task.py 50.2 KB
Newer Older
1
import abc
2
import ast
lintangsutawika's avatar
lintangsutawika committed
3
import logging
4
import random
5
6
7
8
import re
from collections.abc import Callable
from dataclasses import asdict, dataclass
from typing import Any, List, Literal, Tuple, Union
9
10
11
12
13

import datasets
import numpy as np

from lm_eval import utils
14
from lm_eval.api import samplers
haileyschoelkopf's avatar
haileyschoelkopf committed
15
from lm_eval.api.instance import Instance
lintangsutawika's avatar
lintangsutawika committed
16
from lm_eval.api.metrics import (
17
    bits_per_byte,
lintangsutawika's avatar
lintangsutawika committed
18
19
20
21
    mean,
    weighted_perplexity,
)
from lm_eval.api.registry import (
22
23
    AGGREGATION_REGISTRY,
    DEFAULT_METRIC_REGISTRY,
haileyschoelkopf's avatar
haileyschoelkopf committed
24
    get_aggregation,
25
    get_metric,
26
    get_metric_aggregation,
haileyschoelkopf's avatar
haileyschoelkopf committed
27
    is_higher_better,
lintangsutawika's avatar
lintangsutawika committed
28
)
29
30
31
from lm_eval.filters import build_filter_ensemble
from lm_eval.prompts import get_prompt

32

33
34
35
36
ALL_OUTPUT_TYPES = [
    "loglikelihood",
    "multiple_choice",
    "loglikelihood_rolling",
37
    "generate_until",
38
39
]

lintangsutawika's avatar
lintangsutawika committed
40

41
eval_logger = logging.getLogger("lm-eval")
42

lintangsutawika's avatar
lintangsutawika committed
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
@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
58

59
60
@dataclass
class TaskConfig(dict):
61
    # task naming/registry
62
    task: str = None
lintangsutawika's avatar
lintangsutawika committed
63
    task_alias: str = None
64
    group: Union[str, list] = None
lintangsutawika's avatar
lintangsutawika committed
65
    group_alias: Union[str, list] = None
66
67
68
    # HF dataset options.
    # which dataset to use,
    # and what splits for what purpose
69
70
    dataset_path: str = None
    dataset_name: str = None
71
    dataset_kwargs: dict = None
72
73
74
    training_split: str = None
    validation_split: str = None
    test_split: str = None
lintangsutawika's avatar
lintangsutawika committed
75
    fewshot_split: str = None  # TODO: assert that this not None if num_fewshot > 0. (?) assert if this is same split as one evaling (?)
76
77
    # formatting / prompting options.
    # see docs/advanced_task_guide.md for more info
78
    process_docs: Callable = None
79
80
    doc_to_text: Union[Callable, str] = None
    doc_to_target: Union[Callable, str] = None
lintangsutawika's avatar
lintangsutawika committed
81
    doc_to_choice: Union[Callable, str, dict, list] = None
lintangsutawika's avatar
lintangsutawika committed
82
    process_results: Union[Callable, str] = None
83
    use_prompt: str = None
84
    description: str = ""
85
86
    target_delimiter: str = " "
    fewshot_delimiter: str = "\n\n"
haileyschoelkopf's avatar
haileyschoelkopf committed
87
    fewshot_config: dict = None
88
    # runtime configuration options
89
    num_fewshot: int = None
90
    # scoring options
91
    metric_list: list = None
92
    output_type: str = "generate_until"
93
    generation_kwargs: dict = None
94
    repeats: int = 1
lintangsutawika's avatar
lintangsutawika committed
95
    filter_list: Union[str, list] = None
96
97
    should_decontaminate: bool = False
    doc_to_decontamination_query: str = None
98

lintangsutawika's avatar
lintangsutawika committed
99
100
101
    metadata: Union[
        str, list
    ] = None  # by default, not used in the code. allows for users to pass arbitrary info to tasks
102

Ethan Smith's avatar
Ethan Smith committed
103
    def __post_init__(self) -> None:
Lintang Sutawika's avatar
Lintang Sutawika committed
104
        if self.generation_kwargs is not None:
105
            if self.output_type != "generate_until":
Lintang Sutawika's avatar
Lintang Sutawika committed
106
                eval_logger.warning(
107
                    f"[{self.task}] passed `generation_kwargs`, but not using `output_type: generate_until`!"
Lintang Sutawika's avatar
Lintang Sutawika committed
108
                )
109
                assert self.output_type != "generate_until"
Lintang Sutawika's avatar
Lintang Sutawika committed
110
111
112
113
114
115
116

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

            if "until" not in self.generation_kwargs:
117
                self.generation_kwargs["until"] = [self.fewshot_delimiter]
Lintang Sutawika's avatar
Lintang Sutawika committed
118
        else:
119
            if self.output_type == "generate_until":
Lintang Sutawika's avatar
Lintang Sutawika committed
120
121
                # ensure that we greedily generate in absence of explicit arguments otherwise
                self.generation_kwargs = {
Lintang Sutawika's avatar
Lintang Sutawika committed
122
                    "until": None
123
124
                    if self.fewshot_delimiter is None
                    else [self.fewshot_delimiter],
Lintang Sutawika's avatar
Lintang Sutawika committed
125
126
                    "do_sample": False,
                }
127

haileyschoelkopf's avatar
haileyschoelkopf committed
128
        # TODO: how to make TaskConfigs be de- and re-serializable, even when using the !function constructor?
129
130
        # if self.dataset_kwargs is None:
        #     self.dataset_kwargs = {"trust_remote_code": True}
haileyschoelkopf's avatar
haileyschoelkopf committed
131

132
133
134
    def __getitem__(self, item):
        return getattr(self, item)

135
136
137
    def __setitem__(self, item, value):
        return setattr(self, item, value)

Lintang Sutawika's avatar
Lintang Sutawika committed
138
    def to_dict(self, keep_callable=False):
139
140
        """dumps the current config as a dictionary object, as a printable format.
        null fields will not be printed.
haileyschoelkopf's avatar
haileyschoelkopf committed
141
        Used for dumping results alongside full task configuration
142

haileyschoelkopf's avatar
haileyschoelkopf committed
143
144
145
146
147
148
149
150
151
152
        :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)
haileyschoelkopf's avatar
haileyschoelkopf committed
153
            elif isinstance(v, Callable):
Lintang Sutawika's avatar
Lintang Sutawika committed
154
155
156
157
158
                if keep_callable:
                    cfg_dict[k] = v
                else:
                    # TODO: this should handle Promptsource template objects as a separate case?
                    cfg_dict[k] = str(v)
haileyschoelkopf's avatar
haileyschoelkopf committed
159
        return cfg_dict
160

161
162
163
164
165
166
167
168
169
170
171
172

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)
    """

    VERSION = None
173

174
175
176
177
178
179
180
181
    # The name of the `Task` benchmark as denoted in the HuggingFace datasets Hub
    # or a path to a custom `datasets` loading script.
    DATASET_PATH: str = None

    # The name of a subset within `DATASET_PATH`.
    DATASET_NAME: str = None

    OUTPUT_TYPE: str = None
lintangsutawika's avatar
lintangsutawika committed
182

183
184
185
186
187
188
    def __init__(
        self,
        data_dir=None,
        cache_dir=None,
        download_mode=None,
        config=None,
Ethan Smith's avatar
Ethan Smith committed
189
    ) -> None:
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
        """
        :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)
        self._training_docs = None
        self._fewshot_docs = None
        self._instances = None

lintangsutawika's avatar
lintangsutawika committed
216
        self._config = TaskConfig({**config}) if config else TaskConfig()
217

lintangsutawika's avatar
lintangsutawika committed
218
        self._filters = [build_filter_ensemble("none", [["take_first", None]])]
219

Ethan Smith's avatar
Ethan Smith committed
220
    def download(self, data_dir=None, cache_dir=None, download_mode=None) -> None:
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
        """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.
        """
245
246
247
248
249
250
251
        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,
        )
252

253
254
255
256
257
    @property
    def config(self):
        """Returns the TaskConfig associated with this class."""
        return self._config

258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
    @abc.abstractmethod
    def has_training_docs(self):
        """Whether the task has a training set"""
        pass

    @abc.abstractmethod
    def has_validation_docs(self):
        """Whether the task has a validation set"""
        pass

    @abc.abstractmethod
    def has_test_docs(self):
        """Whether the task has a test set"""
        pass

    def training_docs(self):
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

    def validation_docs(self):
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

    def test_docs(self):
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

294
295
296
297
298
299
300
301
302
303
    def fewshot_docs(self):
        """
        :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
304
            eval_logger.warning(
305
                "has_training_docs and has_validation_docs are False"
306
                ", using test_docs as fewshot_docs but this is not recommended."
lintangsutawika's avatar
lintangsutawika committed
307
            )
308
309
            return self.test_docs()

310
311
312
313
314
315
316
317
318
319
    def _process_doc(self, doc):
        """
        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
320

321
322
323
324
325
326
327
328
329
330
331
332
333
    @property
    def instances(self):
        """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)

Ethan Smith's avatar
Ethan Smith committed
334
    def doc_to_decontamination_query(self, doc) -> None:
335
336
337
338
339
340
341
342
343
344
345
346
347
        print(
            "Override doc_to_decontamination_query with document specific decontamination query."
        )
        assert False

    @abc.abstractmethod
    def doc_to_text(self, doc):
        pass

    @abc.abstractmethod
    def doc_to_target(self, doc):
        pass

Ethan Smith's avatar
Ethan Smith committed
348
    def build_all_requests(self, limit=None, rank=None, world_size=None) -> None:
349
350
351
352
353
354
        """Build a set of Instances for a task, and store them in task.instances"""
        if self.has_test_docs():
            docs = self.test_docs()
        elif self.has_validation_docs():
            docs = self.validation_docs()
        else:
355
            assert False, f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"
356

357
        eval_logger.info(f"Building contexts for task on rank {rank}...")
358

359
        instances = []
360
361
        for doc_id, doc in utils.create_iterator(
            enumerate(docs), rank, world_size, limit
lintangsutawika's avatar
lintangsutawika committed
362
        ):
363
            # sample fewshot context #TODO: need to offset doc_id by rank now!
364
            fewshot_ctx = self.fewshot_context(
365
                doc,
366
                0 if self.config.num_fewshot is None else self.config.num_fewshot,
367
            )
368

369
            # TODO: we should override self.config.repeats if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
370
371
372
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
373
                metadata=(self.config["task"], doc_id, self.config.repeats),
lintangsutawika's avatar
lintangsutawika committed
374
            )
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399

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

            instances.extend(inst)

        self._instances = instances
        assert len(self._instances) != 0, "task.build_requests() did not find any docs!"

    @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
400
            The number of times each instance in a dataset is inferred on. Defaults to 1,
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
            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

haileyschoelkopf's avatar
haileyschoelkopf committed
436
437
438
439
440
441
442
443
444
445
    @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))

446
    @utils.positional_deprecated
lintangsutawika's avatar
lintangsutawika committed
447
    def fewshot_context(
448
449
450
451
452
        self,
        doc,
        num_fewshot,
        rnd=random.Random(1234),
        description=None,
lintangsutawika's avatar
lintangsutawika committed
453
    ):
454
455
456
457
458
459
460
        """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
461
462
463
464
465
        :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.
466
467
468
        :returns: str
            The fewshot context.
        """
lintangsutawika's avatar
lintangsutawika committed
469
470
471
472
        assert (
            rnd is not None
        ), "A `random.Random` generator argument must be provided to `rnd`"

473
        description = description if description else ""
474
475

        if num_fewshot == 0:
lintangsutawika's avatar
lintangsutawika committed
476
            labeled_examples = ""
477
        else:
lintangsutawika's avatar
lintangsutawika committed
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
            # 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
502
            )
503
504

        example = self.doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
505
        return description + labeled_examples + example
506
507

    def apply_filters(self):
lintangsutawika's avatar
lintangsutawika committed
508
509
        if hasattr(self, "_filters"):
            for f in self._filters:
lintangsutawika's avatar
lintangsutawika committed
510
                f.apply(self._instances, None)
lintangsutawika's avatar
lintangsutawika committed
511
512
513
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances
514

baberabb's avatar
baberabb committed
515
    def dump_config(self) -> dict:
516
        """Returns a dictionary representing the task's config.
517
518
519
520
521

        :returns: str
            The fewshot context.
        """
        # TODO: this should only return the overrides applied to a non-YAML task's configuration.
522
        # (num_fewshot)
523
        return self.config.to_dict()
524

525
526

class ConfigurableTask(Task):
527
    VERSION = "Yaml"
528
    OUTPUT_TYPE = None
529
    CONFIG = None
530
531
532

    def __init__(
        self, data_dir=None, cache_dir=None, download_mode=None, config: dict = None
Ethan Smith's avatar
Ethan Smith committed
533
    ) -> None:  # TODO no super() call here
534
        # Get pre-configured attributes
535
        self._config = self.CONFIG
536

537
        # Use new configurations if there was no preconfiguration
538
        if self.config is None:
539
            self._config = TaskConfig(**config)
540
541
        # Overwrite configs
        else:
lintangsutawika's avatar
lintangsutawika committed
542
            if config is not None:
543
                self._config.__dict__.update(config)
544

545
        if self.config is None:
lintangsutawika's avatar
lintangsutawika committed
546
547
548
            raise ValueError(
                "Must pass a config to ConfigurableTask, either in cls.CONFIG or `config` kwarg"
            )
549

550
551
552
553
        if isinstance(self.config.metadata, dict):
            if "version" in self.config.metadata:
                self.VERSION = self.config.metadata["version"]

554
555
556
        if self.config.output_type is not None:
            assert self.config.output_type in ALL_OUTPUT_TYPES
            self.OUTPUT_TYPE = self.config.output_type
557

558
559
        if self.config.dataset_path is not None:
            self.DATASET_PATH = self.config.dataset_path
560

561
562
        if self.config.dataset_name is not None:
            self.DATASET_NAME = self.config.dataset_name
563

564
565
566
567
        self._metric_fn_list = {}
        self._metric_fn_kwargs = {}
        self._aggregation_list = {}
        self._higher_is_better = {}
568

569
        if self.config.metric_list is None:
570
            # TODO: handle this in TaskConfig.__post_init__ ?
571
572
            _metric_list = DEFAULT_METRIC_REGISTRY[self.config.output_type]

573
            for metric_name in _metric_list:
haileyschoelkopf's avatar
haileyschoelkopf committed
574
                self._metric_fn_list[metric_name] = get_metric(metric_name)
lintangsutawika's avatar
lintangsutawika committed
575
                self._metric_fn_kwargs[metric_name] = {}
576
577
578
                self._aggregation_list[metric_name] = get_metric_aggregation(
                    metric_name
                )
haileyschoelkopf's avatar
haileyschoelkopf committed
579
                self._higher_is_better[metric_name] = is_higher_better(metric_name)
580
        else:
581
            for metric_config in self.config.metric_list:
582
583
584
585
586
                assert "metric" in metric_config
                metric_name = metric_config["metric"]
                kwargs = {
                    key: metric_config[key]
                    for key in metric_config
Chris's avatar
Chris committed
587
588
                    if key
                    not in ["metric", "aggregation", "higher_is_better", "hf_evaluate"]
589
                }
Chris's avatar
Chris committed
590
591
592
593
                hf_evaluate_metric = (
                    "hf_evaluate" in metric_config
                    and metric_config["hf_evaluate"] is True
                )
594

595
                if self.config.process_results is not None:
596
597
                    self._metric_fn_list[metric_name] = None
                    self._metric_fn_kwargs[metric_name] = {}
598
599
600
601
602
603
                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
604
605
606
                    self._metric_fn_list[metric_name] = get_metric(
                        metric_name, hf_evaluate_metric
                    )
607
                    self._metric_fn_kwargs[metric_name] = kwargs
lintangsutawika's avatar
lintangsutawika committed
608

609
                if "aggregation" in metric_config:
610
                    agg_name = metric_config["aggregation"]
611
                    if isinstance(agg_name, str):
haileyschoelkopf's avatar
haileyschoelkopf committed
612
                        self._aggregation_list[metric_name] = get_aggregation(agg_name)
613
                    elif callable(agg_name):  # noqa: E721
614
615
616
                        self._aggregation_list[metric_name] = metric_config[
                            "aggregation"
                        ]
617
                else:
618
                    INV_AGG_REGISTRY = {v: k for k, v in AGGREGATION_REGISTRY.items()}
lintangsutawika's avatar
lintangsutawika committed
619
                    metric_agg = get_metric_aggregation(metric_name)
620
                    eval_logger.warning(
baberabb's avatar
baberabb committed
621
                        f"[Task: {self._config.task}] metric {metric_name} is defined, but aggregation is not. "
622
623
                        f"using default "
                        f"aggregation={INV_AGG_REGISTRY[metric_agg]}"
624
                    )
625
                    self._aggregation_list[metric_name] = metric_agg
lintangsutawika's avatar
lintangsutawika committed
626

627
628
629
630
631
632
                if "higher_is_better" in metric_config:
                    self._higher_is_better[metric_name] = metric_config[
                        "higher_is_better"
                    ]
                else:
                    eval_logger.warning(
baberabb's avatar
baberabb committed
633
                        f"[Task: {self._config.task}] metric {metric_name} is defined, but higher_is_better is not. "
634
                        f"using default "
haileyschoelkopf's avatar
haileyschoelkopf committed
635
                        f"higher_is_better={is_higher_better(metric_name)}"
636
                    )
haileyschoelkopf's avatar
haileyschoelkopf committed
637
                    self._higher_is_better[metric_name] = is_higher_better(metric_name)
638

639
        self.download(self.config.dataset_kwargs)
640
641
642
        self._training_docs = None
        self._fewshot_docs = None

643
        if self.config.filter_list is not None:
lintangsutawika's avatar
lintangsutawika committed
644
            self._filters = []
645
            for filter_config in self.config.filter_list:
lintangsutawika's avatar
lintangsutawika committed
646
647
648
649
650
651
652
                for filter_pipeline in filter_config:
                    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"
lintangsutawika's avatar
lintangsutawika committed
653
654
655
                        }
                        components.append([function["function"], kwargs])
                    filter_pipeline = build_filter_ensemble(filter_name, components)
lintangsutawika's avatar
lintangsutawika committed
656
                self._filters.append(filter_pipeline)
lintangsutawika's avatar
lintangsutawika committed
657
        else:
658
            self._filters = [build_filter_ensemble("none", [["take_first", None]])]
659

660
661
        if self.config.use_prompt is not None:
            eval_logger.info(f"loading prompt {self.config.use_prompt}")
662
            self.prompt = get_prompt(
663
                self.config.use_prompt, self.DATASET_PATH, self.DATASET_NAME
lintangsutawika's avatar
lintangsutawika committed
664
            )
665
666
667
        else:
            self.prompt = None

lintangsutawika's avatar
lintangsutawika committed
668
        if self.fewshot_docs() is not None:
haileyschoelkopf's avatar
haileyschoelkopf committed
669
            self.sampler = samplers.get_sampler(
haileyschoelkopf's avatar
haileyschoelkopf committed
670
671
672
                self.config.fewshot_config.get("sampler", "default")
                if self.config.fewshot_config
                else "default"
haileyschoelkopf's avatar
haileyschoelkopf committed
673
            )(list(self.fewshot_docs()), self, rnd=random.Random(1234))
674

675
        if self.has_test_docs():
676
            self.task_docs = self.test_docs()
677
        elif self.has_validation_docs():
678
            self.task_docs = self.validation_docs()
679
        else:
680
            assert False, f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"
681

682
        # Test One Doc
683
        self.features = list(self.task_docs.features.keys())
684
685
        self.multiple_input = 0
        self.multiple_target = 0
686
        test_doc = self.task_docs[0]
687
        test_text = self.doc_to_text(test_doc)
688
        test_target = self.doc_to_target(test_doc)
lintangsutawika's avatar
lintangsutawika committed
689

690
        if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
691
            test_choice = self.doc_to_choice(test_doc)
692
            if not isinstance(test_choice, list):
lintangsutawika's avatar
lintangsutawika committed
693
                eval_logger.error("doc_to_choice must return list")
694
695
            else:
                num_choice = len(test_choice)
696

697
            if isinstance(test_text, int):
698
                self.multiple_input = num_choice
699
700
        else:
            test_choice = None
701

702
        if isinstance(test_target, list):
703
            self.multiple_target = len(test_target)
704
        else:
705
            if (isinstance(test_target, int)) and (test_choice is not None):
lintangsutawika's avatar
lintangsutawika committed
706
                test_target = test_choice[test_target]
707
            else:
lintangsutawika's avatar
lintangsutawika committed
708
                test_target = str(test_target)
709

710
711
712
        if test_choice is not None:
            check_choices = test_choice
        else:
lintangsutawika's avatar
lintangsutawika committed
713
            check_choices = [test_target]
714
715
716
717
        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 = (
718
719
                    True
                    if self.config.target_delimiter.rstrip()
720
                    != self.config.target_delimiter
721
                    else False
722
                )
723

724
                if delimiter_has_whitespace and choice_has_whitespace:
725
726
                    eval_logger.debug(
                        f'Both target_delimiter "{self.config.target_delimiter}" and target choice: "{choice}" have whitespace'
727
728
                    )
                elif (not delimiter_has_whitespace) and (not choice_has_whitespace):
729
                    eval_logger.debug(
730
                        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'
731
732
                    )

Ethan Smith's avatar
Ethan Smith committed
733
    def download(self, dataset_kwargs=None) -> None:
734
735
736
737
738
739
        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
740
    def has_training_docs(self) -> bool:
741
        if self.config.training_split is not None:
742
743
744
745
            return True
        else:
            return False

baberabb's avatar
baberabb committed
746
    def has_validation_docs(self) -> bool:
747
        if self.config.validation_split is not None:
748
749
750
751
            return True
        else:
            return False

baberabb's avatar
baberabb committed
752
    def has_test_docs(self) -> bool:
753
        if self.config.test_split is not None:
754
755
756
757
            return True
        else:
            return False

baberabb's avatar
baberabb committed
758
    def training_docs(self) -> datasets.Dataset:
759
        if self.has_training_docs():
760
761
762
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.training_split]
763
                )
764
            return self.dataset[self.config.training_split]
765

baberabb's avatar
baberabb committed
766
    def validation_docs(self) -> datasets.Dataset:
767
        if self.has_validation_docs():
768
769
770
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.validation_split]
771
                )
772
            return self.dataset[self.config.validation_split]
773

baberabb's avatar
baberabb committed
774
    def test_docs(self) -> datasets.Dataset:
775
        if self.has_test_docs():
776
777
778
            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]
779

780
    def fewshot_docs(self):
781
        if self.config.fewshot_split is not None:
782
783
            if self.config.process_docs is not None:
                return self.config.process_docs(self.dataset[self.config.fewshot_split])
784
            return self.dataset[self.config.fewshot_split]
785
        else:
786
            if (self.config.num_fewshot is not None) and (self.config.num_fewshot > 0):
787
                eval_logger.warning(
788
                    f"Task '{self.config.task}': "
789
790
791
792
                    "num_fewshot > 0 but fewshot_split is None. "
                    "using preconfigured rule."
                )
            return super().fewshot_docs()
793

lintangsutawika's avatar
lintangsutawika committed
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
    @utils.positional_deprecated
    def fewshot_context(self, doc, num_fewshot):
        """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.
        """

        if num_fewshot == 0:
            # always prepend the (possibly empty) task description
            labeled_examples = self.config.description
        else:
            labeled_examples = self.config.description + self.sampler.get_context(
                doc, num_fewshot
            )

        example = self.doc_to_text(doc)
816
817
818
819
820
821
822
823
824
825
826
827
828
        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
829

830
831
832
833
834
835
836
837
    def apply_filters(self):
        if hasattr(self, "_filters"):
            for f in self._filters:
                f.apply(self._instances, self.task_docs)
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances

838
    def should_decontaminate(self):
839
        return self.config.should_decontaminate
840
841

    def doc_to_decontamination_query(self, doc):
842
        if self.config.should_decontaminate:
843
844
            if self.config.doc_to_decontamination_query is None:
                return self.doc_to_text(doc)
845
            else:
846
847
848
849
850
851
852
853
854
855
856
                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
                        )
                    )
857

858
859
860
861
862
863
864
865
866
867
868
869
    def _process_doc(self, doc):
        """
        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):
870
871
        if self.prompt is not None:
            doc_to_text = self.prompt
872
        else:
873
            doc_to_text = self.config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
874

875
        if isinstance(doc_to_text, int):
876
            return doc_to_text
877
        elif isinstance(doc_to_text, str):
878
            if doc_to_text in self.features:
879
                # if self.config.doc_to_choice is not None:
880
881
                #     return self.doc_to_choice(doc)[doc[doc_to_text]]
                # else:
882
883
                return doc[doc_to_text]
            else:
lintangsutawika's avatar
lintangsutawika committed
884
                text_string = utils.apply_template(doc_to_text, doc)
lintangsutawika's avatar
lintangsutawika committed
885
                if text_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
886
887
888
                    return ast.literal_eval(text_string)
                else:
                    return text_string
889
        elif callable(doc_to_text):
890
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
891
        # Used when applying a Promptsource template
892
        elif hasattr(doc_to_text, "apply"):
893
894
895
896
897
            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")
898
                return self.config.fewshot_delimiter
899
        else:
900
            print(type(doc_to_text))
901
            raise TypeError
902

903
    def doc_to_target(self, doc: dict) -> Union[int, str, list]:
904
905
        if self.prompt is not None:
            doc_to_target = self.prompt
906
        else:
907
            doc_to_target = self.config.doc_to_target
908

909
        if isinstance(doc_to_target, int):
910
            return doc_to_target
911
        elif isinstance(doc_to_target, str):
912
            if doc_to_target in self.features:
913
                # if self.config.doc_to_choice is not None:
914
915
916
                #     return self.doc_to_choice(doc)[doc[doc_to_target]]
                # else:
                return doc[doc_to_target]
917
            else:
lintangsutawika's avatar
lintangsutawika committed
918
                target_string = utils.apply_template(doc_to_target, doc)
lintangsutawika's avatar
lintangsutawika committed
919
                if target_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
920
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
921
922
923
924
925
                elif (
                    len(target_string) >= 2
                    and (target_string[0] == "[")
                    and (target_string[-1] == "]")
                ):
baberabb's avatar
baberabb committed
926
927
928
929
                    try:
                        return ast.literal_eval(target_string)
                    except (SyntaxError, ValueError):
                        return target_string
lintangsutawika's avatar
lintangsutawika committed
930
931
                else:
                    return target_string
932
        elif isinstance(doc_to_target, list):
933
            return doc_to_target
934
        elif callable(doc_to_target):
935
            return doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
936
        # Used when applying a Promptsource template
937
        elif hasattr(doc_to_target, "apply"):
938
            applied_prompt = doc_to_target.apply(doc)
939
940
941
942
            if len(applied_prompt) == 2:
                return applied_prompt[1]
            else:
                eval_logger.warning("Applied prompt returns empty string")
943
                return self.config.fewshot_delimiter
944
945
        else:
            raise TypeError
946

baberabb's avatar
baberabb committed
947
    def doc_to_choice(self, doc: Any) -> List[str]:
948
949
        if self.prompt is not None:
            doc_to_choice = self.prompt
950
        elif self.config.doc_to_choice is None:
951
952
            eval_logger.error("doc_to_choice was called but not set in config")
        else:
953
            doc_to_choice = self.config.doc_to_choice
954

955
        if isinstance(doc_to_choice, str):
956
957
958
959
            if doc_to_choice in self.features:
                return doc[doc_to_choice]
            else:
                return ast.literal_eval(utils.apply_template(doc_to_choice, doc))
960
        elif isinstance(doc_to_choice, list):
961
            return doc_to_choice
962
        elif isinstance(doc_to_choice, dict):
963
964
965
966
967
968
969
            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
970

baberabb's avatar
baberabb committed
971
972
973
    def construct_requests(
        self, doc: dict, ctx: str, **kwargs
    ) -> Union[List[Instance], Instance]:
974
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
975
            arguments = (ctx, self.doc_to_target(doc))
976
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
977
            arguments = (self.doc_to_target(doc),)
978
        elif self.OUTPUT_TYPE == "multiple_choice":
979
            choices = self.doc_to_choice(doc)
980
            target_delimiter = self.config.target_delimiter
981
982
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
983
                cont = self.doc_to_target(doc)
984
985
986
                arguments = [
                    (ctx + choice, f"{target_delimiter}{cont}") for choice in choices
                ]
987
            else:
988
                # Otherwise they are placed in the continuation
989
                arguments = [(ctx, f"{target_delimiter}{cont}") for cont in choices]
990

991
            request_list = [
992
993
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
994
                    doc=doc,
995
                    arguments=arg,
996
                    idx=i,
997
998
                    **kwargs,
                )
999
                for i, arg in enumerate(arguments)
1000
            ]
1001
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
1002
            if "acc_mutual_info" in self._metric_fn_list.keys():
1003
1004
1005
                # if we are calculating multiple choice accuracy
                # using mutual information instead of raw loglikelihood as metric, need unconditional lls.

lintangsutawika's avatar
lintangsutawika committed
1006
                # here mutual info refers to calculating
1007
1008
1009
1010
1011
1012
                # 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
1013
                            doc=doc,
1014
                            arguments=("", "{}".format(choice)),
1015
1016
1017
                            idx=i,
                            **kwargs,
                        )
lintangsutawika's avatar
lintangsutawika committed
1018
                        for i, choice in enumerate(choices)
1019
1020
1021
                    ]
                )
            return request_list
lintangsutawika's avatar
lintangsutawika committed
1022

1023
        elif self.OUTPUT_TYPE == "generate_until":
1024
            arguments = (ctx, self.config.generation_kwargs)
lintangsutawika's avatar
lintangsutawika committed
1025
1026

        return Instance(
lintangsutawika's avatar
lintangsutawika committed
1027
1028
            request_type=self.OUTPUT_TYPE, doc=doc, arguments=arguments, idx=0, **kwargs
        )
1029
1030

    def process_results(self, doc, results):
1031
1032
        if callable(self.config.process_results):
            return self.config.process_results(doc, results)
lintangsutawika's avatar
lintangsutawika committed
1033

1034
        result_dict = {}
1035
        use_metric = list(self._metric_fn_list.keys())
1036
1037
1038
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
1039
1040
1041
1042
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
1043
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
1044
            (loglikelihood,) = results
1045
1046
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
haileyschoelkopf's avatar
haileyschoelkopf committed
1047
            return {
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
                **(
                    {"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
1063
            }
1064
        elif self.OUTPUT_TYPE == "multiple_choice":
1065
            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
1066

1067
            # retrieve choices in List[str] form, to compute choice lengths, etc.
1068
            choices = self.doc_to_choice(doc)
1069
1070
            completion_len = np.array([float(len(i)) for i in choices])

1071
1072
            if (
                2 * len(choices) == len(lls)
1073
                and "acc_mutual_info" in self._metric_fn_list.keys()
1074
1075
1076
1077
1078
1079
1080
            ):
                # then we are doing mutual info.
                # this stores the "dryrun" / unconditional answer loglikelihoods
                lls_unconditional = lls[1::2]
                assert len(lls_unconditional) == len(choices)
                # and this stores our "regular" conditional loglikelihoods
                lls = lls[::2]
1081

1082
1083
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
1084

1085
1086
            if self.multiple_input:
                gold = self.doc_to_text(doc)
1087
            else:
1088
                gold = self.doc_to_target(doc)
1089
1090

            gold_index_error = False
1091
            if isinstance(gold, list):
Lintang Sutawika's avatar
Lintang Sutawika committed
1092
1093
                gold = [i if i < len(choices) else -100 for i in gold]
                if -100 in gold:
1094
1095
                    gold_index_error = True
            else:
1096
                if isinstance(gold, int):
Lintang Sutawika's avatar
Lintang Sutawika committed
1097
                    gold = gold if gold < len(choices) else -100
1098
                elif isinstance(gold, str):
Lintang Sutawika's avatar
Lintang Sutawika committed
1099
                    gold = choices.index(gold) if gold in choices else -100
lintangsutawika's avatar
lintangsutawika committed
1100

Lintang Sutawika's avatar
Lintang Sutawika committed
1101
                if gold == -100:
1102
1103
1104
1105
                    gold_index_error = True

            if gold_index_error:
                eval_logger.warning(
lintangsutawika's avatar
lintangsutawika committed
1106
                    f"Label index was not in within range of available choices,"
1107
1108
                    f"Sample:\n\n{doc}\n\n"
                )
lintangsutawika's avatar
lintangsutawika committed
1109

1110
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
1111
1112
                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
1113
                exact_match = int(any([is_greedy[i] if i != -100 else 0 for i in gold]))
lintangsutawika's avatar
lintangsutawika committed
1114
1115
1116
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
1117
                # 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
1118
                exact_match = int(is_greedy[gold]) if gold != -100 else 0
1119
1120

            result_dict = {
1121
                **({"acc": acc} if "acc" in use_metric else {}),
1122
1123
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
1124
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
1125
                **({"exact_match": exact_match} if "exact_match" in use_metric else {}),
1126
1127
            }

1128
            if "acc_mutual_info" in use_metric:
lintangsutawika's avatar
lintangsutawika committed
1129
1130
1131
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
1132
1133
1134
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

1135
        elif self.OUTPUT_TYPE == "generate_until":
1136
            gold = self.doc_to_target(doc)
Chris's avatar
Chris committed
1137
            result = results[0]
1138
            if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1139
                # If you set doc_to_choice,
lintangsutawika's avatar
lintangsutawika committed
1140
                # it assumes that doc_to_target returns a number.
1141
1142
                choices = self.doc_to_choice(doc)
                gold = choices[gold]
1143
1144
            # we expect multiple_targets to be a list.
            elif self.multiple_target:
baberabb's avatar
baberabb committed
1145
                gold = list(gold)
Chris's avatar
Chris committed
1146
1147
1148
            elif type(gold) != type(result):
                # cast gold to the same type as result
                gold = type(result)(gold)
1149

lintangsutawika's avatar
lintangsutawika committed
1150
            for metric in self._metric_fn_list.keys():
haileyschoelkopf's avatar
haileyschoelkopf committed
1151
1152
1153
1154
1155
                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
1156
1157
1158
1159
                    if not isinstance(gold, list):
                        # sometimes, a multiple_target dataset has exceptions where one doc has only one string answer
                        # print(gold)
                        gold = [gold]
1160
1161
1162
1163
1164
1165
1166
1167
                    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
1168
                    else:
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
                        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
1190
                else:
1191
                    try:
1192
                        result_score = self._metric_fn_list[metric](
1193
1194
                            references=[gold],
                            predictions=[result],
1195
                            **self._metric_fn_kwargs[metric],
1196
                        )
1197
                    except TypeError:  # needed for now in order to use a different interface between our own metrics and HF Evaluate metrics
1198
                        result_score = self._metric_fn_list[metric]([gold, result])
1199
1200
1201
1202
                    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
1203
        else:
lintangsutawika's avatar
lintangsutawika committed
1204
1205
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
1206
                "'loglikelihood', 'loglikelihood_rolling', 'generate_until' or 'multiple_choice'",
1207
            )
1208
1209
1210
1211
1212
1213
1214

        return result_dict

    def aggregation(self):
        return self._aggregation_list

    def higher_is_better(self):
haileyschoelkopf's avatar
haileyschoelkopf committed
1215
        return self._higher_is_better
1216
1217
1218
1219
1220


class MultipleChoiceTask(Task):
    OUTPUT_TYPE: str = "loglikelihood"

baberabb's avatar
baberabb committed
1221
    def doc_to_target(self, doc: dict) -> str:
1222
1223
        return " " + doc["choices"][doc["gold"]]

baberabb's avatar
baberabb committed
1224
    def construct_requests(self, doc: dict, ctx: str, **kwargs) -> List[Instance]:
1225
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1226
1227
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1228
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1229
                doc=doc,
1230
                arguments=(ctx, " {}".format(choice)),
1231
                idx=i,
1232
1233
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1234
1235
            for i, choice in enumerate(doc["choices"])
        ]
1236

baberabb's avatar
baberabb committed
1237
    def process_results(self, doc: dict, results: List[Tuple[float, bool]]) -> dict:
lintangsutawika's avatar
lintangsutawika committed
1238
1239
1240
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
        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
1252
    def higher_is_better(self) -> dict:
1253
1254
1255
1256
1257
        return {
            "acc": True,
            "acc_norm": True,
        }

baberabb's avatar
baberabb committed
1258
    def aggregation(self) -> dict:
1259
1260
1261
1262
1263
1264
        return {
            "acc": mean,
            "acc_norm": mean,
        }


lintangsutawika's avatar
lintangsutawika committed
1265
class PerplexityTask(Task):
1266
1267
    OUTPUT_TYPE = "loglikelihood_rolling"

baberabb's avatar
baberabb committed
1268
    def has_training_docs(self) -> bool:
1269
1270
        return False

baberabb's avatar
baberabb committed
1271
    def fewshot_examples(self, k: int, rnd) -> List:
1272
1273
1274
        assert k == 0
        return []

baberabb's avatar
baberabb committed
1275
    def fewshot_context(self, doc: dict, num_fewshot: int) -> Literal[""]:
1276
1277
1278
1279
1280
1281
        assert (
            num_fewshot == 0
        ), "The number of fewshot examples must be 0 for perplexity tasks."

        return ""

baberabb's avatar
baberabb committed
1282
    def higher_is_better(self) -> dict:
1283
1284
1285
1286
1287
1288
1289
1290
1291
        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
1292
    def doc_to_text(self, doc) -> str:
1293
1294
1295
1296
1297
        return ""

    def doc_to_target(self, doc):
        return doc

baberabb's avatar
baberabb committed
1298
    def construct_requests(self, doc: dict, ctx: Union[str, None], **kwargs):
1299
1300
        assert not ctx

lintangsutawika's avatar
lintangsutawika committed
1301
1302
1303
1304
1305
1306
1307
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1308

baberabb's avatar
baberabb committed
1309
    def process_results(self, doc: dict, results: float) -> dict:
1310
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1311
1312
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1313
1314
1315
1316
1317
1318
        return {
            "word_perplexity": (loglikelihood, words),
            "byte_perplexity": (loglikelihood, bytes_),
            "bits_per_byte": (loglikelihood, bytes_),
        }

baberabb's avatar
baberabb committed
1319
    def aggregation(self) -> dict:
1320
1321
1322
1323
1324
1325
1326
        return {
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
            "bits_per_byte": bits_per_byte,
        }

    @classmethod
baberabb's avatar
baberabb committed
1327
    def count_bytes(cls, doc) -> int:
1328
1329
1330
        return len(doc.encode("utf-8"))

    @classmethod
baberabb's avatar
baberabb committed
1331
    def count_words(cls, doc) -> int:
1332
1333
        """Downstream tasks with custom word boundaries should override this!"""
        return len(re.split(r"\s+", doc))