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
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108

    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()):
            if v is None:
                cfg_dict.pop(k)
            elif callable(v):
                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
109
110
111
112
113
class ConfigurableGroup(abc.ABC):
    def __init__(
        self,
        config: Optional[dict] = None,
    ) -> None:
lintangsutawika's avatar
lintangsutawika committed
114
        # Create a unique identifier ID
lintangsutawika's avatar
lintangsutawika committed
115
        self._task_id = shortuuid.uuid()[:8]
lintangsutawika's avatar
lintangsutawika committed
116
117
118
119
120
        self._config = GroupConfig(**config)

    @property
    def group(self):
        return self._config.group
121

lintangsutawika's avatar
lintangsutawika committed
122
123
124
    @property
    def group_alias(self):
        return self._config.group_alias
125
126
127
128
129

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

lintangsutawika's avatar
lintangsutawika committed
130
131
132
133
    @property
    def config(self):
        return self._config.to_dict()

lintangsutawika's avatar
lintangsutawika committed
134
135
    @property
    def task_id(self) -> Any:
lintangsutawika's avatar
lintangsutawika committed
136
137
138
139
140
        return "-".join((self.group_name, self._task_id))

    @property
    def group_name(self) -> Any:
        return self._config.group
lintangsutawika's avatar
lintangsutawika committed
141

lintangsutawika's avatar
lintangsutawika committed
142
143
    def __repr__(self):
        return (
144
            f"ConfigurableGroup(group={self.group}," f"group_alias={self.group_alias})"
lintangsutawika's avatar
lintangsutawika committed
145
146
        )

147

148
149
@dataclass
class TaskConfig(dict):
150
    # task naming/registry
151
152
    task: Optional[str] = None
    task_alias: Optional[str] = None
lintangsutawika's avatar
lintangsutawika committed
153
    tag: Optional[Union[str, list]] = None
154
155
    group: Optional[Union[str, list]] = None
    group_alias: Optional[Union[str, list]] = None
156
157
158
    # HF dataset options.
    # which dataset to use,
    # and what splits for what purpose
159
160
161
162
163
164
165
166
167
    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 (?)
168
169
    # formatting / prompting options.
    # see docs/advanced_task_guide.md for more info
170
171
172
173
174
175
    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
176
    description: str = ""
177
178
    target_delimiter: str = " "
    fewshot_delimiter: str = "\n\n"
179
    fewshot_config: Optional[dict] = None
180
    # runtime configuration options
181
    num_fewshot: Optional[int] = None
182
    # scoring options
183
184
185
    metric_list: Optional[list] = None
    output_type: OutputType = "generate_until"
    generation_kwargs: Optional[dict] = None
186
    repeats: int = 1
187
    filter_list: Optional[Union[str, list]] = None
188
    should_decontaminate: bool = False
189
190
191
192
    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
193

Ethan Smith's avatar
Ethan Smith committed
194
    def __post_init__(self) -> None:
Lintang Sutawika's avatar
Lintang Sutawika committed
195
        if self.generation_kwargs is not None:
196
            if self.output_type != "generate_until":
197
                eval_logger.warning(
198
                    f"[{self.task}] passed `generation_kwargs`, but not using `output_type: generate_until`!"
Lintang Sutawika's avatar
Lintang Sutawika committed
199
200
201
202
203
204
205
206
                )

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

            if "until" not in self.generation_kwargs:
207
                self.generation_kwargs["until"] = [self.fewshot_delimiter]
Lintang Sutawika's avatar
Lintang Sutawika committed
208
        else:
209
            if self.output_type == "generate_until":
Lintang Sutawika's avatar
Lintang Sutawika committed
210
211
                # ensure that we greedily generate in absence of explicit arguments otherwise
                self.generation_kwargs = {
212
213
214
215
216
                    "until": (
                        None
                        if self.fewshot_delimiter is None
                        else [self.fewshot_delimiter]
                    ),
Lintang Sutawika's avatar
Lintang Sutawika committed
217
218
                    "do_sample": False,
                }
219

220
221
222
    def __getitem__(self, item):
        return getattr(self, item)

223
224
225
    def __setitem__(self, item, value):
        return setattr(self, item, value)

226
    def to_dict(self, keep_callable: bool = False) -> dict:
227
228
        """dumps the current config as a dictionary object, as a printable format.
        null fields will not be printed.
haileyschoelkopf's avatar
haileyschoelkopf committed
229
        Used for dumping results alongside full task configuration
230

haileyschoelkopf's avatar
haileyschoelkopf committed
231
232
233
234
235
236
237
238
239
240
        :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)
241
242
243
244
245
246
247
248
249
250
            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
251
        return cfg_dict
252

253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
    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)

269
270
271
272
273
274
275
276
277
278
279

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)
    """

280
    VERSION: Optional[Union[int, str]] = None
281

282
283
    # The name of the `Task` benchmark as denoted in the HuggingFace datasets Hub
    # or a path to a custom `datasets` loading script.
284
    DATASET_PATH: Optional[str] = None
285
286

    # The name of a subset within `DATASET_PATH`.
287
    DATASET_NAME: Optional[str] = None
288

289
    OUTPUT_TYPE: Optional[OutputType] = None
lintangsutawika's avatar
lintangsutawika committed
290

291
292
    def __init__(
        self,
293
294
295
296
        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
297
    ) -> None:
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
        """
        :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)
320
321
322
        self._training_docs: Optional[list] = None
        self._fewshot_docs: Optional[list] = None
        self._instances: Optional[List[Instance]] = None
323

324
        # Create a unique identifier ID
lintangsutawika's avatar
lintangsutawika committed
325
        self._task_id = shortuuid.uuid()[:8]
326
        self._config: TaskConfig = TaskConfig({**config}) if config else TaskConfig()
327

lintangsutawika's avatar
lintangsutawika committed
328
        self._filters = [build_filter_ensemble("none", [["take_first", None]])]
329
330
331
        self.fewshot_rnd: Optional[
            random.Random
        ] = None  # purposely induce errors in case of improper usage
332

333
334
335
336
337
338
    def download(
        self,
        data_dir: Optional[str] = None,
        cache_dir: Optional[str] = None,
        download_mode=None,
    ) -> None:
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
        """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.
        """
363
364
365
366
367
368
369
        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,
        )
370

371
    @property
372
    def config(self) -> TaskConfig:
373
374
375
        """Returns the TaskConfig associated with this class."""
        return self._config

376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
    @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

391
    def training_docs(self) -> Iterable:
392
393
394
395
396
397
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

398
    def validation_docs(self) -> Iterable:
399
400
401
402
403
404
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

405
    def test_docs(self) -> Iterable:
406
407
408
409
410
411
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

412
    def fewshot_docs(self) -> Iterable:
413
414
415
416
417
418
419
420
421
        """
        :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
422
            eval_logger.warning(
423
                f"[Task: {self.config.task}] has_training_docs and has_validation_docs are False"
424
                ", using test_docs as fewshot_docs but this is not recommended."
lintangsutawika's avatar
lintangsutawika committed
425
            )
426
427
            return self.test_docs()

428
    def _process_doc(self, doc: dict) -> dict:
429
430
431
432
433
434
435
436
437
        """
        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
438

439
    @property
440
    def instances(self) -> List[Instance]:
441
442
443
444
445
446
447
448
449
450
451
        """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)

452
453
    def doc_to_decontamination_query(self, doc):
        raise NotImplementedError(
454
455
456
457
458
459
460
461
462
463
464
            "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

465
466
    def build_all_requests(
        self,
467
        *,
468
469
470
471
472
473
        limit=None,
        rank=None,
        world_size=None,
        cache_requests=False,
        rewrite_requests_cache=False,
    ) -> None:
474
        """Build a set of Instances for a task, and store them in task.instances"""
475
476
477
478

        # used with caching
        og_limit = limit

479
        cache_key = f"requests-{self._config.task}-{self.config.num_fewshot}shot-rank{rank}-world_size{world_size}"
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494

        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
495
        eval_logger.info(f"Building contexts for {self.config.task} on rank {rank}...")
496

497
        instances = []
498
499
500
501
502
503
504
505
506
507

        # 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(
508
            self.doc_iterator(rank=rank, limit=limit, world_size=world_size)
509
510
511
512
513
514
515
        )

        num_docs = len(doc_id_docs)

        for doc_id, doc in tqdm(
            doc_id_docs,
            total=num_docs,
lintangsutawika's avatar
lintangsutawika committed
516
        ):
517
            # sample fewshot context #TODO: need to offset doc_id by rank now!
518
            fewshot_ctx = self.fewshot_context(
519
                doc,
520
                0 if self.config.num_fewshot is None else self.config.num_fewshot,
521
            )
522

523
            # TODO: we should override self.config.repeats if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
524
525
526
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
527
                metadata=(self.config["task"], doc_id, self.config.repeats),
lintangsutawika's avatar
lintangsutawika committed
528
            )
529
530
531
532

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

533
534
535
536
537
538
539
540
541
542
543
544
545
            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
546

547
548
        if len(self._instances) == 0:
            raise ValueError("task.build_requests() did not find any docs!")
549

550
551
552
        if cache_requests and (not cached_instances or rewrite_requests_cache):
            save_to_cache(file_name=cache_key, obj=instances)

553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
    @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
569
            The number of times each instance in a dataset is inferred on. Defaults to 1,
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
603
604
            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

605
606
607
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

haileyschoelkopf's avatar
haileyschoelkopf committed
608
609
610
611
612
613
614
615
616
617
    @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))

618
    @utils.positional_deprecated
lintangsutawika's avatar
lintangsutawika committed
619
    def fewshot_context(
620
621
622
        self,
        doc,
        num_fewshot,
623
        rnd=None,
624
        description=None,
lintangsutawika's avatar
lintangsutawika committed
625
    ):
626
627
628
629
630
631
632
        """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
633
634
635
636
637
        :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.
638
639
640
        :returns: str
            The fewshot context.
        """
641
        if rnd is None:
642
643
644
645
646
647
            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
648

649
        description = description if description else ""
650
651

        if num_fewshot == 0:
lintangsutawika's avatar
lintangsutawika committed
652
            labeled_examples = ""
653
        else:
lintangsutawika's avatar
lintangsutawika committed
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
            # 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
678
            )
679
680

        example = self.doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
681
        return description + labeled_examples + example
682

683
    def apply_filters(self) -> Optional[List[Instance]]:
Baber Abbasi's avatar
Baber Abbasi committed
684
        """Iterates over FilterEnsembles and applies them to instances"""
lintangsutawika's avatar
lintangsutawika committed
685
686
        if hasattr(self, "_filters"):
            for f in self._filters:
687
                f.apply(self._instances)
lintangsutawika's avatar
lintangsutawika committed
688
689
690
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances
691

baberabb's avatar
baberabb committed
692
    def dump_config(self) -> dict:
Baber Abbasi's avatar
Baber Abbasi committed
693
        """Returns the config as a dictionary."""
694
        # TODO: this should only return the overrides applied to a non-YAML task's configuration.
695
        # (num_fewshot)
696
        return self.config.to_dict()
697

Baber Abbasi's avatar
Baber Abbasi committed
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
736
737
    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)

738
739
740
741
742
    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

743
744
745
746
747
748
749
    @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:
750
751
752
            raise ValueError(
                f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"
            )
753
754
755
756
757
758
759
760
761
762
763
764
765

    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

766
767
768
    @property
    def task_id(self) -> Any:
        return self._task_id
769

770

771
class ConfigurableTask(Task):
772
    VERSION = "Yaml"
773
    OUTPUT_TYPE = None
774
    CONFIG = None
775
776

    def __init__(
777
778
779
780
781
        self,
        data_dir=None,
        cache_dir=None,
        download_mode=None,
        config: Optional[dict] = None,
Ethan Smith's avatar
Ethan Smith committed
782
    ) -> None:  # TODO no super() call here
lintangsutawika's avatar
lintangsutawika committed
783
        # Create a unique identifier ID
lintangsutawika's avatar
lintangsutawika committed
784
        self._task_id = shortuuid.uuid()[:8]
lintangsutawika's avatar
lintangsutawika committed
785

786
        # Get pre-configured attributes
787
        self._config = self.CONFIG
788

789
        # Use new configurations if there was no preconfiguration
790
        if self.config is None:
791
            self._config = TaskConfig(**config)
792
793
        # Overwrite configs
        else:
lintangsutawika's avatar
lintangsutawika committed
794
            if config is not None:
795
                self._config.__dict__.update(config)
796

797
        if self.config is None:
lintangsutawika's avatar
lintangsutawika committed
798
799
800
            raise ValueError(
                "Must pass a config to ConfigurableTask, either in cls.CONFIG or `config` kwarg"
            )
801

802
803
804
805
        if isinstance(self.config.metadata, dict):
            if "version" in self.config.metadata:
                self.VERSION = self.config.metadata["version"]

806
        if self.config.output_type is not None:
807
808
809
810
            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)}'"
                )
811
            self.OUTPUT_TYPE = self.config.output_type
812

813
814
        if self.config.dataset_path is not None:
            self.DATASET_PATH = self.config.dataset_path
815

816
817
        if self.config.dataset_name is not None:
            self.DATASET_NAME = self.config.dataset_name
818

819
820
821
822
        self._metric_fn_list = {}
        self._metric_fn_kwargs = {}
        self._aggregation_list = {}
        self._higher_is_better = {}
823

824
        if self.config.metric_list is None:
825
            # TODO: handle this in TaskConfig.__post_init__ ?
826
827
            _metric_list = DEFAULT_METRIC_REGISTRY[self.config.output_type]

828
            for metric_name in _metric_list:
haileyschoelkopf's avatar
haileyschoelkopf committed
829
                self._metric_fn_list[metric_name] = get_metric(metric_name)
lintangsutawika's avatar
lintangsutawika committed
830
                self._metric_fn_kwargs[metric_name] = {}
831
832
833
                self._aggregation_list[metric_name] = get_metric_aggregation(
                    metric_name
                )
haileyschoelkopf's avatar
haileyschoelkopf committed
834
                self._higher_is_better[metric_name] = is_higher_better(metric_name)
835
        else:
836
            for metric_config in self.config.metric_list:
837
838
839
840
                if "metric" not in metric_config:
                    raise ValueError(
                        "'metric' key not provided for an entry in 'metric_list', must be specified!"
                    )
841
842
843
844
                metric_name = metric_config["metric"]
                kwargs = {
                    key: metric_config[key]
                    for key in metric_config
Chris's avatar
Chris committed
845
846
                    if key
                    not in ["metric", "aggregation", "higher_is_better", "hf_evaluate"]
847
                }
Chris's avatar
Chris committed
848
849
850
851
                hf_evaluate_metric = (
                    "hf_evaluate" in metric_config
                    and metric_config["hf_evaluate"] is True
                )
852

853
                if self.config.process_results is not None:
854
855
                    self._metric_fn_list[metric_name] = None
                    self._metric_fn_kwargs[metric_name] = {}
856
857
858
859
860
861
                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
862
863
864
                    self._metric_fn_list[metric_name] = get_metric(
                        metric_name, hf_evaluate_metric
                    )
865
                    self._metric_fn_kwargs[metric_name] = kwargs
lintangsutawika's avatar
lintangsutawika committed
866

867
                if "aggregation" in metric_config:
868
                    agg_name = metric_config["aggregation"]
869
                    if isinstance(agg_name, str):
haileyschoelkopf's avatar
haileyschoelkopf committed
870
                        self._aggregation_list[metric_name] = get_aggregation(agg_name)
871
                    elif callable(agg_name):  # noqa: E721
872
873
874
                        self._aggregation_list[metric_name] = metric_config[
                            "aggregation"
                        ]
875
                else:
876
                    INV_AGG_REGISTRY = {v: k for k, v in AGGREGATION_REGISTRY.items()}
lintangsutawika's avatar
lintangsutawika committed
877
                    metric_agg = get_metric_aggregation(metric_name)
878
                    eval_logger.warning(
879
                        f"[Task: {self.config.task}] metric {metric_name} is defined, but aggregation is not. "
880
881
                        f"using default "
                        f"aggregation={INV_AGG_REGISTRY[metric_agg]}"
882
                    )
883
                    self._aggregation_list[metric_name] = metric_agg
lintangsutawika's avatar
lintangsutawika committed
884

885
886
887
888
889
890
                if "higher_is_better" in metric_config:
                    self._higher_is_better[metric_name] = metric_config[
                        "higher_is_better"
                    ]
                else:
                    eval_logger.warning(
891
                        f"[Task: {self.config.task}] metric {metric_name} is defined, but higher_is_better is not. "
892
                        f"using default "
haileyschoelkopf's avatar
haileyschoelkopf committed
893
                        f"higher_is_better={is_higher_better(metric_name)}"
894
                    )
haileyschoelkopf's avatar
haileyschoelkopf committed
895
                    self._higher_is_better[metric_name] = is_higher_better(metric_name)
896

897
        self.download(self.config.dataset_kwargs)
898
899
900
        self._training_docs = None
        self._fewshot_docs = None

901
        if self.config.filter_list is not None:
lintangsutawika's avatar
lintangsutawika committed
902
            self._filters = []
903
            for filter_config in self.config.filter_list:
904
905
906
907
908
909
910
911
912
                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
913
                self._filters.append(filter_pipeline)
lintangsutawika's avatar
lintangsutawika committed
914
        else:
915
            self._filters = [build_filter_ensemble("none", [["take_first", None]])]
916

917
918
        if self.config.use_prompt is not None:
            eval_logger.info(f"loading prompt {self.config.use_prompt}")
919
            self.prompt = get_prompt(
920
                self.config.use_prompt, self.DATASET_PATH, self.DATASET_NAME
lintangsutawika's avatar
lintangsutawika committed
921
            )
922
923
924
        else:
            self.prompt = None

lintangsutawika's avatar
lintangsutawika committed
925
        if self.fewshot_docs() is not None:
926
927
928
929
            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
930
931
932
                self.config.fewshot_config.get("sampler", "default")
                if self.config.fewshot_config
                else "default"
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
            )
            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)}"
                )
949

950
        self.task_docs = self.eval_docs
951

952
        # Test One Doc
953
        self.features = list(self.task_docs.features.keys())
954
955
        self.multiple_input = 0
        self.multiple_target = 0
956
        test_doc = self.task_docs[0]
957
        test_text = self.doc_to_text(test_doc)
958
        test_target = self.doc_to_target(test_doc)
lintangsutawika's avatar
lintangsutawika committed
959

960
        if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
961
            test_choice = self.doc_to_choice(test_doc)
962
            if not isinstance(test_choice, list):
lintangsutawika's avatar
lintangsutawika committed
963
                eval_logger.error("doc_to_choice must return list")
964
965
            else:
                num_choice = len(test_choice)
966

967
            if isinstance(test_text, int):
968
                self.multiple_input = num_choice
969
970
        else:
            test_choice = None
971

972
        if isinstance(test_target, list):
973
            self.multiple_target = len(test_target)
974
        else:
975
            if (isinstance(test_target, int)) and (test_choice is not None):
lintangsutawika's avatar
lintangsutawika committed
976
                test_target = test_choice[test_target]
977
            else:
lintangsutawika's avatar
lintangsutawika committed
978
                test_target = str(test_target)
979

980
981
982
        if test_choice is not None:
            check_choices = test_choice
        else:
lintangsutawika's avatar
lintangsutawika committed
983
            check_choices = [test_target]
984
985
986
987
        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 = (
988
989
                    True
                    if self.config.target_delimiter.rstrip()
990
                    != self.config.target_delimiter
991
                    else False
992
                )
993

994
                if delimiter_has_whitespace and choice_has_whitespace:
995
996
                    eval_logger.debug(
                        f'Both target_delimiter "{self.config.target_delimiter}" and target choice: "{choice}" have whitespace'
997
998
                    )
                elif (not delimiter_has_whitespace) and (not choice_has_whitespace):
999
                    eval_logger.debug(
1000
                        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'
1001
1002
                    )

1003
    def download(self, dataset_kwargs: Optional[Dict[str, Any]] = None) -> None:
1004
1005
1006
1007
1008
1009
        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
1010
    def has_training_docs(self) -> bool:
1011
        if self.config.training_split is not None:
1012
1013
1014
1015
            return True
        else:
            return False

baberabb's avatar
baberabb committed
1016
    def has_validation_docs(self) -> bool:
1017
        if self.config.validation_split is not None:
1018
1019
1020
1021
            return True
        else:
            return False

baberabb's avatar
baberabb committed
1022
    def has_test_docs(self) -> bool:
1023
        if self.config.test_split is not None:
1024
1025
1026
1027
            return True
        else:
            return False

baberabb's avatar
baberabb committed
1028
    def training_docs(self) -> datasets.Dataset:
1029
        if self.has_training_docs():
1030
1031
1032
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.training_split]
1033
                )
1034
            return self.dataset[self.config.training_split]
1035

baberabb's avatar
baberabb committed
1036
    def validation_docs(self) -> datasets.Dataset:
1037
        if self.has_validation_docs():
1038
1039
1040
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.validation_split]
1041
                )
1042
            return self.dataset[self.config.validation_split]
1043

baberabb's avatar
baberabb committed
1044
    def test_docs(self) -> datasets.Dataset:
1045
        if self.has_test_docs():
1046
1047
1048
            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]
1049

1050
    def fewshot_docs(self):
1051
        if self.config.fewshot_split is not None:
1052
1053
            if self.config.process_docs is not None:
                return self.config.process_docs(self.dataset[self.config.fewshot_split])
1054
            return self.dataset[self.config.fewshot_split]
1055
        else:
1056
            if (self.config.num_fewshot is not None) and (self.config.num_fewshot > 0):
1057
                eval_logger.warning(
lintangsutawika's avatar
lintangsutawika committed
1058
                    f"[Task: {self.config.task}] "
1059
1060
1061
1062
                    "num_fewshot > 0 but fewshot_split is None. "
                    "using preconfigured rule."
                )
            return super().fewshot_docs()
1063

lintangsutawika's avatar
lintangsutawika committed
1064
    @utils.positional_deprecated
1065
    def fewshot_context(self, doc: str, num_fewshot: int) -> str:
lintangsutawika's avatar
lintangsutawika committed
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
        """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.
        """
1076
1077
        if description := self.config.description:
            description = utils.apply_template(self.config.description, doc)
lintangsutawika's avatar
lintangsutawika committed
1078
1079
1080

        if num_fewshot == 0:
            # always prepend the (possibly empty) task description
1081
            labeled_examples = description
lintangsutawika's avatar
lintangsutawika committed
1082
        else:
1083
            labeled_examples = description + self.sampler.get_context(doc, num_fewshot)
lintangsutawika's avatar
lintangsutawika committed
1084
1085

        example = self.doc_to_text(doc)
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
        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
1099

1100
    def apply_filters(self):
Baber Abbasi's avatar
Baber Abbasi committed
1101
        """Iterates over FilterEnsembles and applies them to instances"""
1102
1103
        if hasattr(self, "_filters"):
            for f in self._filters:
1104
                f.apply(self._instances)
1105
1106
1107
1108
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances

1109
    def should_decontaminate(self):
1110
        return self.config.should_decontaminate
1111
1112

    def doc_to_decontamination_query(self, doc):
1113
        if self.config.should_decontaminate:
1114
1115
            if self.config.doc_to_decontamination_query is None:
                return self.doc_to_text(doc)
1116
            else:
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
                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
                        )
                    )
1128

1129
    def _process_doc(self, doc: dict) -> dict:
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
        """
        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):
1141
1142
        if self.prompt is not None:
            doc_to_text = self.prompt
1143
        else:
1144
            doc_to_text = self.config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
1145

1146
        if isinstance(doc_to_text, int):
1147
            return doc_to_text
1148
        elif isinstance(doc_to_text, str):
1149
            if doc_to_text in self.features:
1150
                # if self.config.doc_to_choice is not None:
1151
1152
                #     return self.doc_to_choice(doc)[doc[doc_to_text]]
                # else:
1153
1154
                return doc[doc_to_text]
            else:
lintangsutawika's avatar
lintangsutawika committed
1155
                text_string = utils.apply_template(doc_to_text, doc)
lintangsutawika's avatar
lintangsutawika committed
1156
                if text_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1157
1158
1159
                    return ast.literal_eval(text_string)
                else:
                    return text_string
1160
        elif callable(doc_to_text):
1161
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
1162
        # Used when applying a Promptsource template
1163
        elif hasattr(doc_to_text, "apply"):
1164
1165
1166
1167
1168
            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")
1169
                return self.config.fewshot_delimiter
1170
        else:
1171
            print(type(doc_to_text))
1172
            raise TypeError
1173

1174
    def doc_to_target(self, doc: Mapping) -> Union[int, str, list]:
1175
1176
        if self.prompt is not None:
            doc_to_target = self.prompt
1177
        else:
1178
            doc_to_target = self.config.doc_to_target
1179

1180
        if isinstance(doc_to_target, int):
1181
            return doc_to_target
1182
        elif isinstance(doc_to_target, str):
1183
            if doc_to_target in self.features:
1184
                # if self.config.doc_to_choice is not None:
1185
1186
1187
                #     return self.doc_to_choice(doc)[doc[doc_to_target]]
                # else:
                return doc[doc_to_target]
1188
            else:
lintangsutawika's avatar
lintangsutawika committed
1189
                target_string = utils.apply_template(doc_to_target, doc)
lintangsutawika's avatar
lintangsutawika committed
1190
                if target_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1191
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
1192
1193
1194
1195
1196
                elif (
                    len(target_string) >= 2
                    and (target_string[0] == "[")
                    and (target_string[-1] == "]")
                ):
baberabb's avatar
baberabb committed
1197
1198
1199
1200
                    try:
                        return ast.literal_eval(target_string)
                    except (SyntaxError, ValueError):
                        return target_string
lintangsutawika's avatar
lintangsutawika committed
1201
1202
                else:
                    return target_string
1203
        elif isinstance(doc_to_target, list):
1204
            return doc_to_target
1205
        elif callable(doc_to_target):
1206
            return doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
1207
        # Used when applying a Promptsource template
1208
        elif hasattr(doc_to_target, "apply"):
1209
            applied_prompt = doc_to_target.apply(doc)
1210
1211
1212
1213
            if len(applied_prompt) == 2:
                return applied_prompt[1]
            else:
                eval_logger.warning("Applied prompt returns empty string")
1214
                return self.config.fewshot_delimiter
1215
1216
        else:
            raise TypeError
1217

baberabb's avatar
baberabb committed
1218
    def doc_to_choice(self, doc: Any) -> List[str]:
1219
1220
        if self.prompt is not None:
            doc_to_choice = self.prompt
1221
        elif self.config.doc_to_choice is None:
1222
1223
            eval_logger.error("doc_to_choice was called but not set in config")
        else:
1224
            doc_to_choice = self.config.doc_to_choice
1225

1226
        if isinstance(doc_to_choice, str):
1227
1228
1229
1230
            if doc_to_choice in self.features:
                return doc[doc_to_choice]
            else:
                return ast.literal_eval(utils.apply_template(doc_to_choice, doc))
1231
        elif isinstance(doc_to_choice, list):
1232
            return doc_to_choice
1233
        elif isinstance(doc_to_choice, dict):
1234
1235
1236
1237
1238
1239
1240
            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
1241

baberabb's avatar
baberabb committed
1242
1243
1244
    def construct_requests(
        self, doc: dict, ctx: str, **kwargs
    ) -> Union[List[Instance], Instance]:
1245
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
1246
            arguments = (ctx, self.doc_to_target(doc))
1247
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
1248
            arguments = (self.doc_to_target(doc),)
1249
        elif self.OUTPUT_TYPE == "multiple_choice":
1250
            choices = self.doc_to_choice(doc)
1251
            target_delimiter = self.config.target_delimiter
1252
1253
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
1254
                cont = self.doc_to_target(doc)
1255
1256
1257
                arguments = [
                    (ctx + choice, f"{target_delimiter}{cont}") for choice in choices
                ]
1258
            else:
1259
                # Otherwise they are placed in the continuation
1260
                arguments = [(ctx, f"{target_delimiter}{cont}") for cont in choices]
1261

1262
            request_list = [
1263
1264
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1265
                    doc=doc,
1266
                    arguments=arg,
1267
                    idx=i,
1268
1269
                    **kwargs,
                )
1270
                for i, arg in enumerate(arguments)
1271
            ]
1272
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
1273
            if "acc_mutual_info" in self._metric_fn_list.keys():
1274
1275
1276
                # if we are calculating multiple choice accuracy
                # using mutual information instead of raw loglikelihood as metric, need unconditional lls.

lintangsutawika's avatar
lintangsutawika committed
1277
                # here mutual info refers to calculating
1278
1279
1280
1281
1282
1283
                # 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
1284
                            doc=doc,
1285
                            arguments=("", "{}".format(choice)),
1286
1287
1288
                            idx=i,
                            **kwargs,
                        )
lintangsutawika's avatar
lintangsutawika committed
1289
                        for i, choice in enumerate(choices)
1290
1291
1292
                    ]
                )
            return request_list
lintangsutawika's avatar
lintangsutawika committed
1293

1294
        elif self.OUTPUT_TYPE == "generate_until":
1295
            arguments = (ctx, deepcopy(self.config.generation_kwargs))
lintangsutawika's avatar
lintangsutawika committed
1296
1297

        return Instance(
lintangsutawika's avatar
lintangsutawika committed
1298
1299
            request_type=self.OUTPUT_TYPE, doc=doc, arguments=arguments, idx=0, **kwargs
        )
1300
1301

    def process_results(self, doc, results):
1302
1303
        if callable(self.config.process_results):
            return self.config.process_results(doc, results)
lintangsutawika's avatar
lintangsutawika committed
1304

1305
        result_dict = {}
1306
        use_metric = list(self._metric_fn_list.keys())
1307
1308
1309
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
1310
1311
1312
1313
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
1314
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
1315
            (loglikelihood,) = results
1316
1317
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
haileyschoelkopf's avatar
haileyschoelkopf committed
1318
            return {
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
                **(
                    {"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
1334
            }
1335
        elif self.OUTPUT_TYPE == "multiple_choice":
1336
            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
1337

1338
            # retrieve choices in List[str] form, to compute choice lengths, etc.
1339
            choices = self.doc_to_choice(doc)
1340
1341
            completion_len = np.array([float(len(i)) for i in choices])

1342
1343
            if (
                2 * len(choices) == len(lls)
1344
                and "acc_mutual_info" in self._metric_fn_list.keys()
1345
1346
1347
1348
            ):
                # then we are doing mutual info.
                # this stores the "dryrun" / unconditional answer loglikelihoods
                lls_unconditional = lls[1::2]
1349
1350
                if len(lls_unconditional) != len(choices):
                    raise ValueError
1351
1352
                # and this stores our "regular" conditional loglikelihoods
                lls = lls[::2]
1353

1354
1355
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
1356

1357
1358
            if self.multiple_input:
                gold = self.doc_to_text(doc)
1359
            else:
1360
                gold = self.doc_to_target(doc)
1361
1362

            gold_index_error = False
1363
            if isinstance(gold, list):
Lintang Sutawika's avatar
Lintang Sutawika committed
1364
1365
                gold = [i if i < len(choices) else -100 for i in gold]
                if -100 in gold:
1366
1367
                    gold_index_error = True
            else:
1368
                if isinstance(gold, int):
Lintang Sutawika's avatar
Lintang Sutawika committed
1369
                    gold = gold if gold < len(choices) else -100
1370
                elif isinstance(gold, str):
Lintang Sutawika's avatar
Lintang Sutawika committed
1371
                    gold = choices.index(gold) if gold in choices else -100
lintangsutawika's avatar
lintangsutawika committed
1372

Lintang Sutawika's avatar
Lintang Sutawika committed
1373
                if gold == -100:
1374
1375
1376
1377
                    gold_index_error = True

            if gold_index_error:
                eval_logger.warning(
lintangsutawika's avatar
lintangsutawika committed
1378
                    f"Label index was not in within range of available choices,"
1379
1380
                    f"Sample:\n\n{doc}\n\n"
                )
lintangsutawika's avatar
lintangsutawika committed
1381

1382
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
1383
1384
                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
1385
                exact_match = int(any([is_greedy[i] if i != -100 else 0 for i in gold]))
lintangsutawika's avatar
lintangsutawika committed
1386
1387
1388
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
1389
                # 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
1390
                exact_match = int(is_greedy[gold]) if gold != -100 else 0
1391

Lintang Sutawika's avatar
Lintang Sutawika committed
1392
1393
1394
1395
            prob_norm = utils.softmax(lls)

            # TODO use keyword arguments to the metric?
            # gold, pred, norm stuff, the original lls,
1396
            result_dict = {
1397
                **({"acc": acc} if "acc" in use_metric else {}),
1398
1399
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
1400
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
1401
                **({"exact_match": exact_match} if "exact_match" in use_metric else {}),
Lintang Sutawika's avatar
Lintang Sutawika committed
1402
1403
1404
1405
1406
                **(
                    {"brier_score": (gold, prob_norm)}
                    if "brier_score" in use_metric
                    else {}
                ),
1407
1408
            }

1409
            if "acc_mutual_info" in use_metric:
lintangsutawika's avatar
lintangsutawika committed
1410
1411
1412
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
1413
1414
1415
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

1416
        elif self.OUTPUT_TYPE == "generate_until":
1417
            gold = self.doc_to_target(doc)
Chris's avatar
Chris committed
1418
            result = results[0]
1419
            if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1420
                # If you set doc_to_choice,
lintangsutawika's avatar
lintangsutawika committed
1421
                # it assumes that doc_to_target returns a number.
1422
1423
                choices = self.doc_to_choice(doc)
                gold = choices[gold]
1424
1425
            # we expect multiple_targets to be a list.
            elif self.multiple_target:
baberabb's avatar
baberabb committed
1426
                gold = list(gold)
Chris's avatar
Chris committed
1427
1428
1429
            elif type(gold) != type(result):
                # cast gold to the same type as result
                gold = type(result)(gold)
1430

lintangsutawika's avatar
lintangsutawika committed
1431
            for metric in self._metric_fn_list.keys():
haileyschoelkopf's avatar
haileyschoelkopf committed
1432
1433
1434
1435
1436
                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
1437
1438
1439
1440
                    if not isinstance(gold, list):
                        # sometimes, a multiple_target dataset has exceptions where one doc has only one string answer
                        # print(gold)
                        gold = [gold]
1441
1442
1443
1444
1445
1446
1447
1448
                    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
1449
                    else:
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
                        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
1471
                else:
1472
                    try:
1473
                        result_score = self._metric_fn_list[metric](
1474
1475
                            references=[gold],
                            predictions=[result],
1476
                            **self._metric_fn_kwargs[metric],
1477
                        )
1478
                    except TypeError:  # needed for now in order to use a different interface between our own metrics and HF Evaluate metrics
1479
                        result_score = self._metric_fn_list[metric]([gold, result])
1480
1481
1482
1483
                    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
1484
        else:
lintangsutawika's avatar
lintangsutawika committed
1485
1486
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
1487
                "'loglikelihood', 'loglikelihood_rolling', 'generate_until' or 'multiple_choice'",
1488
            )
1489
1490
1491

        return result_dict

Baber Abbasi's avatar
Baber Abbasi committed
1492
    def aggregation(self) -> dict:
1493
1494
        return self._aggregation_list

Baber Abbasi's avatar
Baber Abbasi committed
1495
    def higher_is_better(self) -> dict:
haileyschoelkopf's avatar
haileyschoelkopf committed
1496
        return self._higher_is_better
1497

Baber Abbasi's avatar
Baber Abbasi committed
1498
1499
1500
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

lintangsutawika's avatar
lintangsutawika committed
1501
1502
    @property
    def task_id(self) -> Any:
lintangsutawika's avatar
lintangsutawika committed
1503
1504
1505
1506
1507
        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
1508

1509
1510
1511
1512
1513
1514
1515
1516
    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)})"
        )

1517
1518

class MultipleChoiceTask(Task):
1519
    OUTPUT_TYPE = "loglikelihood"
1520

baberabb's avatar
baberabb committed
1521
    def doc_to_target(self, doc: dict) -> str:
1522
1523
        return " " + doc["choices"][doc["gold"]]

baberabb's avatar
baberabb committed
1524
    def construct_requests(self, doc: dict, ctx: str, **kwargs) -> List[Instance]:
1525
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1526
1527
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1528
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1529
                doc=doc,
1530
                arguments=(ctx, " {}".format(choice)),
1531
                idx=i,
1532
1533
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1534
1535
            for i, choice in enumerate(doc["choices"])
        ]
1536

1537
    def process_results(self, doc: dict, results: Iterable[Tuple[float, bool]]) -> dict:
lintangsutawika's avatar
lintangsutawika committed
1538
1539
1540
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
        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
1552
    def higher_is_better(self) -> dict:
1553
1554
1555
1556
1557
        return {
            "acc": True,
            "acc_norm": True,
        }

baberabb's avatar
baberabb committed
1558
    def aggregation(self) -> dict:
1559
1560
1561
1562
1563
1564
        return {
            "acc": mean,
            "acc_norm": mean,
        }


lintangsutawika's avatar
lintangsutawika committed
1565
class PerplexityTask(Task):
1566
1567
    OUTPUT_TYPE = "loglikelihood_rolling"

baberabb's avatar
baberabb committed
1568
    def has_training_docs(self) -> bool:
1569
1570
        return False

baberabb's avatar
baberabb committed
1571
    def fewshot_examples(self, k: int, rnd) -> List:
1572
1573
1574
1575
        if k != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1576
1577
        return []

baberabb's avatar
baberabb committed
1578
    def fewshot_context(self, doc: dict, num_fewshot: int) -> Literal[""]:
1579
1580
1581
1582
        if num_fewshot != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1583
1584
1585

        return ""

baberabb's avatar
baberabb committed
1586
    def higher_is_better(self) -> dict:
1587
1588
1589
1590
1591
1592
1593
1594
1595
        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
1596
    def doc_to_text(self, doc) -> str:
1597
1598
1599
1600
1601
        return ""

    def doc_to_target(self, doc):
        return doc

1602
1603
1604
    def construct_requests(self, doc: dict, ctx: Optional[str], **kwargs):
        if bool(ctx):
            raise ValueError
1605

lintangsutawika's avatar
lintangsutawika committed
1606
1607
1608
1609
1610
1611
1612
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1613

1614
    def process_results(self, doc: dict, results: Tuple[float]) -> dict:
1615
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1616
1617
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1618
1619
1620
1621
1622
1623
        return {
            "word_perplexity": (loglikelihood, words),
            "byte_perplexity": (loglikelihood, bytes_),
            "bits_per_byte": (loglikelihood, bytes_),
        }

baberabb's avatar
baberabb committed
1624
    def aggregation(self) -> dict:
1625
1626
1627
1628
1629
1630
1631
        return {
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
            "bits_per_byte": bits_per_byte,
        }

    @classmethod
baberabb's avatar
baberabb committed
1632
    def count_bytes(cls, doc) -> int:
1633
1634
1635
        return len(doc.encode("utf-8"))

    @classmethod
baberabb's avatar
baberabb committed
1636
    def count_words(cls, doc) -> int:
1637
1638
        """Downstream tasks with custom word boundaries should override this!"""
        return len(re.split(r"\s+", doc))