task.py 40.6 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
from tqdm import tqdm
12
13
14
15

import datasets
import numpy as np

16
17
from typing import Union
from collections.abc import Callable
18

19
from lm_eval import utils
20
from lm_eval.api import samplers
haileyschoelkopf's avatar
haileyschoelkopf committed
21
from lm_eval.api.instance import Instance
lintangsutawika's avatar
lintangsutawika committed
22
from lm_eval.api.filter import FilterEnsemble
23
24
25
26

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
27
28
29
30
from lm_eval.api.metrics import (
    mean,
    weighted_perplexity,
    bits_per_byte,
lintangsutawika's avatar
lintangsutawika committed
31
    metric_max_over_ground_truths,
lintangsutawika's avatar
lintangsutawika committed
32
33
)
from lm_eval.api.registry import (
haileyschoelkopf's avatar
haileyschoelkopf committed
34
35
36
37
    get_metric,
    get_aggregation,
    get_default_aggregation,
    is_higher_better,
38
39
    DEFAULT_METRIC_REGISTRY,
    OUTPUT_TYPE_REGISTRY,
lintangsutawika's avatar
lintangsutawika committed
40
41
    AGGREGATION_REGISTRY,
)
42

43
44
45
46
47
48
49
ALL_OUTPUT_TYPES = [
    "loglikelihood",
    "multiple_choice",
    "loglikelihood_rolling",
    "greedy_until",
]

50
51
52

@dataclass
class TaskConfig(dict):
53
    # task naming/registry
54
    task: str = None
55
    group: Union[str, list] = None
56
57
58
    # HF dataset options.
    # which dataset to use,
    # and what splits for what purpose
59
60
    dataset_path: str = None
    dataset_name: str = None
61
    dataset_kwargs: dict = None
62
63
64
    training_split: str = None
    validation_split: str = None
    test_split: str = None
lintangsutawika's avatar
lintangsutawika committed
65
    fewshot_split: str = None  # TODO: assert that this not None if num_fewshot > 0. (?) assert if this is same split as one evaling (?)
66
67
    # formatting / prompting options.
    # see docs/advanced_task_guide.md for more info
68
    template_aliases: Union[str, list] = None
69
70
    doc_to_text: Union[Callable, str] = None
    doc_to_target: Union[Callable, str] = None
lintangsutawika's avatar
lintangsutawika committed
71
    doc_to_choice: Union[Callable, str, dict, list] = None
72
    gold_alias: Union[Callable, str] = None
73
    use_prompt: str = None
74
    description: str = ""
75
76
    target_delimiter: str = " "
    fewshot_delimiter: str = "\n\n"
77
    # runtime configuration options
78
    num_fewshot: int = 0
79
    # scoring options
80
81
    metric_list: str = None
    output_type: str = "greedy_until"
82
    generation_kwargs: dict = None
83
    repeats: int = 1
lintangsutawika's avatar
lintangsutawika committed
84
    filter_list: Union[str, list] = None
85
86
    should_decontaminate: bool = False
    doc_to_decontamination_query: str = None
87

lintangsutawika's avatar
lintangsutawika committed
88
    metadata: str = None  # by default, not used in the code. allows for users to pass arbitrary info to tasks
89

90
91
92
93
    def __post_init__(self):
        # allow user-specified aliases so that users can
        # force prompt-compatibility for some prompt regardless of
        # field names in prompt
94
        if self.template_aliases:
95
96
            if type(self.doc_to_text) == str:
                self.doc_to_text = self.template_aliases + self.doc_to_text
97

98
99
            if type(self.doc_to_target) == str:
                self.doc_to_target = self.template_aliases + self.doc_to_target
100

101
            if type(self.gold_alias) == str:
lintangsutawika's avatar
lintangsutawika committed
102
                self.gold_alias = self.template_aliases + self.gold_alias
103

Lintang Sutawika's avatar
Lintang Sutawika committed
104
105
106
107
108
109
110
111
112
113
114
115
        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:
116
                self.generation_kwargs["until"] = [self.fewshot_delimiter]
Lintang Sutawika's avatar
Lintang Sutawika committed
117
118
119
120
        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
121
                    "until": None
122
123
                    if self.fewshot_delimiter is None
                    else [self.fewshot_delimiter],
Lintang Sutawika's avatar
Lintang Sutawika committed
124
125
126
                    "do_sample": False,
                    "temperature": 0.0,
                }
127

haileyschoelkopf's avatar
haileyschoelkopf committed
128
129
        # TODO: how to make TaskConfigs be de- and re-serializable, even when using the !function constructor?

130
131
132
    def __getitem__(self, item):
        return getattr(self, item)

133
    def to_dict(self):
134
135
        """dumps the current config as a dictionary object, as a printable format.
        null fields will not be printed.
haileyschoelkopf's avatar
haileyschoelkopf committed
136
        Used for dumping results alongside full task configuration
137

haileyschoelkopf's avatar
haileyschoelkopf committed
138
139
140
141
142
143
144
145
146
147
        :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
148
149
150
            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
151
        return cfg_dict
152

153
154
155
156
157
158
159
160
161
162
163
164

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
165

166
167
168
169
170
171
172
173
    # 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
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
208
    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
209
        self._config = TaskConfig(**config) if config else TaskConfig()
210
211
212

        if not hasattr(self, "_filters"):
            self._filters = []
lintangsutawika's avatar
lintangsutawika committed
213
            for name, components in self._config.get(
214
                "filters", [["none", [["take_first", None]]]]
lintangsutawika's avatar
lintangsutawika committed
215
            ):
216
217
218
                filter_pipeline = build_filter_ensemble(name, components)
                self._filters.append(filter_pipeline)

lintangsutawika's avatar
lintangsutawika committed
219
        self.sampler = samplers.Sampler(
220
221
            list(self.fewshot_docs()), self, rnd=random.Random(1234)
        )
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
247

    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.
        """
248
249
250
251
252
253
254
        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,
        )
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
291

    @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 []

292
293
294
295
296
297
298
299
300
301
    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
302
            eval_logger.warning(
303
                "has_training_docs and has_validation_docs are False"
304
                ", using test_docs as fewshot_docs but this is not recommended."
lintangsutawika's avatar
lintangsutawika committed
305
            )
306
307
            return self.test_docs()

308
309
310
311
312
313
314
315
316
317
    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
318

319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
    @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

346
    def build_all_requests(self, limit=None, rank=None, world_size=None):
347
348
349
350
351
352
353
354
355
356
        """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!"

357
358
359
360
        eval_logger.info(
            f"Building contexts for task '{self._config.task}' on rank {rank}..."
        )

361
        instances = []
362
363
        for doc_id, doc in utils.create_iterator(
            enumerate(docs), rank, world_size, limit
lintangsutawika's avatar
lintangsutawika committed
364
        ):
365
            # sample fewshot context #TODO: need to offset doc_id by rank now!
366
            fewshot_ctx = self.fewshot_context(
367
368
                doc,
                self._config.num_fewshot,
369
            )
370

haileyschoelkopf's avatar
haileyschoelkopf committed
371
            # TODO: we should override self._config.repeats if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
372
373
374
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
375
                metadata=(self._config["task"], doc_id, self._config.repeats),
lintangsutawika's avatar
lintangsutawika committed
376
            )
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401

            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
402
            The number of times each instance in a dataset is inferred on. Defaults to 1,
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
436
437
            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
438
439
440
441
442
443
444
445
446
447
    @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))

448
    @utils.positional_deprecated
449
    def fewshot_context(self, doc, num_fewshot):
450
451
452
453
454
455
456
457
458
459
460
461
        """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:
462
463
            # always prepend the (possibly empty) task description
            labeled_examples = self._config.description
464
        else:
lintangsutawika's avatar
lintangsutawika committed
465
466
467
            labeled_examples = self._config.description + self.sampler.get_context(
                doc, num_fewshot
            )
468
469

        example = self.doc_to_text(doc)
470
471
472
473
        if type(example) == str:
            return labeled_examples + example
        elif type(example) == list:
            return [labeled_examples + ex for ex in example]
474
475
476
        elif type(example) == int:
            choices = self.doc_to_choice(doc)
            return labeled_examples + choices[example]
477
478
479

    def apply_filters(self):

lintangsutawika's avatar
lintangsutawika committed
480
481
482
483
484
485
        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
486

487
    def dump_config(self):
488
        """Returns a dictionary representing the task's config.
489
490
491
492
493

        :returns: str
            The fewshot context.
        """
        # TODO: this should only return the overrides applied to a non-YAML task's configuration.
494
        # (num_fewshot)
495
496
        return self._config.to_dict()

497
498
499

class ConfigurableTask(Task):

500
    VERSION = "Yaml"
501
    OUTPUT_TYPE = None
502
    CONFIG = None
503
504
505
506

    def __init__(
        self, data_dir=None, cache_dir=None, download_mode=None, config: dict = None
    ):
507
        # Get pre-configured attributes
508
        self._config = self.CONFIG
509

510
511
        # Use new configurations if there was no preconfiguration
        if self._config is None:
512
            self._config = TaskConfig(**config)
513
514
        # Overwrite configs
        else:
lintangsutawika's avatar
lintangsutawika committed
515
            if config is not None:
516
                self._config.__dict__.update(config)
517

518
        if self._config is None:
lintangsutawika's avatar
lintangsutawika committed
519
520
521
            raise ValueError(
                "Must pass a config to ConfigurableTask, either in cls.CONFIG or `config` kwarg"
            )
522
523

        if self._config.output_type is not None:
524
            assert self._config.output_type in ALL_OUTPUT_TYPES
525
526
            self.OUTPUT_TYPE = self._config.output_type

527
528
529
530
531
532
        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

533
534
535
536
        self._metric_fn_list = {}
        self._metric_fn_kwargs = {}
        self._aggregation_list = {}
        self._higher_is_better = {}
537

538
        _metric_list = DEFAULT_METRIC_REGISTRY[self._config.output_type]
539
        if self._config.metric_list is None:
540
            # TODO: handle this in TaskConfig.__post_init__ ?
541
            for metric_name in _metric_list:
haileyschoelkopf's avatar
haileyschoelkopf committed
542
543
                self._metric_fn_list[metric_name] = get_metric(metric_name)
                self._aggregation_list[metric_name] = get_default_aggregation(
544
                    metric_name
haileyschoelkopf's avatar
haileyschoelkopf committed
545
546
                )
                self._higher_is_better[metric_name] = is_higher_better(metric_name)
547
548
549
550
551
552
553
554
555
        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
556
557
                self._metric_fn_list[metric_name] = get_metric(metric_name)
                self._metric_fn_kwargs[metric_name] = kwargs
lintangsutawika's avatar
lintangsutawika committed
558

559
                if "aggregation" in metric_config:
560
                    agg_name = metric_config["aggregation"]
561
                    if type(agg_name) == str:
haileyschoelkopf's avatar
haileyschoelkopf committed
562
                        self._aggregation_list[metric_name] = get_aggregation(agg_name)
563
564
565
566
                    elif callable(agg_name):
                        self._aggregation_list[metric_name] = metric_config[
                            "aggregation"
                        ]
567
                else:
568
569

                    INV_AGG_REGISTRY = {v: k for k, v in AGGREGATION_REGISTRY.items()}
haileyschoelkopf's avatar
haileyschoelkopf committed
570
                    metric_agg = get_default_aggregation(metric_name)
571
                    eval_logger.warning(
572
573
574
                        f"metric {metric_name} is defined, but aggregation is not. "
                        f"using default "
                        f"aggregation={INV_AGG_REGISTRY[metric_agg]}"
575
                    )
576
                    self._aggregation_list[metric_name] = metric_agg
lintangsutawika's avatar
lintangsutawika committed
577

578
579
580
581
582
583
                if "higher_is_better" in metric_config:
                    self._higher_is_better[metric_name] = metric_config[
                        "higher_is_better"
                    ]
                else:
                    eval_logger.warning(
584
585
                        f"metric {metric_name} is defined, but higher_is_better is not. "
                        f"using default "
haileyschoelkopf's avatar
haileyschoelkopf committed
586
                        f"higher_is_better={is_higher_better(metric_name)}"
587
                    )
haileyschoelkopf's avatar
haileyschoelkopf committed
588
                    self._higher_is_better[metric_name] = is_higher_better(metric_name)
589

590
        self.download(self._config.dataset_kwargs)
591
592
593
        self._training_docs = None
        self._fewshot_docs = None

lintangsutawika's avatar
lintangsutawika committed
594
        if self._config.filter_list is not None:
lintangsutawika's avatar
lintangsutawika committed
595
            self._filters = []
lintangsutawika's avatar
lintangsutawika committed
596
597
598
599
600
601
602
603
            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
604
605
606
                        }
                        components.append([function["function"], kwargs])
                    filter_pipeline = build_filter_ensemble(filter_name, components)
lintangsutawika's avatar
lintangsutawika committed
607
                self._filters.append(filter_pipeline)
lintangsutawika's avatar
lintangsutawika committed
608
        else:
609
            self._filters = [build_filter_ensemble("none", [["take_first", None]])]
610
611

        if self._config.use_prompt is not None:
lintangsutawika's avatar
lintangsutawika committed
612
            eval_logger.info(f"loading prompt {self._config.use_prompt}")
613
            self.prompt = get_prompt(
lintangsutawika's avatar
lintangsutawika committed
614
615
                self._config.use_prompt, self.DATASET_PATH, self.DATASET_NAME
            )
616
617
618
        else:
            self.prompt = None

lintangsutawika's avatar
lintangsutawika committed
619
620
        if self.fewshot_docs() is not None:
            self.sampler = samplers.Sampler(
621
                list(self.fewshot_docs()), self, rnd=random.Random(1234)
622
            )
623

624
625
626
627
        if self._config.template_aliases is not None:
            for key, alias in self._config.template_aliases:
                self.dataset.rename_column(key, alias)

628
629
630
631
632
633
634
635
636
        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!"

637
        # Test One Doc
638
639
640
        self.features = list(docs.features.keys())
        self.multiple_input = 0
        self.multiple_target = 0
641
642
        test_doc = docs[0]
        test_text = self.doc_to_text(test_doc)
643
        test_target = self.doc_to_target(test_doc)
lintangsutawika's avatar
lintangsutawika committed
644
645
646
647
648

        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")
649
650
            else:
                num_choice = len(test_choice)
651

652
653
            if type(test_text) is int:
                self.multiple_input = num_choice
654

655
        if type(test_target) is list:
656
657
            self.multiple_target = len(test_target)

658
659
660
661
662
663
664
665
    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 {},
        )

666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
    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]

696
    def fewshot_docs(self):
697
        if self._config.fewshot_split is not None:
698
            return self.dataset[self._config.fewshot_split]
699
700
701
        else:
            if self._config.num_fewshot > 0:
                eval_logger.warning(
haileyschoelkopf's avatar
haileyschoelkopf committed
702
                    f"Task '{self._config.task}': "
703
704
705
706
                    "num_fewshot > 0 but fewshot_split is None. "
                    "using preconfigured rule."
                )
            return super().fewshot_docs()
707

708
709
710
711
712
    def should_decontaminate(self):
        return self._config.should_decontaminate

    def doc_to_decontamination_query(self, doc):
        if self._config.should_decontaminate:
713
714
715
716
717
718
            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)
                )
719

720
721
722
723
724
725
726
727
728
729
730
731
    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):
732
733
734

        if self.prompt is not None:
            doc_to_text = self.prompt
735
736
        else:
            doc_to_text = self._config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
737

738
739
740
        if type(doc_to_text) == int:
            return doc_to_text
        elif type(doc_to_text) == str:
741
            if doc_to_text in self.features:
742
743
744
                # if self._config.doc_to_choice is not None:
                #     return self.doc_to_choice(doc)[doc[doc_to_text]]
                # else:
745
746
                return doc[doc_to_text]
            else:
lintangsutawika's avatar
lintangsutawika committed
747
748
749
750
751
                text_string = utils.apply_template(doc_to_text, doc)
                if text_string.isdigit():
                    return ast.literal_eval(text_string)
                else:
                    return text_string
752
        elif callable(doc_to_text):
753
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
754
        # Used when applying a Promptsource template
755
        elif hasattr(doc_to_text, "apply"):
756
            return doc_to_text.apply(doc)[0]
757
        else:
758
            print(type(doc_to_text))
759
            raise TypeError
760
761

    def doc_to_target(self, doc):
762
763
764

        if self.prompt is not None:
            doc_to_target = self.prompt
765
766
767
        else:
            doc_to_target = self._config.doc_to_target

768
769
770
        if type(doc_to_target) == int:
            return doc_to_target
        elif type(doc_to_target) == str:
771
            if doc_to_target in self.features:
772
773
774
775
                # 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]
776
            else:
lintangsutawika's avatar
lintangsutawika committed
777
778
779
780
781
                target_string = utils.apply_template(doc_to_target, doc)
                if target_string.isdigit():
                    return ast.literal_eval(target_string)
                else:
                    return target_string
782
        elif callable(doc_to_target):
783
            return doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
784
        # Used when applying a Promptsource template
785
786
        elif hasattr(doc_to_target, "apply"):
            return doc_to_target.apply(doc)[1]
787
788
        else:
            raise TypeError
789
790
791
792
793

    def doc_to_choice(self, doc):

        if self.prompt is not None:
            doc_to_choice = self.prompt
lintangsutawika's avatar
lintangsutawika committed
794
        elif self._config.doc_to_choice is None:
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
            eval_logger.error("doc_to_choice was called but not set in config")
        else:
            doc_to_choice = self._config.doc_to_choice

        if type(doc_to_choice) == str:
            return ast.literal_eval(utils.apply_template(doc_to_choice, doc))
        elif type(doc_to_choice) == list:
            return doc_to_choice
        elif type(doc_to_choice) == dict:
            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
811

812
    def gold_alias(self, doc):
813
814
815
816
817
        # 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
818
        if self._config.gold_alias is not None:
819
820
            doc_to_target = self._config.gold_alias
        else:
lintangsutawika's avatar
lintangsutawika committed
821
            return self.doc_to_target(doc)
822
823
824
825
826
827
828
829
830
831

        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

832
833
    def construct_requests(self, doc, ctx, **kwargs):

834
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
835
            arguments = (ctx, self.doc_to_target(doc))
836
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
837
            arguments = (self.doc_to_target(doc),)
838
        elif self.OUTPUT_TYPE == "multiple_choice":
839
840
841
842

            choices = self.doc_to_choice(doc)
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
843
                cont = self.doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
844
                arguments = [(ctx, " {}".format(cont)) for ctx in choices]
845
            else:
846
                # Otherwise they are placed in the continuation
lintangsutawika's avatar
lintangsutawika committed
847
                arguments = [(ctx, " {}".format(cont)) for cont in choices]
848

849
            request_list = [
850
851
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
852
                    doc=doc,
853
                    arguments=arg,
854
                    idx=i,
855
856
                    **kwargs,
                )
857
                for i, arg in enumerate(arguments)
858
            ]
859
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
860
            if "acc_mutual_info" in self._metric_fn_list.keys():
861
862
863
                # if we are calculating multiple choice accuracy
                # using mutual information instead of raw loglikelihood as metric, need unconditional lls.

lintangsutawika's avatar
lintangsutawika committed
864
                # here mutual info refers to calculating
865
866
867
868
869
870
                # 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
871
                            doc=doc,
872
                            arguments=("", "{}".format(choice)),
873
874
875
                            idx=i,
                            **kwargs,
                        )
lintangsutawika's avatar
lintangsutawika committed
876
                        for i, choice in enumerate(choices)
877
878
879
                    ]
                )
            return request_list
lintangsutawika's avatar
lintangsutawika committed
880

881
        elif self.OUTPUT_TYPE == "greedy_until":
882
            arguments = (ctx, self._config.generation_kwargs)
lintangsutawika's avatar
lintangsutawika committed
883
884

        return Instance(
lintangsutawika's avatar
lintangsutawika committed
885
886
            request_type=self.OUTPUT_TYPE, doc=doc, arguments=arguments, idx=0, **kwargs
        )
887
888
889

    def process_results(self, doc, results):

lintangsutawika's avatar
lintangsutawika committed
890
891
892
        # if callable(self._config.process_results):
        #     return self._config.process_results(doc, results)

893
        result_dict = {}
894
        use_metric = list(self._metric_fn_list.keys())
895
896
897
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
898
899
900
901
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
902
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
903
            (loglikelihood,) = results
904
905
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
haileyschoelkopf's avatar
haileyschoelkopf committed
906
            return {
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
                **(
                    {"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
922
            }
923
        elif self.OUTPUT_TYPE == "multiple_choice":
924
925

            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
926

927
            # retrieve choices in List[str] form, to compute choice lengths, etc.
928
            choices = self.doc_to_choice(doc)
929
930
            completion_len = np.array([float(len(i)) for i in choices])

931
932
            if (
                2 * len(choices) == len(lls)
933
                and "acc_mutual_info" in self._metric_fn_list.keys()
934
935
936
937
938
939
940
            ):
                # 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]
941

942
943
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
944

945
946
            if self.multiple_input:
                gold = self.doc_to_text(doc)
947
            else:
948
                gold = self.doc_to_target(doc)
949
950
                if type(gold) is str:
                    gold = choices.index(gold)
lintangsutawika's avatar
lintangsutawika committed
951

952
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
953
954
                acc = 1.0 if pred in gold else 0.0
                acc_norm = 1.0 if pred_norm in gold else 0.0
955
                exact_match = int(any([is_greedy[i] for i in gold]))
lintangsutawika's avatar
lintangsutawika committed
956
957
958
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
959
960
                # TODO: this gets score of 0 on arc_challenge for pythia-70m. need to test that this works properly
                exact_match = int(is_greedy[gold])
961
962

            result_dict = {
963
                **({"acc": acc} if "acc" in use_metric else {}),
964
965
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
966
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
967
                **({"exact_match": exact_match} if "exact_match" in use_metric else {}),
968
969
            }

970
            if "acc_mutual_info" in use_metric:
lintangsutawika's avatar
lintangsutawika committed
971
972
973
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
974
975
976
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

977
978
        elif self.OUTPUT_TYPE == "greedy_until":

979
            gold = self.doc_to_target(doc)
980

981
            for key, result in zip(self._metric_fn_list.keys(), results):
haileyschoelkopf's avatar
haileyschoelkopf committed
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
                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 = []
                    for gold_option in gold:
                        res = self._metric_fn_list[key](
                            references=[gold_option],
                            predictions=[result],
                            **self._metric_fn_kwargs[key],
                        )
                        if isinstance(res, dict):
                            # TODO: this handles the case where HF evaluate returns a dict.
                            res = res[key]
                        scores.append(res)
                    if any(scores):
                        result = 1.0
                    else:
                        result = 0.0
                else:
                    result = self._metric_fn_list[key](
                        references=[gold],
                        predictions=[result],
                        **self._metric_fn_kwargs[key],
                    )
1007

haileyschoelkopf's avatar
haileyschoelkopf committed
1008
1009
1010
1011
                if isinstance(result, dict):
                    result_dict.update(result)
                else:
                    result_dict[key] = result
1012
        else:
lintangsutawika's avatar
lintangsutawika committed
1013
1014
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
1015
                "'loglikelihood', 'loglikelihood_rolling', 'greedy_until' or 'multiple_choice'",
1016
            )
1017
1018
1019
1020
1021
1022
1023

        return result_dict

    def aggregation(self):
        return self._aggregation_list

    def higher_is_better(self):
haileyschoelkopf's avatar
haileyschoelkopf committed
1024
        return self._higher_is_better
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034


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):
1035
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1036
1037
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1038
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1039
                doc=doc,
1040
                arguments=(ctx, " {}".format(choice)),
1041
                idx=i,
1042
1043
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1044
1045
            for i, choice in enumerate(doc["choices"])
        ]
1046
1047

    def process_results(self, doc, results):
lintangsutawika's avatar
lintangsutawika committed
1048
1049
1050
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
        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
1075
class PerplexityTask(Task):
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085

    OUTPUT_TYPE = "loglikelihood_rolling"

    def has_training_docs(self):
        return False

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

1086
    def fewshot_context(self, doc, num_fewshot):
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
        assert (
            num_fewshot == 0
        ), "The number of fewshot examples must be 0 for perplexity tasks."

        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
1112
1113
1114
1115
1116
1117
1118
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1119
1120
1121

    def process_results(self, doc, results):
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1122
1123
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
        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))