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._config = GroupConfig(**config)
142
        self._task_id = self._config.group
lintangsutawika's avatar
lintangsutawika committed
143
144
145
146

    @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:
162
163
164
165
166
        return self._task_id

    @task_id.setter
    def task_id(self, value):
        self._task_id = value
lintangsutawika's avatar
lintangsutawika committed
167
168
169
170

    @property
    def group_name(self) -> Any:
        return self._config.group
lintangsutawika's avatar
lintangsutawika committed
171

lintangsutawika's avatar
lintangsutawika committed
172
173
    def __repr__(self):
        return (
174
            f"ConfigurableGroup(group={self.group}," f"group_alias={self.group_alias})"
lintangsutawika's avatar
lintangsutawika committed
175
176
        )

177

178
179
@dataclass
class TaskConfig(dict):
180
    # task naming/registry
181
182
    task: Optional[str] = None
    task_alias: Optional[str] = None
lintangsutawika's avatar
lintangsutawika committed
183
    tag: Optional[Union[str, list]] = None
184
    group: Optional[Union[str, list]] = None
185
186
187
    # HF dataset options.
    # which dataset to use,
    # and what splits for what purpose
188
189
190
191
192
193
194
195
196
    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 (?)
197
198
    # formatting / prompting options.
    # see docs/advanced_task_guide.md for more info
199
200
201
202
203
204
    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
205
    description: str = ""
206
207
    target_delimiter: str = " "
    fewshot_delimiter: str = "\n\n"
208
    fewshot_config: Optional[dict] = None
209
    # runtime configuration options
210
    num_fewshot: Optional[int] = None
211
    # scoring options
212
213
214
    metric_list: Optional[list] = None
    output_type: OutputType = "generate_until"
    generation_kwargs: Optional[dict] = None
215
    repeats: int = 1
216
    filter_list: Optional[Union[str, list]] = None
217
    should_decontaminate: bool = False
218
219
220
221
    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
222

Ethan Smith's avatar
Ethan Smith committed
223
    def __post_init__(self) -> None:
224
225
226
227
228
229
230
231
232
233
234
235
        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
236
        if self.generation_kwargs is not None:
237
            if self.output_type != "generate_until":
238
                eval_logger.warning(
239
                    f"[{self.task}] passed `generation_kwargs`, but not using `output_type: generate_until`!"
Lintang Sutawika's avatar
Lintang Sutawika committed
240
241
242
243
244
245
246
247
                )

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

            if "until" not in self.generation_kwargs:
248
                self.generation_kwargs["until"] = [self.fewshot_delimiter]
Lintang Sutawika's avatar
Lintang Sutawika committed
249
        else:
250
            if self.output_type == "generate_until":
Lintang Sutawika's avatar
Lintang Sutawika committed
251
252
                # ensure that we greedily generate in absence of explicit arguments otherwise
                self.generation_kwargs = {
253
254
255
256
257
                    "until": (
                        None
                        if self.fewshot_delimiter is None
                        else [self.fewshot_delimiter]
                    ),
Lintang Sutawika's avatar
Lintang Sutawika committed
258
259
                    "do_sample": False,
                }
260

261
262
263
    def __getitem__(self, item):
        return getattr(self, item)

264
265
266
    def __setitem__(self, item, value):
        return setattr(self, item, value)

267
    def to_dict(self, keep_callable: bool = False) -> dict:
268
269
        """dumps the current config as a dictionary object, as a printable format.
        null fields will not be printed.
haileyschoelkopf's avatar
haileyschoelkopf committed
270
        Used for dumping results alongside full task configuration
271

haileyschoelkopf's avatar
haileyschoelkopf committed
272
273
274
275
276
277
278
279
280
281
        :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)
282
283
284
285
286
287
288
289
290
291
            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
292
        return cfg_dict
293

294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
    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)

310
311
312
313
314
315
316
317
318
319
320

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)
    """

321
    VERSION: Optional[Union[int, str]] = None
322

323
324
    # The name of the `Task` benchmark as denoted in the HuggingFace datasets Hub
    # or a path to a custom `datasets` loading script.
325
    DATASET_PATH: Optional[str] = None
326
327

    # The name of a subset within `DATASET_PATH`.
328
    DATASET_NAME: Optional[str] = None
329

330
    OUTPUT_TYPE: Optional[OutputType] = None
lintangsutawika's avatar
lintangsutawika committed
331

332
333
    def __init__(
        self,
334
335
336
337
        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
338
    ) -> None:
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
        """
        :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)
361
362
363
        self._training_docs: Optional[list] = None
        self._fewshot_docs: Optional[list] = None
        self._instances: Optional[List[Instance]] = None
364

365
        # Create a unique identifier ID
lintangsutawika's avatar
lintangsutawika committed
366
        self._task_id = shortuuid.uuid()[:8]
367
        self._config: TaskConfig = TaskConfig({**config}) if config else TaskConfig()
368

lintangsutawika's avatar
lintangsutawika committed
369
        self._filters = [build_filter_ensemble("none", [["take_first", None]])]
370
371
372
        self.fewshot_rnd: Optional[
            random.Random
        ] = None  # purposely induce errors in case of improper usage
373

374
375
376
377
378
379
    def download(
        self,
        data_dir: Optional[str] = None,
        cache_dir: Optional[str] = None,
        download_mode=None,
    ) -> None:
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
        """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.
        """
404
405
406
407
408
409
410
        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,
        )
411

412
    @property
413
    def config(self) -> TaskConfig:
414
415
416
        """Returns the TaskConfig associated with this class."""
        return self._config

417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
    @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

432
    def training_docs(self) -> Iterable:
433
434
435
436
437
438
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

439
    def validation_docs(self) -> Iterable:
440
441
442
443
444
445
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

446
    def test_docs(self) -> Iterable:
447
448
449
450
451
452
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

453
    def fewshot_docs(self) -> Iterable:
454
455
456
457
458
459
460
461
462
        """
        :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
463
            eval_logger.warning(
464
                f"[Task: {self.config.task}] has_training_docs and has_validation_docs are False"
465
                ", using test_docs as fewshot_docs but this is not recommended."
lintangsutawika's avatar
lintangsutawika committed
466
            )
467
468
            return self.test_docs()

469
    def _process_doc(self, doc: dict) -> dict:
470
471
472
473
474
475
476
477
478
        """
        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
479

480
    @property
481
    def instances(self) -> List[Instance]:
482
483
484
485
486
487
488
489
490
491
492
        """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)

493
494
    def doc_to_decontamination_query(self, doc):
        raise NotImplementedError(
495
496
497
498
499
500
501
502
503
504
505
            "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

506
507
    def build_all_requests(
        self,
508
        *,
509
510
511
512
513
        limit=None,
        rank=None,
        world_size=None,
        cache_requests=False,
        rewrite_requests_cache=False,
KonradSzafer's avatar
KonradSzafer committed
514
515
516
517
        system_instruction=None,
        apply_chat_template=False,
        fewshot_as_multiturn=False,
        lm=None,
518
    ) -> None:
519
        """Build a set of Instances for a task, and store them in task.instances"""
520
521
522
523

        # used with caching
        og_limit = limit

524
        cache_key = f"requests-{self._config.task}-{self.config.num_fewshot}shot-rank{rank}-world_size{world_size}"
KonradSzafer's avatar
KonradSzafer committed
525
526
527
528
529
530
531
532
        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 ""
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547

        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
548
        eval_logger.info(f"Building contexts for {self.config.task} on rank {rank}...")
549

550
        instances = []
551
552
553
554
555
556
557
558
559
560

        # 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(
561
            self.doc_iterator(rank=rank, limit=limit, world_size=world_size)
562
563
564
565
566
567
568
        )

        num_docs = len(doc_id_docs)

        for doc_id, doc in tqdm(
            doc_id_docs,
            total=num_docs,
lintangsutawika's avatar
lintangsutawika committed
569
        ):
570
            # sample fewshot context #TODO: need to offset doc_id by rank now!
571
            fewshot_ctx = self.fewshot_context(
572
                doc,
573
                0 if self.config.num_fewshot is None else self.config.num_fewshot,
KonradSzafer's avatar
KonradSzafer committed
574
575
576
577
                system_instruction,
                apply_chat_template,
                fewshot_as_multiturn,
                lm,
578
            )
579

580
            # TODO: we should override self.config.repeats if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
581
582
583
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
584
                metadata=(self.config["task"], doc_id, self.config.repeats),
lintangsutawika's avatar
lintangsutawika committed
585
            )
586
587
588
589

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

590
591
592
593
594
595
596
597
598
599
600
601
602
            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
603

604
605
        if len(self._instances) == 0:
            raise ValueError("task.build_requests() did not find any docs!")
606

607
608
609
        if cache_requests and (not cached_instances or rewrite_requests_cache):
            save_to_cache(file_name=cache_key, obj=instances)

610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
    @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
626
            The number of times each instance in a dataset is inferred on. Defaults to 1,
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
            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

662
663
664
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

haileyschoelkopf's avatar
haileyschoelkopf committed
665
666
667
668
669
670
671
672
673
674
    @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))

675
    @utils.positional_deprecated
lintangsutawika's avatar
lintangsutawika committed
676
    def fewshot_context(
677
678
679
        self,
        doc,
        num_fewshot,
680
        rnd=None,
681
        description=None,
lintangsutawika's avatar
lintangsutawika committed
682
    ):
683
684
685
686
687
688
689
        """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
690
691
692
693
694
        :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.
695
696
697
        :returns: str
            The fewshot context.
        """
698
        if rnd is None:
699
700
701
702
703
704
            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
705

706
        description = description if description else ""
707
708

        if num_fewshot == 0:
lintangsutawika's avatar
lintangsutawika committed
709
            labeled_examples = ""
710
        else:
lintangsutawika's avatar
lintangsutawika committed
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
            # 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
735
            )
736
737

        example = self.doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
738
        return description + labeled_examples + example
739

740
    def apply_filters(self) -> Optional[List[Instance]]:
Baber Abbasi's avatar
Baber Abbasi committed
741
        """Iterates over FilterEnsembles and applies them to instances"""
lintangsutawika's avatar
lintangsutawika committed
742
743
        if hasattr(self, "_filters"):
            for f in self._filters:
744
                f.apply(self._instances)
lintangsutawika's avatar
lintangsutawika committed
745
746
747
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances
748

baberabb's avatar
baberabb committed
749
    def dump_config(self) -> dict:
Baber Abbasi's avatar
Baber Abbasi committed
750
        """Returns the config as a dictionary."""
751
        # TODO: this should only return the overrides applied to a non-YAML task's configuration.
752
        # (num_fewshot)
753
        return self.config.to_dict()
754

Baber Abbasi's avatar
Baber Abbasi committed
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
791
792
793
794
    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)

795
796
797
798
799
    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

800
801
802
803
804
805
806
    @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:
807
808
809
            raise ValueError(
                f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"
            )
810
811
812
813
814
815
816
817
818
819
820
821
822

    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

823
824
825
    @property
    def task_id(self) -> Any:
        return self._task_id
826

827
828
829
830
    @task_id.setter
    def task_id(self, value):
        self._task_id = value

831

832
class ConfigurableTask(Task):
833
    VERSION = "Yaml"
834
    OUTPUT_TYPE = None
835
    CONFIG = None
836
837

    def __init__(
838
839
840
841
842
        self,
        data_dir=None,
        cache_dir=None,
        download_mode=None,
        config: Optional[dict] = None,
Ethan Smith's avatar
Ethan Smith committed
843
    ) -> None:  # TODO no super() call here
lintangsutawika's avatar
lintangsutawika committed
844
        # Create a unique identifier ID
lintangsutawika's avatar
lintangsutawika committed
845
        self._task_id = shortuuid.uuid()[:8]
lintangsutawika's avatar
lintangsutawika committed
846

847
        # Get pre-configured attributes
848
        self._config = self.CONFIG
849

850
        # Use new configurations if there was no preconfiguration
851
        if self.config is None:
852
            self._config = TaskConfig(**config)
853
854
        # Overwrite configs
        else:
lintangsutawika's avatar
lintangsutawika committed
855
            if config is not None:
856
                self._config.__dict__.update(config)
857

858
        if self.config is None:
lintangsutawika's avatar
lintangsutawika committed
859
860
861
            raise ValueError(
                "Must pass a config to ConfigurableTask, either in cls.CONFIG or `config` kwarg"
            )
862

863
864
865
866
        if isinstance(self.config.metadata, dict):
            if "version" in self.config.metadata:
                self.VERSION = self.config.metadata["version"]

867
        if self.config.output_type is not None:
868
869
870
871
            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)}'"
                )
872
            self.OUTPUT_TYPE = self.config.output_type
873

874
875
        if self.config.dataset_path is not None:
            self.DATASET_PATH = self.config.dataset_path
876

877
878
        if self.config.dataset_name is not None:
            self.DATASET_NAME = self.config.dataset_name
879

880
881
882
883
        self._metric_fn_list = {}
        self._metric_fn_kwargs = {}
        self._aggregation_list = {}
        self._higher_is_better = {}
884

885
        if self.config.metric_list is None:
886
            # TODO: handle this in TaskConfig.__post_init__ ?
887
888
            _metric_list = DEFAULT_METRIC_REGISTRY[self.config.output_type]

889
            for metric_name in _metric_list:
haileyschoelkopf's avatar
haileyschoelkopf committed
890
                self._metric_fn_list[metric_name] = get_metric(metric_name)
lintangsutawika's avatar
lintangsutawika committed
891
                self._metric_fn_kwargs[metric_name] = {}
892
893
894
                self._aggregation_list[metric_name] = get_metric_aggregation(
                    metric_name
                )
haileyschoelkopf's avatar
haileyschoelkopf committed
895
                self._higher_is_better[metric_name] = is_higher_better(metric_name)
896
        else:
897
            for metric_config in self.config.metric_list:
898
899
900
901
                if "metric" not in metric_config:
                    raise ValueError(
                        "'metric' key not provided for an entry in 'metric_list', must be specified!"
                    )
902
903
904
905
                metric_name = metric_config["metric"]
                kwargs = {
                    key: metric_config[key]
                    for key in metric_config
Chris's avatar
Chris committed
906
907
                    if key
                    not in ["metric", "aggregation", "higher_is_better", "hf_evaluate"]
908
                }
Chris's avatar
Chris committed
909
910
911
912
                hf_evaluate_metric = (
                    "hf_evaluate" in metric_config
                    and metric_config["hf_evaluate"] is True
                )
913

914
                if self.config.process_results is not None:
915
916
                    self._metric_fn_list[metric_name] = None
                    self._metric_fn_kwargs[metric_name] = {}
917
918
919
920
921
922
                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
923
924
925
                    self._metric_fn_list[metric_name] = get_metric(
                        metric_name, hf_evaluate_metric
                    )
926
                    self._metric_fn_kwargs[metric_name] = kwargs
lintangsutawika's avatar
lintangsutawika committed
927

928
                if "aggregation" in metric_config:
929
                    agg_name = metric_config["aggregation"]
930
                    if isinstance(agg_name, str):
haileyschoelkopf's avatar
haileyschoelkopf committed
931
                        self._aggregation_list[metric_name] = get_aggregation(agg_name)
932
                    elif callable(agg_name):  # noqa: E721
933
934
935
                        self._aggregation_list[metric_name] = metric_config[
                            "aggregation"
                        ]
936
                else:
937
                    INV_AGG_REGISTRY = {v: k for k, v in AGGREGATION_REGISTRY.items()}
lintangsutawika's avatar
lintangsutawika committed
938
                    metric_agg = get_metric_aggregation(metric_name)
939
                    eval_logger.warning(
940
                        f"[Task: {self.config.task}] metric {metric_name} is defined, but aggregation is not. "
941
942
                        f"using default "
                        f"aggregation={INV_AGG_REGISTRY[metric_agg]}"
943
                    )
944
                    self._aggregation_list[metric_name] = metric_agg
lintangsutawika's avatar
lintangsutawika committed
945

946
947
948
949
950
951
                if "higher_is_better" in metric_config:
                    self._higher_is_better[metric_name] = metric_config[
                        "higher_is_better"
                    ]
                else:
                    eval_logger.warning(
952
                        f"[Task: {self.config.task}] metric {metric_name} is defined, but higher_is_better is not. "
953
                        f"using default "
haileyschoelkopf's avatar
haileyschoelkopf committed
954
                        f"higher_is_better={is_higher_better(metric_name)}"
955
                    )
haileyschoelkopf's avatar
haileyschoelkopf committed
956
                    self._higher_is_better[metric_name] = is_higher_better(metric_name)
957

958
        self.download(self.config.dataset_kwargs)
959
960
961
        self._training_docs = None
        self._fewshot_docs = None

962
        if self.config.filter_list is not None:
lintangsutawika's avatar
lintangsutawika committed
963
            self._filters = []
964
            for filter_config in self.config.filter_list:
965
966
967
968
969
970
971
972
973
                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
974
                self._filters.append(filter_pipeline)
lintangsutawika's avatar
lintangsutawika committed
975
        else:
976
            self._filters = [build_filter_ensemble("none", [["take_first", None]])]
977

978
979
        if self.config.use_prompt is not None:
            eval_logger.info(f"loading prompt {self.config.use_prompt}")
980
            self.prompt = get_prompt(
981
                self.config.use_prompt, self.DATASET_PATH, self.DATASET_NAME
lintangsutawika's avatar
lintangsutawika committed
982
            )
983
984
985
        else:
            self.prompt = None

lintangsutawika's avatar
lintangsutawika committed
986
        if self.fewshot_docs() is not None:
987
988
989
990
            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
991
992
993
                self.config.fewshot_config.get("sampler", "default")
                if self.config.fewshot_config
                else "default"
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
            )
            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)}"
                )
1010

1011
        self.task_docs = self.eval_docs
1012

1013
        # Test One Doc
1014
        self.features = list(self.task_docs.features.keys())
1015
1016
        self.multiple_input = 0
        self.multiple_target = 0
1017
        test_doc = self.task_docs[0]
1018
        test_text = self.doc_to_text(test_doc)
1019
        test_target = self.doc_to_target(test_doc)
lintangsutawika's avatar
lintangsutawika committed
1020

1021
        if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1022
            test_choice = self.doc_to_choice(test_doc)
1023
            if not isinstance(test_choice, list):
lintangsutawika's avatar
lintangsutawika committed
1024
                eval_logger.error("doc_to_choice must return list")
1025
1026
            else:
                num_choice = len(test_choice)
1027

1028
            if isinstance(test_text, int):
1029
                self.multiple_input = num_choice
1030
1031
        else:
            test_choice = None
1032

1033
        if isinstance(test_target, list):
1034
            self.multiple_target = len(test_target)
1035
        else:
1036
            if (isinstance(test_target, int)) and (test_choice is not None):
lintangsutawika's avatar
lintangsutawika committed
1037
                test_target = test_choice[test_target]
1038
            else:
lintangsutawika's avatar
lintangsutawika committed
1039
                test_target = str(test_target)
1040

1041
1042
1043
        if test_choice is not None:
            check_choices = test_choice
        else:
lintangsutawika's avatar
lintangsutawika committed
1044
            check_choices = [test_target]
1045
1046
1047
1048
        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 = (
1049
1050
                    True
                    if self.config.target_delimiter.rstrip()
1051
                    != self.config.target_delimiter
1052
                    else False
1053
                )
1054

1055
                if delimiter_has_whitespace and choice_has_whitespace:
1056
1057
                    eval_logger.debug(
                        f'Both target_delimiter "{self.config.target_delimiter}" and target choice: "{choice}" have whitespace'
1058
1059
                    )
                elif (not delimiter_has_whitespace) and (not choice_has_whitespace):
1060
                    eval_logger.debug(
1061
                        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'
1062
1063
                    )

1064
    def download(self, dataset_kwargs: Optional[Dict[str, Any]] = None) -> None:
1065
1066
1067
1068
1069
1070
        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
1071
    def has_training_docs(self) -> bool:
1072
        if self.config.training_split is not None:
1073
1074
1075
1076
            return True
        else:
            return False

baberabb's avatar
baberabb committed
1077
    def has_validation_docs(self) -> bool:
1078
        if self.config.validation_split is not None:
1079
1080
1081
1082
            return True
        else:
            return False

baberabb's avatar
baberabb committed
1083
    def has_test_docs(self) -> bool:
1084
        if self.config.test_split is not None:
1085
1086
1087
1088
            return True
        else:
            return False

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

baberabb's avatar
baberabb committed
1097
    def validation_docs(self) -> datasets.Dataset:
1098
        if self.has_validation_docs():
1099
1100
1101
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.validation_split]
1102
                )
1103
            return self.dataset[self.config.validation_split]
1104

baberabb's avatar
baberabb committed
1105
    def test_docs(self) -> datasets.Dataset:
1106
        if self.has_test_docs():
1107
1108
1109
            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]
1110

1111
    def fewshot_docs(self):
1112
        if self.config.fewshot_split is not None:
1113
1114
            if self.config.process_docs is not None:
                return self.config.process_docs(self.dataset[self.config.fewshot_split])
1115
            return self.dataset[self.config.fewshot_split]
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
        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."
                )
1128
        else:
1129
            if (self.config.num_fewshot is not None) and (self.config.num_fewshot > 0):
1130
                eval_logger.warning(
lintangsutawika's avatar
lintangsutawika committed
1131
                    f"[Task: {self.config.task}] "
1132
1133
1134
1135
                    "num_fewshot > 0 but fewshot_split is None. "
                    "using preconfigured rule."
                )
            return super().fewshot_docs()
1136

KonradSzafer's avatar
KonradSzafer committed
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
    @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
1158
    @utils.positional_deprecated
KonradSzafer's avatar
KonradSzafer committed
1159
1160
1161
1162
1163
1164
1165
1166
1167
    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
1168
1169
1170
1171
1172
1173
1174
        """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
1175
1176
1177
1178
1179
1180
1181
1182
        :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
1183
1184
1185
        :returns: str
            The fewshot context.
        """
KonradSzafer's avatar
KonradSzafer committed
1186
1187
1188
1189
1190
1191
1192

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

        # get task description
1193
1194
        if description := self.config.description:
            description = utils.apply_template(self.config.description, doc)
lintangsutawika's avatar
lintangsutawika committed
1195

KonradSzafer's avatar
KonradSzafer committed
1196
1197
1198
1199
1200
1201
1202
1203
1204
        # 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
1205
        else:
KonradSzafer's avatar
KonradSzafer committed
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
            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
1225
1226

        example = self.doc_to_text(doc)
KonradSzafer's avatar
KonradSzafer committed
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
        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)
1256
        else:
KonradSzafer's avatar
KonradSzafer committed
1257
1258
            if self.multiple_input:
                return labeled_examples
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
            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
1269

1270
    def apply_filters(self):
Baber Abbasi's avatar
Baber Abbasi committed
1271
        """Iterates over FilterEnsembles and applies them to instances"""
1272
1273
        if hasattr(self, "_filters"):
            for f in self._filters:
1274
                f.apply(self._instances)
1275
1276
1277
1278
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances

1279
    def should_decontaminate(self):
1280
        return self.config.should_decontaminate
1281
1282

    def doc_to_decontamination_query(self, doc):
1283
        if self.config.should_decontaminate:
1284
1285
            if self.config.doc_to_decontamination_query is None:
                return self.doc_to_text(doc)
1286
            else:
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
                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
                        )
                    )
1298

1299
    def _process_doc(self, doc: dict) -> dict:
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
        """
        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):
1311
1312
        if self.prompt is not None:
            doc_to_text = self.prompt
1313
        else:
1314
            doc_to_text = self.config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
1315

1316
        if isinstance(doc_to_text, int):
1317
            return doc_to_text
1318
        elif isinstance(doc_to_text, str):
1319
            if doc_to_text in self.features:
1320
                # if self.config.doc_to_choice is not None:
1321
1322
                #     return self.doc_to_choice(doc)[doc[doc_to_text]]
                # else:
1323
1324
                return doc[doc_to_text]
            else:
lintangsutawika's avatar
lintangsutawika committed
1325
                text_string = utils.apply_template(doc_to_text, doc)
lintangsutawika's avatar
lintangsutawika committed
1326
                if text_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1327
1328
1329
                    return ast.literal_eval(text_string)
                else:
                    return text_string
1330
        elif callable(doc_to_text):
1331
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
1332
        # Used when applying a Promptsource template
1333
        elif hasattr(doc_to_text, "apply"):
1334
1335
1336
1337
1338
            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")
1339
                return self.config.fewshot_delimiter
1340
        else:
1341
            print(type(doc_to_text))
1342
            raise TypeError
1343

1344
    def doc_to_target(self, doc: Mapping) -> Union[int, str, list]:
1345
1346
        if self.prompt is not None:
            doc_to_target = self.prompt
1347
        else:
1348
            doc_to_target = self.config.doc_to_target
1349

1350
        if isinstance(doc_to_target, int):
1351
            return doc_to_target
1352
        elif isinstance(doc_to_target, str):
1353
            if doc_to_target in self.features:
1354
                # if self.config.doc_to_choice is not None:
1355
1356
1357
                #     return self.doc_to_choice(doc)[doc[doc_to_target]]
                # else:
                return doc[doc_to_target]
1358
            else:
lintangsutawika's avatar
lintangsutawika committed
1359
                target_string = utils.apply_template(doc_to_target, doc)
lintangsutawika's avatar
lintangsutawika committed
1360
                if target_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1361
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
1362
1363
1364
1365
1366
                elif (
                    len(target_string) >= 2
                    and (target_string[0] == "[")
                    and (target_string[-1] == "]")
                ):
baberabb's avatar
baberabb committed
1367
1368
1369
1370
                    try:
                        return ast.literal_eval(target_string)
                    except (SyntaxError, ValueError):
                        return target_string
lintangsutawika's avatar
lintangsutawika committed
1371
1372
                else:
                    return target_string
1373
        elif isinstance(doc_to_target, list):
1374
            return doc_to_target
1375
        elif callable(doc_to_target):
1376
            return doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
1377
        # Used when applying a Promptsource template
1378
        elif hasattr(doc_to_target, "apply"):
1379
            applied_prompt = doc_to_target.apply(doc)
1380
1381
1382
1383
            if len(applied_prompt) == 2:
                return applied_prompt[1]
            else:
                eval_logger.warning("Applied prompt returns empty string")
1384
                return self.config.fewshot_delimiter
1385
1386
        else:
            raise TypeError
1387

baberabb's avatar
baberabb committed
1388
    def doc_to_choice(self, doc: Any) -> List[str]:
1389
1390
        if self.prompt is not None:
            doc_to_choice = self.prompt
1391
        elif self.config.doc_to_choice is None:
1392
1393
            eval_logger.error("doc_to_choice was called but not set in config")
        else:
1394
            doc_to_choice = self.config.doc_to_choice
1395

1396
        if isinstance(doc_to_choice, str):
1397
1398
1399
1400
            if doc_to_choice in self.features:
                return doc[doc_to_choice]
            else:
                return ast.literal_eval(utils.apply_template(doc_to_choice, doc))
1401
        elif isinstance(doc_to_choice, list):
1402
            return doc_to_choice
1403
        elif isinstance(doc_to_choice, dict):
1404
1405
1406
1407
1408
1409
1410
            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
1411

baberabb's avatar
baberabb committed
1412
1413
1414
    def construct_requests(
        self, doc: dict, ctx: str, **kwargs
    ) -> Union[List[Instance], Instance]:
1415
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
1416
            arguments = (ctx, self.doc_to_target(doc))
1417
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
1418
            arguments = (self.doc_to_target(doc),)
1419
        elif self.OUTPUT_TYPE == "multiple_choice":
1420
            choices = self.doc_to_choice(doc)
1421
            target_delimiter = self.config.target_delimiter
1422
1423
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
1424
                cont = self.doc_to_target(doc)
1425
1426
1427
                arguments = [
                    (ctx + choice, f"{target_delimiter}{cont}") for choice in choices
                ]
1428
            else:
1429
                # Otherwise they are placed in the continuation
1430
                arguments = [(ctx, f"{target_delimiter}{cont}") for cont in choices]
1431

1432
            request_list = [
1433
1434
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1435
                    doc=doc,
1436
                    arguments=arg,
1437
                    idx=i,
1438
1439
                    **kwargs,
                )
1440
                for i, arg in enumerate(arguments)
1441
            ]
1442
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
1443
            if "acc_mutual_info" in self._metric_fn_list.keys():
1444
1445
1446
                # if we are calculating multiple choice accuracy
                # using mutual information instead of raw loglikelihood as metric, need unconditional lls.

lintangsutawika's avatar
lintangsutawika committed
1447
                # here mutual info refers to calculating
1448
1449
1450
1451
1452
1453
                # 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
1454
                            doc=doc,
1455
                            arguments=("", "{}".format(choice)),
1456
1457
1458
                            idx=i,
                            **kwargs,
                        )
lintangsutawika's avatar
lintangsutawika committed
1459
                        for i, choice in enumerate(choices)
1460
1461
1462
                    ]
                )
            return request_list
lintangsutawika's avatar
lintangsutawika committed
1463

1464
        elif self.OUTPUT_TYPE == "generate_until":
1465
            arguments = (ctx, deepcopy(self.config.generation_kwargs))
lintangsutawika's avatar
lintangsutawika committed
1466
1467

        return Instance(
lintangsutawika's avatar
lintangsutawika committed
1468
1469
            request_type=self.OUTPUT_TYPE, doc=doc, arguments=arguments, idx=0, **kwargs
        )
1470
1471

    def process_results(self, doc, results):
1472
1473
        if callable(self.config.process_results):
            return self.config.process_results(doc, results)
lintangsutawika's avatar
lintangsutawika committed
1474

1475
        result_dict = {}
1476
        use_metric = list(self._metric_fn_list.keys())
1477
1478
1479
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
1480
1481
1482
1483
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
1484
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
1485
            (loglikelihood,) = results
1486
1487
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
haileyschoelkopf's avatar
haileyschoelkopf committed
1488
            return {
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
                **(
                    {"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
1504
            }
1505
        elif self.OUTPUT_TYPE == "multiple_choice":
1506
            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
1507

1508
            # retrieve choices in List[str] form, to compute choice lengths, etc.
1509
            choices = self.doc_to_choice(doc)
1510
1511
            completion_len = np.array([float(len(i)) for i in choices])

1512
1513
            if (
                2 * len(choices) == len(lls)
1514
                and "acc_mutual_info" in self._metric_fn_list.keys()
1515
1516
1517
1518
            ):
                # then we are doing mutual info.
                # this stores the "dryrun" / unconditional answer loglikelihoods
                lls_unconditional = lls[1::2]
1519
1520
                if len(lls_unconditional) != len(choices):
                    raise ValueError
1521
1522
                # and this stores our "regular" conditional loglikelihoods
                lls = lls[::2]
1523

1524
1525
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
1526

1527
1528
            if self.multiple_input:
                gold = self.doc_to_text(doc)
1529
            else:
1530
                gold = self.doc_to_target(doc)
1531
1532

            gold_index_error = False
1533
            if isinstance(gold, list):
Lintang Sutawika's avatar
Lintang Sutawika committed
1534
1535
                gold = [i if i < len(choices) else -100 for i in gold]
                if -100 in gold:
1536
1537
                    gold_index_error = True
            else:
1538
                if isinstance(gold, int):
Lintang Sutawika's avatar
Lintang Sutawika committed
1539
                    gold = gold if gold < len(choices) else -100
1540
                elif isinstance(gold, str):
Lintang Sutawika's avatar
Lintang Sutawika committed
1541
                    gold = choices.index(gold) if gold in choices else -100
lintangsutawika's avatar
lintangsutawika committed
1542

Lintang Sutawika's avatar
Lintang Sutawika committed
1543
                if gold == -100:
1544
1545
1546
1547
                    gold_index_error = True

            if gold_index_error:
                eval_logger.warning(
lintangsutawika's avatar
lintangsutawika committed
1548
                    f"Label index was not in within range of available choices,"
1549
1550
                    f"Sample:\n\n{doc}\n\n"
                )
lintangsutawika's avatar
lintangsutawika committed
1551

1552
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
1553
1554
                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
1555
                exact_match = int(any([is_greedy[i] if i != -100 else 0 for i in gold]))
lintangsutawika's avatar
lintangsutawika committed
1556
1557
1558
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
1559
                # 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
1560
                exact_match = int(is_greedy[gold]) if gold != -100 else 0
1561

Lintang Sutawika's avatar
Lintang Sutawika committed
1562
1563
1564
1565
            prob_norm = utils.softmax(lls)

            # TODO use keyword arguments to the metric?
            # gold, pred, norm stuff, the original lls,
1566
            result_dict = {
1567
                **({"acc": acc} if "acc" in use_metric else {}),
1568
1569
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
1570
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
1571
                **({"exact_match": exact_match} if "exact_match" in use_metric else {}),
Lintang Sutawika's avatar
Lintang Sutawika committed
1572
1573
1574
1575
1576
                **(
                    {"brier_score": (gold, prob_norm)}
                    if "brier_score" in use_metric
                    else {}
                ),
1577
1578
            }

1579
            if "acc_mutual_info" in use_metric:
lintangsutawika's avatar
lintangsutawika committed
1580
1581
1582
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
1583
1584
1585
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

1586
        elif self.OUTPUT_TYPE == "generate_until":
1587
            gold = self.doc_to_target(doc)
Chris's avatar
Chris committed
1588
            result = results[0]
1589
            if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1590
                # If you set doc_to_choice,
lintangsutawika's avatar
lintangsutawika committed
1591
                # it assumes that doc_to_target returns a number.
1592
1593
                choices = self.doc_to_choice(doc)
                gold = choices[gold]
1594
1595
            # we expect multiple_targets to be a list.
            elif self.multiple_target:
baberabb's avatar
baberabb committed
1596
                gold = list(gold)
Chris's avatar
Chris committed
1597
1598
1599
            elif type(gold) != type(result):
                # cast gold to the same type as result
                gold = type(result)(gold)
1600

lintangsutawika's avatar
lintangsutawika committed
1601
            for metric in self._metric_fn_list.keys():
haileyschoelkopf's avatar
haileyschoelkopf committed
1602
1603
1604
1605
1606
                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
1607
1608
1609
1610
                    if not isinstance(gold, list):
                        # sometimes, a multiple_target dataset has exceptions where one doc has only one string answer
                        # print(gold)
                        gold = [gold]
1611
1612
1613
1614
1615
1616
1617
1618
                    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
1619
                    else:
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
                        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
1641
                else:
1642
                    try:
1643
                        result_score = self._metric_fn_list[metric](
1644
1645
                            references=[gold],
                            predictions=[result],
1646
                            **self._metric_fn_kwargs[metric],
1647
                        )
1648
                    except TypeError:  # needed for now in order to use a different interface between our own metrics and HF Evaluate metrics
1649
                        result_score = self._metric_fn_list[metric]([gold, result])
1650
1651
1652
1653
                    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
1654
        else:
lintangsutawika's avatar
lintangsutawika committed
1655
1656
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
1657
                "'loglikelihood', 'loglikelihood_rolling', 'generate_until' or 'multiple_choice'",
1658
            )
1659
1660
1661

        return result_dict

Baber Abbasi's avatar
Baber Abbasi committed
1662
    def aggregation(self) -> dict:
1663
1664
        return self._aggregation_list

Baber Abbasi's avatar
Baber Abbasi committed
1665
    def higher_is_better(self) -> dict:
haileyschoelkopf's avatar
haileyschoelkopf committed
1666
        return self._higher_is_better
1667

Baber Abbasi's avatar
Baber Abbasi committed
1668
1669
1670
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

lintangsutawika's avatar
lintangsutawika committed
1671
1672
1673
    @property
    def task_name(self) -> Any:
        return getattr(self.config, "task", None)
lintangsutawika's avatar
lintangsutawika committed
1674

1675
1676
1677
1678
1679
1680
1681
1682
    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)})"
        )

1683
1684

class MultipleChoiceTask(Task):
1685
    OUTPUT_TYPE = "loglikelihood"
1686

baberabb's avatar
baberabb committed
1687
    def doc_to_target(self, doc: dict) -> str:
1688
1689
        return " " + doc["choices"][doc["gold"]]

baberabb's avatar
baberabb committed
1690
    def construct_requests(self, doc: dict, ctx: str, **kwargs) -> List[Instance]:
1691
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1692
1693
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1694
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1695
                doc=doc,
1696
                arguments=(ctx, " {}".format(choice)),
1697
                idx=i,
1698
1699
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1700
1701
            for i, choice in enumerate(doc["choices"])
        ]
1702

1703
    def process_results(self, doc: dict, results: Iterable[Tuple[float, bool]]) -> dict:
lintangsutawika's avatar
lintangsutawika committed
1704
1705
1706
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
        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
1718
    def higher_is_better(self) -> dict:
1719
1720
1721
1722
1723
        return {
            "acc": True,
            "acc_norm": True,
        }

baberabb's avatar
baberabb committed
1724
    def aggregation(self) -> dict:
1725
1726
1727
1728
1729
1730
        return {
            "acc": mean,
            "acc_norm": mean,
        }


lintangsutawika's avatar
lintangsutawika committed
1731
class PerplexityTask(Task):
1732
1733
    OUTPUT_TYPE = "loglikelihood_rolling"

baberabb's avatar
baberabb committed
1734
    def has_training_docs(self) -> bool:
1735
1736
        return False

baberabb's avatar
baberabb committed
1737
    def fewshot_examples(self, k: int, rnd) -> List:
1738
1739
1740
1741
        if k != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1742
1743
        return []

baberabb's avatar
baberabb committed
1744
    def fewshot_context(self, doc: dict, num_fewshot: int) -> Literal[""]:
1745
1746
1747
1748
        if num_fewshot != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1749
1750
1751

        return ""

baberabb's avatar
baberabb committed
1752
    def higher_is_better(self) -> dict:
1753
1754
1755
1756
1757
1758
1759
1760
1761
        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
1762
    def doc_to_text(self, doc) -> str:
1763
1764
1765
1766
1767
        return ""

    def doc_to_target(self, doc):
        return doc

1768
1769
1770
    def construct_requests(self, doc: dict, ctx: Optional[str], **kwargs):
        if bool(ctx):
            raise ValueError
1771

lintangsutawika's avatar
lintangsutawika committed
1772
1773
1774
1775
1776
1777
1778
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1779

1780
    def process_results(self, doc: dict, results: Tuple[float]) -> dict:
1781
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1782
1783
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1784
1785
1786
1787
1788
1789
        return {
            "word_perplexity": (loglikelihood, words),
            "byte_perplexity": (loglikelihood, bytes_),
            "bits_per_byte": (loglikelihood, bytes_),
        }

baberabb's avatar
baberabb committed
1790
    def aggregation(self) -> dict:
1791
1792
1793
1794
1795
1796
1797
        return {
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
            "bits_per_byte": bits_per_byte,
        }

    @classmethod
baberabb's avatar
baberabb committed
1798
    def count_bytes(cls, doc) -> int:
1799
1800
1801
        return len(doc.encode("utf-8"))

    @classmethod
baberabb's avatar
baberabb committed
1802
    def count_words(cls, doc) -> int:
1803
1804
        """Downstream tasks with custom word boundaries should override this!"""
        return len(re.split(r"\s+", doc))