task.py 68.1 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
lintangsutawika's avatar
lintangsutawika committed
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

55
56
@dataclass
class AggMetricConfig(dict):
57
    metric: Optional[str] = None
58
59
    aggregation: Optional[str] = "mean"
    weight_by_size: Optional[str] = False
60
    # list of filter names which should be incorporated into the aggregated metric.
lintangsutawika's avatar
lintangsutawika committed
61
    filter_list: Optional[Union[str, list]] = "none"
62
63

    def __post_init__(self):
64
65
66
67
68
        if self.aggregation != "mean":
            raise ValueError(
                f"Currently, only 'mean' is supported for automatically aggregating scores across groups' subtasks. Got '{self.aggregation}'."
            )

69
70
        if isinstance(self.filter_list, str):
            self.filter_list = [self.filter_list]
lintangsutawika's avatar
lintangsutawika committed
71

lintangsutawika's avatar
lintangsutawika committed
72

lintangsutawika's avatar
lintangsutawika committed
73
74
@dataclass
class GroupConfig(dict):
lintangsutawika's avatar
lintangsutawika committed
75
76
77
    group: Optional[str] = None
    group_alias: Optional[str] = None
    task: Optional[Union[str, list]] = None
78
    aggregate_metric_list: Optional[
79
80
        Union[List[AggMetricConfig], AggMetricConfig, dict]
    ] = None
lintangsutawika's avatar
lintangsutawika committed
81
82
83
    metadata: Optional[
        dict
    ] = None  # by default, not used in the code. allows for users to pass arbitrary info to tasks
lintangsutawika's avatar
lintangsutawika committed
84
85
86
87
88
89
90

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

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

91
    def __post_init__(self):
92
93
94
        if self.aggregate_metric_list is not None:
            if isinstance(self.aggregate_metric_list, dict):
                self.aggregate_metric_list = [self.aggregate_metric_list]
95

96
            self.aggregate_metric_list = [
97
                AggMetricConfig(**item) if isinstance(item, dict) else item
98
                for item in self.aggregate_metric_list
99
100
            ]

lintangsutawika's avatar
lintangsutawika committed
101
102
103
104
105
106
107
108
109
110
111
112
113
    def to_dict(self, keep_callable: bool = False) -> dict:
        """dumps the current config as a dictionary object, as a printable format.
        null fields will not be printed.
        Used for dumping results alongside full task configuration

        :return: dict
            A printable dictionary version of the TaskConfig object.

        # TODO: should any default value in the TaskConfig not be printed?
        """
        cfg_dict = asdict(self)
        # remove values that are `None`
        for k, v in list(cfg_dict.items()):
lintangsutawika's avatar
lintangsutawika committed
114
            if callable(v):
lintangsutawika's avatar
lintangsutawika committed
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
                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
135
136
137
138
139
class ConfigurableGroup(abc.ABC):
    def __init__(
        self,
        config: Optional[dict] = None,
    ) -> None:
lintangsutawika's avatar
lintangsutawika committed
140
        # Create a unique identifier ID
lintangsutawika's avatar
lintangsutawika committed
141
        self._task_id = shortuuid.uuid()[:8]
lintangsutawika's avatar
lintangsutawika committed
142
143
144
145
146
        self._config = GroupConfig(**config)

    @property
    def group(self):
        return self._config.group
147

lintangsutawika's avatar
lintangsutawika committed
148
149
150
    @property
    def group_alias(self):
        return self._config.group_alias
151
152
153
154
155

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

lintangsutawika's avatar
lintangsutawika committed
156
157
158
159
    @property
    def config(self):
        return self._config.to_dict()

lintangsutawika's avatar
lintangsutawika committed
160
161
    @property
    def task_id(self) -> Any:
lintangsutawika's avatar
lintangsutawika committed
162
163
164
165
166
        return "-".join((self.group_name, self._task_id))

    @property
    def group_name(self) -> Any:
        return self._config.group
lintangsutawika's avatar
lintangsutawika committed
167

lintangsutawika's avatar
lintangsutawika committed
168
169
    def __repr__(self):
        return (
170
            f"ConfigurableGroup(group={self.group}," f"group_alias={self.group_alias})"
lintangsutawika's avatar
lintangsutawika committed
171
172
        )

173

174
175
@dataclass
class TaskConfig(dict):
176
    # task naming/registry
177
178
    task: Optional[str] = None
    task_alias: Optional[str] = None
lintangsutawika's avatar
lintangsutawika committed
179
    tag: Optional[Union[str, list]] = None
180
    group: Optional[Union[str, list]] = None
181
182
183
    # HF dataset options.
    # which dataset to use,
    # and what splits for what purpose
184
185
186
187
188
189
190
191
192
    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 (?)
193
194
    # formatting / prompting options.
    # see docs/advanced_task_guide.md for more info
195
196
197
198
199
200
    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
201
    description: str = ""
202
203
    target_delimiter: str = " "
    fewshot_delimiter: str = "\n\n"
204
    fewshot_config: Optional[dict] = None
205
    # runtime configuration options
206
    num_fewshot: Optional[int] = None
207
    # scoring options
208
209
210
    metric_list: Optional[list] = None
    output_type: OutputType = "generate_until"
    generation_kwargs: Optional[dict] = None
211
    repeats: int = 1
212
    filter_list: Optional[Union[str, list]] = None
213
    should_decontaminate: bool = False
214
215
216
217
    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
218

Ethan Smith's avatar
Ethan Smith committed
219
    def __post_init__(self) -> None:
220
221
222
223
224
225
226
227
228
229
230
231
        if self.group is not None:
            eval_logger.warning(
                "A task YAML file was found to contain a `group` key. Groups which provide aggregate scores over several subtasks now require a separate config file--if not aggregating, you may want to use the `tag` config option instead within your config. Setting `group` within a TaskConfig will be deprecated in v0.4.4. Please see https://github.com/EleutherAI/lm-evaluation-harness/blob/main/docs/task_guide.md for more information."
            )

            if self.tag is None:
                self.tag = self.group
            else:
                raise ValueError(
                    "Got both a `group` and `tag` entry within a TaskConfig. Please use one or the other--`group` values will be deprecated in v0.4.4."
                )

Lintang Sutawika's avatar
Lintang Sutawika committed
232
        if self.generation_kwargs is not None:
233
            if self.output_type != "generate_until":
234
                eval_logger.warning(
235
                    f"[{self.task}] passed `generation_kwargs`, but not using `output_type: generate_until`!"
Lintang Sutawika's avatar
Lintang Sutawika committed
236
237
238
239
240
241
242
243
                )

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

            if "until" not in self.generation_kwargs:
244
                self.generation_kwargs["until"] = [self.fewshot_delimiter]
Lintang Sutawika's avatar
Lintang Sutawika committed
245
        else:
246
            if self.output_type == "generate_until":
Lintang Sutawika's avatar
Lintang Sutawika committed
247
248
                # ensure that we greedily generate in absence of explicit arguments otherwise
                self.generation_kwargs = {
249
250
251
252
253
                    "until": (
                        None
                        if self.fewshot_delimiter is None
                        else [self.fewshot_delimiter]
                    ),
Lintang Sutawika's avatar
Lintang Sutawika committed
254
255
                    "do_sample": False,
                }
256

257
258
259
    def __getitem__(self, item):
        return getattr(self, item)

260
261
262
    def __setitem__(self, item, value):
        return setattr(self, item, value)

263
    def to_dict(self, keep_callable: bool = False) -> dict:
264
265
        """dumps the current config as a dictionary object, as a printable format.
        null fields will not be printed.
haileyschoelkopf's avatar
haileyschoelkopf committed
266
        Used for dumping results alongside full task configuration
267

haileyschoelkopf's avatar
haileyschoelkopf committed
268
269
270
271
272
273
274
275
276
277
        :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)
278
279
280
281
282
283
284
285
286
287
            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
288
        return cfg_dict
289

290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
    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)

306
307
308
309
310
311
312
313
314
315
316

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)
    """

317
    VERSION: Optional[Union[int, str]] = None
318

319
320
    # The name of the `Task` benchmark as denoted in the HuggingFace datasets Hub
    # or a path to a custom `datasets` loading script.
321
    DATASET_PATH: Optional[str] = None
322
323

    # The name of a subset within `DATASET_PATH`.
324
    DATASET_NAME: Optional[str] = None
325

326
    OUTPUT_TYPE: Optional[OutputType] = None
lintangsutawika's avatar
lintangsutawika committed
327

328
329
    def __init__(
        self,
330
331
332
333
        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
334
    ) -> None:
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
        """
        :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)
357
358
359
        self._training_docs: Optional[list] = None
        self._fewshot_docs: Optional[list] = None
        self._instances: Optional[List[Instance]] = None
360

361
        # Create a unique identifier ID
lintangsutawika's avatar
lintangsutawika committed
362
        self._task_id = shortuuid.uuid()[:8]
363
        self._config: TaskConfig = TaskConfig({**config}) if config else TaskConfig()
364

lintangsutawika's avatar
lintangsutawika committed
365
        self._filters = [build_filter_ensemble("none", [["take_first", None]])]
366
367
368
        self.fewshot_rnd: Optional[
            random.Random
        ] = None  # purposely induce errors in case of improper usage
369

370
371
372
373
374
375
    def download(
        self,
        data_dir: Optional[str] = None,
        cache_dir: Optional[str] = None,
        download_mode=None,
    ) -> None:
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
        """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.
        """
400
401
402
403
404
405
406
        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,
        )
407

408
    @property
409
    def config(self) -> TaskConfig:
410
411
412
        """Returns the TaskConfig associated with this class."""
        return self._config

413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
    @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

428
    def training_docs(self) -> Iterable:
429
430
431
432
433
434
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

435
    def validation_docs(self) -> Iterable:
436
437
438
439
440
441
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

442
    def test_docs(self) -> Iterable:
443
444
445
446
447
448
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

449
    def fewshot_docs(self) -> Iterable:
450
451
452
453
454
455
456
457
458
        """
        :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
459
            eval_logger.warning(
460
                f"[Task: {self.config.task}] has_training_docs and has_validation_docs are False"
461
                ", using test_docs as fewshot_docs but this is not recommended."
lintangsutawika's avatar
lintangsutawika committed
462
            )
463
464
            return self.test_docs()

465
    def _process_doc(self, doc: dict) -> dict:
466
467
468
469
470
471
472
473
474
        """
        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
475

476
    @property
477
    def instances(self) -> List[Instance]:
478
479
480
481
482
483
484
485
486
487
488
        """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)

489
490
    def doc_to_decontamination_query(self, doc):
        raise NotImplementedError(
491
492
493
494
495
496
497
498
499
500
501
            "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

502
503
    def build_all_requests(
        self,
504
        *,
505
506
507
508
509
        limit=None,
        rank=None,
        world_size=None,
        cache_requests=False,
        rewrite_requests_cache=False,
KonradSzafer's avatar
KonradSzafer committed
510
511
512
513
        system_instruction=None,
        apply_chat_template=False,
        fewshot_as_multiturn=False,
        lm=None,
514
    ) -> None:
515
        """Build a set of Instances for a task, and store them in task.instances"""
516
517
518
519

        # used with caching
        og_limit = limit

520
        cache_key = f"requests-{self._config.task}-{self.config.num_fewshot}shot-rank{rank}-world_size{world_size}"
KonradSzafer's avatar
KonradSzafer committed
521
522
523
524
525
526
527
528
        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 ""
        )
        cache_key += f"-tokenizer{lm.tokenizer_name}" if apply_chat_template else ""
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543

        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
544
        eval_logger.info(f"Building contexts for {self.config.task} on rank {rank}...")
545

546
        instances = []
547
548
549
550
551
552
553
554
555
556

        # 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(
557
            self.doc_iterator(rank=rank, limit=limit, world_size=world_size)
558
559
560
561
562
563
564
        )

        num_docs = len(doc_id_docs)

        for doc_id, doc in tqdm(
            doc_id_docs,
            total=num_docs,
lintangsutawika's avatar
lintangsutawika committed
565
        ):
566
            # sample fewshot context #TODO: need to offset doc_id by rank now!
567
            fewshot_ctx = self.fewshot_context(
568
                doc,
569
                0 if self.config.num_fewshot is None else self.config.num_fewshot,
KonradSzafer's avatar
KonradSzafer committed
570
571
572
573
                system_instruction,
                apply_chat_template,
                fewshot_as_multiturn,
                lm,
574
            )
575

576
            # TODO: we should override self.config.repeats if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
577
578
579
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
580
                metadata=(self.config["task"], doc_id, self.config.repeats),
lintangsutawika's avatar
lintangsutawika committed
581
            )
582
583
584
585

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

586
587
588
589
590
591
592
593
594
595
596
597
598
            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
599

600
601
        if len(self._instances) == 0:
            raise ValueError("task.build_requests() did not find any docs!")
602

603
604
605
        if cache_requests and (not cached_instances or rewrite_requests_cache):
            save_to_cache(file_name=cache_key, obj=instances)

606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
    @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
622
            The number of times each instance in a dataset is inferred on. Defaults to 1,
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
            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

658
659
660
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

haileyschoelkopf's avatar
haileyschoelkopf committed
661
662
663
664
665
666
667
668
669
670
    @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))

671
    @utils.positional_deprecated
lintangsutawika's avatar
lintangsutawika committed
672
    def fewshot_context(
673
674
675
        self,
        doc,
        num_fewshot,
676
        rnd=None,
677
        description=None,
lintangsutawika's avatar
lintangsutawika committed
678
    ):
679
680
681
682
683
684
685
        """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
686
687
688
689
690
        :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.
691
692
693
        :returns: str
            The fewshot context.
        """
694
        if rnd is None:
695
696
697
698
699
700
            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
701

702
        description = description if description else ""
703
704

        if num_fewshot == 0:
lintangsutawika's avatar
lintangsutawika committed
705
            labeled_examples = ""
706
        else:
lintangsutawika's avatar
lintangsutawika committed
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
            # 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
731
            )
732
733

        example = self.doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
734
        return description + labeled_examples + example
735

736
    def apply_filters(self) -> Optional[List[Instance]]:
Baber Abbasi's avatar
Baber Abbasi committed
737
        """Iterates over FilterEnsembles and applies them to instances"""
lintangsutawika's avatar
lintangsutawika committed
738
739
        if hasattr(self, "_filters"):
            for f in self._filters:
740
                f.apply(self._instances)
lintangsutawika's avatar
lintangsutawika committed
741
742
743
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances
744

baberabb's avatar
baberabb committed
745
    def dump_config(self) -> dict:
Baber Abbasi's avatar
Baber Abbasi committed
746
        """Returns the config as a dictionary."""
747
        # TODO: this should only return the overrides applied to a non-YAML task's configuration.
748
        # (num_fewshot)
749
        return self.config.to_dict()
750

Baber Abbasi's avatar
Baber Abbasi committed
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
    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)

791
792
793
794
795
    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

796
797
798
799
800
801
802
    @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:
803
804
805
            raise ValueError(
                f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"
            )
806
807
808
809
810
811
812
813
814
815
816
817
818

    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

819
820
821
    @property
    def task_id(self) -> Any:
        return self._task_id
822

823

824
class ConfigurableTask(Task):
825
    VERSION = "Yaml"
826
    OUTPUT_TYPE = None
827
    CONFIG = None
828
829

    def __init__(
830
831
832
833
834
        self,
        data_dir=None,
        cache_dir=None,
        download_mode=None,
        config: Optional[dict] = None,
Ethan Smith's avatar
Ethan Smith committed
835
    ) -> None:  # TODO no super() call here
lintangsutawika's avatar
lintangsutawika committed
836
        # Create a unique identifier ID
lintangsutawika's avatar
lintangsutawika committed
837
        self._task_id = shortuuid.uuid()[:8]
lintangsutawika's avatar
lintangsutawika committed
838

839
        # Get pre-configured attributes
840
        self._config = self.CONFIG
841

842
        # Use new configurations if there was no preconfiguration
843
        if self.config is None:
844
            self._config = TaskConfig(**config)
845
846
        # Overwrite configs
        else:
lintangsutawika's avatar
lintangsutawika committed
847
            if config is not None:
848
                self._config.__dict__.update(config)
849

850
        if self.config is None:
lintangsutawika's avatar
lintangsutawika committed
851
852
853
            raise ValueError(
                "Must pass a config to ConfigurableTask, either in cls.CONFIG or `config` kwarg"
            )
854

855
856
857
858
        if isinstance(self.config.metadata, dict):
            if "version" in self.config.metadata:
                self.VERSION = self.config.metadata["version"]

859
        if self.config.output_type is not None:
860
861
862
863
            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)}'"
                )
864
            self.OUTPUT_TYPE = self.config.output_type
865

866
867
        if self.config.dataset_path is not None:
            self.DATASET_PATH = self.config.dataset_path
868

869
870
        if self.config.dataset_name is not None:
            self.DATASET_NAME = self.config.dataset_name
871

872
873
874
875
        self._metric_fn_list = {}
        self._metric_fn_kwargs = {}
        self._aggregation_list = {}
        self._higher_is_better = {}
876

877
        if self.config.metric_list is None:
878
            # TODO: handle this in TaskConfig.__post_init__ ?
879
880
            _metric_list = DEFAULT_METRIC_REGISTRY[self.config.output_type]

881
            for metric_name in _metric_list:
haileyschoelkopf's avatar
haileyschoelkopf committed
882
                self._metric_fn_list[metric_name] = get_metric(metric_name)
lintangsutawika's avatar
lintangsutawika committed
883
                self._metric_fn_kwargs[metric_name] = {}
884
885
886
                self._aggregation_list[metric_name] = get_metric_aggregation(
                    metric_name
                )
haileyschoelkopf's avatar
haileyschoelkopf committed
887
                self._higher_is_better[metric_name] = is_higher_better(metric_name)
888
        else:
889
            for metric_config in self.config.metric_list:
890
891
892
893
                if "metric" not in metric_config:
                    raise ValueError(
                        "'metric' key not provided for an entry in 'metric_list', must be specified!"
                    )
894
895
896
897
                metric_name = metric_config["metric"]
                kwargs = {
                    key: metric_config[key]
                    for key in metric_config
Chris's avatar
Chris committed
898
899
                    if key
                    not in ["metric", "aggregation", "higher_is_better", "hf_evaluate"]
900
                }
Chris's avatar
Chris committed
901
902
903
904
                hf_evaluate_metric = (
                    "hf_evaluate" in metric_config
                    and metric_config["hf_evaluate"] is True
                )
905

906
                if self.config.process_results is not None:
907
908
                    self._metric_fn_list[metric_name] = None
                    self._metric_fn_kwargs[metric_name] = {}
909
910
911
912
913
914
                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
915
916
917
                    self._metric_fn_list[metric_name] = get_metric(
                        metric_name, hf_evaluate_metric
                    )
918
                    self._metric_fn_kwargs[metric_name] = kwargs
lintangsutawika's avatar
lintangsutawika committed
919

920
                if "aggregation" in metric_config:
921
                    agg_name = metric_config["aggregation"]
922
                    if isinstance(agg_name, str):
haileyschoelkopf's avatar
haileyschoelkopf committed
923
                        self._aggregation_list[metric_name] = get_aggregation(agg_name)
924
                    elif callable(agg_name):  # noqa: E721
925
926
927
                        self._aggregation_list[metric_name] = metric_config[
                            "aggregation"
                        ]
928
                else:
929
                    INV_AGG_REGISTRY = {v: k for k, v in AGGREGATION_REGISTRY.items()}
lintangsutawika's avatar
lintangsutawika committed
930
                    metric_agg = get_metric_aggregation(metric_name)
931
                    eval_logger.warning(
932
                        f"[Task: {self.config.task}] metric {metric_name} is defined, but aggregation is not. "
933
934
                        f"using default "
                        f"aggregation={INV_AGG_REGISTRY[metric_agg]}"
935
                    )
936
                    self._aggregation_list[metric_name] = metric_agg
lintangsutawika's avatar
lintangsutawika committed
937

938
939
940
941
942
943
                if "higher_is_better" in metric_config:
                    self._higher_is_better[metric_name] = metric_config[
                        "higher_is_better"
                    ]
                else:
                    eval_logger.warning(
944
                        f"[Task: {self.config.task}] metric {metric_name} is defined, but higher_is_better is not. "
945
                        f"using default "
haileyschoelkopf's avatar
haileyschoelkopf committed
946
                        f"higher_is_better={is_higher_better(metric_name)}"
947
                    )
haileyschoelkopf's avatar
haileyschoelkopf committed
948
                    self._higher_is_better[metric_name] = is_higher_better(metric_name)
949

950
        self.download(self.config.dataset_kwargs)
951
952
953
        self._training_docs = None
        self._fewshot_docs = None

954
        if self.config.filter_list is not None:
lintangsutawika's avatar
lintangsutawika committed
955
            self._filters = []
956
            for filter_config in self.config.filter_list:
957
958
959
960
961
962
963
964
965
                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
966
                self._filters.append(filter_pipeline)
lintangsutawika's avatar
lintangsutawika committed
967
        else:
968
            self._filters = [build_filter_ensemble("none", [["take_first", None]])]
969

970
971
        if self.config.use_prompt is not None:
            eval_logger.info(f"loading prompt {self.config.use_prompt}")
972
            self.prompt = get_prompt(
973
                self.config.use_prompt, self.DATASET_PATH, self.DATASET_NAME
lintangsutawika's avatar
lintangsutawika committed
974
            )
975
976
977
        else:
            self.prompt = None

lintangsutawika's avatar
lintangsutawika committed
978
        if self.fewshot_docs() is not None:
979
980
981
982
            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
983
984
985
                self.config.fewshot_config.get("sampler", "default")
                if self.config.fewshot_config
                else "default"
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
            )
            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)}"
                )
1002

1003
        self.task_docs = self.eval_docs
1004

1005
        # Test One Doc
1006
        self.features = list(self.task_docs.features.keys())
1007
1008
        self.multiple_input = 0
        self.multiple_target = 0
1009
        test_doc = self.task_docs[0]
1010
        test_text = self.doc_to_text(test_doc)
1011
        test_target = self.doc_to_target(test_doc)
lintangsutawika's avatar
lintangsutawika committed
1012

1013
        if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1014
            test_choice = self.doc_to_choice(test_doc)
1015
            if not isinstance(test_choice, list):
lintangsutawika's avatar
lintangsutawika committed
1016
                eval_logger.error("doc_to_choice must return list")
1017
1018
            else:
                num_choice = len(test_choice)
1019

1020
            if isinstance(test_text, int):
1021
                self.multiple_input = num_choice
1022
1023
        else:
            test_choice = None
1024

1025
        if isinstance(test_target, list):
1026
            self.multiple_target = len(test_target)
1027
        else:
1028
            if (isinstance(test_target, int)) and (test_choice is not None):
lintangsutawika's avatar
lintangsutawika committed
1029
                test_target = test_choice[test_target]
1030
            else:
lintangsutawika's avatar
lintangsutawika committed
1031
                test_target = str(test_target)
1032

1033
1034
1035
        if test_choice is not None:
            check_choices = test_choice
        else:
lintangsutawika's avatar
lintangsutawika committed
1036
            check_choices = [test_target]
1037
1038
1039
1040
        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 = (
1041
1042
                    True
                    if self.config.target_delimiter.rstrip()
1043
                    != self.config.target_delimiter
1044
                    else False
1045
                )
1046

1047
                if delimiter_has_whitespace and choice_has_whitespace:
1048
1049
                    eval_logger.debug(
                        f'Both target_delimiter "{self.config.target_delimiter}" and target choice: "{choice}" have whitespace'
1050
1051
                    )
                elif (not delimiter_has_whitespace) and (not choice_has_whitespace):
1052
                    eval_logger.debug(
1053
                        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'
1054
1055
                    )

1056
    def download(self, dataset_kwargs: Optional[Dict[str, Any]] = None) -> None:
1057
1058
1059
1060
1061
1062
        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
1063
    def has_training_docs(self) -> bool:
1064
        if self.config.training_split is not None:
1065
1066
1067
1068
            return True
        else:
            return False

baberabb's avatar
baberabb committed
1069
    def has_validation_docs(self) -> bool:
1070
        if self.config.validation_split is not None:
1071
1072
1073
1074
            return True
        else:
            return False

baberabb's avatar
baberabb committed
1075
    def has_test_docs(self) -> bool:
1076
        if self.config.test_split is not None:
1077
1078
1079
1080
            return True
        else:
            return False

baberabb's avatar
baberabb committed
1081
    def training_docs(self) -> datasets.Dataset:
1082
        if self.has_training_docs():
1083
1084
1085
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.training_split]
1086
                )
1087
            return self.dataset[self.config.training_split]
1088

baberabb's avatar
baberabb committed
1089
    def validation_docs(self) -> datasets.Dataset:
1090
        if self.has_validation_docs():
1091
1092
1093
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.validation_split]
1094
                )
1095
            return self.dataset[self.config.validation_split]
1096

baberabb's avatar
baberabb committed
1097
    def test_docs(self) -> datasets.Dataset:
1098
        if self.has_test_docs():
1099
1100
1101
            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]
1102

1103
    def fewshot_docs(self):
1104
        if self.config.fewshot_split is not None:
1105
1106
            if self.config.process_docs is not None:
                return self.config.process_docs(self.dataset[self.config.fewshot_split])
1107
            return self.dataset[self.config.fewshot_split]
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
        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."
                )
1120
        else:
1121
            if (self.config.num_fewshot is not None) and (self.config.num_fewshot > 0):
1122
                eval_logger.warning(
lintangsutawika's avatar
lintangsutawika committed
1123
                    f"[Task: {self.config.task}] "
1124
1125
1126
1127
                    "num_fewshot > 0 but fewshot_split is None. "
                    "using preconfigured rule."
                )
            return super().fewshot_docs()
1128

KonradSzafer's avatar
KonradSzafer committed
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
    @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
1150
    @utils.positional_deprecated
KonradSzafer's avatar
KonradSzafer committed
1151
1152
1153
1154
1155
1156
1157
1158
1159
    def fewshot_context(
        self,
        doc: str,
        num_fewshot: int,
        system_instruction: Optional[str] = None,
        apply_chat_template: bool = False,
        fewshot_as_multiturn: bool = False,
        lm=None,
    ) -> str:
lintangsutawika's avatar
lintangsutawika committed
1160
1161
1162
1163
1164
1165
1166
        """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
1167
1168
1169
1170
1171
1172
1173
1174
        :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.
        :param lm:
            Language model with definition of the tokenizer/function to use for applying the chat template.
lintangsutawika's avatar
lintangsutawika committed
1175
1176
1177
        :returns: str
            The fewshot context.
        """
KonradSzafer's avatar
KonradSzafer committed
1178
1179
1180
1181
1182
1183
1184

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

        # get task description
1185
1186
        if description := self.config.description:
            description = utils.apply_template(self.config.description, doc)
lintangsutawika's avatar
lintangsutawika committed
1187

KonradSzafer's avatar
KonradSzafer committed
1188
1189
1190
1191
1192
1193
1194
1195
1196
        # 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
1197
        else:
KonradSzafer's avatar
KonradSzafer committed
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
            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
1217
1218

        example = self.doc_to_text(doc)
KonradSzafer's avatar
KonradSzafer committed
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
        if apply_chat_template:
            if self.multiple_input:
                return lm.apply_chat_template(labeled_examples)
            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)
                    labeled_examples_list.append(lm.apply_chat_template(chat))
                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)
            return lm.apply_chat_template(labeled_examples)
1248
        else:
KonradSzafer's avatar
KonradSzafer committed
1249
1250
            if self.multiple_input:
                return labeled_examples
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
            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
1261

1262
    def apply_filters(self):
Baber Abbasi's avatar
Baber Abbasi committed
1263
        """Iterates over FilterEnsembles and applies them to instances"""
1264
1265
        if hasattr(self, "_filters"):
            for f in self._filters:
1266
                f.apply(self._instances)
1267
1268
1269
1270
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances

1271
    def should_decontaminate(self):
1272
        return self.config.should_decontaminate
1273
1274

    def doc_to_decontamination_query(self, doc):
1275
        if self.config.should_decontaminate:
1276
1277
            if self.config.doc_to_decontamination_query is None:
                return self.doc_to_text(doc)
1278
            else:
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
                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
                        )
                    )
1290

1291
    def _process_doc(self, doc: dict) -> dict:
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
        """
        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):
1303
1304
        if self.prompt is not None:
            doc_to_text = self.prompt
1305
        else:
1306
            doc_to_text = self.config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
1307

1308
        if isinstance(doc_to_text, int):
1309
            return doc_to_text
1310
        elif isinstance(doc_to_text, str):
1311
            if doc_to_text in self.features:
1312
                # if self.config.doc_to_choice is not None:
1313
1314
                #     return self.doc_to_choice(doc)[doc[doc_to_text]]
                # else:
1315
1316
                return doc[doc_to_text]
            else:
lintangsutawika's avatar
lintangsutawika committed
1317
                text_string = utils.apply_template(doc_to_text, doc)
lintangsutawika's avatar
lintangsutawika committed
1318
                if text_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1319
1320
1321
                    return ast.literal_eval(text_string)
                else:
                    return text_string
1322
        elif callable(doc_to_text):
1323
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
1324
        # Used when applying a Promptsource template
1325
        elif hasattr(doc_to_text, "apply"):
1326
1327
1328
1329
1330
            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")
1331
                return self.config.fewshot_delimiter
1332
        else:
1333
            print(type(doc_to_text))
1334
            raise TypeError
1335

1336
    def doc_to_target(self, doc: Mapping) -> Union[int, str, list]:
1337
1338
        if self.prompt is not None:
            doc_to_target = self.prompt
1339
        else:
1340
            doc_to_target = self.config.doc_to_target
1341

1342
        if isinstance(doc_to_target, int):
1343
            return doc_to_target
1344
        elif isinstance(doc_to_target, str):
1345
            if doc_to_target in self.features:
1346
                # if self.config.doc_to_choice is not None:
1347
1348
1349
                #     return self.doc_to_choice(doc)[doc[doc_to_target]]
                # else:
                return doc[doc_to_target]
1350
            else:
lintangsutawika's avatar
lintangsutawika committed
1351
                target_string = utils.apply_template(doc_to_target, doc)
lintangsutawika's avatar
lintangsutawika committed
1352
                if target_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1353
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
1354
1355
1356
1357
1358
                elif (
                    len(target_string) >= 2
                    and (target_string[0] == "[")
                    and (target_string[-1] == "]")
                ):
baberabb's avatar
baberabb committed
1359
1360
1361
1362
                    try:
                        return ast.literal_eval(target_string)
                    except (SyntaxError, ValueError):
                        return target_string
lintangsutawika's avatar
lintangsutawika committed
1363
1364
                else:
                    return target_string
1365
        elif isinstance(doc_to_target, list):
1366
            return doc_to_target
1367
        elif callable(doc_to_target):
1368
            return doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
1369
        # Used when applying a Promptsource template
1370
        elif hasattr(doc_to_target, "apply"):
1371
            applied_prompt = doc_to_target.apply(doc)
1372
1373
1374
1375
            if len(applied_prompt) == 2:
                return applied_prompt[1]
            else:
                eval_logger.warning("Applied prompt returns empty string")
1376
                return self.config.fewshot_delimiter
1377
1378
        else:
            raise TypeError
1379

baberabb's avatar
baberabb committed
1380
    def doc_to_choice(self, doc: Any) -> List[str]:
1381
1382
        if self.prompt is not None:
            doc_to_choice = self.prompt
1383
        elif self.config.doc_to_choice is None:
1384
1385
            eval_logger.error("doc_to_choice was called but not set in config")
        else:
1386
            doc_to_choice = self.config.doc_to_choice
1387

1388
        if isinstance(doc_to_choice, str):
1389
1390
1391
1392
            if doc_to_choice in self.features:
                return doc[doc_to_choice]
            else:
                return ast.literal_eval(utils.apply_template(doc_to_choice, doc))
1393
        elif isinstance(doc_to_choice, list):
1394
            return doc_to_choice
1395
        elif isinstance(doc_to_choice, dict):
1396
1397
1398
1399
1400
1401
1402
            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
1403

baberabb's avatar
baberabb committed
1404
1405
1406
    def construct_requests(
        self, doc: dict, ctx: str, **kwargs
    ) -> Union[List[Instance], Instance]:
1407
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
1408
            arguments = (ctx, self.doc_to_target(doc))
1409
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
1410
            arguments = (self.doc_to_target(doc),)
1411
        elif self.OUTPUT_TYPE == "multiple_choice":
1412
            choices = self.doc_to_choice(doc)
1413
            target_delimiter = self.config.target_delimiter
1414
1415
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
1416
                cont = self.doc_to_target(doc)
1417
1418
1419
                arguments = [
                    (ctx + choice, f"{target_delimiter}{cont}") for choice in choices
                ]
1420
            else:
1421
                # Otherwise they are placed in the continuation
1422
                arguments = [(ctx, f"{target_delimiter}{cont}") for cont in choices]
1423

1424
            request_list = [
1425
1426
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1427
                    doc=doc,
1428
                    arguments=arg,
1429
                    idx=i,
1430
1431
                    **kwargs,
                )
1432
                for i, arg in enumerate(arguments)
1433
            ]
1434
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
1435
            if "acc_mutual_info" in self._metric_fn_list.keys():
1436
1437
1438
                # if we are calculating multiple choice accuracy
                # using mutual information instead of raw loglikelihood as metric, need unconditional lls.

lintangsutawika's avatar
lintangsutawika committed
1439
                # here mutual info refers to calculating
1440
1441
1442
1443
1444
1445
                # 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
1446
                            doc=doc,
1447
                            arguments=("", "{}".format(choice)),
1448
1449
1450
                            idx=i,
                            **kwargs,
                        )
lintangsutawika's avatar
lintangsutawika committed
1451
                        for i, choice in enumerate(choices)
1452
1453
1454
                    ]
                )
            return request_list
lintangsutawika's avatar
lintangsutawika committed
1455

1456
        elif self.OUTPUT_TYPE == "generate_until":
1457
            arguments = (ctx, deepcopy(self.config.generation_kwargs))
lintangsutawika's avatar
lintangsutawika committed
1458
1459

        return Instance(
lintangsutawika's avatar
lintangsutawika committed
1460
1461
            request_type=self.OUTPUT_TYPE, doc=doc, arguments=arguments, idx=0, **kwargs
        )
1462
1463

    def process_results(self, doc, results):
1464
1465
        if callable(self.config.process_results):
            return self.config.process_results(doc, results)
lintangsutawika's avatar
lintangsutawika committed
1466

1467
        result_dict = {}
1468
        use_metric = list(self._metric_fn_list.keys())
1469
1470
1471
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
1472
1473
1474
1475
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
1476
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
1477
            (loglikelihood,) = results
1478
1479
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
haileyschoelkopf's avatar
haileyschoelkopf committed
1480
            return {
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
                **(
                    {"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
1496
            }
1497
        elif self.OUTPUT_TYPE == "multiple_choice":
1498
            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
1499

1500
            # retrieve choices in List[str] form, to compute choice lengths, etc.
1501
            choices = self.doc_to_choice(doc)
1502
1503
            completion_len = np.array([float(len(i)) for i in choices])

1504
1505
            if (
                2 * len(choices) == len(lls)
1506
                and "acc_mutual_info" in self._metric_fn_list.keys()
1507
1508
1509
1510
            ):
                # then we are doing mutual info.
                # this stores the "dryrun" / unconditional answer loglikelihoods
                lls_unconditional = lls[1::2]
1511
1512
                if len(lls_unconditional) != len(choices):
                    raise ValueError
1513
1514
                # and this stores our "regular" conditional loglikelihoods
                lls = lls[::2]
1515

1516
1517
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
1518

1519
1520
            if self.multiple_input:
                gold = self.doc_to_text(doc)
1521
            else:
1522
                gold = self.doc_to_target(doc)
1523
1524

            gold_index_error = False
1525
            if isinstance(gold, list):
Lintang Sutawika's avatar
Lintang Sutawika committed
1526
1527
                gold = [i if i < len(choices) else -100 for i in gold]
                if -100 in gold:
1528
1529
                    gold_index_error = True
            else:
1530
                if isinstance(gold, int):
Lintang Sutawika's avatar
Lintang Sutawika committed
1531
                    gold = gold if gold < len(choices) else -100
1532
                elif isinstance(gold, str):
Lintang Sutawika's avatar
Lintang Sutawika committed
1533
                    gold = choices.index(gold) if gold in choices else -100
lintangsutawika's avatar
lintangsutawika committed
1534

Lintang Sutawika's avatar
Lintang Sutawika committed
1535
                if gold == -100:
1536
1537
1538
1539
                    gold_index_error = True

            if gold_index_error:
                eval_logger.warning(
lintangsutawika's avatar
lintangsutawika committed
1540
                    f"Label index was not in within range of available choices,"
1541
1542
                    f"Sample:\n\n{doc}\n\n"
                )
lintangsutawika's avatar
lintangsutawika committed
1543

1544
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
1545
1546
                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
1547
                exact_match = int(any([is_greedy[i] if i != -100 else 0 for i in gold]))
lintangsutawika's avatar
lintangsutawika committed
1548
1549
1550
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
1551
                # 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
1552
                exact_match = int(is_greedy[gold]) if gold != -100 else 0
1553

Lintang Sutawika's avatar
Lintang Sutawika committed
1554
1555
1556
1557
            prob_norm = utils.softmax(lls)

            # TODO use keyword arguments to the metric?
            # gold, pred, norm stuff, the original lls,
1558
            result_dict = {
1559
                **({"acc": acc} if "acc" in use_metric else {}),
1560
1561
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
1562
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
1563
                **({"exact_match": exact_match} if "exact_match" in use_metric else {}),
Lintang Sutawika's avatar
Lintang Sutawika committed
1564
1565
1566
1567
1568
                **(
                    {"brier_score": (gold, prob_norm)}
                    if "brier_score" in use_metric
                    else {}
                ),
1569
1570
            }

1571
            if "acc_mutual_info" in use_metric:
lintangsutawika's avatar
lintangsutawika committed
1572
1573
1574
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
1575
1576
1577
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

1578
        elif self.OUTPUT_TYPE == "generate_until":
1579
            gold = self.doc_to_target(doc)
Chris's avatar
Chris committed
1580
            result = results[0]
1581
            if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1582
                # If you set doc_to_choice,
lintangsutawika's avatar
lintangsutawika committed
1583
                # it assumes that doc_to_target returns a number.
1584
1585
                choices = self.doc_to_choice(doc)
                gold = choices[gold]
1586
1587
            # we expect multiple_targets to be a list.
            elif self.multiple_target:
baberabb's avatar
baberabb committed
1588
                gold = list(gold)
Chris's avatar
Chris committed
1589
1590
1591
            elif type(gold) != type(result):
                # cast gold to the same type as result
                gold = type(result)(gold)
1592

lintangsutawika's avatar
lintangsutawika committed
1593
            for metric in self._metric_fn_list.keys():
haileyschoelkopf's avatar
haileyschoelkopf committed
1594
1595
1596
1597
1598
                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
1599
1600
1601
1602
                    if not isinstance(gold, list):
                        # sometimes, a multiple_target dataset has exceptions where one doc has only one string answer
                        # print(gold)
                        gold = [gold]
1603
1604
1605
1606
1607
1608
1609
1610
                    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
1611
                    else:
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
                        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
1633
                else:
1634
                    try:
1635
                        result_score = self._metric_fn_list[metric](
1636
1637
                            references=[gold],
                            predictions=[result],
1638
                            **self._metric_fn_kwargs[metric],
1639
                        )
1640
                    except TypeError:  # needed for now in order to use a different interface between our own metrics and HF Evaluate metrics
1641
                        result_score = self._metric_fn_list[metric]([gold, result])
1642
1643
1644
1645
                    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
1646
        else:
lintangsutawika's avatar
lintangsutawika committed
1647
1648
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
1649
                "'loglikelihood', 'loglikelihood_rolling', 'generate_until' or 'multiple_choice'",
1650
            )
1651
1652
1653

        return result_dict

Baber Abbasi's avatar
Baber Abbasi committed
1654
    def aggregation(self) -> dict:
1655
1656
        return self._aggregation_list

Baber Abbasi's avatar
Baber Abbasi committed
1657
    def higher_is_better(self) -> dict:
haileyschoelkopf's avatar
haileyschoelkopf committed
1658
        return self._higher_is_better
1659

Baber Abbasi's avatar
Baber Abbasi committed
1660
1661
1662
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

lintangsutawika's avatar
lintangsutawika committed
1663
1664
    @property
    def task_id(self) -> Any:
lintangsutawika's avatar
lintangsutawika committed
1665
1666
1667
1668
1669
        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
1670

1671
1672
1673
1674
1675
1676
1677
1678
    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)})"
        )

1679
1680

class MultipleChoiceTask(Task):
1681
    OUTPUT_TYPE = "loglikelihood"
1682

baberabb's avatar
baberabb committed
1683
    def doc_to_target(self, doc: dict) -> str:
1684
1685
        return " " + doc["choices"][doc["gold"]]

baberabb's avatar
baberabb committed
1686
    def construct_requests(self, doc: dict, ctx: str, **kwargs) -> List[Instance]:
1687
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1688
1689
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1690
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1691
                doc=doc,
1692
                arguments=(ctx, " {}".format(choice)),
1693
                idx=i,
1694
1695
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1696
1697
            for i, choice in enumerate(doc["choices"])
        ]
1698

1699
    def process_results(self, doc: dict, results: Iterable[Tuple[float, bool]]) -> dict:
lintangsutawika's avatar
lintangsutawika committed
1700
1701
1702
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
        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
1714
    def higher_is_better(self) -> dict:
1715
1716
1717
1718
1719
        return {
            "acc": True,
            "acc_norm": True,
        }

baberabb's avatar
baberabb committed
1720
    def aggregation(self) -> dict:
1721
1722
1723
1724
1725
1726
        return {
            "acc": mean,
            "acc_norm": mean,
        }


lintangsutawika's avatar
lintangsutawika committed
1727
class PerplexityTask(Task):
1728
1729
    OUTPUT_TYPE = "loglikelihood_rolling"

baberabb's avatar
baberabb committed
1730
    def has_training_docs(self) -> bool:
1731
1732
        return False

baberabb's avatar
baberabb committed
1733
    def fewshot_examples(self, k: int, rnd) -> List:
1734
1735
1736
1737
        if k != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1738
1739
        return []

baberabb's avatar
baberabb committed
1740
    def fewshot_context(self, doc: dict, num_fewshot: int) -> Literal[""]:
1741
1742
1743
1744
        if num_fewshot != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1745
1746
1747

        return ""

baberabb's avatar
baberabb committed
1748
    def higher_is_better(self) -> dict:
1749
1750
1751
1752
1753
1754
1755
1756
1757
        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
1758
    def doc_to_text(self, doc) -> str:
1759
1760
1761
1762
1763
        return ""

    def doc_to_target(self, doc):
        return doc

1764
1765
1766
    def construct_requests(self, doc: dict, ctx: Optional[str], **kwargs):
        if bool(ctx):
            raise ValueError
1767

lintangsutawika's avatar
lintangsutawika committed
1768
1769
1770
1771
1772
1773
1774
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1775

1776
    def process_results(self, doc: dict, results: Tuple[float]) -> dict:
1777
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1778
1779
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1780
1781
1782
1783
1784
1785
        return {
            "word_perplexity": (loglikelihood, words),
            "byte_perplexity": (loglikelihood, bytes_),
            "bits_per_byte": (loglikelihood, bytes_),
        }

baberabb's avatar
baberabb committed
1786
    def aggregation(self) -> dict:
1787
1788
1789
1790
1791
1792
1793
        return {
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
            "bits_per_byte": bits_per_byte,
        }

    @classmethod
baberabb's avatar
baberabb committed
1794
    def count_bytes(cls, doc) -> int:
1795
1796
1797
        return len(doc.encode("utf-8"))

    @classmethod
baberabb's avatar
baberabb committed
1798
    def count_words(cls, doc) -> int:
1799
1800
        """Downstream tasks with custom word boundaries should override this!"""
        return len(re.split(r"\s+", doc))