task.py 65.6 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
25
from tqdm import tqdm
26
27

from lm_eval import utils
28
from lm_eval.api import samplers
29
30
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
31
from lm_eval.api.registry import (
32
33
    AGGREGATION_REGISTRY,
    DEFAULT_METRIC_REGISTRY,
haileyschoelkopf's avatar
haileyschoelkopf committed
34
    get_aggregation,
35
    get_metric,
36
    get_metric_aggregation,
haileyschoelkopf's avatar
haileyschoelkopf committed
37
    is_higher_better,
lintangsutawika's avatar
lintangsutawika committed
38
)
39
from lm_eval.caching.cache import load_from_cache, save_to_cache
40
41
42
from lm_eval.filters import build_filter_ensemble
from lm_eval.prompts import get_prompt

43

44
45
46
47
ALL_OUTPUT_TYPES = [
    "loglikelihood",
    "multiple_choice",
    "loglikelihood_rolling",
48
    "generate_until",
49
50
]

51
eval_logger = logging.getLogger("lm-eval")
52

lintangsutawika's avatar
lintangsutawika committed
53

54
55
@dataclass
class TaskConfig(dict):
56
    # task naming/registry
57
58
    task: Optional[str] = None
    task_alias: Optional[str] = None
Lintang Sutawika's avatar
Lintang Sutawika committed
59
    tag: Optional[Union[str, list]] = None
60
61
62
    # HF dataset options.
    # which dataset to use,
    # and what splits for what purpose
63
64
65
66
67
68
    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
69
    fewshot_split: Optional[str] = (
Baber Abbasi's avatar
Baber Abbasi committed
70
        None  # TODO: assert that this not None if num_fewshot > 0. (?) assert if this is same split as one evaluating (?)
71
    )
72
73
    # formatting / prompting options.
    # see docs/advanced_task_guide.md for more info
74
75
76
    process_docs: Optional[Callable] = None
    doc_to_text: Optional[Union[Callable, str]] = None
    doc_to_target: Optional[Union[Callable, str]] = None
77
    doc_to_image: Union[Callable, str] = None
78
79
80
    doc_to_choice: Optional[Union[Callable, str, dict, list]] = None
    process_results: Optional[Union[Callable, str]] = None
    use_prompt: Optional[str] = None
81
    description: str = ""
82
83
    target_delimiter: str = " "
    fewshot_delimiter: str = "\n\n"
84
    fewshot_config: Optional[dict] = None
85
    # runtime configuration options
86
    num_fewshot: Optional[int] = None
87
    # scoring options
88
89
90
    metric_list: Optional[list] = None
    output_type: OutputType = "generate_until"
    generation_kwargs: Optional[dict] = None
91
    repeats: int = 1
92
    filter_list: Optional[Union[str, list]] = None
93
    should_decontaminate: bool = False
94
    doc_to_decontamination_query: Optional[str] = None
95
96
97
    metadata: Optional[dict] = (
        None  # by default, not used in the code. allows for users to pass arbitrary info to tasks
    )
98

Ethan Smith's avatar
Ethan Smith committed
99
    def __post_init__(self) -> None:
Lintang Sutawika's avatar
Lintang Sutawika committed
100
        if self.generation_kwargs is not None:
101
            if self.output_type != "generate_until":
102
                eval_logger.warning(
103
                    f"[{self.task}] passed `generation_kwargs`, but not using `output_type: generate_until`!"
Lintang Sutawika's avatar
Lintang Sutawika committed
104
105
106
107
108
109
110
111
                )

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

            if "until" not in self.generation_kwargs:
112
                self.generation_kwargs["until"] = [self.fewshot_delimiter]
Lintang Sutawika's avatar
Lintang Sutawika committed
113
        else:
114
            if self.output_type == "generate_until":
Lintang Sutawika's avatar
Lintang Sutawika committed
115
116
                # ensure that we greedily generate in absence of explicit arguments otherwise
                self.generation_kwargs = {
117
118
119
120
121
                    "until": (
                        None
                        if self.fewshot_delimiter is None
                        else [self.fewshot_delimiter]
                    ),
Lintang Sutawika's avatar
Lintang Sutawika committed
122
123
                    "do_sample": False,
                }
124

125
126
127
    def __getitem__(self, item):
        return getattr(self, item)

128
129
130
    def __setitem__(self, item, value):
        return setattr(self, item, value)

131
    def to_dict(self, keep_callable: bool = False) -> dict:
132
133
        """dumps the current config as a dictionary object, as a printable format.
        null fields will not be printed.
haileyschoelkopf's avatar
haileyschoelkopf committed
134
        Used for dumping results alongside full task configuration
135

haileyschoelkopf's avatar
haileyschoelkopf committed
136
137
138
139
140
141
142
143
144
145
        :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)
146
147
148
149
150
151
152
153
154
155
            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
156
        return cfg_dict
157

158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
    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)

174
175
176
177
178
179
180
181
182
183
184

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)
    """

185
    VERSION: Optional[Union[int, str]] = None
186

187
188
    # The name of the `Task` benchmark as denoted in the HuggingFace datasets Hub
    # or a path to a custom `datasets` loading script.
189
    DATASET_PATH: Optional[str] = None
190
191

    # The name of a subset within `DATASET_PATH`.
192
    DATASET_NAME: Optional[str] = None
193

194
    OUTPUT_TYPE: Optional[OutputType] = None
lintangsutawika's avatar
lintangsutawika committed
195

196
197
    def __init__(
        self,
198
199
200
201
        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
202
    ) -> None:
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
        """
        :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)
225
226
227
        self._training_docs: Optional[list] = None
        self._fewshot_docs: Optional[list] = None
        self._instances: Optional[List[Instance]] = None
228

229
        self._config: TaskConfig = TaskConfig({**config}) if config else TaskConfig()
230

lintangsutawika's avatar
lintangsutawika committed
231
        self._filters = [build_filter_ensemble("none", [["take_first", None]])]
232
233
234
        self.fewshot_rnd: Optional[random.Random] = (
            None  # purposely induce errors in case of improper usage
        )
235

236
237
238
239
240
241
    def download(
        self,
        data_dir: Optional[str] = None,
        cache_dir: Optional[str] = None,
        download_mode=None,
    ) -> None:
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
        """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.
        """
266
267
268
269
270
271
272
        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,
        )
273

274
    @property
275
    def config(self) -> TaskConfig:
276
277
278
        """Returns the TaskConfig associated with this class."""
        return self._config

279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
    @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

294
    def training_docs(self) -> Iterable:
295
296
297
298
299
300
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

301
    def validation_docs(self) -> Iterable:
302
303
304
305
306
307
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

308
    def test_docs(self) -> Iterable:
309
310
311
312
313
314
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

315
    def fewshot_docs(self) -> Iterable:
316
317
318
319
320
321
322
323
324
        """
        :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
325
            eval_logger.warning(
326
                f"[Task: {self.config.task}] has_training_docs and has_validation_docs are False"
327
                ", using test_docs as fewshot_docs but this is not recommended."
lintangsutawika's avatar
lintangsutawika committed
328
            )
329
330
            return self.test_docs()

331
    def _process_doc(self, doc: dict) -> dict:
332
333
334
335
336
337
338
339
340
        """
        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
341

342
    @property
343
    def instances(self) -> List[Instance]:
344
345
346
347
348
349
350
351
352
353
354
        """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)

355
356
    def doc_to_decontamination_query(self, doc):
        raise NotImplementedError(
357
358
359
360
361
362
363
364
365
366
367
            "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

368
369
370
371
    # not an abstractmethod because not every language-only task has to implement this
    def doc_to_image(self, doc):
        raise NotImplementedError

372
373
    def build_all_requests(
        self,
374
        *,
375
376
377
378
379
380
381
382
383
384
        limit: Union[int, None] = None,
        rank: int = 0,
        world_size: int = 1,
        cache_requests: bool = False,
        rewrite_requests_cache: bool = False,
        system_instruction: Optional[str] = None,
        apply_chat_template: bool = False,
        fewshot_as_multiturn: bool = False,
        chat_template: Optional[Callable] = None,
        tokenizer_name: str = "",
385
    ) -> None:
386
        """Build a set of Instances for a task, and store them in task.instances"""
387
388
389
390

        # used with caching
        og_limit = limit

391
        cache_key = f"requests-{self._config.task}-{self.config.num_fewshot}shot-rank{rank}-world_size{world_size}"
KonradSzafer's avatar
KonradSzafer committed
392
393
394
395
396
397
398
        cache_key += "-chat_template" if apply_chat_template else ""
        cache_key += "-fewshot_as_multiturn" if fewshot_as_multiturn else ""
        cache_key += (
            f"-system_prompt_hash{utils.hash_string(system_instruction)}"
            if system_instruction is not None
            else ""
        )
399
        cache_key += f"-tokenizer{tokenizer_name}"
400

Baber Abbasi's avatar
Baber Abbasi committed
401
        cached_instances = load_from_cache(file_name=cache_key, cache=cache_requests)
402
403
404
405
406
407
408
409
410
411
412
413
414

        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
415
        eval_logger.info(f"Building contexts for {self.config.task} on rank {rank}...")
416

417
        instances = []
418
419
420
421
422
423
424
425
426
427

        # 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(
428
            self.doc_iterator(rank=rank, limit=limit, world_size=world_size)
429
430
431
432
433
434
435
        )

        num_docs = len(doc_id_docs)

        for doc_id, doc in tqdm(
            doc_id_docs,
            total=num_docs,
lintangsutawika's avatar
lintangsutawika committed
436
        ):
437
            # sample fewshot context #TODO: need to offset doc_id by rank now!
438
            fewshot_ctx = self.fewshot_context(
439
                doc,
440
                0 if self.config.num_fewshot is None else self.config.num_fewshot,
KonradSzafer's avatar
KonradSzafer committed
441
442
443
                system_instruction,
                apply_chat_template,
                fewshot_as_multiturn,
444
                chat_template,
445
            )
446

447
            # TODO: we should override self.config.repeats if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
448
449
450
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
451
                metadata=(self.config["task"], doc_id, self.config.repeats),
452
                apply_chat_template=apply_chat_template,
lintangsutawika's avatar
lintangsutawika committed
453
            )
454
455
456
457

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

458
459
460
461
462
463
464
465
466
467
468
469
470
            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
471

472
473
        if len(self._instances) == 0:
            raise ValueError("task.build_requests() did not find any docs!")
474

475
476
477
        if cache_requests and (not cached_instances or rewrite_requests_cache):
            save_to_cache(file_name=cache_key, obj=instances)

478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
    @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
494
            The number of times each instance in a dataset is inferred on. Defaults to 1,
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
            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

530
531
532
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

haileyschoelkopf's avatar
haileyschoelkopf committed
533
534
535
536
537
538
539
540
541
542
    @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))

543
    @utils.positional_deprecated
lintangsutawika's avatar
lintangsutawika committed
544
    def fewshot_context(
545
546
547
        self,
        doc,
        num_fewshot,
548
        rnd=None,
549
        description=None,
lintangsutawika's avatar
lintangsutawika committed
550
    ):
551
552
553
554
555
556
557
        """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
558
559
560
561
562
        :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.
563
564
565
        :returns: str
            The fewshot context.
        """
566
        if rnd is None:
567
568
569
570
571
572
            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
573

574
        description = description if description else ""
575
576

        if num_fewshot == 0:
lintangsutawika's avatar
lintangsutawika committed
577
            labeled_examples = ""
578
        else:
lintangsutawika's avatar
lintangsutawika committed
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
            # 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
603
            )
604
605

        example = self.doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
606
        return description + labeled_examples + example
607

608
    def apply_filters(self) -> Optional[List[Instance]]:
Baber Abbasi's avatar
Baber Abbasi committed
609
        """Iterates over FilterEnsembles and applies them to instances"""
lintangsutawika's avatar
lintangsutawika committed
610
611
        if hasattr(self, "_filters"):
            for f in self._filters:
612
                f.apply(self._instances)
lintangsutawika's avatar
lintangsutawika committed
613
614
615
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances
616

baberabb's avatar
baberabb committed
617
    def dump_config(self) -> dict:
Baber Abbasi's avatar
Baber Abbasi committed
618
        """Returns the config as a dictionary."""
619
        # TODO: this should only return the overrides applied to a non-YAML task's configuration.
620
        # (num_fewshot)
621
        return self.config.to_dict()
622

Baber Abbasi's avatar
Baber Abbasi committed
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
    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)

663
664
665
666
667
    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

668
669
670
671
672
673
674
    @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:
675
676
677
            raise ValueError(
                f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"
            )
678
679
680
681
682
683
684
685
686
687
688
689
690

    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

691
692

class ConfigurableTask(Task):
693
    VERSION = "Yaml"
694
    OUTPUT_TYPE = None
695
    CONFIG = None
696
697

    def __init__(
698
699
700
701
702
        self,
        data_dir=None,
        cache_dir=None,
        download_mode=None,
        config: Optional[dict] = None,
Ethan Smith's avatar
Ethan Smith committed
703
    ) -> None:  # TODO no super() call here
704
        # Get pre-configured attributes
705
        self._config = self.CONFIG
706

707
        # Use new configurations if there was no preconfiguration
708
        if self.config is None:
709
            self._config = TaskConfig(**config)
710
711
        # Overwrite configs
        else:
lintangsutawika's avatar
lintangsutawika committed
712
            if config is not None:
713
                self._config.__dict__.update(config)
714

715
        if self.config is None:
lintangsutawika's avatar
lintangsutawika committed
716
717
718
            raise ValueError(
                "Must pass a config to ConfigurableTask, either in cls.CONFIG or `config` kwarg"
            )
719

720
721
722
723
        if isinstance(self.config.metadata, dict):
            if "version" in self.config.metadata:
                self.VERSION = self.config.metadata["version"]

724
        if self.config.output_type is not None:
725
726
727
728
            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)}'"
                )
729
            self.OUTPUT_TYPE = self.config.output_type
730

731
732
733
734
        if self.config.doc_to_image is not None:
            # mark the task as requiring multimodality.
            self.MULTIMODAL = True

735
736
        if self.config.dataset_path is not None:
            self.DATASET_PATH = self.config.dataset_path
737

738
739
        if self.config.dataset_name is not None:
            self.DATASET_NAME = self.config.dataset_name
740

741
742
743
744
        self._metric_fn_list = {}
        self._metric_fn_kwargs = {}
        self._aggregation_list = {}
        self._higher_is_better = {}
745

746
        if self.config.metric_list is None:
747
            # TODO: handle this in TaskConfig.__post_init__ ?
748
749
            _metric_list = DEFAULT_METRIC_REGISTRY[self.config.output_type]

750
            for metric_name in _metric_list:
haileyschoelkopf's avatar
haileyschoelkopf committed
751
                self._metric_fn_list[metric_name] = get_metric(metric_name)
lintangsutawika's avatar
lintangsutawika committed
752
                self._metric_fn_kwargs[metric_name] = {}
753
754
755
                self._aggregation_list[metric_name] = get_metric_aggregation(
                    metric_name
                )
haileyschoelkopf's avatar
haileyschoelkopf committed
756
                self._higher_is_better[metric_name] = is_higher_better(metric_name)
757
        else:
758
            for metric_config in self.config.metric_list:
759
760
761
762
                if "metric" not in metric_config:
                    raise ValueError(
                        "'metric' key not provided for an entry in 'metric_list', must be specified!"
                    )
763
764
765
766
                metric_name = metric_config["metric"]
                kwargs = {
                    key: metric_config[key]
                    for key in metric_config
Chris's avatar
Chris committed
767
768
                    if key
                    not in ["metric", "aggregation", "higher_is_better", "hf_evaluate"]
769
                }
Chris's avatar
Chris committed
770
771
772
773
                hf_evaluate_metric = (
                    "hf_evaluate" in metric_config
                    and metric_config["hf_evaluate"] is True
                )
774

775
                if self.config.process_results is not None:
776
777
                    self._metric_fn_list[metric_name] = None
                    self._metric_fn_kwargs[metric_name] = {}
778
779
780
781
782
783
                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
784
785
786
                    self._metric_fn_list[metric_name] = get_metric(
                        metric_name, hf_evaluate_metric
                    )
787
                    self._metric_fn_kwargs[metric_name] = kwargs
lintangsutawika's avatar
lintangsutawika committed
788

789
                if "aggregation" in metric_config:
790
                    agg_name = metric_config["aggregation"]
791
                    if isinstance(agg_name, str):
haileyschoelkopf's avatar
haileyschoelkopf committed
792
                        self._aggregation_list[metric_name] = get_aggregation(agg_name)
793
                    elif callable(agg_name):  # noqa: E721
794
795
796
                        self._aggregation_list[metric_name] = metric_config[
                            "aggregation"
                        ]
797
                else:
798
                    INV_AGG_REGISTRY = {v: k for k, v in AGGREGATION_REGISTRY.items()}
lintangsutawika's avatar
lintangsutawika committed
799
                    metric_agg = get_metric_aggregation(metric_name)
800
                    eval_logger.warning(
801
                        f"[Task: {self.config.task}] metric {metric_name} is defined, but aggregation is not. "
802
803
                        f"using default "
                        f"aggregation={INV_AGG_REGISTRY[metric_agg]}"
804
                    )
805
                    self._aggregation_list[metric_name] = metric_agg
lintangsutawika's avatar
lintangsutawika committed
806

807
808
809
810
811
812
                if "higher_is_better" in metric_config:
                    self._higher_is_better[metric_name] = metric_config[
                        "higher_is_better"
                    ]
                else:
                    eval_logger.warning(
813
                        f"[Task: {self.config.task}] metric {metric_name} is defined, but higher_is_better is not. "
814
                        f"using default "
haileyschoelkopf's avatar
haileyschoelkopf committed
815
                        f"higher_is_better={is_higher_better(metric_name)}"
816
                    )
haileyschoelkopf's avatar
haileyschoelkopf committed
817
                    self._higher_is_better[metric_name] = is_higher_better(metric_name)
818

819
        self.download(self.config.dataset_kwargs)
820
821
822
        self._training_docs = None
        self._fewshot_docs = None

823
        if self.config.filter_list is not None:
lintangsutawika's avatar
lintangsutawika committed
824
            self._filters = []
825
            for filter_config in self.config.filter_list:
826
827
828
829
830
831
832
833
834
                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
835
                self._filters.append(filter_pipeline)
lintangsutawika's avatar
lintangsutawika committed
836
        else:
837
            self._filters = [build_filter_ensemble("none", [["take_first", None]])]
838

839
840
        if self.config.use_prompt is not None:
            eval_logger.info(f"loading prompt {self.config.use_prompt}")
841
            self.prompt = get_prompt(
842
                self.config.use_prompt, self.DATASET_PATH, self.DATASET_NAME
lintangsutawika's avatar
lintangsutawika committed
843
            )
844
845
846
        else:
            self.prompt = None

lintangsutawika's avatar
lintangsutawika committed
847
        if self.fewshot_docs() is not None:
848
849
850
851
            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
852
853
854
                self.config.fewshot_config.get("sampler", "default")
                if self.config.fewshot_config
                else "default"
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
            )
            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)}"
                )
871

872
        self.task_docs = self.eval_docs
873

874
        # Test One Doc
875
        self.features = list(self.task_docs.features.keys())
876
877
        self.multiple_input = 0
        self.multiple_target = 0
878
        test_doc = self.task_docs[0]
879
        test_text = self.doc_to_text(test_doc)
880
        test_target = self.doc_to_target(test_doc)
lintangsutawika's avatar
lintangsutawika committed
881

882
        if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
883
            test_choice = self.doc_to_choice(test_doc)
884
            if not isinstance(test_choice, list):
lintangsutawika's avatar
lintangsutawika committed
885
                eval_logger.error("doc_to_choice must return list")
886
887
            else:
                num_choice = len(test_choice)
888

889
            if isinstance(test_text, int):
890
                self.multiple_input = num_choice
891
892
        else:
            test_choice = None
893

894
        if isinstance(test_target, list):
895
            self.multiple_target = len(test_target)
896
        else:
897
            if (isinstance(test_target, int)) and (test_choice is not None):
lintangsutawika's avatar
lintangsutawika committed
898
                test_target = test_choice[test_target]
899
            else:
lintangsutawika's avatar
lintangsutawika committed
900
                test_target = str(test_target)
901

902
903
904
        if test_choice is not None:
            check_choices = test_choice
        else:
lintangsutawika's avatar
lintangsutawika committed
905
            check_choices = [test_target]
906
907
908
909
        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 = (
910
911
                    True
                    if self.config.target_delimiter.rstrip()
912
                    != self.config.target_delimiter
913
                    else False
914
                )
915

916
                if delimiter_has_whitespace and choice_has_whitespace:
917
918
                    eval_logger.debug(
                        f'Both target_delimiter "{self.config.target_delimiter}" and target choice: "{choice}" have whitespace'
919
920
                    )
                elif (not delimiter_has_whitespace) and (not choice_has_whitespace):
921
                    eval_logger.debug(
922
                        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'
923
924
                    )

925
    def download(self, dataset_kwargs: Optional[Dict[str, Any]] = None) -> None:
926
927
928
929
930
931
        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
932
    def has_training_docs(self) -> bool:
933
        if self.config.training_split is not None:
934
935
936
937
            return True
        else:
            return False

baberabb's avatar
baberabb committed
938
    def has_validation_docs(self) -> bool:
939
        if self.config.validation_split is not None:
940
941
942
943
            return True
        else:
            return False

baberabb's avatar
baberabb committed
944
    def has_test_docs(self) -> bool:
945
        if self.config.test_split is not None:
946
947
948
949
            return True
        else:
            return False

baberabb's avatar
baberabb committed
950
    def training_docs(self) -> datasets.Dataset:
951
        if self.has_training_docs():
952
953
954
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.training_split]
955
                )
956
            return self.dataset[self.config.training_split]
957

baberabb's avatar
baberabb committed
958
    def validation_docs(self) -> datasets.Dataset:
959
        if self.has_validation_docs():
960
961
962
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.validation_split]
963
                )
964
            return self.dataset[self.config.validation_split]
965

baberabb's avatar
baberabb committed
966
    def test_docs(self) -> datasets.Dataset:
967
        if self.has_test_docs():
968
969
970
            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]
971

972
    def fewshot_docs(self):
973
        if self.config.fewshot_split is not None:
974
975
            if self.config.process_docs is not None:
                return self.config.process_docs(self.dataset[self.config.fewshot_split])
976
            return self.dataset[self.config.fewshot_split]
977
978
979
980
981
982
983
984
985
986
987
988
        elif (
            self.config.fewshot_config is not None
            and self.config.fewshot_config.get("samples", None) is not None
        ):
            if isinstance(self.config.fewshot_config["samples"], list):
                return self.config.fewshot_config["samples"]
            elif callable(self.config.fewshot_config["samples"]):
                return self.config.fewshot_config["samples"]()
            else:
                raise Exception(
                    "`fewshot_config['samples']` was incorrectly defined in the configuration. It should be either a list of samples as a dict, or function returning this list."
                )
989
        else:
990
            if (self.config.num_fewshot is not None) and (self.config.num_fewshot > 0):
991
                eval_logger.warning(
Lintang Sutawika's avatar
Lintang Sutawika committed
992
                    f"[Task: {self.config.task}] "
993
994
995
996
                    "num_fewshot > 0 but fewshot_split is None. "
                    "using preconfigured rule."
                )
            return super().fewshot_docs()
997

KonradSzafer's avatar
KonradSzafer committed
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
    @staticmethod
    def append_target_question(
        labeled_examples: List[Dict[str, str]],
        question: str,
        fewshot_as_multiturn: bool = False,
    ) -> None:
        """Adds a target question to the labeled examples list.
        If fewshot_as_multiturn is True, or labeled_examples is empty, or the last entry is a system turn, appends the question as a new user entry.
        Otherwise, it is appended to the last user entry, ensuring that the conversation alternates between the user and the assistant.
        """
        if not fewshot_as_multiturn:
            # if no messages or last message is system, append as new user entry
            if len(labeled_examples) == 0 or labeled_examples[-1]["role"] == "system":
                labeled_examples.append({"role": "user", "content": question})
            # if last message is user, append to it to avoid two user messages in a row
            else:
                labeled_examples[-1]["content"] += question
        else:
            # if fewshot_as_multiturn is True, append as next user entry (last is always assistant)
            labeled_examples.append({"role": "user", "content": question})

lintangsutawika's avatar
lintangsutawika committed
1019
    @utils.positional_deprecated
KonradSzafer's avatar
KonradSzafer committed
1020
1021
1022
1023
1024
1025
1026
    def fewshot_context(
        self,
        doc: str,
        num_fewshot: int,
        system_instruction: Optional[str] = None,
        apply_chat_template: bool = False,
        fewshot_as_multiturn: bool = False,
1027
        chat_template: Optional[Callable] = None,
KonradSzafer's avatar
KonradSzafer committed
1028
    ) -> str:
lintangsutawika's avatar
lintangsutawika committed
1029
1030
1031
1032
1033
1034
1035
        """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.
KonradSzafer's avatar
KonradSzafer committed
1036
1037
1038
1039
1040
1041
        :param  system_instruction: str
            System instruction to be applied to the prompt.
        :param apply_chat_template: bool
            Whether to apply the chat template to the fewshot context.
        :param fewshot_as_multiturn: bool
            Whether to provide the fewshot examples as a multiturn conversation or a single user turn.
1042
1043
        :param chat_template:
            callable (from lm.apply_chat_template) that takes in a list[Dict] chat transcript and renders it into a string.
lintangsutawika's avatar
lintangsutawika committed
1044
1045
1046
        :returns: str
            The fewshot context.
        """
KonradSzafer's avatar
KonradSzafer committed
1047
1048
1049
1050
1051
1052
1053

        if apply_chat_template:
            labeled_examples = []
        else:
            labeled_examples = ""

        # get task description
1054
1055
        if description := self.config.description:
            description = utils.apply_template(self.config.description, doc)
lintangsutawika's avatar
lintangsutawika committed
1056

KonradSzafer's avatar
KonradSzafer committed
1057
1058
1059
1060
1061
1062
1063
1064
1065
        # create system prompt based on the provided system instruction and description
        if system_instruction is not None and description:
            system_prompt = (
                f"{system_instruction}{self.sampler.fewshot_delimiter}{description}"
            )
        elif system_instruction is not None:
            system_prompt = system_instruction
        elif description:
            system_prompt = description
lintangsutawika's avatar
lintangsutawika committed
1066
        else:
KonradSzafer's avatar
KonradSzafer committed
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
            system_prompt = ""

        # add system prompt if specified
        if system_prompt:
            if apply_chat_template:
                labeled_examples.append({"role": "system", "content": system_prompt})
            else:
                labeled_examples = system_prompt

        # if few-shot - append examples after the system prompt
        if num_fewshot > 0:
            if apply_chat_template:
                labeled_examples.extend(
                    self.sampler.get_chat_context(
                        doc, num_fewshot, fewshot_as_multiturn
                    )
                )
            else:
                labeled_examples += self.sampler.get_context(doc, num_fewshot)
lintangsutawika's avatar
lintangsutawika committed
1086
1087

        example = self.doc_to_text(doc)
KonradSzafer's avatar
KonradSzafer committed
1088
1089
        if apply_chat_template:
            if self.multiple_input:
1090
                return chat_template(labeled_examples)
KonradSzafer's avatar
KonradSzafer committed
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
            if isinstance(example, str):
                self.append_target_question(
                    labeled_examples, example, fewshot_as_multiturn
                )
            # for loglikelihood create a list of questions with appended choices
            elif isinstance(example, list):
                labeled_examples_list = []
                # copy chat history for each example and append the answer
                for ex in example:
                    chat = deepcopy(labeled_examples)
                    self.append_target_question(chat, ex, fewshot_as_multiturn)
1102
                    labeled_examples_list.append(chat_template(chat))
KonradSzafer's avatar
KonradSzafer committed
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
                return labeled_examples_list
            # if example is an integer, append the choice or convert to string
            elif isinstance(example, int):
                if self.config.doc_to_choice is not None:
                    choices = self.doc_to_choice(doc)
                    self.append_target_question(
                        labeled_examples, choices[example], fewshot_as_multiturn
                    )
                else:
                    self.append_target_question(
                        labeled_examples, str(example), fewshot_as_multiturn
                    )
                # return lm.apply_chat_template(labeled_examples)
1116
            return chat_template(labeled_examples)
1117
        else:
KonradSzafer's avatar
KonradSzafer committed
1118
1119
            if self.multiple_input:
                return labeled_examples
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
            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
1130

1131
    def apply_filters(self):
Baber Abbasi's avatar
Baber Abbasi committed
1132
        """Iterates over FilterEnsembles and applies them to instances"""
1133
1134
        if hasattr(self, "_filters"):
            for f in self._filters:
1135
                f.apply(self._instances)
1136
1137
1138
1139
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances

1140
    def should_decontaminate(self):
1141
        return self.config.should_decontaminate
1142
1143

    def doc_to_decontamination_query(self, doc):
1144
        if self.config.should_decontaminate:
1145
1146
            if self.config.doc_to_decontamination_query is None:
                return self.doc_to_text(doc)
1147
            else:
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
                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
                        )
                    )
1159

1160
    def _process_doc(self, doc: dict) -> dict:
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
        """
        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

Yu Shi Jie's avatar
Yu Shi Jie committed
1171
    def doc_to_text(self, doc, doc_to_text=None):
1172
1173
        if self.prompt is not None:
            doc_to_text = self.prompt
Yu Shi Jie's avatar
Yu Shi Jie committed
1174
1175
        elif doc_to_text is not None:
            doc_to_text = doc_to_text
1176
        else:
1177
            doc_to_text = self.config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
1178

1179
        if isinstance(doc_to_text, int):
1180
            return doc_to_text
1181
        elif isinstance(doc_to_text, str):
1182
            if doc_to_text in self.features:
1183
                # if self.config.doc_to_choice is not None:
1184
1185
                #     return self.doc_to_choice(doc)[doc[doc_to_text]]
                # else:
1186
1187
                return doc[doc_to_text]
            else:
lintangsutawika's avatar
lintangsutawika committed
1188
                text_string = utils.apply_template(doc_to_text, doc)
lintangsutawika's avatar
lintangsutawika committed
1189
                if text_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1190
1191
1192
                    return ast.literal_eval(text_string)
                else:
                    return text_string
1193
        elif callable(doc_to_text):
1194
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
1195
        # Used when applying a Promptsource template
1196
        elif hasattr(doc_to_text, "apply"):
1197
1198
1199
1200
1201
            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")
1202
                return self.config.fewshot_delimiter
1203
        else:
1204
            print(type(doc_to_text))
1205
            raise TypeError
1206

Yu Shi Jie's avatar
Yu Shi Jie committed
1207
    def doc_to_target(self, doc: Mapping, doc_to_target=None) -> Union[int, str, list]:
1208
1209
        if self.prompt is not None:
            doc_to_target = self.prompt
Yu Shi Jie's avatar
Yu Shi Jie committed
1210
1211
        elif doc_to_target is not None:
            doc_to_target = doc_to_target
1212
        else:
1213
            doc_to_target = self.config.doc_to_target
1214

1215
        if isinstance(doc_to_target, int):
1216
            return doc_to_target
1217
        elif isinstance(doc_to_target, str):
1218
            if doc_to_target in self.features:
1219
                # if self.config.doc_to_choice is not None:
1220
1221
1222
                #     return self.doc_to_choice(doc)[doc[doc_to_target]]
                # else:
                return doc[doc_to_target]
1223
            else:
lintangsutawika's avatar
lintangsutawika committed
1224
                target_string = utils.apply_template(doc_to_target, doc)
lintangsutawika's avatar
lintangsutawika committed
1225
                if target_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1226
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
1227
1228
1229
1230
1231
                elif (
                    len(target_string) >= 2
                    and (target_string[0] == "[")
                    and (target_string[-1] == "]")
                ):
baberabb's avatar
baberabb committed
1232
1233
1234
1235
                    try:
                        return ast.literal_eval(target_string)
                    except (SyntaxError, ValueError):
                        return target_string
lintangsutawika's avatar
lintangsutawika committed
1236
1237
                else:
                    return target_string
1238
        elif isinstance(doc_to_target, list):
1239
            return doc_to_target
1240
        elif callable(doc_to_target):
1241
            return doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
1242
        # Used when applying a Promptsource template
1243
        elif hasattr(doc_to_target, "apply"):
1244
            applied_prompt = doc_to_target.apply(doc)
1245
1246
1247
1248
            if len(applied_prompt) == 2:
                return applied_prompt[1]
            else:
                eval_logger.warning("Applied prompt returns empty string")
1249
                return self.config.fewshot_delimiter
1250
1251
        else:
            raise TypeError
1252

Yu Shi Jie's avatar
Yu Shi Jie committed
1253
    def doc_to_choice(self, doc: Any, doc_to_choice=None) -> List[str]:
1254
1255
        if self.prompt is not None:
            doc_to_choice = self.prompt
Yu Shi Jie's avatar
Yu Shi Jie committed
1256
1257
        elif doc_to_choice is not None:
            doc_to_choice = doc_to_choice
1258
        elif self.config.doc_to_choice is None:
1259
1260
            eval_logger.error("doc_to_choice was called but not set in config")
        else:
1261
            doc_to_choice = self.config.doc_to_choice
1262

1263
        if isinstance(doc_to_choice, str):
1264
1265
1266
1267
            if doc_to_choice in self.features:
                return doc[doc_to_choice]
            else:
                return ast.literal_eval(utils.apply_template(doc_to_choice, doc))
1268
        elif isinstance(doc_to_choice, list):
1269
            return doc_to_choice
1270
        elif isinstance(doc_to_choice, dict):
1271
1272
1273
1274
1275
1276
1277
            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
1278

1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
    def doc_to_image(self, doc: Any, doc_to_image=None) -> Union[int, str, list]:
        if doc_to_image is not None:
            doc_to_image = doc_to_image
        elif self.config.doc_to_image is not None:
            doc_to_image = self.config.doc_to_image
        else:
            return None

        if isinstance(doc_to_image, list):
            image_feature = [
                self.doc_to_image(doc, feature) for feature in doc_to_image
            ]
            return [feature for feature in image_feature if feature is not None]
        elif isinstance(doc_to_image, str):
            if doc_to_image in self.features:
                return doc[doc_to_image]
            else:
                return ast.literal_eval(utils.apply_template(doc_to_image, doc))
        elif callable(doc_to_image):
            return doc_to_image(doc)
        else:
            return None

baberabb's avatar
baberabb committed
1302
1303
1304
    def construct_requests(
        self, doc: dict, ctx: str, **kwargs
    ) -> Union[List[Instance], Instance]:
1305
1306
        apply_chat_template = kwargs.pop("apply_chat_template", False)

1307
1308
        aux_arguments = None

1309
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
1310
            arguments = (ctx, self.doc_to_target(doc))
1311
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
1312
            arguments = (self.doc_to_target(doc),)
1313
        elif self.OUTPUT_TYPE == "multiple_choice":
1314
            choices = self.doc_to_choice(doc)
1315
            target_delimiter = self.config.target_delimiter
1316
1317
            if apply_chat_template:
                target_delimiter = ""
1318
1319
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
1320
                cont = self.doc_to_target(doc)
1321
1322
1323
                arguments = [
                    (ctx + choice, f"{target_delimiter}{cont}") for choice in choices
                ]
1324
            else:
1325
                # Otherwise they are placed in the continuation
1326
                arguments = [(ctx, f"{target_delimiter}{cont}") for cont in choices]
1327

1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
            if "acc_mutual_info" in self._metric_fn_list.keys():
                # if we are calculating multiple choice accuracy
                # using mutual information instead of raw loglikelihood as metric, need unconditional lls.

                # here mutual info refers to calculating
                # 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.
                aux_arguments = [("", f"{choice}") for choice in choices]

                arguments.extend(aux_arguments)

        elif self.OUTPUT_TYPE == "generate_until":
            arguments = (ctx, deepcopy(self.config.generation_kwargs))

        multimodal_arg = {}
        if (
            self.config.doc_to_image
        ):  # TODO: ensure that non-multimodal tasks aren't getting visual args
            multimodal_arg = {
                **multimodal_arg,
                **{"visual": self.doc_to_image(doc)},
            }

        if bool(multimodal_arg):
            if isinstance(arguments, list):
                arguments = [arg + (multimodal_arg,) for arg in arguments]
            else:
                arguments = arguments + (multimodal_arg,)

        if self.OUTPUT_TYPE == "multiple_choice":
1359
            request_list = [
1360
1361
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1362
                    doc=doc,
1363
                    arguments=arg,
1364
                    idx=i,
1365
1366
                    **kwargs,
                )
1367
                for i, arg in enumerate(arguments)
1368
            ]
1369
1370

            return request_list
lintangsutawika's avatar
lintangsutawika committed
1371

lintangsutawika's avatar
lintangsutawika committed
1372
        return Instance(
1373
1374
1375
1376
1377
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=arguments,
            idx=0,
            **kwargs,
lintangsutawika's avatar
lintangsutawika committed
1378
        )
1379
1380

    def process_results(self, doc, results):
1381
1382
        if callable(self.config.process_results):
            return self.config.process_results(doc, results)
lintangsutawika's avatar
lintangsutawika committed
1383

1384
        result_dict = {}
1385
        use_metric = list(self._metric_fn_list.keys())
1386
1387
1388
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
1389
1390
1391
1392
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
1393
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
1394
            (loglikelihood,) = results
1395
1396
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
haileyschoelkopf's avatar
haileyschoelkopf committed
1397
            return {
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
                **(
                    {"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
1413
            }
1414
        elif self.OUTPUT_TYPE == "multiple_choice":
1415
            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
1416

1417
            # retrieve choices in List[str] form, to compute choice lengths, etc.
1418
            choices = self.doc_to_choice(doc)
1419
1420
            completion_len = np.array([float(len(i)) for i in choices])

1421
1422
            if (
                2 * len(choices) == len(lls)
1423
                and "acc_mutual_info" in self._metric_fn_list.keys()
1424
1425
1426
1427
            ):
                # then we are doing mutual info.
                # this stores the "dryrun" / unconditional answer loglikelihoods
                lls_unconditional = lls[1::2]
1428
1429
                if len(lls_unconditional) != len(choices):
                    raise ValueError
1430
1431
                # and this stores our "regular" conditional loglikelihoods
                lls = lls[::2]
1432

1433
1434
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
1435

1436
1437
            if self.multiple_input:
                gold = self.doc_to_text(doc)
1438
            else:
1439
                gold = self.doc_to_target(doc)
1440
1441

            gold_index_error = False
1442
            if isinstance(gold, list):
Lintang Sutawika's avatar
Lintang Sutawika committed
1443
1444
                gold = [i if i < len(choices) else -100 for i in gold]
                if -100 in gold:
1445
1446
                    gold_index_error = True
            else:
1447
                if isinstance(gold, int):
Lintang Sutawika's avatar
Lintang Sutawika committed
1448
                    gold = gold if gold < len(choices) else -100
1449
                elif isinstance(gold, str):
Lintang Sutawika's avatar
Lintang Sutawika committed
1450
                    gold = choices.index(gold) if gold in choices else -100
lintangsutawika's avatar
lintangsutawika committed
1451

Lintang Sutawika's avatar
Lintang Sutawika committed
1452
                if gold == -100:
1453
1454
1455
1456
                    gold_index_error = True

            if gold_index_error:
                eval_logger.warning(
lintangsutawika's avatar
lintangsutawika committed
1457
                    f"Label index was not in within range of available choices,"
1458
1459
                    f"Sample:\n\n{doc}\n\n"
                )
lintangsutawika's avatar
lintangsutawika committed
1460

1461
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
1462
1463
                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
1464
                exact_match = int(any([is_greedy[i] if i != -100 else 0 for i in gold]))
lintangsutawika's avatar
lintangsutawika committed
1465
1466
1467
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
1468
                # 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
1469
                exact_match = int(is_greedy[gold]) if gold != -100 else 0
1470

Lintang Sutawika's avatar
Lintang Sutawika committed
1471
1472
1473
1474
            prob_norm = utils.softmax(lls)

            # TODO use keyword arguments to the metric?
            # gold, pred, norm stuff, the original lls,
1475
            result_dict = {
1476
                **({"acc": acc} if "acc" in use_metric else {}),
1477
1478
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
1479
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
1480
                **({"exact_match": exact_match} if "exact_match" in use_metric else {}),
Lintang Sutawika's avatar
Lintang Sutawika committed
1481
1482
1483
1484
1485
                **(
                    {"brier_score": (gold, prob_norm)}
                    if "brier_score" in use_metric
                    else {}
                ),
1486
1487
            }

1488
            if "acc_mutual_info" in use_metric:
lintangsutawika's avatar
lintangsutawika committed
1489
1490
1491
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
1492
1493
1494
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

1495
        elif self.OUTPUT_TYPE == "generate_until":
1496
            gold = self.doc_to_target(doc)
Chris's avatar
Chris committed
1497
            result = results[0]
1498
            if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1499
                # If you set doc_to_choice,
lintangsutawika's avatar
lintangsutawika committed
1500
                # it assumes that doc_to_target returns a number.
1501
1502
                choices = self.doc_to_choice(doc)
                gold = choices[gold]
1503
1504
            # we expect multiple_targets to be a list.
            elif self.multiple_target:
baberabb's avatar
baberabb committed
1505
                gold = list(gold)
1506
1507
1508
1509
            elif (
                type(gold) is not type(result)
                and "bypass" not in self._metric_fn_list.keys()
            ):
Chris's avatar
Chris committed
1510
1511
                # cast gold to the same type as result
                gold = type(result)(gold)
1512

lintangsutawika's avatar
lintangsutawika committed
1513
            for metric in self._metric_fn_list.keys():
haileyschoelkopf's avatar
haileyschoelkopf committed
1514
1515
1516
1517
1518
                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
1519
1520
1521
1522
                    if not isinstance(gold, list):
                        # sometimes, a multiple_target dataset has exceptions where one doc has only one string answer
                        # print(gold)
                        gold = [gold]
1523
1524
1525
1526
1527
1528
1529
1530
                    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
1531
                    else:
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
                        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
1553
                else:
1554
                    try:
1555
                        result_score = self._metric_fn_list[metric](
1556
1557
                            references=[gold],
                            predictions=[result],
1558
                            **self._metric_fn_kwargs[metric],
1559
                        )
1560
                    except TypeError:  # needed for now in order to use a different interface between our own metrics and HF Evaluate metrics
1561
                        result_score = self._metric_fn_list[metric]([gold, result])
1562
1563
1564
1565
                    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
1566
        else:
lintangsutawika's avatar
lintangsutawika committed
1567
1568
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
1569
                "'loglikelihood', 'loglikelihood_rolling', 'generate_until' or 'multiple_choice'",
1570
            )
1571
1572
1573

        return result_dict

Baber Abbasi's avatar
Baber Abbasi committed
1574
    def aggregation(self) -> dict:
1575
1576
        return self._aggregation_list

Baber Abbasi's avatar
Baber Abbasi committed
1577
    def higher_is_better(self) -> dict:
haileyschoelkopf's avatar
haileyschoelkopf committed
1578
        return self._higher_is_better
1579

Baber Abbasi's avatar
Baber Abbasi committed
1580
1581
1582
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

Lintang Sutawika's avatar
Lintang Sutawika committed
1583
1584
1585
1586
    @property
    def task_name(self) -> Any:
        return getattr(self.config, "task", None)

1587
1588
1589
1590
1591
    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)},"
Baber Abbasi's avatar
Baber Abbasi committed
1592
            f"num_samples={len(self.eval_docs)})"
1593
1594
        )

1595
1596

class MultipleChoiceTask(Task):
1597
    OUTPUT_TYPE = "loglikelihood"
1598

baberabb's avatar
baberabb committed
1599
    def doc_to_target(self, doc: dict) -> str:
1600
1601
        return " " + doc["choices"][doc["gold"]]

baberabb's avatar
baberabb committed
1602
    def construct_requests(self, doc: dict, ctx: str, **kwargs) -> List[Instance]:
1603
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1604
1605
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1606
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1607
                doc=doc,
1608
                arguments=(ctx, " {}".format(choice)),
1609
                idx=i,
1610
1611
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1612
1613
            for i, choice in enumerate(doc["choices"])
        ]
1614

1615
    def process_results(self, doc: dict, results: Iterable[Tuple[float, bool]]) -> dict:
lintangsutawika's avatar
lintangsutawika committed
1616
1617
1618
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
        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
1630
    def higher_is_better(self) -> dict:
1631
1632
1633
1634
1635
        return {
            "acc": True,
            "acc_norm": True,
        }

baberabb's avatar
baberabb committed
1636
    def aggregation(self) -> dict:
1637
1638
1639
1640
1641
1642
        return {
            "acc": mean,
            "acc_norm": mean,
        }


lintangsutawika's avatar
lintangsutawika committed
1643
class PerplexityTask(Task):
1644
1645
    OUTPUT_TYPE = "loglikelihood_rolling"

baberabb's avatar
baberabb committed
1646
    def has_training_docs(self) -> bool:
1647
1648
        return False

baberabb's avatar
baberabb committed
1649
    def fewshot_examples(self, k: int, rnd) -> List:
1650
1651
1652
1653
        if k != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1654
1655
        return []

baberabb's avatar
baberabb committed
1656
    def fewshot_context(self, doc: dict, num_fewshot: int) -> Literal[""]:
1657
1658
1659
1660
        if num_fewshot != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1661
1662
1663

        return ""

baberabb's avatar
baberabb committed
1664
    def higher_is_better(self) -> dict:
1665
1666
1667
1668
1669
1670
1671
1672
1673
        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
1674
    def doc_to_text(self, doc) -> str:
1675
1676
1677
1678
1679
        return ""

    def doc_to_target(self, doc):
        return doc

1680
1681
1682
    def construct_requests(self, doc: dict, ctx: Optional[str], **kwargs):
        if bool(ctx):
            raise ValueError
1683

lintangsutawika's avatar
lintangsutawika committed
1684
1685
1686
1687
1688
1689
1690
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1691

1692
    def process_results(self, doc: dict, results: Tuple[float]) -> dict:
1693
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1694
1695
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1696
1697
1698
1699
1700
1701
        return {
            "word_perplexity": (loglikelihood, words),
            "byte_perplexity": (loglikelihood, bytes_),
            "bits_per_byte": (loglikelihood, bytes_),
        }

baberabb's avatar
baberabb committed
1702
    def aggregation(self) -> dict:
1703
1704
1705
1706
1707
1708
1709
        return {
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
            "bits_per_byte": bits_per_byte,
        }

    @classmethod
baberabb's avatar
baberabb committed
1710
    def count_bytes(cls, doc) -> int:
1711
1712
1713
        return len(doc.encode("utf-8"))

    @classmethod
baberabb's avatar
baberabb committed
1714
    def count_words(cls, doc) -> int:
1715
        """Downstream tasks with custom word boundaries should override this!"""
Lintang Sutawika's avatar
Lintang Sutawika committed
1716
        return len(re.split(r"\s+", doc))