"comfyui/server.py" did not exist on "f7adb62a30a98b04317e79bf772acdd5f692877d"
task.py 57.1 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
8
from dataclasses import asdict, dataclass
9
from inspect import getsource
10
11
12
13
14
15
16
17
18
19
20
21
from typing import (
    Any,
    Dict,
    Iterable,
    Iterator,
    List,
    Literal,
    Mapping,
    Optional,
    Tuple,
    Union,
)
22
23
24

import datasets
import numpy as np
25
from tqdm import tqdm
26
27

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

43

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

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

lintangsutawika's avatar
lintangsutawika committed
53

lintangsutawika's avatar
lintangsutawika committed
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
@dataclass
class GroupConfig(dict):
    group: str = None
    task: Union[str, list] = None
    weight_by_size: bool = False

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

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

    def to_dict(self):
        return asdict(self)

lintangsutawika's avatar
lintangsutawika committed
69

70
71
@dataclass
class TaskConfig(dict):
72
    # task naming/registry
73
74
75
76
    task: Optional[str] = None
    task_alias: Optional[str] = None
    group: Optional[Union[str, list]] = None
    group_alias: Optional[Union[str, list]] = None
77
78
79
    # HF dataset options.
    # which dataset to use,
    # and what splits for what purpose
80
81
82
83
84
85
86
87
88
    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 (?)
89
90
    # formatting / prompting options.
    # see docs/advanced_task_guide.md for more info
91
92
93
94
95
96
    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
97
    description: str = ""
98
99
    target_delimiter: str = " "
    fewshot_delimiter: str = "\n\n"
100
    fewshot_config: Optional[dict] = None
101
    # runtime configuration options
102
    num_fewshot: Optional[int] = None
103
    # scoring options
104
105
106
    metric_list: Optional[list] = None
    output_type: OutputType = "generate_until"
    generation_kwargs: Optional[dict] = None
107
    repeats: int = 1
108
    filter_list: Optional[Union[str, list]] = None
109
    should_decontaminate: bool = False
110
111
112
113
    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
114

Ethan Smith's avatar
Ethan Smith committed
115
    def __post_init__(self) -> None:
Lintang Sutawika's avatar
Lintang Sutawika committed
116
        if self.generation_kwargs is not None:
117
            if self.output_type != "generate_until":
118
                raise ValueError(
119
                    f"[{self.task}] passed `generation_kwargs`, but not using `output_type: generate_until`!"
Lintang Sutawika's avatar
Lintang Sutawika committed
120
121
122
123
124
125
126
127
                )

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

            if "until" not in self.generation_kwargs:
128
                self.generation_kwargs["until"] = [self.fewshot_delimiter]
Lintang Sutawika's avatar
Lintang Sutawika committed
129
        else:
130
            if self.output_type == "generate_until":
Lintang Sutawika's avatar
Lintang Sutawika committed
131
132
                # ensure that we greedily generate in absence of explicit arguments otherwise
                self.generation_kwargs = {
133
134
135
136
137
                    "until": (
                        None
                        if self.fewshot_delimiter is None
                        else [self.fewshot_delimiter]
                    ),
Lintang Sutawika's avatar
Lintang Sutawika committed
138
139
                    "do_sample": False,
                }
140

141
142
143
    def __getitem__(self, item):
        return getattr(self, item)

144
145
146
    def __setitem__(self, item, value):
        return setattr(self, item, value)

147
    def to_dict(self, keep_callable: bool = False) -> dict:
148
149
        """dumps the current config as a dictionary object, as a printable format.
        null fields will not be printed.
haileyschoelkopf's avatar
haileyschoelkopf committed
150
        Used for dumping results alongside full task configuration
151

haileyschoelkopf's avatar
haileyschoelkopf committed
152
153
154
155
156
157
158
159
160
161
        :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)
162
163
164
165
166
167
168
169
170
171
            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
172
        return cfg_dict
173

174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
    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)

190
191
192
193
194
195
196
197
198
199
200

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)
    """

201
    VERSION: Optional[Union[int, str]] = None
202

203
204
    # The name of the `Task` benchmark as denoted in the HuggingFace datasets Hub
    # or a path to a custom `datasets` loading script.
205
    DATASET_PATH: Optional[str] = None
206
207

    # The name of a subset within `DATASET_PATH`.
208
    DATASET_NAME: Optional[str] = None
209

210
    OUTPUT_TYPE: Optional[OutputType] = None
lintangsutawika's avatar
lintangsutawika committed
211

212
213
    def __init__(
        self,
214
215
216
217
        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
218
    ) -> None:
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
        """
        :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)
241
242
243
        self._training_docs: Optional[list] = None
        self._fewshot_docs: Optional[list] = None
        self._instances: Optional[List[Instance]] = None
244

245
        self._config: TaskConfig = TaskConfig({**config}) if config else TaskConfig()
246

lintangsutawika's avatar
lintangsutawika committed
247
        self._filters = [build_filter_ensemble("none", [["take_first", None]])]
248

249
250
251
252
253
254
    def download(
        self,
        data_dir: Optional[str] = None,
        cache_dir: Optional[str] = None,
        download_mode=None,
    ) -> None:
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
        """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.
        """
279
280
281
282
283
284
285
        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,
        )
286

287
    @property
288
    def config(self) -> TaskConfig:
289
290
291
        """Returns the TaskConfig associated with this class."""
        return self._config

292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
    @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

307
    def training_docs(self) -> Iterable:
308
309
310
311
312
313
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

314
    def validation_docs(self) -> Iterable:
315
316
317
318
319
320
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

321
    def test_docs(self) -> Iterable:
322
323
324
325
326
327
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

328
    def fewshot_docs(self) -> Iterable:
329
330
331
332
333
334
335
336
337
        """
        :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
338
            eval_logger.warning(
339
                f"[Task: {self.config.task}] has_training_docs and has_validation_docs are False"
340
                ", using test_docs as fewshot_docs but this is not recommended."
lintangsutawika's avatar
lintangsutawika committed
341
            )
342
343
            return self.test_docs()

344
    def _process_doc(self, doc: dict) -> dict:
345
346
347
348
349
350
351
352
353
        """
        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
354

355
    @property
356
    def instances(self) -> List[Instance]:
357
358
359
360
361
362
363
364
365
366
367
        """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)

368
369
    def doc_to_decontamination_query(self, doc):
        raise NotImplementedError(
370
371
372
373
374
375
376
377
378
379
380
            "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

381
382
    def build_all_requests(
        self,
383
        *,
384
385
386
387
388
389
        limit=None,
        rank=None,
        world_size=None,
        cache_requests=False,
        rewrite_requests_cache=False,
    ) -> None:
390
        """Build a set of Instances for a task, and store them in task.instances"""
391
392
393
394

        # used with caching
        og_limit = limit

395
        cache_key = f"requests-{self._config.task}-{self.config.num_fewshot}shot-rank{rank}-world_size{world_size}"
396
397
398
399
400
401
402
403
404
405
406
407
408
409

        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
410

Baber Abbasi's avatar
Baber Abbasi committed
411
        eval_logger.info(f"Building contexts for {self.config.task} on rank {rank}...")
412

413
        instances = []
414
415
416
417
418
419
420
421
422
423

        # 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(
424
            self.doc_iterator(rank=rank, limit=limit, world_size=world_size)
425
426
427
428
429
430
431
        )

        num_docs = len(doc_id_docs)

        for doc_id, doc in tqdm(
            doc_id_docs,
            total=num_docs,
lintangsutawika's avatar
lintangsutawika committed
432
        ):
433
            # sample fewshot context #TODO: need to offset doc_id by rank now!
434
            fewshot_ctx = self.fewshot_context(
435
                doc,
436
                0 if self.config.num_fewshot is None else self.config.num_fewshot,
437
            )
438

439
            # TODO: we should override self.config.repeats if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
440
441
442
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
443
                metadata=(self.config["task"], doc_id, self.config.repeats),
lintangsutawika's avatar
lintangsutawika committed
444
            )
445
446
447
448

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

449
450
451
452
453
454
455
456
457
458
459
460
461
            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
462

463
464
        if len(self._instances) == 0:
            raise ValueError("task.build_requests() did not find any docs!")
465

466
467
468
        if cache_requests and (not cached_instances or rewrite_requests_cache):
            save_to_cache(file_name=cache_key, obj=instances)

469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
    @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
485
            The number of times each instance in a dataset is inferred on. Defaults to 1,
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
            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

521
522
523
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

haileyschoelkopf's avatar
haileyschoelkopf committed
524
525
526
527
528
529
530
531
532
533
    @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))

534
    @utils.positional_deprecated
lintangsutawika's avatar
lintangsutawika committed
535
    def fewshot_context(
536
537
538
539
540
        self,
        doc,
        num_fewshot,
        rnd=random.Random(1234),
        description=None,
lintangsutawika's avatar
lintangsutawika committed
541
    ):
542
543
544
545
546
547
548
        """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
549
550
551
552
553
        :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.
554
555
556
        :returns: str
            The fewshot context.
        """
557
558
559
560
        if rnd is None:
            raise ValueError(
                "A `random.Random` generator argument must be provided to `rnd`"
            )
lintangsutawika's avatar
lintangsutawika committed
561

562
        description = description if description else ""
563
564

        if num_fewshot == 0:
lintangsutawika's avatar
lintangsutawika committed
565
            labeled_examples = ""
566
        else:
lintangsutawika's avatar
lintangsutawika committed
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
            # 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
591
            )
592
593

        example = self.doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
594
        return description + labeled_examples + example
595

596
    def apply_filters(self) -> Optional[List[Instance]]:
Baber Abbasi's avatar
Baber Abbasi committed
597
        """Iterates over FilterEnsembles and applies them to instances"""
lintangsutawika's avatar
lintangsutawika committed
598
599
        if hasattr(self, "_filters"):
            for f in self._filters:
600
                f.apply(self._instances)
lintangsutawika's avatar
lintangsutawika committed
601
602
603
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances
604

baberabb's avatar
baberabb committed
605
    def dump_config(self) -> dict:
Baber Abbasi's avatar
Baber Abbasi committed
606
        """Returns the config as a dictionary."""
607
        # TODO: this should only return the overrides applied to a non-YAML task's configuration.
608
        # (num_fewshot)
609
        return self.config.to_dict()
610

Baber Abbasi's avatar
Baber Abbasi committed
611
612
613
614
615
616
617
618
619
620
621
622
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
    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)

651
652
653
654
655
656
657
    @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:
658
659
660
            raise ValueError(
                f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"
            )
661
662
663
664
665
666
667
668
669
670
671
672
673

    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

674
675

class ConfigurableTask(Task):
676
    VERSION = "Yaml"
677
    OUTPUT_TYPE = None
678
    CONFIG = None
679
680

    def __init__(
681
682
683
684
685
        self,
        data_dir=None,
        cache_dir=None,
        download_mode=None,
        config: Optional[dict] = None,
Ethan Smith's avatar
Ethan Smith committed
686
    ) -> None:  # TODO no super() call here
687
        # Get pre-configured attributes
688
        self._config = self.CONFIG
689

690
        # Use new configurations if there was no preconfiguration
691
        if self.config is None:
692
            self._config = TaskConfig(**config)
693
694
        # Overwrite configs
        else:
lintangsutawika's avatar
lintangsutawika committed
695
            if config is not None:
696
                self._config.__dict__.update(config)
697

698
        if self.config is None:
lintangsutawika's avatar
lintangsutawika committed
699
700
701
            raise ValueError(
                "Must pass a config to ConfigurableTask, either in cls.CONFIG or `config` kwarg"
            )
702

703
704
705
706
        if isinstance(self.config.metadata, dict):
            if "version" in self.config.metadata:
                self.VERSION = self.config.metadata["version"]

707
        if self.config.output_type is not None:
708
709
710
711
            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)}'"
                )
712
            self.OUTPUT_TYPE = self.config.output_type
713

714
715
        if self.config.dataset_path is not None:
            self.DATASET_PATH = self.config.dataset_path
716

717
718
        if self.config.dataset_name is not None:
            self.DATASET_NAME = self.config.dataset_name
719

720
721
722
723
        self._metric_fn_list = {}
        self._metric_fn_kwargs = {}
        self._aggregation_list = {}
        self._higher_is_better = {}
724

725
        if self.config.metric_list is None:
726
            # TODO: handle this in TaskConfig.__post_init__ ?
727
728
            _metric_list = DEFAULT_METRIC_REGISTRY[self.config.output_type]

729
            for metric_name in _metric_list:
haileyschoelkopf's avatar
haileyschoelkopf committed
730
                self._metric_fn_list[metric_name] = get_metric(metric_name)
lintangsutawika's avatar
lintangsutawika committed
731
                self._metric_fn_kwargs[metric_name] = {}
732
733
734
                self._aggregation_list[metric_name] = get_metric_aggregation(
                    metric_name
                )
haileyschoelkopf's avatar
haileyschoelkopf committed
735
                self._higher_is_better[metric_name] = is_higher_better(metric_name)
736
        else:
737
            for metric_config in self.config.metric_list:
738
739
740
741
                if "metric" not in metric_config:
                    raise ValueError(
                        "'metric' key not provided for an entry in 'metric_list', must be specified!"
                    )
742
743
744
745
                metric_name = metric_config["metric"]
                kwargs = {
                    key: metric_config[key]
                    for key in metric_config
Chris's avatar
Chris committed
746
747
                    if key
                    not in ["metric", "aggregation", "higher_is_better", "hf_evaluate"]
748
                }
Chris's avatar
Chris committed
749
750
751
752
                hf_evaluate_metric = (
                    "hf_evaluate" in metric_config
                    and metric_config["hf_evaluate"] is True
                )
753

754
                if self.config.process_results is not None:
755
756
                    self._metric_fn_list[metric_name] = None
                    self._metric_fn_kwargs[metric_name] = {}
757
758
759
760
761
762
                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
763
764
765
                    self._metric_fn_list[metric_name] = get_metric(
                        metric_name, hf_evaluate_metric
                    )
766
                    self._metric_fn_kwargs[metric_name] = kwargs
lintangsutawika's avatar
lintangsutawika committed
767

768
                if "aggregation" in metric_config:
769
                    agg_name = metric_config["aggregation"]
770
                    if isinstance(agg_name, str):
haileyschoelkopf's avatar
haileyschoelkopf committed
771
                        self._aggregation_list[metric_name] = get_aggregation(agg_name)
772
                    elif callable(agg_name):  # noqa: E721
773
774
775
                        self._aggregation_list[metric_name] = metric_config[
                            "aggregation"
                        ]
776
                else:
777
                    INV_AGG_REGISTRY = {v: k for k, v in AGGREGATION_REGISTRY.items()}
lintangsutawika's avatar
lintangsutawika committed
778
                    metric_agg = get_metric_aggregation(metric_name)
779
                    eval_logger.warning(
780
                        f"[Task: {self.config.task}] metric {metric_name} is defined, but aggregation is not. "
781
782
                        f"using default "
                        f"aggregation={INV_AGG_REGISTRY[metric_agg]}"
783
                    )
784
                    self._aggregation_list[metric_name] = metric_agg
lintangsutawika's avatar
lintangsutawika committed
785

786
787
788
789
790
791
                if "higher_is_better" in metric_config:
                    self._higher_is_better[metric_name] = metric_config[
                        "higher_is_better"
                    ]
                else:
                    eval_logger.warning(
792
                        f"[Task: {self.config.task}] metric {metric_name} is defined, but higher_is_better is not. "
793
                        f"using default "
haileyschoelkopf's avatar
haileyschoelkopf committed
794
                        f"higher_is_better={is_higher_better(metric_name)}"
795
                    )
haileyschoelkopf's avatar
haileyschoelkopf committed
796
                    self._higher_is_better[metric_name] = is_higher_better(metric_name)
797

798
        self.download(self.config.dataset_kwargs)
799
800
801
        self._training_docs = None
        self._fewshot_docs = None

802
        if self.config.filter_list is not None:
lintangsutawika's avatar
lintangsutawika committed
803
            self._filters = []
804
            for filter_config in self.config.filter_list:
805
806
807
808
809
810
811
812
813
                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
814
                self._filters.append(filter_pipeline)
lintangsutawika's avatar
lintangsutawika committed
815
        else:
816
            self._filters = [build_filter_ensemble("none", [["take_first", None]])]
817

818
819
        if self.config.use_prompt is not None:
            eval_logger.info(f"loading prompt {self.config.use_prompt}")
820
            self.prompt = get_prompt(
821
                self.config.use_prompt, self.DATASET_PATH, self.DATASET_NAME
lintangsutawika's avatar
lintangsutawika committed
822
            )
823
824
825
        else:
            self.prompt = None

lintangsutawika's avatar
lintangsutawika committed
826
        if self.fewshot_docs() is not None:
haileyschoelkopf's avatar
haileyschoelkopf committed
827
            self.sampler = samplers.get_sampler(
haileyschoelkopf's avatar
haileyschoelkopf committed
828
829
830
                self.config.fewshot_config.get("sampler", "default")
                if self.config.fewshot_config
                else "default"
haileyschoelkopf's avatar
haileyschoelkopf committed
831
            )(list(self.fewshot_docs()), self, rnd=random.Random(1234))
832

833
        self.task_docs = self.eval_docs
834

835
        # Test One Doc
836
        self.features = list(self.task_docs.features.keys())
837
838
        self.multiple_input = 0
        self.multiple_target = 0
839
        test_doc = self.task_docs[0]
840
        test_text = self.doc_to_text(test_doc)
841
        test_target = self.doc_to_target(test_doc)
lintangsutawika's avatar
lintangsutawika committed
842

843
        if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
844
            test_choice = self.doc_to_choice(test_doc)
845
            if not isinstance(test_choice, list):
lintangsutawika's avatar
lintangsutawika committed
846
                eval_logger.error("doc_to_choice must return list")
847
848
            else:
                num_choice = len(test_choice)
849

850
            if isinstance(test_text, int):
851
                self.multiple_input = num_choice
852
853
        else:
            test_choice = None
854

855
        if isinstance(test_target, list):
856
            self.multiple_target = len(test_target)
857
        else:
858
            if (isinstance(test_target, int)) and (test_choice is not None):
lintangsutawika's avatar
lintangsutawika committed
859
                test_target = test_choice[test_target]
860
            else:
lintangsutawika's avatar
lintangsutawika committed
861
                test_target = str(test_target)
862

863
864
865
        if test_choice is not None:
            check_choices = test_choice
        else:
lintangsutawika's avatar
lintangsutawika committed
866
            check_choices = [test_target]
867
868
869
870
        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 = (
871
872
                    True
                    if self.config.target_delimiter.rstrip()
873
                    != self.config.target_delimiter
874
                    else False
875
                )
876

877
                if delimiter_has_whitespace and choice_has_whitespace:
878
879
                    eval_logger.debug(
                        f'Both target_delimiter "{self.config.target_delimiter}" and target choice: "{choice}" have whitespace'
880
881
                    )
                elif (not delimiter_has_whitespace) and (not choice_has_whitespace):
882
                    eval_logger.debug(
883
                        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'
884
885
                    )

886
    def download(self, dataset_kwargs: Optional[Dict[str, Any]] = None) -> None:
887
888
889
890
891
892
        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
893
    def has_training_docs(self) -> bool:
894
        if self.config.training_split is not None:
895
896
897
898
            return True
        else:
            return False

baberabb's avatar
baberabb committed
899
    def has_validation_docs(self) -> bool:
900
        if self.config.validation_split is not None:
901
902
903
904
            return True
        else:
            return False

baberabb's avatar
baberabb committed
905
    def has_test_docs(self) -> bool:
906
        if self.config.test_split is not None:
907
908
909
910
            return True
        else:
            return False

baberabb's avatar
baberabb committed
911
    def training_docs(self) -> datasets.Dataset:
912
        if self.has_training_docs():
913
914
915
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.training_split]
916
                )
917
            return self.dataset[self.config.training_split]
918

baberabb's avatar
baberabb committed
919
    def validation_docs(self) -> datasets.Dataset:
920
        if self.has_validation_docs():
921
922
923
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.validation_split]
924
                )
925
            return self.dataset[self.config.validation_split]
926

baberabb's avatar
baberabb committed
927
    def test_docs(self) -> datasets.Dataset:
928
        if self.has_test_docs():
929
930
931
            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]
932

933
    def fewshot_docs(self):
934
        if self.config.fewshot_split is not None:
935
936
            if self.config.process_docs is not None:
                return self.config.process_docs(self.dataset[self.config.fewshot_split])
937
            return self.dataset[self.config.fewshot_split]
938
        else:
939
            if (self.config.num_fewshot is not None) and (self.config.num_fewshot > 0):
940
                eval_logger.warning(
941
                    f"Task '{self.config.task}': "
942
943
944
945
                    "num_fewshot > 0 but fewshot_split is None. "
                    "using preconfigured rule."
                )
            return super().fewshot_docs()
946

lintangsutawika's avatar
lintangsutawika committed
947
    @utils.positional_deprecated
948
    def fewshot_context(self, doc: str, num_fewshot: int) -> str:
lintangsutawika's avatar
lintangsutawika committed
949
950
951
952
953
954
955
956
957
958
        """Returns a fewshot context string that is made up of a prepended description
        (if provided), the `num_fewshot` number of examples, and an appended prompt example.

        :param doc: str
            The document as returned from training_docs, validation_docs, or test_docs.
        :param num_fewshot: int
            The number of fewshot examples to provide in the returned context string.
        :returns: str
            The fewshot context.
        """
959
960
        if description := self.config.description:
            description = utils.apply_template(self.config.description, doc)
lintangsutawika's avatar
lintangsutawika committed
961
962
963

        if num_fewshot == 0:
            # always prepend the (possibly empty) task description
964
            labeled_examples = description
lintangsutawika's avatar
lintangsutawika committed
965
        else:
966
            labeled_examples = description + self.sampler.get_context(doc, num_fewshot)
lintangsutawika's avatar
lintangsutawika committed
967
968

        example = self.doc_to_text(doc)
969
970
971
972
973
974
975
976
977
978
979
980
981
        if self.multiple_input:
            return labeled_examples
        else:
            if isinstance(example, str):
                return labeled_examples + example
            elif isinstance(example, list):
                return [labeled_examples + ex for ex in example]
            elif isinstance(example, int):
                if self.config.doc_to_choice is not None:
                    choices = self.doc_to_choice(doc)
                    return labeled_examples + choices[example]
                else:
                    return labeled_examples + str(example)
lintangsutawika's avatar
lintangsutawika committed
982

983
    def apply_filters(self):
Baber Abbasi's avatar
Baber Abbasi committed
984
        """Iterates over FilterEnsembles and applies them to instances"""
985
986
        if hasattr(self, "_filters"):
            for f in self._filters:
987
                f.apply(self._instances)
988
989
990
991
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances

992
    def should_decontaminate(self):
993
        return self.config.should_decontaminate
994
995

    def doc_to_decontamination_query(self, doc):
996
        if self.config.should_decontaminate:
997
998
            if self.config.doc_to_decontamination_query is None:
                return self.doc_to_text(doc)
999
            else:
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
                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
                        )
                    )
1011

1012
    def _process_doc(self, doc: dict) -> dict:
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
        """
        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):
1024
1025
        if self.prompt is not None:
            doc_to_text = self.prompt
1026
        else:
1027
            doc_to_text = self.config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
1028

1029
        if isinstance(doc_to_text, int):
1030
            return doc_to_text
1031
        elif isinstance(doc_to_text, str):
1032
            if doc_to_text in self.features:
1033
                # if self.config.doc_to_choice is not None:
1034
1035
                #     return self.doc_to_choice(doc)[doc[doc_to_text]]
                # else:
1036
1037
                return doc[doc_to_text]
            else:
lintangsutawika's avatar
lintangsutawika committed
1038
                text_string = utils.apply_template(doc_to_text, doc)
lintangsutawika's avatar
lintangsutawika committed
1039
                if text_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1040
1041
1042
                    return ast.literal_eval(text_string)
                else:
                    return text_string
1043
        elif callable(doc_to_text):
1044
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
1045
        # Used when applying a Promptsource template
1046
        elif hasattr(doc_to_text, "apply"):
1047
1048
1049
1050
1051
            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")
1052
                return self.config.fewshot_delimiter
1053
        else:
1054
            print(type(doc_to_text))
1055
            raise TypeError
1056

1057
    def doc_to_target(self, doc: Mapping) -> Union[int, str, list]:
1058
1059
        if self.prompt is not None:
            doc_to_target = self.prompt
1060
        else:
1061
            doc_to_target = self.config.doc_to_target
1062

1063
        if isinstance(doc_to_target, int):
1064
            return doc_to_target
1065
        elif isinstance(doc_to_target, str):
1066
            if doc_to_target in self.features:
1067
                # if self.config.doc_to_choice is not None:
1068
1069
1070
                #     return self.doc_to_choice(doc)[doc[doc_to_target]]
                # else:
                return doc[doc_to_target]
1071
            else:
lintangsutawika's avatar
lintangsutawika committed
1072
                target_string = utils.apply_template(doc_to_target, doc)
lintangsutawika's avatar
lintangsutawika committed
1073
                if target_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1074
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
1075
1076
1077
1078
1079
                elif (
                    len(target_string) >= 2
                    and (target_string[0] == "[")
                    and (target_string[-1] == "]")
                ):
baberabb's avatar
baberabb committed
1080
1081
1082
1083
                    try:
                        return ast.literal_eval(target_string)
                    except (SyntaxError, ValueError):
                        return target_string
lintangsutawika's avatar
lintangsutawika committed
1084
1085
                else:
                    return target_string
1086
        elif isinstance(doc_to_target, list):
1087
            return doc_to_target
1088
        elif callable(doc_to_target):
1089
            return doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
1090
        # Used when applying a Promptsource template
1091
        elif hasattr(doc_to_target, "apply"):
1092
            applied_prompt = doc_to_target.apply(doc)
1093
1094
1095
1096
            if len(applied_prompt) == 2:
                return applied_prompt[1]
            else:
                eval_logger.warning("Applied prompt returns empty string")
1097
                return self.config.fewshot_delimiter
1098
1099
        else:
            raise TypeError
1100

baberabb's avatar
baberabb committed
1101
    def doc_to_choice(self, doc: Any) -> List[str]:
1102
1103
        if self.prompt is not None:
            doc_to_choice = self.prompt
1104
        elif self.config.doc_to_choice is None:
1105
1106
            eval_logger.error("doc_to_choice was called but not set in config")
        else:
1107
            doc_to_choice = self.config.doc_to_choice
1108

1109
        if isinstance(doc_to_choice, str):
1110
1111
1112
1113
            if doc_to_choice in self.features:
                return doc[doc_to_choice]
            else:
                return ast.literal_eval(utils.apply_template(doc_to_choice, doc))
1114
        elif isinstance(doc_to_choice, list):
1115
            return doc_to_choice
1116
        elif isinstance(doc_to_choice, dict):
1117
1118
1119
1120
1121
1122
1123
            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
1124

baberabb's avatar
baberabb committed
1125
1126
1127
    def construct_requests(
        self, doc: dict, ctx: str, **kwargs
    ) -> Union[List[Instance], Instance]:
1128
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
1129
            arguments = (ctx, self.doc_to_target(doc))
1130
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
1131
            arguments = (self.doc_to_target(doc),)
1132
        elif self.OUTPUT_TYPE == "multiple_choice":
1133
            choices = self.doc_to_choice(doc)
1134
            target_delimiter = self.config.target_delimiter
1135
1136
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
1137
                cont = self.doc_to_target(doc)
1138
1139
1140
                arguments = [
                    (ctx + choice, f"{target_delimiter}{cont}") for choice in choices
                ]
1141
            else:
1142
                # Otherwise they are placed in the continuation
1143
                arguments = [(ctx, f"{target_delimiter}{cont}") for cont in choices]
1144

1145
            request_list = [
1146
1147
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1148
                    doc=doc,
1149
                    arguments=arg,
1150
                    idx=i,
1151
1152
                    **kwargs,
                )
1153
                for i, arg in enumerate(arguments)
1154
            ]
1155
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
1156
            if "acc_mutual_info" in self._metric_fn_list.keys():
1157
1158
1159
                # if we are calculating multiple choice accuracy
                # using mutual information instead of raw loglikelihood as metric, need unconditional lls.

lintangsutawika's avatar
lintangsutawika committed
1160
                # here mutual info refers to calculating
1161
1162
1163
1164
1165
1166
                # 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
1167
                            doc=doc,
1168
                            arguments=("", "{}".format(choice)),
1169
1170
1171
                            idx=i,
                            **kwargs,
                        )
lintangsutawika's avatar
lintangsutawika committed
1172
                        for i, choice in enumerate(choices)
1173
1174
1175
                    ]
                )
            return request_list
lintangsutawika's avatar
lintangsutawika committed
1176

1177
        elif self.OUTPUT_TYPE == "generate_until":
1178
            arguments = (ctx, deepcopy(self.config.generation_kwargs))
lintangsutawika's avatar
lintangsutawika committed
1179
1180

        return Instance(
lintangsutawika's avatar
lintangsutawika committed
1181
1182
            request_type=self.OUTPUT_TYPE, doc=doc, arguments=arguments, idx=0, **kwargs
        )
1183
1184

    def process_results(self, doc, results):
1185
        if callable(self.config.process_results):
Lintang Sutawika's avatar
Lintang Sutawika committed
1186
            return self.config.process_results(doc, results)
lintangsutawika's avatar
lintangsutawika committed
1187

1188
        result_dict = {}
1189
        use_metric = list(self._metric_fn_list.keys())
1190
1191
1192
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
1193
1194
            prob_norm = np.exp(ll)

1195
1196
1197
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
1198
                **(
lintangsutawika's avatar
lintangsutawika committed
1199
                    {"brier_score": (0, [prob_norm])}  # Gold is Index 0
1200
1201
1202
                    if "brier_score" in use_metric
                    else {}
                ),
1203
            }
1204
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
1205
            (loglikelihood,) = results
1206
1207
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
haileyschoelkopf's avatar
haileyschoelkopf committed
1208
            return {
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
                **(
                    {"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
1224
            }
1225
        elif self.OUTPUT_TYPE == "multiple_choice":
1226
            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
1227

1228
            # retrieve choices in List[str] form, to compute choice lengths, etc.
1229
            choices = self.doc_to_choice(doc)
1230
1231
            completion_len = np.array([float(len(i)) for i in choices])

1232
1233
            if (
                2 * len(choices) == len(lls)
1234
                and "acc_mutual_info" in self._metric_fn_list.keys()
1235
1236
1237
1238
            ):
                # then we are doing mutual info.
                # this stores the "dryrun" / unconditional answer loglikelihoods
                lls_unconditional = lls[1::2]
1239
1240
                if len(lls_unconditional) != len(choices):
                    raise ValueError
1241
1242
                # and this stores our "regular" conditional loglikelihoods
                lls = lls[::2]
1243

1244
1245
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
1246

1247
1248
            if self.multiple_input:
                gold = self.doc_to_text(doc)
1249
            else:
1250
                gold = self.doc_to_target(doc)
1251
1252

            gold_index_error = False
1253
            if isinstance(gold, list):
Lintang Sutawika's avatar
Lintang Sutawika committed
1254
1255
                gold = [i if i < len(choices) else -100 for i in gold]
                if -100 in gold:
1256
1257
                    gold_index_error = True
            else:
1258
                if isinstance(gold, int):
Lintang Sutawika's avatar
Lintang Sutawika committed
1259
                    gold = gold if gold < len(choices) else -100
1260
                elif isinstance(gold, str):
Lintang Sutawika's avatar
Lintang Sutawika committed
1261
                    gold = choices.index(gold) if gold in choices else -100
lintangsutawika's avatar
lintangsutawika committed
1262

Lintang Sutawika's avatar
Lintang Sutawika committed
1263
                if gold == -100:
1264
1265
1266
1267
                    gold_index_error = True

            if gold_index_error:
                eval_logger.warning(
lintangsutawika's avatar
lintangsutawika committed
1268
                    f"Label index was not in within range of available choices,"
1269
1270
                    f"Sample:\n\n{doc}\n\n"
                )
lintangsutawika's avatar
lintangsutawika committed
1271

1272
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
1273
1274
                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
1275
                exact_match = int(any([is_greedy[i] if i != -100 else 0 for i in gold]))
lintangsutawika's avatar
lintangsutawika committed
1276
1277
1278
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
1279
                # 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
1280
                exact_match = int(is_greedy[gold]) if gold != -100 else 0
1281

lintangsutawika's avatar
lintangsutawika committed
1282
            prob_norm = utils.softmax(lls)
lintangsutawika's avatar
lintangsutawika committed
1283

lintangsutawika's avatar
lintangsutawika committed
1284
            # TODO use keyword arguments to the metric?
lintangsutawika's avatar
format  
lintangsutawika committed
1285
            # gold, pred, norm stuff, the original lls,
1286
            result_dict = {
1287
                **({"acc": acc} if "acc" in use_metric else {}),
1288
1289
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
1290
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
1291
                **({"exact_match": exact_match} if "exact_match" in use_metric else {}),
lintangsutawika's avatar
format  
lintangsutawika committed
1292
                **(
lintangsutawika's avatar
lintangsutawika committed
1293
                    # {"brier_score": (gold, prob_norm)}
1294
                    {"brier_score": [np.eye(len(prob_norm))[gold], prob_norm]}
lintangsutawika's avatar
format  
lintangsutawika committed
1295
1296
1297
                    if "brier_score" in use_metric
                    else {}
                ),
1298
1299
            }

1300
            if "acc_mutual_info" in use_metric:
lintangsutawika's avatar
lintangsutawika committed
1301
1302
1303
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
1304
1305
1306
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

1307
        elif self.OUTPUT_TYPE == "generate_until":
1308
            gold = self.doc_to_target(doc)
Chris's avatar
Chris committed
1309
            result = results[0]
1310
            if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1311
                # If you set doc_to_choice,
lintangsutawika's avatar
lintangsutawika committed
1312
                # it assumes that doc_to_target returns a number.
1313
1314
                choices = self.doc_to_choice(doc)
                gold = choices[gold]
1315
1316
            # we expect multiple_targets to be a list.
            elif self.multiple_target:
baberabb's avatar
baberabb committed
1317
                gold = list(gold)
Chris's avatar
Chris committed
1318
1319
1320
            elif type(gold) != type(result):
                # cast gold to the same type as result
                gold = type(result)(gold)
1321

lintangsutawika's avatar
lintangsutawika committed
1322
            for metric in self._metric_fn_list.keys():
haileyschoelkopf's avatar
haileyschoelkopf committed
1323
1324
1325
1326
1327
                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
1328
1329
1330
1331
                    if not isinstance(gold, list):
                        # sometimes, a multiple_target dataset has exceptions where one doc has only one string answer
                        # print(gold)
                        gold = [gold]
1332
1333
1334
1335
1336
1337
1338
1339
                    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
1340
                    else:
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
                        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
1362
                else:
1363
                    try:
1364
                        result_score = self._metric_fn_list[metric](
1365
1366
                            references=[gold],
                            predictions=[result],
1367
                            **self._metric_fn_kwargs[metric],
1368
                        )
1369
                    except TypeError:  # needed for now in order to use a different interface between our own metrics and HF Evaluate metrics
1370
                        result_score = self._metric_fn_list[metric]([gold, result])
1371
1372
1373
1374
                    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
1375
        else:
lintangsutawika's avatar
lintangsutawika committed
1376
1377
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
1378
                "'loglikelihood', 'loglikelihood_rolling', 'generate_until' or 'multiple_choice'",
1379
            )
1380
1381
1382

        return result_dict

Baber Abbasi's avatar
Baber Abbasi committed
1383
    def aggregation(self) -> dict:
1384
1385
        return self._aggregation_list

Baber Abbasi's avatar
Baber Abbasi committed
1386
    def higher_is_better(self) -> dict:
haileyschoelkopf's avatar
haileyschoelkopf committed
1387
        return self._higher_is_better
1388

Baber Abbasi's avatar
Baber Abbasi committed
1389
1390
1391
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

1392
1393
1394
1395
1396
1397
1398
1399
1400
    def __repr__(self):
        return (
            f"ConfigurableTask(task_name={getattr(self.config, 'task', None)},"
            f"group_name={getattr(self.config, 'group', None)},"
            f"output_type={self.OUTPUT_TYPE},"
            f"num_fewshot={getattr(self.config, 'num_fewshot', None)},"
            f"num_samples={len(self.eval_docs)})"
        )

1401
1402

class MultipleChoiceTask(Task):
1403
    OUTPUT_TYPE = "loglikelihood"
1404

baberabb's avatar
baberabb committed
1405
    def doc_to_target(self, doc: dict) -> str:
1406
1407
        return " " + doc["choices"][doc["gold"]]

baberabb's avatar
baberabb committed
1408
    def construct_requests(self, doc: dict, ctx: str, **kwargs) -> List[Instance]:
1409
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1410
1411
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1412
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1413
                doc=doc,
1414
                arguments=(ctx, " {}".format(choice)),
1415
                idx=i,
1416
1417
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1418
1419
            for i, choice in enumerate(doc["choices"])
        ]
1420

1421
    def process_results(self, doc: dict, results: Iterable[Tuple[float, bool]]) -> dict:
lintangsutawika's avatar
lintangsutawika committed
1422
1423
1424
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
        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
1436
    def higher_is_better(self) -> dict:
1437
1438
1439
1440
1441
        return {
            "acc": True,
            "acc_norm": True,
        }

baberabb's avatar
baberabb committed
1442
    def aggregation(self) -> dict:
1443
1444
1445
1446
1447
1448
        return {
            "acc": mean,
            "acc_norm": mean,
        }


lintangsutawika's avatar
lintangsutawika committed
1449
class PerplexityTask(Task):
1450
1451
    OUTPUT_TYPE = "loglikelihood_rolling"

baberabb's avatar
baberabb committed
1452
    def has_training_docs(self) -> bool:
1453
1454
        return False

baberabb's avatar
baberabb committed
1455
    def fewshot_examples(self, k: int, rnd) -> List:
1456
1457
1458
1459
        if k != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1460
1461
        return []

baberabb's avatar
baberabb committed
1462
    def fewshot_context(self, doc: dict, num_fewshot: int) -> Literal[""]:
1463
1464
1465
1466
        if num_fewshot != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1467
1468
1469

        return ""

baberabb's avatar
baberabb committed
1470
    def higher_is_better(self) -> dict:
1471
1472
1473
1474
1475
1476
1477
1478
1479
        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
1480
    def doc_to_text(self, doc) -> str:
1481
1482
1483
1484
1485
        return ""

    def doc_to_target(self, doc):
        return doc

1486
1487
1488
    def construct_requests(self, doc: dict, ctx: Optional[str], **kwargs):
        if bool(ctx):
            raise ValueError
1489

lintangsutawika's avatar
lintangsutawika committed
1490
1491
1492
1493
1494
1495
1496
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1497

1498
    def process_results(self, doc: dict, results: Tuple[float]) -> dict:
1499
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1500
1501
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1502
1503
1504
1505
1506
1507
        return {
            "word_perplexity": (loglikelihood, words),
            "byte_perplexity": (loglikelihood, bytes_),
            "bits_per_byte": (loglikelihood, bytes_),
        }

baberabb's avatar
baberabb committed
1508
    def aggregation(self) -> dict:
1509
1510
1511
1512
1513
1514
1515
        return {
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
            "bits_per_byte": bits_per_byte,
        }

    @classmethod
baberabb's avatar
baberabb committed
1516
    def count_bytes(cls, doc) -> int:
1517
1518
1519
        return len(doc.encode("utf-8"))

    @classmethod
baberabb's avatar
baberabb committed
1520
    def count_words(cls, doc) -> int:
1521
1522
        """Downstream tasks with custom word boundaries should override this!"""
        return len(re.split(r"\s+", doc))