task.py 41.2 KB
Newer Older
1
import abc
2
from dataclasses import dataclass, field, asdict
3
4

import re
5
import ast
lintangsutawika's avatar
lintangsutawika committed
6
import yaml
7
8
9
import evaluate
import random
import itertools
10
import functools
11
12
13
14

import datasets
import numpy as np

15
16
from typing import Union
from collections.abc import Callable
17

18
from lm_eval import utils
19
from lm_eval.api import samplers
haileyschoelkopf's avatar
haileyschoelkopf committed
20
from lm_eval.api.instance import Instance
lintangsutawika's avatar
lintangsutawika committed
21
from lm_eval.api.filter import FilterEnsemble
22
23
24
25

from lm_eval.logger import eval_logger
from lm_eval.prompts import get_prompt
from lm_eval.filters import build_filter_ensemble
lintangsutawika's avatar
lintangsutawika committed
26
27
28
29
30
31
from lm_eval.api.metrics import (
    mean,
    weighted_perplexity,
    bits_per_byte,
)
from lm_eval.api.registry import (
haileyschoelkopf's avatar
haileyschoelkopf committed
32
33
34
35
    get_metric,
    get_aggregation,
    get_default_aggregation,
    is_higher_better,
36
37
    DEFAULT_METRIC_REGISTRY,
    OUTPUT_TYPE_REGISTRY,
lintangsutawika's avatar
lintangsutawika committed
38
39
    AGGREGATION_REGISTRY,
)
40

41
42
43
44
45
ALL_OUTPUT_TYPES = [
    "loglikelihood",
    "multiple_choice",
    "loglikelihood_rolling",
    "greedy_until",
lintangsutawika's avatar
lintangsutawika committed
46
    "winograd_schema",
47
48
]

49
50
51

@dataclass
class TaskConfig(dict):
52
    # task naming/registry
53
    task: str = None
54
    group: Union[str, list] = None
55
56
57
    # HF dataset options.
    # which dataset to use,
    # and what splits for what purpose
58
59
    dataset_path: str = None
    dataset_name: str = None
60
    dataset_kwargs: dict = None
61
62
63
    training_split: str = None
    validation_split: str = None
    test_split: str = None
lintangsutawika's avatar
lintangsutawika committed
64
    fewshot_split: str = None  # TODO: assert that this not None if num_fewshot > 0. (?) assert if this is same split as one evaling (?)
65
66
    # formatting / prompting options.
    # see docs/advanced_task_guide.md for more info
67
    template_aliases: Union[str, list] = None
68
69
    doc_to_text: Union[Callable, str] = None
    doc_to_target: Union[Callable, str] = None
lintangsutawika's avatar
lintangsutawika committed
70
    doc_to_choice: Union[Callable, str, dict, list] = None
71
    gold_alias: Union[Callable, str] = None
72
    use_prompt: str = None
73
    description: str = ""
74
75
    target_delimiter: str = " "
    fewshot_delimiter: str = "\n\n"
76
    # runtime configuration options
77
    num_fewshot: int = 0
78
    # scoring options
79
80
    metric_list: str = None
    output_type: str = "greedy_until"
81
    generation_kwargs: dict = None
82
    repeats: int = 1
lintangsutawika's avatar
lintangsutawika committed
83
    filter_list: Union[str, list] = None
84
85
    should_decontaminate: bool = False
    doc_to_decontamination_query: str = None
86

lintangsutawika's avatar
lintangsutawika committed
87
    metadata: str = None  # by default, not used in the code. allows for users to pass arbitrary info to tasks
88

89
90
91
92
    def __post_init__(self):
        # allow user-specified aliases so that users can
        # force prompt-compatibility for some prompt regardless of
        # field names in prompt
93
94
95
        if self.template_aliases is not None:
            if type(self.doc_to_text) == str:
                self.doc_to_text = self.template_aliases + self.doc_to_text
96

97
98
            if type(self.doc_to_target) == str:
                self.doc_to_target = self.template_aliases + self.doc_to_target
99

100
            if type(self.gold_alias) == str:
lintangsutawika's avatar
lintangsutawika committed
101
                self.gold_alias = self.template_aliases + self.gold_alias
102

Lintang Sutawika's avatar
Lintang Sutawika committed
103
104
105
106
107
108
109
110
111
112
113
114
        if self.generation_kwargs is not None:
            if self.output_type != "greedy_until":
                eval_logger.warning(
                    "passed `generation_kwargs`, but not using a generation request type!"
                )

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

            if "until" not in self.generation_kwargs:
115
                self.generation_kwargs["until"] = [self.fewshot_delimiter]
Lintang Sutawika's avatar
Lintang Sutawika committed
116
117
118
119
        else:
            if self.output_type == "greedy_until":
                # ensure that we greedily generate in absence of explicit arguments otherwise
                self.generation_kwargs = {
Lintang Sutawika's avatar
Lintang Sutawika committed
120
                    "until": None
121
122
                    if self.fewshot_delimiter is None
                    else [self.fewshot_delimiter],
Lintang Sutawika's avatar
Lintang Sutawika committed
123
124
125
                    "do_sample": False,
                    "temperature": 0.0,
                }
126

haileyschoelkopf's avatar
haileyschoelkopf committed
127
128
        # TODO: how to make TaskConfigs be de- and re-serializable, even when using the !function constructor?

129
130
131
    def __getitem__(self, item):
        return getattr(self, item)

132
    def to_dict(self):
133
134
        """dumps the current config as a dictionary object, as a printable format.
        null fields will not be printed.
haileyschoelkopf's avatar
haileyschoelkopf committed
135
        Used for dumping results alongside full task configuration
136

haileyschoelkopf's avatar
haileyschoelkopf committed
137
138
139
140
141
142
143
144
145
146
        :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
147
148
149
            elif isinstance(v, Callable):
                # TODO: this should handle Promptsource template objects as a separate case?
                cfg_dict[k] = str(v)
haileyschoelkopf's avatar
haileyschoelkopf committed
150
        return cfg_dict
151

152
153
154
155
156
157
158
159
160
161
162
163

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
164

165
166
167
168
169
170
171
172
    # 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
173

174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
    def __init__(
        self,
        data_dir=None,
        cache_dir=None,
        download_mode=None,
        config=None,
    ):
        """
        :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.
        """
        self.download(data_dir, cache_dir, download_mode)
        self._training_docs = None
        self._fewshot_docs = None
        self._instances = None

haileyschoelkopf's avatar
haileyschoelkopf committed
208
        self._config = TaskConfig(**config) if config else TaskConfig()
209
210
211

        if not hasattr(self, "_filters"):
            self._filters = []
lintangsutawika's avatar
lintangsutawika committed
212
            for name, components in self._config.get(
213
                "filters", [["none", [["take_first", None]]]]
lintangsutawika's avatar
lintangsutawika committed
214
            ):
215
216
217
                filter_pipeline = build_filter_ensemble(name, components)
                self._filters.append(filter_pipeline)

lintangsutawika's avatar
lintangsutawika committed
218
219
220
        self.sampler = samplers.Sampler(
            list(self.fewshot_docs()), self, rnd=random.Random()
        )  # TODO: pass the correct docs in here
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246

    def download(self, data_dir=None, cache_dir=None, download_mode=None):
        """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.
        """
247
248
249
250
251
252
253
        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,
        )
254
255
256
257
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

    @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 []

291
292
293
294
295
296
297
298
299
300
    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
301
            eval_logger.warning(
302
                "has_training_docs and has_validation_docs are False"
303
                ", using test_docs as fewshot_docs but this is not recommended."
lintangsutawika's avatar
lintangsutawika committed
304
            )
305
306
            return self.test_docs()

307
308
309
310
311
312
313
314
315
316
    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
317

318
319
    def doc_to_choice(self, doc):
        if self._config.doc_to_choice is None:
lintangsutawika's avatar
typo  
lintangsutawika committed
320
            eval_logger.error("doc_to_choice was called but not set in config")
lintangsutawika's avatar
lintangsutawika committed
321
322
        elif type(self._config.doc_to_choice) == list:
            return self._config.doc_to_choice
323
324
        elif type(self._config.doc_to_choice) == dict:
            return list(self._config.doc_to_choice.values())
325
        elif type(self._config.doc_to_choice) == str:
326
327
328
            return ast.literal_eval(
                utils.apply_template(self._config.doc_to_choice, doc)
            )
329
        else:
330
            return self._config.doc_to_choice(doc)
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358

    @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)

    def doc_to_decontamination_query(self, doc):
        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

359
    def build_all_requests(self, limit=None, rank=None, world_size=None):
360
361
362
363
364
365
366
367
368
369
370
        """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:
            assert (
                False
            ), f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"

        instances = []
371
372
        for doc_id, doc in utils.create_iterator(
            enumerate(docs), rank, world_size, limit
lintangsutawika's avatar
lintangsutawika committed
373
        ):
374
            # sample fewshot context #TODO: need to offset doc_id by rank now!
375
376
377
            fewshot_ctx = self.fewshot_context(
                doc, self._config.num_fewshot, rnd=random.Random()
            )
378

haileyschoelkopf's avatar
haileyschoelkopf committed
379
            # TODO: we should override self._config.repeats if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
380
381
382
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
383
                metadata=(self._config["task"], doc_id, self._config.repeats),
lintangsutawika's avatar
lintangsutawika committed
384
            )
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409

            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
410
            The number of times each instance in a dataset is inferred on. Defaults to 1,
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
436
437
438
439
440
441
442
443
444
445
            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
446
447
448
449
450
451
452
453
454
455
    @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))

456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
    @utils.positional_deprecated
    def fewshot_context(self, doc, num_fewshot, rnd=None):
        """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.
        :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`.
        :returns: str
            The fewshot context.
        """
        assert (
            rnd is not None
        ), "A `random.Random` generator argument must be provided to `rnd`"

        if num_fewshot == 0:
476
477
            # always prepend the (possibly empty) task description
            labeled_examples = self._config.description
478
        else:
lintangsutawika's avatar
lintangsutawika committed
479
480
481
            labeled_examples = self._config.description + self.sampler.get_context(
                doc, num_fewshot
            )
482
483

        example = self.doc_to_text(doc)
484
485
486
487
        if type(example) == str:
            return labeled_examples + example
        elif type(example) == list:
            return [labeled_examples + ex for ex in example]
488
489
490

    def apply_filters(self):

lintangsutawika's avatar
lintangsutawika committed
491
492
493
494
495
496
        if hasattr(self, "_filters"):
            for f in self._filters:
                f.apply(self._instances)
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances
497

498
    def dump_config(self):
499
        """Returns a dictionary representing the task's config.
500
501
502
503
504

        :returns: str
            The fewshot context.
        """
        # TODO: this should only return the overrides applied to a non-YAML task's configuration.
505
        # (num_fewshot)
506
507
        return self._config.to_dict()

508
509
510

class ConfigurableTask(Task):

511
    VERSION = "Yaml"
512
    OUTPUT_TYPE = None
513
    CONFIG = None
514
515
516
517

    def __init__(
        self, data_dir=None, cache_dir=None, download_mode=None, config: dict = None
    ):
518
        # Get pre-configured attributes
519
        self._config = self.CONFIG
520

521
522
        # Use new configurations if there was no preconfiguration
        if self._config is None:
523
            self._config = TaskConfig(**config)
524
525
        # Overwrite configs
        else:
lintangsutawika's avatar
lintangsutawika committed
526
            if config is not None:
527
                self._config.__dict__.update(config)
528

529
        if self._config is None:
lintangsutawika's avatar
lintangsutawika committed
530
531
532
            raise ValueError(
                "Must pass a config to ConfigurableTask, either in cls.CONFIG or `config` kwarg"
            )
533
534

        if self._config.output_type is not None:
535
            assert self._config.output_type in ALL_OUTPUT_TYPES
536
537
            self.OUTPUT_TYPE = self._config.output_type

538
539
540
541
542
543
        if self._config.dataset_path is not None:
            self.DATASET_PATH = self._config.dataset_path

        if self._config.dataset_name is not None:
            self.DATASET_NAME = self._config.dataset_name

544
545
546
547
        self._metric_fn_list = {}
        self._metric_fn_kwargs = {}
        self._aggregation_list = {}
        self._higher_is_better = {}
548

549
        _metric_list = DEFAULT_METRIC_REGISTRY[self._config.output_type]
550
        if self._config.metric_list is None:
551
            # TODO: handle this in TaskConfig.__post_init__ ?
552
            for metric_name in _metric_list:
haileyschoelkopf's avatar
haileyschoelkopf committed
553
554
                self._metric_fn_list[metric_name] = get_metric(metric_name)
                self._aggregation_list[metric_name] = get_default_aggregation(
555
                    metric_name
haileyschoelkopf's avatar
haileyschoelkopf committed
556
557
                )
                self._higher_is_better[metric_name] = is_higher_better(metric_name)
558
559
560
561
562
563
564
565
566
        else:
            for metric_config in self._config.metric_list:
                assert "metric" in metric_config
                metric_name = metric_config["metric"]
                kwargs = {
                    key: metric_config[key]
                    for key in metric_config
                    if key not in ["metric", "aggregation", "higher_is_better"]
                }
haileyschoelkopf's avatar
haileyschoelkopf committed
567
568
                self._metric_fn_list[metric_name] = get_metric(metric_name)
                self._metric_fn_kwargs[metric_name] = kwargs
lintangsutawika's avatar
lintangsutawika committed
569

570
                if "aggregation" in metric_config:
571
                    agg_name = metric_config["aggregation"]
572
                    if type(agg_name) == str:
haileyschoelkopf's avatar
haileyschoelkopf committed
573
                        self._aggregation_list[metric_name] = get_aggregation(agg_name)
574
575
576
577
                    elif callable(agg_name):
                        self._aggregation_list[metric_name] = metric_config[
                            "aggregation"
                        ]
578
                else:
579
580

                    INV_AGG_REGISTRY = {v: k for k, v in AGGREGATION_REGISTRY.items()}
haileyschoelkopf's avatar
haileyschoelkopf committed
581
                    metric_agg = get_default_aggregation(metric_name)
582
                    eval_logger.warning(
583
584
585
                        f"metric {metric_name} is defined, but aggregation is not. "
                        f"using default "
                        f"aggregation={INV_AGG_REGISTRY[metric_agg]}"
586
                    )
587
                    self._aggregation_list[metric_name] = metric_agg
lintangsutawika's avatar
lintangsutawika committed
588

589
590
591
592
593
594
                if "higher_is_better" in metric_config:
                    self._higher_is_better[metric_name] = metric_config[
                        "higher_is_better"
                    ]
                else:
                    eval_logger.warning(
595
596
                        f"metric {metric_name} is defined, but higher_is_better is not. "
                        f"using default "
haileyschoelkopf's avatar
haileyschoelkopf committed
597
                        f"higher_is_better={is_higher_better(metric_name)}"
598
                    )
haileyschoelkopf's avatar
haileyschoelkopf committed
599
                    self._higher_is_better[metric_name] = is_higher_better(metric_name)
600

601
        self.download(self._config.dataset_kwargs)
602
603
604
        self._training_docs = None
        self._fewshot_docs = None

lintangsutawika's avatar
lintangsutawika committed
605
        if self._config.filter_list is not None:
lintangsutawika's avatar
lintangsutawika committed
606
            self._filters = []
lintangsutawika's avatar
lintangsutawika committed
607
608
609
610
611
612
613
614
            for filter_config in self._config.filter_list:
                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
615
616
617
                        }
                        components.append([function["function"], kwargs])
                    filter_pipeline = build_filter_ensemble(filter_name, components)
lintangsutawika's avatar
lintangsutawika committed
618
                self._filters.append(filter_pipeline)
lintangsutawika's avatar
lintangsutawika committed
619
        else:
620
            self._filters = [build_filter_ensemble("none", [["take_first", None]])]
621
622

        if self._config.use_prompt is not None:
lintangsutawika's avatar
lintangsutawika committed
623
            eval_logger.info(f"loading prompt {self._config.use_prompt}")
624
            self.prompt = get_prompt(
lintangsutawika's avatar
lintangsutawika committed
625
626
                self._config.use_prompt, self.DATASET_PATH, self.DATASET_NAME
            )
627
628
629
        else:
            self.prompt = None

lintangsutawika's avatar
lintangsutawika committed
630
631
632
        if self.fewshot_docs() is not None:
            self.sampler = samplers.Sampler(
                list(self.fewshot_docs()), self, rnd=random.Random()
633
            )
634

635
636
637
638
        if self._config.template_aliases is not None:
            for key, alias in self._config.template_aliases:
                self.dataset.rename_column(key, alias)

639
640
641
642
643
644
645
646
647
        if self.has_test_docs():
            docs = self.test_docs()
        elif self.has_validation_docs():
            docs = self.validation_docs()
        else:
            assert (
                False
            ), f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"

648
        # Test One Doc
649
        test_doc = docs[0]
650
        self.features = list(test_doc.keys())
651
        test_text = self.doc_to_text(test_doc)
652
        test_target = self.doc_to_target(test_doc)
lintangsutawika's avatar
lintangsutawika committed
653
654
655
656
657

        if self._config.doc_to_choice is not None:
            test_choice = self.doc_to_choice(test_doc)
            if type(test_choice) is not list:
                eval_logger.error("doc_to_choice must return list")
658

659
        if self._config.output_type == "multiple_choice":
660
            if type(test_text) is list:
lintangsutawika's avatar
lintangsutawika committed
661
                self.multiple_input = len(test_text)
662
            elif type(test_text) is str:
lintangsutawika's avatar
lintangsutawika committed
663
                self.multiple_input = 0
664

665
666
667
668
        if type(test_target) is list:
            self.multiple_output = len(test_target)
        else:
            self.multiple_output = 0
669

670
671
672
673
674
675
676
677
    def download(self, dataset_kwargs=None):

        self.dataset = datasets.load_dataset(
            path=self.DATASET_PATH,
            name=self.DATASET_NAME,
            **dataset_kwargs if dataset_kwargs is not None else {},
        )

678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
    def has_training_docs(self):
        if self._config.training_split is not None:
            return True
        else:
            return False

    def has_validation_docs(self):
        if self._config.validation_split is not None:
            return True
        else:
            return False

    def has_test_docs(self):
        if self._config.test_split is not None:
            return True
        else:
            return False

    def training_docs(self):
        if self._config.training_split is not None:
            return self.dataset[self._config.training_split]

    def validation_docs(self):
        if self._config.validation_split is not None:
            return self.dataset[self._config.validation_split]

    def test_docs(self):
        if self._config.test_split is not None:
            return self.dataset[self._config.test_split]

708
    def fewshot_docs(self):
709
        if self._config.fewshot_split is not None:
710
            return self.dataset[self._config.fewshot_split]
711
712
713
        else:
            if self._config.num_fewshot > 0:
                eval_logger.warning(
haileyschoelkopf's avatar
haileyschoelkopf committed
714
                    f"Task '{self._config.task}': "
715
716
717
718
                    "num_fewshot > 0 but fewshot_split is None. "
                    "using preconfigured rule."
                )
            return super().fewshot_docs()
719

720
721
722
723
724
    def should_decontaminate(self):
        return self._config.should_decontaminate

    def doc_to_decontamination_query(self, doc):
        if self._config.should_decontaminate:
725
726
727
728
729
730
            if self._config.doc_to_decontamination_query in self.features:
                return doc[self._config.doc_to_decontamination_query]
            else:
                return ast.literal_eval(
                    utils.apply_template(self._config.doc_to_decontamination_query, doc)
                )
731

732
733
734
735
736
737
738
739
740
741
742
743
    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):
744
745
746

        if self.prompt is not None:
            doc_to_text = self.prompt
747
748
        else:
            doc_to_text = self._config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
749

750
        if type(doc_to_text) == str:
751
752
753
754
            if doc_to_text in self.features:
                return doc[doc_to_text]
            else:
                return utils.apply_template(doc_to_text, doc)
755
        elif callable(doc_to_text):
756
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
757
        # Used when applying a Promptsource template
758
        elif hasattr(doc_to_text, "apply"):
759
            return doc_to_text.apply(doc)[0]
760
        else:
761
            print(type(doc_to_text))
762
            raise TypeError
763
764

    def doc_to_target(self, doc):
765
766
767

        if self.prompt is not None:
            doc_to_target = self.prompt
768
769
770
        else:
            doc_to_target = self._config.doc_to_target

771
        if type(doc_to_target) == str:
772
            if doc_to_target in self.features:
lintangsutawika's avatar
lintangsutawika committed
773
774
775
776
                if self._config.doc_to_choice is not None:
                    return self.doc_to_choice(doc)[doc[doc_to_target]]
                else:
                    return doc[doc_to_target]
777
778
            else:
                return utils.apply_template(doc_to_target, doc)
779
        elif callable(doc_to_target):
780
            return doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
781
        # Used when applying a Promptsource template
782
783
        elif hasattr(doc_to_target, "apply"):
            return doc_to_target.apply(doc)[1]
784
785
        else:
            raise TypeError
786

787
    def gold_alias(self, doc):
788
789
790
791
792
        # returns a version of the gold target answer to a document,
        # which should be passed into metric for scoring as the ground truth.

        # in multiple_choice tasks, this should be castable to an int corresponding to the index
        # within the answer choices, while doc_to_target is the string version of {{answer_choices[gold]}}.
lintangsutawika's avatar
lintangsutawika committed
793
        if self._config.gold_alias is not None:
794
795
            doc_to_target = self._config.gold_alias
        else:
lintangsutawika's avatar
lintangsutawika committed
796
            return self.doc_to_target(doc)
797
798
799
800
801
802
803
804
805
806

        if type(doc_to_target) == str:
            return utils.apply_template(doc_to_target, doc)
        elif callable(doc_to_target):
            return doc_to_target(doc)
        elif hasattr(doc_to_target, "apply"):
            return doc_to_target.apply(doc)[1]
        else:
            raise TypeError

807
808
    def construct_requests(self, doc, ctx, **kwargs):

809
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
810
            arguments = (ctx, self.doc_to_target(doc))
811
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
812
            arguments = (self.doc_to_target(doc),)
813
        elif self.OUTPUT_TYPE == "multiple_choice":
814
            # we pass the user-defined answer_choices var (in aliases) and translate the result to a Python list.
lintangsutawika's avatar
lintangsutawika committed
815
816
            # TODO: any cleaner way to do this?
            if self.multiple_input > 0:
817
                choices = self.doc_to_text(doc)
818
                cont = self.doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
819
                arguments = [(ctx, " {}".format(cont)) for ctx in choices]
820
            else:
821
                choices = self.doc_to_choice(doc)
lintangsutawika's avatar
lintangsutawika committed
822
                arguments = [(ctx, " {}".format(cont)) for cont in choices]
823

824
            request_list = [
825
826
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
827
                    doc=doc,
828
                    arguments=arg,
829
                    idx=i,
830
831
                    **kwargs,
                )
832
                for i, arg in enumerate(arguments)
833
            ]
834
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
835
            if "acc_mutual_info" in self._metric_fn_list.keys():
836
837
838
                # if we are calculating multiple choice accuracy
                # using mutual information instead of raw loglikelihood as metric, need unconditional lls.

lintangsutawika's avatar
lintangsutawika committed
839
                # here mutual info refers to calculating
840
841
842
843
844
845
                # 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
846
                            doc=doc,
847
                            arguments=("", "{}".format(choice)),
848
849
850
                            idx=i,
                            **kwargs,
                        )
lintangsutawika's avatar
lintangsutawika committed
851
                        for i, choice in enumerate(choices)
852
853
854
                    ]
                )
            return request_list
lintangsutawika's avatar
lintangsutawika committed
855

856
        elif self.OUTPUT_TYPE == "greedy_until":
857
            arguments = (ctx, self._config.generation_kwargs)
lintangsutawika's avatar
lintangsutawika committed
858

859
860
861
862
        elif self.OUTPUT_TYPE == "winograd_schema":
            # similar to multiple_choice task type except each request contains
            # multiple differing contexts with the same continuation

863
            contexts = self.doc_to_choice(doc)
864
            choice = self.doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
865

866
867
868
869
870
871
872
873
874
875
            request_list = [
                Instance(
                    request_type="loglikelihood",
                    doc=doc,
                    arguments=(context, " {}".format(choice)),
                    idx=i,
                    **kwargs,
                )
                for i, context in enumerate(contexts)
            ]
lintangsutawika's avatar
lintangsutawika committed
876

877
878
            return request_list

lintangsutawika's avatar
lintangsutawika committed
879
        return Instance(
lintangsutawika's avatar
lintangsutawika committed
880
881
            request_type=self.OUTPUT_TYPE, doc=doc, arguments=arguments, idx=0, **kwargs
        )
882
883
884

    def process_results(self, doc, results):

lintangsutawika's avatar
lintangsutawika committed
885
886
887
        # if callable(self._config.process_results):
        #     return self._config.process_results(doc, results)

888
        result_dict = {}
889
        use_metric = list(self._metric_fn_list.keys())
890
891
892
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
893
894
895
896
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
897
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
898
            (loglikelihood,) = results
899
900
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
haileyschoelkopf's avatar
haileyschoelkopf committed
901
            return {
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
                **(
                    {"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
917
            }
918
        elif self.OUTPUT_TYPE == "multiple_choice":
919
920

            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
921

922
            # retrieve choices in List[str] form, to compute choice lengths, etc.
lintangsutawika's avatar
lintangsutawika committed
923
            if self.multiple_input:
924
                choices = [self.doc_to_target(doc)] * self.multiple_input
lintangsutawika's avatar
lintangsutawika committed
925
            else:
926
                choices = self.doc_to_choice(doc)
927
928
            completion_len = np.array([float(len(i)) for i in choices])

929
930
            if (
                2 * len(choices) == len(lls)
931
                and "acc_mutual_info" in self._metric_fn_list.keys()
932
933
934
935
936
937
938
            ):
                # 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]
939

940
941
            pred_idx = np.argmax(lls)
            pred_idx_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
942

943
            # Gives priority to evaluate base on gold_alias
944
            if self._config.gold_alias is not None:
945
946
                pred = pred_idx
                pred_norm = pred_idx_norm
947
948
                gold_idx = int(self.gold_alias(doc))
                gold = gold_idx
949
            else:
950
951
                pred = choices[pred_idx]
                pred_norm = choices[pred_idx_norm]
952
                gold = self.doc_to_target(doc)
953
                if self.multiple_output > 0:
lintangsutawika's avatar
lintangsutawika committed
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
                    if type(gold[0]) == int:
                        gold_idx = gold
                        gold = [choices[idx] for idx in gold_idx]
                    elif type(gold[0]) == str:
                        gold_idx = [choices.index(g) for g in gold]
                else:
                    if type(gold) == int:
                        gold_idx = gold
                        gold = choices[gold_idx]
                    elif type(gold) == str:
                        gold_idx = choices.index(gold)

            if self.multiple_output:
                acc = 1.0 if pred in gold else 0.0
                acc_norm = 1.0 if pred_norm in gold else 0.0
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
972
973

            result_dict = {
974
                **({"acc": acc} if "acc" in use_metric else {}),
975
976
                **({"f1": (gold_idx, pred_idx)} if "f1" in use_metric else {}),
                **({"mcc": (gold_idx, pred_idx)} if "mcc" in use_metric else {}),
977
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
978
979
            }

980
            if "exact_match" in self._metric_fn_list.keys():
981
                # TODO: this gets score of 0 on arc_challenge for pythia-70m. need to test that this works properly
lintangsutawika's avatar
lintangsutawika committed
982
                is_greedy = is_greedy[gold]  # take value for the gold answer
983
984
                result_dict["exact_match"] = int(is_greedy)

985
            if "acc_mutual_info" in use_metric:
lintangsutawika's avatar
lintangsutawika committed
986
987
988
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
989
990
991
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
        elif self.OUTPUT_TYPE == "winograd_schema":

            lls, is_greedy = zip(*results)
            if self._config.gold_alias is not None:
                gold = int(self.gold_alias(doc))
            else:
                gold = int(self.doc_to_target(doc))

            pred = np.argmax(lls)
            acc = 1.0 if np.argmax(lls) == gold else 0.0

            result_dict = {
                **({"acc": acc} if "acc" in use_metric else {}),
            }

1007
1008
1009
        elif self.OUTPUT_TYPE == "greedy_until":

            if self._config.gold_alias is not None:
1010
                gold = self.gold_alias(doc)
1011
1012
1013
            else:
                gold = self.doc_to_target(doc)

1014
            for key, result in zip(self._metric_fn_list.keys(), results):
haileyschoelkopf's avatar
haileyschoelkopf committed
1015
                _dict = self._metric_fn_list[key](
1016
                    references=gold if self.multiple_output else [gold],
haileyschoelkopf's avatar
haileyschoelkopf committed
1017
1018
                    predictions=[result],
                    **self._metric_fn_kwargs[key],
1019
                )
1020

lintangsutawika's avatar
lintangsutawika committed
1021
                result_dict = {**result_dict, **_dict}
1022
        else:
lintangsutawika's avatar
lintangsutawika committed
1023
1024
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
1025
                "'loglikelihood', 'loglikelihood_rolling', 'greedy_until', 'multiple_choice' or 'winograd_schema' ",
1026
            )
1027
1028
1029
1030
1031
1032
1033

        return result_dict

    def aggregation(self):
        return self._aggregation_list

    def higher_is_better(self):
haileyschoelkopf's avatar
haileyschoelkopf committed
1034
        return self._higher_is_better
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044


class MultipleChoiceTask(Task):

    OUTPUT_TYPE: str = "loglikelihood"

    def doc_to_target(self, doc):
        return " " + doc["choices"][doc["gold"]]

    def construct_requests(self, doc, ctx, **kwargs):
1045
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1046
1047
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1048
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1049
                doc=doc,
1050
                arguments=(ctx, " {}".format(choice)),
1051
                idx=i,
1052
1053
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1054
1055
            for i, choice in enumerate(doc["choices"])
        ]
1056
1057

    def process_results(self, doc, results):
lintangsutawika's avatar
lintangsutawika committed
1058
1059
1060
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
        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,
        }

    def higher_is_better(self):
        return {
            "acc": True,
            "acc_norm": True,
        }

    def aggregation(self):
        return {
            "acc": mean,
            "acc_norm": mean,
        }


lintangsutawika's avatar
lintangsutawika committed
1085
class PerplexityTask(Task):
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095

    OUTPUT_TYPE = "loglikelihood_rolling"

    def has_training_docs(self):
        return False

    def fewshot_examples(self, k, rnd):
        assert k == 0
        return []

lintangsutawika's avatar
lintangsutawika committed
1096
    def fewshot_context(self, doc, num_fewshot, rnd=None):
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
        assert (
            num_fewshot == 0
        ), "The number of fewshot examples must be 0 for perplexity tasks."
        assert (
            rnd is not None
        ), "A `random.Random` generator argument must be provided to `rnd`."

        return ""

    def higher_is_better(self):
        return {
            "word_perplexity": False,
            "byte_perplexity": False,
            "bits_per_byte": False,
        }

    def doc_to_decontamination_query(self, doc):
        return doc

    def doc_to_text(self, doc):
        return ""

    def doc_to_target(self, doc):
        return doc

    def construct_requests(self, doc, ctx, **kwargs):
        assert not ctx

lintangsutawika's avatar
lintangsutawika committed
1125
1126
1127
1128
1129
1130
1131
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1132
1133
1134

    def process_results(self, doc, results):
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1135
1136
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
        return {
            "word_perplexity": (loglikelihood, words),
            "byte_perplexity": (loglikelihood, bytes_),
            "bits_per_byte": (loglikelihood, bytes_),
        }

    def aggregation(self):
        return {
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
            "bits_per_byte": bits_per_byte,
        }

    @classmethod
    def count_bytes(cls, doc):
        return len(doc.encode("utf-8"))

    @classmethod
    def count_words(cls, doc):
        """Downstream tasks with custom word boundaries should override this!"""
        return len(re.split(r"\s+", doc))