

from lm_eval import utils
29
from lm_eval.api import samplers
30
31
from lm_eval.api.instance import Instance, OutputType
from lm_eval.api.metrics import bits_per_byte, mean, weighted_perplexity
lintangsutawika's avatar
lintangsutawika committed
32
from lm_eval.api.registry import (
33
34
    AGGREGATION_REGISTRY,
    DEFAULT_METRIC_REGISTRY,
haileyschoelkopf's avatar
haileyschoelkopf committed
35
    get_aggregation,
36
    get_metric,
37
    get_metric_aggregation,
haileyschoelkopf's avatar
haileyschoelkopf committed
38
    is_higher_better,
lintangsutawika's avatar
lintangsutawika committed
39
)
40
from lm_eval.caching.cache import load_from_cache, save_to_cache
41
42
43
from lm_eval.filters import build_filter_ensemble
from lm_eval.prompts import get_prompt

44

45
46
47
48
ALL_OUTPUT_TYPES = [
    "loglikelihood",
    "multiple_choice",
    "loglikelihood_rolling",
49
    "generate_until",
50
51
]

52
eval_logger = logging.getLogger("lm-eval")
53

lintangsutawika's avatar
lintangsutawika committed
54

lintangsutawika's avatar
lintangsutawika committed
55
56
@dataclass
class GroupConfig(dict):
lintangsutawika's avatar
lintangsutawika committed
57
58
59
    group: Optional[str] = None
    group_alias: Optional[str] = None
    task: Optional[Union[str, list]] = None
60
    tag_to_task: Optional[str] = False
lintangsutawika's avatar
lintangsutawika committed
61
62
63
    aggregate_metric: Optional[str] = False
    aggregate_fn: Optional[str] = "mean"
    weight_by_size: Optional[str] = False
lintangsutawika's avatar
lintangsutawika committed
64
    metric_alias: Optional[str] = None
65
    version: Optional[str] = 0
lintangsutawika's avatar
lintangsutawika committed
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85

    def __getitem__(self, item):
        return getattr(self, item)

    def __setitem__(self, item, value):
        return setattr(self, item, value)

    def to_dict(self, keep_callable: bool = False) -> dict:
        """dumps the current config as a dictionary object, as a printable format.
        null fields will not be printed.
        Used for dumping results alongside full task configuration

        :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()):
lintangsutawika's avatar
lintangsutawika committed
86
            if callable(v):
lintangsutawika's avatar
lintangsutawika committed
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
                cfg_dict[k] = self.serialize_function(v, keep_callable=keep_callable)
        return cfg_dict

    def serialize_function(
        self, value: Union[Callable, str], keep_callable=False
    ) -> Union[Callable, str]:
        """Serializes a given function or string.

        If 'keep_callable' is True, the original callable is returned.
        Otherwise, attempts to return the source code of the callable using 'getsource'.
        """
        if keep_callable:
            return value
        else:
            try:
                return getsource(value)
            except (TypeError, OSError):
                return str(value)


lintangsutawika's avatar
lintangsutawika committed
107
108
109
110
111
class ConfigurableGroup(abc.ABC):
    def __init__(
        self,
        config: Optional[dict] = None,
    ) -> None:
lintangsutawika's avatar
lintangsutawika committed
112
        # Create a unique identifier ID
lintangsutawika's avatar
lintangsutawika committed
113
        self._task_id = shortuuid.uuid()[:8]
lintangsutawika's avatar
lintangsutawika committed
114
115
116
117
118
        self._config = GroupConfig(**config)

    @property
    def group(self):
        return self._config.group
119

lintangsutawika's avatar
lintangsutawika committed
120
121
122
    @property
    def group_alias(self):
        return self._config.group_alias
123
124
125
126
127

    @property
    def version(self):
        return self._config.version

lintangsutawika's avatar
lintangsutawika committed
128
129
130
131
    @property
    def config(self):
        return self._config.to_dict()

lintangsutawika's avatar
lintangsutawika committed
132
133
    @property
    def task_id(self) -> Any:
lintangsutawika's avatar
lintangsutawika committed
134
135
136
137
138
        return "-".join((self.group_name, self._task_id))

    @property
    def group_name(self) -> Any:
        return self._config.group
lintangsutawika's avatar
lintangsutawika committed
139

lintangsutawika's avatar
lintangsutawika committed
140
141
    def __repr__(self):
        return (
142
            f"ConfigurableGroup(group={self.group}," f"group_alias={self.group_alias})"
lintangsutawika's avatar
lintangsutawika committed
143
144
        )

145

146
147
@dataclass
class TaskConfig(dict):
148
    # task naming/registry
149
150
    task: Optional[str] = None
    task_alias: Optional[str] = None
lintangsutawika's avatar
lintangsutawika committed
151
    tag: Optional[Union[str, list]] = None
152
153
    group: Optional[Union[str, list]] = None
    group_alias: Optional[Union[str, list]] = None
154
155
156
    # HF dataset options.
    # which dataset to use,
    # and what splits for what purpose
157
158
159
160
161
162
163
164
165
    dataset_path: Optional[str] = None
    dataset_name: Optional[str] = None
    dataset_kwargs: Optional[dict] = None
    training_split: Optional[str] = None
    validation_split: Optional[str] = None
    test_split: Optional[str] = None
    fewshot_split: Optional[
        str
    ] = None  # TODO: assert that this not None if num_fewshot > 0. (?) assert if this is same split as one evaling (?)
166
167
    # formatting / prompting options.
    # see docs/advanced_task_guide.md for more info
168
169
170
171
172
173
    process_docs: Optional[Callable] = None
    doc_to_text: Optional[Union[Callable, str]] = None
    doc_to_target: Optional[Union[Callable, str]] = None
    doc_to_choice: Optional[Union[Callable, str, dict, list]] = None
    process_results: Optional[Union[Callable, str]] = None
    use_prompt: Optional[str] = None
174
    description: str = ""
175
176
    target_delimiter: str = " "
    fewshot_delimiter: str = "\n\n"
177
    fewshot_config: Optional[dict] = None
178
    # runtime configuration options
179
    num_fewshot: Optional[int] = None
180
    # scoring options
181
182
183
    metric_list: Optional[list] = None
    output_type: OutputType = "generate_until"
    generation_kwargs: Optional[dict] = None
184
    repeats: int = 1
185
    filter_list: Optional[Union[str, list]] = None
186
    should_decontaminate: bool = False
187
188
189
190
    doc_to_decontamination_query: Optional[str] = None
    metadata: Optional[
        dict
    ] = None  # by default, not used in the code. allows for users to pass arbitrary info to tasks
191

Ethan Smith's avatar
Ethan Smith committed
192
    def __post_init__(self) -> None:
Lintang Sutawika's avatar
Lintang Sutawika committed
193
        if self.generation_kwargs is not None:
194
            if self.output_type != "generate_until":
195
                eval_logger.warning(
196
                    f"[{self.task}] passed `generation_kwargs`, but not using `output_type: generate_until`!"
Lintang Sutawika's avatar
Lintang Sutawika committed
197
198
199
200
201
202
203
204
                )

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

            if "until" not in self.generation_kwargs:
205
                self.generation_kwargs["until"] = [self.fewshot_delimiter]
Lintang Sutawika's avatar
Lintang Sutawika committed
206
        else:
207
            if self.output_type == "generate_until":
Lintang Sutawika's avatar
Lintang Sutawika committed
208
209
                # ensure that we greedily generate in absence of explicit arguments otherwise
                self.generation_kwargs = {
210
211
212
213
214
                    "until": (
                        None
                        if self.fewshot_delimiter is None
                        else [self.fewshot_delimiter]
                    ),
Lintang Sutawika's avatar
Lintang Sutawika committed
215
216
                    "do_sample": False,
                }
217

218
219
220
    def __getitem__(self, item):
        return getattr(self, item)

221
222
223
    def __setitem__(self, item, value):
        return setattr(self, item, value)

224
    def to_dict(self, keep_callable: bool = False) -> dict:
225
226
        """dumps the current config as a dictionary object, as a printable format.
        null fields will not be printed.
haileyschoelkopf's avatar
haileyschoelkopf committed
227
        Used for dumping results alongside full task configuration
228

haileyschoelkopf's avatar
haileyschoelkopf committed
229
230
231
232
233
234
235
236
237
238
        :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)
239
240
241
242
243
244
245
246
247
248
            elif k == "metric_list":
                for metric_dict in v:
                    for metric_key, metric_value in metric_dict.items():
                        if callable(metric_value):
                            metric_dict[metric_key] = self.serialize_function(
                                metric_value, keep_callable=keep_callable
                            )
                cfg_dict[k] = v
            elif callable(v):
                cfg_dict[k] = self.serialize_function(v, keep_callable=keep_callable)
haileyschoelkopf's avatar
haileyschoelkopf committed
249
        return cfg_dict
250

251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
    def serialize_function(
        self, value: Union[Callable, str], keep_callable=False
    ) -> Union[Callable, str]:
        """Serializes a given function or string.

        If 'keep_callable' is True, the original callable is returned.
        Otherwise, attempts to return the source code of the callable using 'getsource'.
        """
        if keep_callable:
            return value
        else:
            try:
                return getsource(value)
            except (TypeError, OSError):
                return str(value)

267
268
269
270
271
272
273
274
275
276
277

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)
    """

278
    VERSION: Optional[Union[int, str]] = None
279

280
281
    # The name of the `Task` benchmark as denoted in the HuggingFace datasets Hub
    # or a path to a custom `datasets` loading script.
282
    DATASET_PATH: Optional[str] = None
283
284

    # The name of a subset within `DATASET_PATH`.
285
    DATASET_NAME: Optional[str] = None
286

287
    OUTPUT_TYPE: Optional[OutputType] = None
lintangsutawika's avatar
lintangsutawika committed
288

289
290
    def __init__(
        self,
291
292
293
294
        data_dir: Optional[str] = None,
        cache_dir: Optional[str] = None,
        download_mode: Optional[datasets.DownloadMode] = None,
        config: Optional[Mapping] = None,  # Union[dict, TaskConfig]
Ethan Smith's avatar
Ethan Smith committed
295
    ) -> None:
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
        """
        :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)
318
319
320
        self._training_docs: Optional[list] = None
        self._fewshot_docs: Optional[list] = None
        self._instances: Optional[List[Instance]] = None
321

322
        # Create a unique identifier ID
lintangsutawika's avatar
lintangsutawika committed
323
        self._task_id = shortuuid.uuid()[:8]
324
        self._config: TaskConfig = TaskConfig({**config}) if config else TaskConfig()
325

lintangsutawika's avatar
lintangsutawika committed
326
        self._filters = [build_filter_ensemble("none", [["take_first", None]])]
327
328
329
        self.fewshot_rnd: Optional[
            random.Random
        ] = None  # purposely induce errors in case of improper usage
330

331
332
333
334
335
336
    def download(
        self,
        data_dir: Optional[str] = None,
        cache_dir: Optional[str] = None,
        download_mode=None,
    ) -> None:
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
        """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.
        """
361
362
363
364
365
366
367
        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,
        )
368

369
    @property
370
    def config(self) -> TaskConfig:
371
372
373
        """Returns the TaskConfig associated with this class."""
        return self._config

374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
    @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

389
    def training_docs(self) -> Iterable:
390
391
392
393
394
395
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

396
    def validation_docs(self) -> Iterable:
397
398
399
400
401
402
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

403
    def test_docs(self) -> Iterable:
404
405
406
407
408
409
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

410
    def fewshot_docs(self) -> Iterable:
411
412
413
414
415
416
417
418
419
        """
        :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
420
            eval_logger.warning(
421
                f"[Task: {self.config.task}] has_training_docs and has_validation_docs are False"
422
                ", using test_docs as fewshot_docs but this is not recommended."
lintangsutawika's avatar
lintangsutawika committed
423
            )
424
425
            return self.test_docs()

426
    def _process_doc(self, doc: dict) -> dict:
427
428
429
430
431
432
433
434
435
        """
        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
436

437
    @property
438
    def instances(self) -> List[Instance]:
439
440
441
442
443
444
445
446
447
448
449
        """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)

450
451
    def doc_to_decontamination_query(self, doc):
        raise NotImplementedError(
452
453
454
455
456
457
458
459
460
461
462
            "Override doc_to_decontamination_query with document specific decontamination query."
        )

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

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

463
464
    def build_all_requests(
        self,
465
        *,
466
467
468
469
470
471
        limit=None,
        rank=None,
        world_size=None,
        cache_requests=False,
        rewrite_requests_cache=False,
    ) -> None:
472
        """Build a set of Instances for a task, and store them in task.instances"""
473
474
475
476

        # used with caching
        og_limit = limit

477
        cache_key = f"requests-{self._config.task}-{self.config.num_fewshot}shot-rank{rank}-world_size{world_size}"
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492

        cached_instances = load_from_cache(file_name=cache_key)

        if cache_requests and cached_instances and not rewrite_requests_cache:
            cached_instances = cached_instances[:limit]

            flattened_instances = [
                instance
                for instance_group in cached_instances
                for instance in instance_group
            ]

            self._instances = flattened_instances
            return

Baber Abbasi's avatar
Baber Abbasi committed
493
        eval_logger.info(f"Building contexts for {self.config.task} on rank {rank}...")
494

495
        instances = []
496
497
498
499
500
501
502
503
504
505

        # process all documents when caching is specified for simplicity
        if (
            cache_requests
            and (not cached_instances or rewrite_requests_cache)
            and limit is not None
        ):
            limit = None

        doc_id_docs = list(
506
            self.doc_iterator(rank=rank, limit=limit, world_size=world_size)
507
508
509
510
511
512
513
        )

        num_docs = len(doc_id_docs)

        for doc_id, doc in tqdm(
            doc_id_docs,
            total=num_docs,
lintangsutawika's avatar
lintangsutawika committed
514
        ):
515
            # sample fewshot context #TODO: need to offset doc_id by rank now!
516
            fewshot_ctx = self.fewshot_context(
517
                doc,
518
                0 if self.config.num_fewshot is None else self.config.num_fewshot,
519
            )
520

521
            # TODO: we should override self.config.repeats if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
522
523
524
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
525
                metadata=(self.config["task"], doc_id, self.config.repeats),
lintangsutawika's avatar
lintangsutawika committed
526
            )
527
528
529
530

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

531
532
533
534
535
536
537
538
539
540
541
542
543
            instances.append(inst)

        # now flatten, this is to allow slicing to work with pickles

        sliced_instances = instances[:og_limit]

        flattened_instances = [
            instance
            for instance_group in sliced_instances
            for instance in instance_group
        ]

        self._instances = flattened_instances
544

545
546
        if len(self._instances) == 0:
            raise ValueError("task.build_requests() did not find any docs!")
547

548
549
550
        if cache_requests and (not cached_instances or rewrite_requests_cache):
            save_to_cache(file_name=cache_key, obj=instances)

551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
    @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
567
            The number of times each instance in a dataset is inferred on. Defaults to 1,
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
            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

603
604
605
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

haileyschoelkopf's avatar
haileyschoelkopf committed
606
607
608
609
610
611
612
613
614
615
    @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))

616
    @utils.positional_deprecated
lintangsutawika's avatar
lintangsutawika committed
617
    def fewshot_context(
618
619
620
        self,
        doc,
        num_fewshot,
621
        rnd=None,
622
        description=None,
lintangsutawika's avatar
lintangsutawika committed
623
    ):
624
625
626
627
628
629
630
        """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
631
632
633
634
635
        :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.
636
637
638
        :returns: str
            The fewshot context.
        """
639
        if rnd is None:
640
641
642
643
644
645
            if self.fewshot_rnd is not None:
                rnd = self.fewshot_rnd
            else:
                raise ValueError(
                    "A `random.Random` generator argument must be provided to `rnd`"
                )
lintangsutawika's avatar
lintangsutawika committed
646

647
        description = description if description else ""
648
649

        if num_fewshot == 0:
lintangsutawika's avatar
lintangsutawika committed
650
            labeled_examples = ""
651
        else:
lintangsutawika's avatar
lintangsutawika committed
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
            # 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
676
            )
677
678

        example = self.doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
679
        return description + labeled_examples + example
680

681
    def apply_filters(self) -> Optional[List[Instance]]:
Baber Abbasi's avatar
Baber Abbasi committed
682
        """Iterates over FilterEnsembles and applies them to instances"""
lintangsutawika's avatar
lintangsutawika committed
683
684
        if hasattr(self, "_filters"):
            for f in self._filters:
685
                f.apply(self._instances)
lintangsutawika's avatar
lintangsutawika committed
686
687
688
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances
689

baberabb's avatar
baberabb committed
690
    def dump_config(self) -> dict:
Baber Abbasi's avatar
Baber Abbasi committed
691
        """Returns the config as a dictionary."""
692
        # TODO: this should only return the overrides applied to a non-YAML task's configuration.
693
        # (num_fewshot)
694
        return self.config.to_dict()
695

Baber Abbasi's avatar
Baber Abbasi committed
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
    def set_config(self, key: str, value: Any, update: bool = False) -> None:
        """Set or update the configuration for a given key."""
        if key is None:
            raise ValueError("Key must be provided.")

        if update:
            current_value = getattr(self._config, key, {})
            if not isinstance(current_value, dict):
                raise TypeError(
                    f"Expected a dict for key '{key}', got {type(current_value).__name__} instead."
                )
            current_value.update(value)
        else:
            setattr(self._config, key, value)

    def override_metric(self, metric_name: str) -> None:
        """
        Override the default metrics used for evaluation with custom metrics.

        Parameters:
        - metric_name (str): The name of the custom metric to override. Should be registered in api.metrics.
        """
        (
            self._metric_fn_list,
            self._aggregation_list,
            self._metric_fn_kwargs,
            self._higher_is_better,
        ) = ({}, {}, {}, {})
        self._metric_fn_list[metric_name] = get_metric(metric_name)
        self._aggregation_list[metric_name] = get_metric_aggregation(metric_name)
        self._higher_is_better[metric_name] = is_higher_better(metric_name)
        self._metric_fn_kwargs[metric_name] = {}
        if not isinstance(self, ConfigurableTask):
            self.process_results = lambda x, y: {metric_name: get_metric(metric_name)}
            self.aggregation = lambda: {
                metric_name: get_metric_aggregation(metric_name)
            }
        setattr(self._config, "metric_list", [{"metric": metric_name}])
        setattr(self._config, "process_results", None)

736
737
738
739
740
    def set_fewshot_seed(self, seed: Optional[int] = None) -> None:
        self.fewshot_rnd = random.Random(seed)
        if hasattr(self, "sampler"):
            self.sampler.rnd = self.fewshot_rnd

741
742
743
744
745
746
747
    @property
    def eval_docs(self) -> Union[datasets.Dataset, List[dict]]:
        if self.has_test_docs():
            return self.test_docs()
        elif self.has_validation_docs():
            return self.validation_docs()
        else:
748
749
750
            raise ValueError(
                f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"
            )
751
752
753
754
755
756
757
758
759
760
761
762
763

    def doc_iterator(
        self, *, rank: int = 0, limit: Union[int, None] = None, world_size: int = 1
    ) -> Iterator[Tuple[int, Any]]:
        limit = int(limit) if limit else None
        doc_iterator = utils.create_iterator(
            enumerate(self.eval_docs),
            rank=int(rank),
            limit=limit,
            world_size=int(world_size),
        )
        return doc_iterator

764
765
766
    @property
    def task_id(self) -> Any:
        return self._task_id
767

768

769
class ConfigurableTask(Task):
770
    VERSION = "Yaml"
771
    OUTPUT_TYPE = None
772
    CONFIG = None
773
774

    def __init__(
775
776
777
778
779
        self,
        data_dir=None,
        cache_dir=None,
        download_mode=None,
        config: Optional[dict] = None,
Ethan Smith's avatar
Ethan Smith committed
780
    ) -> None:  # TODO no super() call here
lintangsutawika's avatar
lintangsutawika committed
781
        # Create a unique identifier ID
lintangsutawika's avatar
lintangsutawika committed
782
        self._task_id = shortuuid.uuid()[:8]
lintangsutawika's avatar
lintangsutawika committed
783

784
        # Get pre-configured attributes
785
        self._config = self.CONFIG
786

787
        # Use new configurations if there was no preconfiguration
788
        if self.config is None:
789
            self._config = TaskConfig(**config)
790
791
        # Overwrite configs
        else:
lintangsutawika's avatar
lintangsutawika committed
792
            if config is not None:
793
                self._config.__dict__.update(config)
794

795
        if self.config is None:
lintangsutawika's avatar
lintangsutawika committed
796
797
798
            raise ValueError(
                "Must pass a config to ConfigurableTask, either in cls.CONFIG or `config` kwarg"
            )
799

800
801
802
803
        if isinstance(self.config.metadata, dict):
            if "version" in self.config.metadata:
                self.VERSION = self.config.metadata["version"]

804
        if self.config.output_type is not None:
805
806
807
808
            if self.config.output_type not in ALL_OUTPUT_TYPES:
                raise ValueError(
                    f"Got invalid output_type '{self.config.output_type}', must be in '{','.join(ALL_OUTPUT_TYPES)}'"
                )
809
            self.OUTPUT_TYPE = self.config.output_type
810

811
812
        if self.config.dataset_path is not None:
            self.DATASET_PATH = self.config.dataset_path
813

814
815
        if self.config.dataset_name is not None:
            self.DATASET_NAME = self.config.dataset_name
816

817
818
819
820
        self._metric_fn_list = {}
        self._metric_fn_kwargs = {}
        self._aggregation_list = {}
        self._higher_is_better = {}
821

822
        if self.config.metric_list is None:
823
            # TODO: handle this in TaskConfig.__post_init__ ?
824
825
            _metric_list = DEFAULT_METRIC_REGISTRY[self.config.output_type]

826
            for metric_name in _metric_list:
haileyschoelkopf's avatar
haileyschoelkopf committed
827
                self._metric_fn_list[metric_name] = get_metric(metric_name)
lintangsutawika's avatar
lintangsutawika committed
828
                self._metric_fn_kwargs[metric_name] = {}
829
830
831
                self._aggregation_list[metric_name] = get_metric_aggregation(
                    metric_name
                )
haileyschoelkopf's avatar
haileyschoelkopf committed
832
                self._higher_is_better[metric_name] = is_higher_better(metric_name)
833
        else:
834
            for metric_config in self.config.metric_list:
835
836
837
838
                if "metric" not in metric_config:
                    raise ValueError(
                        "'metric' key not provided for an entry in 'metric_list', must be specified!"
                    )
839
840
841
842
                metric_name = metric_config["metric"]
                kwargs = {
                    key: metric_config[key]
                    for key in metric_config
Chris's avatar
Chris committed
843
844
                    if key
                    not in ["metric", "aggregation", "higher_is_better", "hf_evaluate"]
845
                }
Chris's avatar
Chris committed
846
847
848
849
                hf_evaluate_metric = (
                    "hf_evaluate" in metric_config
                    and metric_config["hf_evaluate"] is True
                )
850

851
                if self.config.process_results is not None:
852
853
                    self._metric_fn_list[metric_name] = None
                    self._metric_fn_kwargs[metric_name] = {}
854
855
856
857
858
859
                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
860
861
862
                    self._metric_fn_list[metric_name] = get_metric(
                        metric_name, hf_evaluate_metric
                    )
863
                    self._metric_fn_kwargs[metric_name] = kwargs
lintangsutawika's avatar
lintangsutawika committed
864

865
                if "aggregation" in metric_config:
866
                    agg_name = metric_config["aggregation"]
867
                    if isinstance(agg_name, str):
haileyschoelkopf's avatar
haileyschoelkopf committed
868
                        self._aggregation_list[metric_name] = get_aggregation(agg_name)
869
                    elif callable(agg_name):  # noqa: E721
870
871
872
                        self._aggregation_list[metric_name] = metric_config[
                            "aggregation"
                        ]
873
                else:
874
                    INV_AGG_REGISTRY = {v: k for k, v in AGGREGATION_REGISTRY.items()}
lintangsutawika's avatar
lintangsutawika committed
875
                    metric_agg = get_metric_aggregation(metric_name)
876
                    eval_logger.warning(
877
                        f"[Task: {self.config.task}] metric {metric_name} is defined, but aggregation is not. "
878
879
                        f"using default "
                        f"aggregation={INV_AGG_REGISTRY[metric_agg]}"
880
                    )
881
                    self._aggregation_list[metric_name] = metric_agg
lintangsutawika's avatar
lintangsutawika committed
882

883
884
885
886
887
888
                if "higher_is_better" in metric_config:
                    self._higher_is_better[metric_name] = metric_config[
                        "higher_is_better"
                    ]
                else:
                    eval_logger.warning(
889
                        f"[Task: {self.config.task}] metric {metric_name} is defined, but higher_is_better is not. "
890
                        f"using default "
haileyschoelkopf's avatar
haileyschoelkopf committed
891
                        f"higher_is_better={is_higher_better(metric_name)}"
892
                    )
haileyschoelkopf's avatar
haileyschoelkopf committed
893
                    self._higher_is_better[metric_name] = is_higher_better(metric_name)
894

895
        self.download(self.config.dataset_kwargs)
896
897
898
        self._training_docs = None
        self._fewshot_docs = None

899
        if self.config.filter_list is not None:
lintangsutawika's avatar
lintangsutawika committed
900
            self._filters = []
901
            for filter_config in self.config.filter_list:
902
903
904
905
906
907
908
909
910
                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"
                    }
                    components.append([function["function"], kwargs])
                filter_pipeline = build_filter_ensemble(filter_name, components)
lintangsutawika's avatar
lintangsutawika committed
911
                self._filters.append(filter_pipeline)
lintangsutawika's avatar
lintangsutawika committed
912
        else:
913
            self._filters = [build_filter_ensemble("none", [["take_first", None]])]
914

915
916
        if self.config.use_prompt is not None:
            eval_logger.info(f"loading prompt {self.config.use_prompt}")
917
            self.prompt = get_prompt(
918
                self.config.use_prompt, self.DATASET_PATH, self.DATASET_NAME
lintangsutawika's avatar
lintangsutawika committed
919
            )
920
921
922
        else:
            self.prompt = None

lintangsutawika's avatar
lintangsutawika committed
923
        if self.fewshot_docs() is not None:
924
925
926
927
            self.fewshot_rnd = (
                random.Random()
            )  # setting with no seed, to be overridden at a later time
            config_sampler: Union[str, Callable] = (
haileyschoelkopf's avatar
haileyschoelkopf committed
928
929
930
                self.config.fewshot_config.get("sampler", "default")
                if self.config.fewshot_config
                else "default"
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
            )
            if isinstance(config_sampler, str):
                self.sampler = samplers.get_sampler(config_sampler)(
                    list(self.fewshot_docs()), self, rnd=self.fewshot_rnd
                )
            elif callable(config_sampler) and issubclass(
                config_sampler, samplers.ContextSampler
            ):
                self.sampler = config_sampler(
                    docs=list(self.fewshot_docs()), task=self, rnd=self.fewshot_rnd
                )
            else:
                raise TypeError(
                    f"fewshot_config.sampler should be a string or callable of ContextSampler type, "
                    f"not {type(config_sampler)}"
                )
947

948
        self.task_docs = self.eval_docs
949

950
        # Test One Doc
951
        self.features = list(self.task_docs.features.keys())
952
953
        self.multiple_input = 0
        self.multiple_target = 0
954
        test_doc = self.task_docs[0]
955
        test_text = self.doc_to_text(test_doc)
956
        test_target = self.doc_to_target(test_doc)
lintangsutawika's avatar
lintangsutawika committed
957

958
        if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
959
            test_choice = self.doc_to_choice(test_doc)
960
            if not isinstance(test_choice, list):
lintangsutawika's avatar
lintangsutawika committed
961
                eval_logger.error("doc_to_choice must return list")
962
963
            else:
                num_choice = len(test_choice)
964

965
            if isinstance(test_text, int):
966
                self.multiple_input = num_choice
967
968
        else:
            test_choice = None
969

970
        if isinstance(test_target, list):
971
            self.multiple_target = len(test_target)
972
        else:
973
            if (isinstance(test_target, int)) and (test_choice is not None):
lintangsutawika's avatar
lintangsutawika committed
974
                test_target = test_choice[test_target]
975
            else:
lintangsutawika's avatar
lintangsutawika committed
976
                test_target = str(test_target)
977

978
979
980
        if test_choice is not None:
            check_choices = test_choice
        else:
lintangsutawika's avatar
lintangsutawika committed
981
            check_choices = [test_target]
982
983
984
985
        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 = (
986
987
                    True
                    if self.config.target_delimiter.rstrip()
988
                    != self.config.target_delimiter
989
                    else False
990
                )
991

992
                if delimiter_has_whitespace and choice_has_whitespace:
993
994
                    eval_logger.debug(
                        f'Both target_delimiter "{self.config.target_delimiter}" and target choice: "{choice}" have whitespace'
995
996
                    )
                elif (not delimiter_has_whitespace) and (not choice_has_whitespace):
997
                    eval_logger.debug(
998
                        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'
999
1000
                    )

1001
    def download(self, dataset_kwargs: Optional[Dict[str, Any]] = None) -> None:
1002
1003
1004
1005
1006
1007
        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
1008
    def has_training_docs(self) -> bool:
1009
        if self.config.training_split is not None:
1010
1011
1012
1013
            return True
        else:
            return False

baberabb's avatar
baberabb committed
1014
    def has_validation_docs(self) -> bool:
1015
        if self.config.validation_split is not None:
1016
1017
1018
1019
            return True
        else:
            return False

baberabb's avatar
baberabb committed
1020
    def has_test_docs(self) -> bool:
1021
        if self.config.test_split is not None:
1022
1023
1024
1025
            return True
        else:
            return False

baberabb's avatar
baberabb committed
1026
    def training_docs(self) -> datasets.Dataset:
1027
        if self.has_training_docs():
1028
1029
1030
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.training_split]
1031
                )
1032
            return self.dataset[self.config.training_split]
1033

baberabb's avatar
baberabb committed
1034
    def validation_docs(self) -> datasets.Dataset:
1035
        if self.has_validation_docs():
1036
1037
1038
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.validation_split]
1039
                )
1040
            return self.dataset[self.config.validation_split]
1041

baberabb's avatar
baberabb committed
1042
    def test_docs(self) -> datasets.Dataset:
1043
        if self.has_test_docs():
1044
1045
1046
            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]
1047

1048
    def fewshot_docs(self):
1049
        if self.config.fewshot_split is not None:
1050
1051
            if self.config.process_docs is not None:
                return self.config.process_docs(self.dataset[self.config.fewshot_split])
1052
            return self.dataset[self.config.fewshot_split]
1053
        else:
1054
            if (self.config.num_fewshot is not None) and (self.config.num_fewshot > 0):
1055
                eval_logger.warning(
lintangsutawika's avatar
lintangsutawika committed
1056
                    f"[Task: {self.config.task}] "
1057
1058
1059
1060
                    "num_fewshot > 0 but fewshot_split is None. "
                    "using preconfigured rule."
                )
            return super().fewshot_docs()
1061

lintangsutawika's avatar
lintangsutawika committed
1062
    @utils.positional_deprecated
1063
    def fewshot_context(self, doc: str, num_fewshot: int) -> str:
lintangsutawika's avatar
lintangsutawika committed
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
        """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.
        """
1074
1075
        if description := self.config.description:
            description = utils.apply_template(self.config.description, doc)
lintangsutawika's avatar
lintangsutawika committed
1076
1077
1078

        if num_fewshot == 0:
            # always prepend the (possibly empty) task description
1079
            labeled_examples = description
lintangsutawika's avatar
lintangsutawika committed
1080
        else:
1081
            labeled_examples = description + self.sampler.get_context(doc, num_fewshot)
lintangsutawika's avatar
lintangsutawika committed
1082
1083

        example = self.doc_to_text(doc)
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
        if self.multiple_input:
            return labeled_examples
        else:
            if isinstance(example, str):
                return labeled_examples + example
            elif isinstance(example, list):
                return [labeled_examples + ex for ex in example]
            elif isinstance(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)
lintangsutawika's avatar
lintangsutawika committed
1097

1098
    def apply_filters(self):
Baber Abbasi's avatar
Baber Abbasi committed
1099
        """Iterates over FilterEnsembles and applies them to instances"""
1100
1101
        if hasattr(self, "_filters"):
            for f in self._filters:
1102
                f.apply(self._instances)
1103
1104
1105
1106
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances

1107
    def should_decontaminate(self):
1108
        return self.config.should_decontaminate
1109
1110

    def doc_to_decontamination_query(self, doc):
1111
        if self.config.should_decontaminate:
1112
1113
            if self.config.doc_to_decontamination_query is None:
                return self.doc_to_text(doc)
1114
            else:
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
                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
                        )
                    )
1126

1127
    def _process_doc(self, doc: dict) -> dict:
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
        """
        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):
1139
1140
        if self.prompt is not None:
            doc_to_text = self.prompt
1141
        else:
1142
            doc_to_text = self.config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
1143

1144
        if isinstance(doc_to_text, int):
1145
            return doc_to_text
1146
        elif isinstance(doc_to_text, str):
1147
            if doc_to_text in self.features:
1148
                # if self.config.doc_to_choice is not None:
1149
1150
                #     return self.doc_to_choice(doc)[doc[doc_to_text]]
                # else:
1151
1152
                return doc[doc_to_text]
            else:
lintangsutawika's avatar
lintangsutawika committed
1153
                text_string = utils.apply_template(doc_to_text, doc)
lintangsutawika's avatar
lintangsutawika committed
1154
                if text_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1155
1156
1157
                    return ast.literal_eval(text_string)
                else:
                    return text_string
1158
        elif callable(doc_to_text):
1159
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
1160
        # Used when applying a Promptsource template
1161
        elif hasattr(doc_to_text, "apply"):
1162
1163
1164
1165
1166
            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")
1167
                return self.config.fewshot_delimiter
1168
        else:
1169
            print(type(doc_to_text))
1170
            raise TypeError
1171

1172
    def doc_to_target(self, doc: Mapping) -> Union[int, str, list]:
1173
1174
        if self.prompt is not None:
            doc_to_target = self.prompt
1175
        else:
1176
            doc_to_target = self.config.doc_to_target
1177

1178
        if isinstance(doc_to_target, int):
1179
            return doc_to_target
1180
        elif isinstance(doc_to_target, str):
1181
            if doc_to_target in self.features:
1182
                # if self.config.doc_to_choice is not None:
1183
1184
1185
                #     return self.doc_to_choice(doc)[doc[doc_to_target]]
                # else:
                return doc[doc_to_target]
1186
            else:
lintangsutawika's avatar
lintangsutawika committed
1187
                target_string = utils.apply_template(doc_to_target, doc)
lintangsutawika's avatar
lintangsutawika committed
1188
                if target_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1189
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
1190
1191
1192
1193
1194
                elif (
                    len(target_string) >= 2
                    and (target_string[0] == "[")
                    and (target_string[-1] == "]")
                ):
baberabb's avatar
baberabb committed
1195
1196
1197
1198
                    try:
                        return ast.literal_eval(target_string)
                    except (SyntaxError, ValueError):
                        return target_string
lintangsutawika's avatar
lintangsutawika committed
1199
1200
                else:
                    return target_string
1201
        elif isinstance(doc_to_target, list):
1202
            return doc_to_target
1203
        elif callable(doc_to_target):
1204
            return doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
1205
        # Used when applying a Promptsource template
1206
        elif hasattr(doc_to_target, "apply"):
1207
            applied_prompt = doc_to_target.apply(doc)
1208
1209
1210
1211
            if len(applied_prompt) == 2:
                return applied_prompt[1]
            else:
                eval_logger.warning("Applied prompt returns empty string")
1212
                return self.config.fewshot_delimiter
1213
1214
        else:
            raise TypeError
1215

baberabb's avatar
baberabb committed
1216
    def doc_to_choice(self, doc: Any) -> List[str]:
1217
1218
        if self.prompt is not None:
            doc_to_choice = self.prompt
1219
        elif self.config.doc_to_choice is None:
1220
1221
            eval_logger.error("doc_to_choice was called but not set in config")
        else:
1222
            doc_to_choice = self.config.doc_to_choice
1223

1224
        if isinstance(doc_to_choice, str):
1225
1226
1227
1228
            if doc_to_choice in self.features:
                return doc[doc_to_choice]
            else:
                return ast.literal_eval(utils.apply_template(doc_to_choice, doc))
1229
        elif isinstance(doc_to_choice, list):
1230
            return doc_to_choice
1231
        elif isinstance(doc_to_choice, dict):
1232
1233
1234
1235
1236
1237
1238
            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
1239

baberabb's avatar
baberabb committed
1240
1241
1242
    def construct_requests(
        self, doc: dict, ctx: str, **kwargs
    ) -> Union[List[Instance], Instance]:
1243
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
1244
            arguments = (ctx, self.doc_to_target(doc))
1245
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
1246
            arguments = (self.doc_to_target(doc),)
1247
        elif self.OUTPUT_TYPE == "multiple_choice":
1248
            choices = self.doc_to_choice(doc)
1249
            target_delimiter = self.config.target_delimiter
1250
1251
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
1252
                cont = self.doc_to_target(doc)
1253
1254
1255
                arguments = [
                    (ctx + choice, f"{target_delimiter}{cont}") for choice in choices
                ]
1256
            else:
1257
                # Otherwise they are placed in the continuation
1258
                arguments = [(ctx, f"{target_delimiter}{cont}") for cont in choices]
1259

1260
            request_list = [
1261
1262
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1263
                    doc=doc,
1264
                    arguments=arg,
1265
                    idx=i,
1266
1267
                    **kwargs,
                )
1268
                for i, arg in enumerate(arguments)
1269
            ]
1270
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
1271
            if "acc_mutual_info" in self._metric_fn_list.keys():
1272
1273
1274
                # if we are calculating multiple choice accuracy
                # using mutual information instead of raw loglikelihood as metric, need unconditional lls.

lintangsutawika's avatar
lintangsutawika committed
1275
                # here mutual info refers to calculating
1276
1277
1278
1279
1280
1281
                # 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
1282
                            doc=doc,
1283
                            arguments=("", "{}".format(choice)),
1284
1285
1286
                            idx=i,
                            **kwargs,
                        )
lintangsutawika's avatar
lintangsutawika committed
1287
                        for i, choice in enumerate(choices)
1288
1289
1290
                    ]
                )
            return request_list
lintangsutawika's avatar
lintangsutawika committed
1291

1292
        elif self.OUTPUT_TYPE == "generate_until":
1293
            arguments = (ctx, deepcopy(self.config.generation_kwargs))
lintangsutawika's avatar
lintangsutawika committed
1294
1295

        return Instance(
lintangsutawika's avatar
lintangsutawika committed
1296
1297
            request_type=self.OUTPUT_TYPE, doc=doc, arguments=arguments, idx=0, **kwargs
        )
1298
1299

    def process_results(self, doc, results):
1300
1301
        if callable(self.config.process_results):
            return self.config.process_results(doc, results)
lintangsutawika's avatar
lintangsutawika committed
1302

1303
        result_dict = {}
1304
        use_metric = list(self._metric_fn_list.keys())
1305
1306
1307
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
1308
1309
1310
1311
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
1312
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
1313
            (loglikelihood,) = results
1314
1315
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
haileyschoelkopf's avatar
haileyschoelkopf committed
1316
            return {
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
                **(
                    {"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
1332
            }
1333
        elif self.OUTPUT_TYPE == "multiple_choice":
1334
            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
1335

1336
            # retrieve choices in List[str] form, to compute choice lengths, etc.
1337
            choices = self.doc_to_choice(doc)
1338
1339
            completion_len = np.array([float(len(i)) for i in choices])

1340
1341
            if (
                2 * len(choices) == len(lls)
1342
                and "acc_mutual_info" in self._metric_fn_list.keys()
1343
1344
1345
1346
            ):
                # then we are doing mutual info.
                # this stores the "dryrun" / unconditional answer loglikelihoods
                lls_unconditional = lls[1::2]
1347
1348
                if len(lls_unconditional) != len(choices):
                    raise ValueError
1349
1350
                # and this stores our "regular" conditional loglikelihoods
                lls = lls[::2]
1351

1352
1353
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
1354

1355
1356
            if self.multiple_input:
                gold = self.doc_to_text(doc)
1357
            else:
1358
                gold = self.doc_to_target(doc)
1359
1360

            gold_index_error = False
1361
            if isinstance(gold, list):
Lintang Sutawika's avatar
Lintang Sutawika committed
1362
1363
                gold = [i if i < len(choices) else -100 for i in gold]
                if -100 in gold:
1364
1365
                    gold_index_error = True
            else:
1366
                if isinstance(gold, int):
Lintang Sutawika's avatar
Lintang Sutawika committed
1367
                    gold = gold if gold < len(choices) else -100
1368
                elif isinstance(gold, str):
Lintang Sutawika's avatar
Lintang Sutawika committed
1369
                    gold = choices.index(gold) if gold in choices else -100
lintangsutawika's avatar
lintangsutawika committed
1370

Lintang Sutawika's avatar
Lintang Sutawika committed
1371
                if gold == -100:
1372
1373
1374
1375
                    gold_index_error = True

            if gold_index_error:
                eval_logger.warning(
lintangsutawika's avatar
lintangsutawika committed
1376
                    f"Label index was not in within range of available choices,"
1377
1378
                    f"Sample:\n\n{doc}\n\n"
                )
lintangsutawika's avatar
lintangsutawika committed
1379

1380
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
1381
1382
                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
1383
                exact_match = int(any([is_greedy[i] if i != -100 else 0 for i in gold]))
lintangsutawika's avatar
lintangsutawika committed
1384
1385
1386
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
1387
                # 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
1388
                exact_match = int(is_greedy[gold]) if gold != -100 else 0
1389

Lintang Sutawika's avatar
Lintang Sutawika committed
1390
1391
1392
1393
            prob_norm = utils.softmax(lls)

            # TODO use keyword arguments to the metric?
            # gold, pred, norm stuff, the original lls,
1394
            result_dict = {
1395
                **({"acc": acc} if "acc" in use_metric else {}),
1396
1397
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
1398
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
1399
                **({"exact_match": exact_match} if "exact_match" in use_metric else {}),
Lintang Sutawika's avatar
Lintang Sutawika committed
1400
1401
1402
1403
1404
                **(
                    {"brier_score": (gold, prob_norm)}
                    if "brier_score" in use_metric
                    else {}
                ),
1405
1406
            }

1407
            if "acc_mutual_info" in use_metric:
lintangsutawika's avatar
lintangsutawika committed
1408
1409
1410
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
1411
1412
1413
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

1414
        elif self.OUTPUT_TYPE == "generate_until":
1415
            gold = self.doc_to_target(doc)
Chris's avatar
Chris committed
1416
            result = results[0]
1417
            if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1418
                # If you set doc_to_choice,
lintangsutawika's avatar
lintangsutawika committed
1419
                # it assumes that doc_to_target returns a number.
1420
1421
                choices = self.doc_to_choice(doc)
                gold = choices[gold]
1422
1423
            # we expect multiple_targets to be a list.
            elif self.multiple_target:
baberabb's avatar
baberabb committed
1424
                gold = list(gold)
Chris's avatar
Chris committed
1425
1426
1427
            elif type(gold) != type(result):
                # cast gold to the same type as result
                gold = type(result)(gold)
1428

lintangsutawika's avatar
lintangsutawika committed
1429
            for metric in self._metric_fn_list.keys():
haileyschoelkopf's avatar
haileyschoelkopf committed
1430
1431
1432
1433
1434
                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
1435
1436
1437
1438
                    if not isinstance(gold, list):
                        # sometimes, a multiple_target dataset has exceptions where one doc has only one string answer
                        # print(gold)
                        gold = [gold]
1439
1440
1441
1442
1443
1444
1445
1446
                    if metric == "exact_match":
                        result = [result for _ in range(len(gold))]
                        scores = self._metric_fn_list[metric](
                            references=gold,
                            predictions=result,
                            **self._metric_fn_kwargs[metric],
                        )[metric]
                        result_score = 1.0 if scores > 0.0 else 0.0
haileyschoelkopf's avatar
haileyschoelkopf committed
1447
                    else:
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
                        for gold_option in gold:
                            try:
                                result_score = self._metric_fn_list[metric](
                                    references=[gold_option],
                                    predictions=[result],
                                    **self._metric_fn_kwargs[metric],
                                )
                            except (
                                TypeError
                            ):  # TODO: this is hacky and I don't want to do it
                                result_score = self._metric_fn_list[metric](
                                    [gold_option, result]
                                )
                            if isinstance(result_score, dict):
                                # TODO: this handles the case where HF evaluate returns a dict.
                                result_score = result_score[metric]
                            scores.append(result_score)
                        if any(scores):
                            result_score = 1.0
                        else:
                            result_score = 0.0
haileyschoelkopf's avatar
haileyschoelkopf committed
1469
                else:
1470
                    try:
1471
                        result_score = self._metric_fn_list[metric](
1472
1473
                            references=[gold],
                            predictions=[result],
1474
                            **self._metric_fn_kwargs[metric],
1475
                        )
1476
                    except TypeError:  # needed for now in order to use a different interface between our own metrics and HF Evaluate metrics
1477
                        result_score = self._metric_fn_list[metric]([gold, result])
1478
1479
1480
1481
                    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
1482
        else:
lintangsutawika's avatar
lintangsutawika committed
1483
1484
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
1485
                "'loglikelihood', 'loglikelihood_rolling', 'generate_until' or 'multiple_choice'",
1486
            )
1487
1488
1489

        return result_dict

Baber Abbasi's avatar
Baber Abbasi committed
1490
    def aggregation(self) -> dict:
1491
1492
        return self._aggregation_list

Baber Abbasi's avatar
Baber Abbasi committed
1493
    def higher_is_better(self) -> dict:
haileyschoelkopf's avatar
haileyschoelkopf committed
1494
        return self._higher_is_better
1495

Baber Abbasi's avatar
Baber Abbasi committed
1496
1497
1498
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

lintangsutawika's avatar
lintangsutawika committed
1499
1500
    @property
    def task_id(self) -> Any:
lintangsutawika's avatar
lintangsutawika committed
1501
1502
1503
1504
1505
        return "-".join((self.task_name, self._task_id))

    @property
    def task_name(self) -> Any:
        return getattr(self.config, "task", None)
lintangsutawika's avatar
lintangsutawika committed
1506

1507
1508
1509
1510
1511
1512
1513
1514
    def __repr__(self):
        return (
            f"ConfigurableTask(task_name={getattr(self.config, 'task', None)},"
            f"output_type={self.OUTPUT_TYPE},"
            f"num_fewshot={getattr(self.config, 'num_fewshot', None)},"
            f"num_samples={len(self.eval_docs)})"
        )

1515
1516

class MultipleChoiceTask(Task):
1517
    OUTPUT_TYPE = "loglikelihood"
1518

baberabb's avatar
baberabb committed
1519
    def doc_to_target(self, doc: dict) -> str:
1520
1521
        return " " + doc["choices"][doc["gold"]]

baberabb's avatar
baberabb committed
1522
    def construct_requests(self, doc: dict, ctx: str, **kwargs) -> List[Instance]:
1523
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1524
1525
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1526
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1527
                doc=doc,
1528
                arguments=(ctx, " {}".format(choice)),
1529
                idx=i,
1530
1531
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1532
1533
            for i, choice in enumerate(doc["choices"])
        ]
1534

1535
    def process_results(self, doc: dict, results: Iterable[Tuple[float, bool]]) -> dict:
lintangsutawika's avatar
lintangsutawika committed
1536
1537
1538
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
        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
1550
    def higher_is_better(self) -> dict:
1551
1552
1553
1554
1555
        return {
            "acc": True,
            "acc_norm": True,
        }

baberabb's avatar
baberabb committed
1556
    def aggregation(self) -> dict:
1557
1558
1559
1560
1561
1562
        return {
            "acc": mean,
            "acc_norm": mean,
        }


lintangsutawika's avatar
lintangsutawika committed
1563
class PerplexityTask(Task):
1564
1565
    OUTPUT_TYPE = "loglikelihood_rolling"

baberabb's avatar
baberabb committed
1566
    def has_training_docs(self) -> bool:
1567
1568
        return False

baberabb's avatar
baberabb committed
1569
    def fewshot_examples(self, k: int, rnd) -> List:
1570
1571
1572
1573
        if k != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1574
1575
        return []

baberabb's avatar
baberabb committed
1576
    def fewshot_context(self, doc: dict, num_fewshot: int) -> Literal[""]:
1577
1578
1579
1580
        if num_fewshot != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1581
1582
1583

        return ""

baberabb's avatar
baberabb committed
1584
    def higher_is_better(self) -> dict:
1585
1586
1587
1588
1589
1590
1591
1592
1593
        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
1594
    def doc_to_text(self, doc) -> str:
1595
1596
1597
1598
1599
        return ""

    def doc_to_target(self, doc):
        return doc

1600
1601
1602
    def construct_requests(self, doc: dict, ctx: Optional[str], **kwargs):
        if bool(ctx):
            raise ValueError
1603

lintangsutawika's avatar
lintangsutawika committed
1604
1605
1606
1607
1608
1609
1610
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1611

1612
    def process_results(self, doc: dict, results: Tuple[float]) -> dict:
1613
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1614
1615
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1616
1617
1618
1619
1620
1621
        return {
            "word_perplexity": (loglikelihood, words),
            "byte_perplexity": (loglikelihood, bytes_),
            "bits_per_byte": (loglikelihood, bytes_),
        }

baberabb's avatar
baberabb committed
1622
    def aggregation(self) -> dict:
1623
1624
1625
1626
1627
1628
1629
        return {
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
            "bits_per_byte": bits_per_byte,
        }

    @classmethod
baberabb's avatar
baberabb committed
1630
    def count_bytes(cls, doc) -> int:
1631
1632
1633
        return len(doc.encode("utf-8"))

    @classmethod
baberabb's avatar
baberabb committed
1634
    def count_words(cls, doc) -> int:
1635
1636
        """Downstream tasks with custom word boundaries should override this!"""
        return len(re.split(r"\s+", doc))