task.py 40.9 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
lintangsutawika's avatar
lintangsutawika committed
73
    process_results: Union[Callable, str] = None
74
    use_prompt: str = None
75
    description: str = ""
76
77
    target_delimiter: str = " "
    fewshot_delimiter: str = "\n\n"
78
    # runtime configuration options
79
    num_fewshot: int = 0
80
    # scoring options
81
82
    metric_list: str = None
    output_type: str = "greedy_until"
83
    generation_kwargs: dict = None
84
    repeats: int = 1
lintangsutawika's avatar
lintangsutawika committed
85
    filter_list: Union[str, list] = None
86
87
    should_decontaminate: bool = False
    doc_to_decontamination_query: str = None
88

lintangsutawika's avatar
lintangsutawika committed
89
    metadata: str = None  # by default, not used in the code. allows for users to pass arbitrary info to tasks
90

91
92
93
94
    def __post_init__(self):
        # allow user-specified aliases so that users can
        # force prompt-compatibility for some prompt regardless of
        # field names in prompt
95
        if self.template_aliases:
96
97
            if type(self.doc_to_text) == str:
                self.doc_to_text = self.template_aliases + self.doc_to_text
98

99
100
            if type(self.doc_to_target) == str:
                self.doc_to_target = self.template_aliases + self.doc_to_target
101

102
            if type(self.gold_alias) == str:
lintangsutawika's avatar
lintangsutawika committed
103
                self.gold_alias = self.template_aliases + self.gold_alias
104

Lintang Sutawika's avatar
Lintang Sutawika committed
105
106
107
108
109
110
111
112
113
114
115
116
        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:
117
                self.generation_kwargs["until"] = [self.fewshot_delimiter]
Lintang Sutawika's avatar
Lintang Sutawika committed
118
119
120
121
        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
122
                    "until": None
123
124
                    if self.fewshot_delimiter is None
                    else [self.fewshot_delimiter],
Lintang Sutawika's avatar
Lintang Sutawika committed
125
126
127
                    "do_sample": False,
                    "temperature": 0.0,
                }
128

haileyschoelkopf's avatar
haileyschoelkopf committed
129
130
        # TODO: how to make TaskConfigs be de- and re-serializable, even when using the !function constructor?

131
132
133
    def __getitem__(self, item):
        return getattr(self, item)

134
    def to_dict(self):
135
136
        """dumps the current config as a dictionary object, as a printable format.
        null fields will not be printed.
haileyschoelkopf's avatar
haileyschoelkopf committed
137
        Used for dumping results alongside full task configuration
138

haileyschoelkopf's avatar
haileyschoelkopf committed
139
140
141
142
143
144
145
146
147
148
        :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
149
150
151
            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
152
        return cfg_dict
153

154
155
156
157
158
159
160
161
162
163
164
165

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
166

167
168
169
170
171
172
173
174
    # 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
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
209
    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
210
        self._config = TaskConfig(**config) if config else TaskConfig()
211
212
213

        if not hasattr(self, "_filters"):
            self._filters = []
lintangsutawika's avatar
lintangsutawika committed
214
            for name, components in self._config.get(
215
                "filters", [["none", [["take_first", None]]]]
lintangsutawika's avatar
lintangsutawika committed
216
            ):
217
218
219
                filter_pipeline = build_filter_ensemble(name, components)
                self._filters.append(filter_pipeline)

lintangsutawika's avatar
lintangsutawika committed
220
        self.sampler = samplers.Sampler(
221
222
            list(self.fewshot_docs()), self, rnd=random.Random(1234)
        )
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
248

    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.
        """
249
250
251
252
253
254
255
        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,
        )
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
292

    @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 []

293
294
295
296
297
298
299
300
301
302
    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
303
            eval_logger.warning(
304
                "has_training_docs and has_validation_docs are False"
305
                ", using test_docs as fewshot_docs but this is not recommended."
lintangsutawika's avatar
lintangsutawika committed
306
            )
307
308
            return self.test_docs()

309
310
311
312
313
314
315
316
317
318
    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
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
346
    @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

347
    def build_all_requests(self, limit=None, rank=None, world_size=None):
348
349
350
351
352
353
354
355
356
357
        """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!"

358
359
360
361
        eval_logger.info(
            f"Building contexts for task '{self._config.task}' on rank {rank}..."
        )

362
        instances = []
363
364
        for doc_id, doc in utils.create_iterator(
            enumerate(docs), rank, world_size, limit
lintangsutawika's avatar
lintangsutawika committed
365
        ):
366
            # sample fewshot context #TODO: need to offset doc_id by rank now!
367
            fewshot_ctx = self.fewshot_context(
368
369
                doc,
                self._config.num_fewshot,
370
            )
371

haileyschoelkopf's avatar
haileyschoelkopf committed
372
            # TODO: we should override self._config.repeats if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
373
374
375
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
376
                metadata=(self._config["task"], doc_id, self._config.repeats),
lintangsutawika's avatar
lintangsutawika committed
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
402

            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
403
            The number of times each instance in a dataset is inferred on. Defaults to 1,
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
438
            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
439
440
441
442
443
444
445
446
447
448
    @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))

449
    @utils.positional_deprecated
450
    def fewshot_context(self, doc, num_fewshot):
451
452
453
454
455
456
457
458
459
460
461
462
        """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:
463
464
            # always prepend the (possibly empty) task description
            labeled_examples = self._config.description
465
        else:
lintangsutawika's avatar
lintangsutawika committed
466
467
468
            labeled_examples = self._config.description + self.sampler.get_context(
                doc, num_fewshot
            )
469
470

        example = self.doc_to_text(doc)
471
472
473
474
        if type(example) == str:
            return labeled_examples + example
        elif type(example) == list:
            return [labeled_examples + ex for ex in example]
475
476
477
        elif type(example) == int:
            choices = self.doc_to_choice(doc)
            return labeled_examples + choices[example]
478
479
480

    def apply_filters(self):

lintangsutawika's avatar
lintangsutawika committed
481
482
483
484
485
486
        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
487

488
    def dump_config(self):
489
        """Returns a dictionary representing the task's config.
490
491
492
493
494

        :returns: str
            The fewshot context.
        """
        # TODO: this should only return the overrides applied to a non-YAML task's configuration.
495
        # (num_fewshot)
496
497
        return self._config.to_dict()

498
499
500

class ConfigurableTask(Task):

501
    VERSION = "Yaml"
502
    OUTPUT_TYPE = None
503
    CONFIG = None
504
505
506
507

    def __init__(
        self, data_dir=None, cache_dir=None, download_mode=None, config: dict = None
    ):
508
        # Get pre-configured attributes
509
        self._config = self.CONFIG
510

511
512
        # Use new configurations if there was no preconfiguration
        if self._config is None:
513
            self._config = TaskConfig(**config)
514
515
        # Overwrite configs
        else:
lintangsutawika's avatar
lintangsutawika committed
516
            if config is not None:
517
                self._config.__dict__.update(config)
518

519
        if self._config is None:
lintangsutawika's avatar
lintangsutawika committed
520
521
522
            raise ValueError(
                "Must pass a config to ConfigurableTask, either in cls.CONFIG or `config` kwarg"
            )
523
524

        if self._config.output_type is not None:
525
            assert self._config.output_type in ALL_OUTPUT_TYPES
526
527
            self.OUTPUT_TYPE = self._config.output_type

528
529
530
531
532
533
        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

534
535
536
537
        self._metric_fn_list = {}
        self._metric_fn_kwargs = {}
        self._aggregation_list = {}
        self._higher_is_better = {}
538

539
        _metric_list = DEFAULT_METRIC_REGISTRY[self._config.output_type]
540
        if self._config.metric_list is None:
541
            # TODO: handle this in TaskConfig.__post_init__ ?
542
            for metric_name in _metric_list:
haileyschoelkopf's avatar
haileyschoelkopf committed
543
544
                self._metric_fn_list[metric_name] = get_metric(metric_name)
                self._aggregation_list[metric_name] = get_default_aggregation(
545
                    metric_name
haileyschoelkopf's avatar
haileyschoelkopf committed
546
547
                )
                self._higher_is_better[metric_name] = is_higher_better(metric_name)
548
549
550
551
552
553
554
555
556
        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"]
                }
557
558
559
560
561
562
563

                if self._config.process_results is None:
                    self._metric_fn_list[metric_name] = get_metric(metric_name)
                    self._metric_fn_kwargs[metric_name] = kwargs
                else:
                    self._metric_fn_list[metric_name] = None
                    self._metric_fn_kwargs[metric_name] = {}
lintangsutawika's avatar
lintangsutawika committed
564

565
                if "aggregation" in metric_config:
566
                    agg_name = metric_config["aggregation"]
567
                    if type(agg_name) == str:
haileyschoelkopf's avatar
haileyschoelkopf committed
568
                        self._aggregation_list[metric_name] = get_aggregation(agg_name)
569
570
571
572
                    elif callable(agg_name):
                        self._aggregation_list[metric_name] = metric_config[
                            "aggregation"
                        ]
573
                else:
574
575

                    INV_AGG_REGISTRY = {v: k for k, v in AGGREGATION_REGISTRY.items()}
haileyschoelkopf's avatar
haileyschoelkopf committed
576
                    metric_agg = get_default_aggregation(metric_name)
577
                    eval_logger.warning(
578
579
580
                        f"metric {metric_name} is defined, but aggregation is not. "
                        f"using default "
                        f"aggregation={INV_AGG_REGISTRY[metric_agg]}"
581
                    )
582
                    self._aggregation_list[metric_name] = metric_agg
lintangsutawika's avatar
lintangsutawika committed
583

584
585
586
587
588
589
                if "higher_is_better" in metric_config:
                    self._higher_is_better[metric_name] = metric_config[
                        "higher_is_better"
                    ]
                else:
                    eval_logger.warning(
590
591
                        f"metric {metric_name} is defined, but higher_is_better is not. "
                        f"using default "
haileyschoelkopf's avatar
haileyschoelkopf committed
592
                        f"higher_is_better={is_higher_better(metric_name)}"
593
                    )
haileyschoelkopf's avatar
haileyschoelkopf committed
594
                    self._higher_is_better[metric_name] = is_higher_better(metric_name)
595

596
        self.download(self._config.dataset_kwargs)
597
598
599
        self._training_docs = None
        self._fewshot_docs = None

lintangsutawika's avatar
lintangsutawika committed
600
        if self._config.filter_list is not None:
lintangsutawika's avatar
lintangsutawika committed
601
            self._filters = []
lintangsutawika's avatar
lintangsutawika committed
602
603
604
605
606
607
608
609
            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
610
611
612
                        }
                        components.append([function["function"], kwargs])
                    filter_pipeline = build_filter_ensemble(filter_name, components)
lintangsutawika's avatar
lintangsutawika committed
613
                self._filters.append(filter_pipeline)
lintangsutawika's avatar
lintangsutawika committed
614
        else:
615
            self._filters = [build_filter_ensemble("none", [["take_first", None]])]
616
617

        if self._config.use_prompt is not None:
lintangsutawika's avatar
lintangsutawika committed
618
            eval_logger.info(f"loading prompt {self._config.use_prompt}")
619
            self.prompt = get_prompt(
lintangsutawika's avatar
lintangsutawika committed
620
621
                self._config.use_prompt, self.DATASET_PATH, self.DATASET_NAME
            )
622
623
624
        else:
            self.prompt = None

lintangsutawika's avatar
lintangsutawika committed
625
626
        if self.fewshot_docs() is not None:
            self.sampler = samplers.Sampler(
627
                list(self.fewshot_docs()), self, rnd=random.Random(1234)
628
            )
629

630
631
632
633
        if self._config.template_aliases is not None:
            for key, alias in self._config.template_aliases:
                self.dataset.rename_column(key, alias)

634
635
636
637
638
639
640
641
642
        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!"

643
        # Test One Doc
644
645
646
        self.features = list(docs.features.keys())
        self.multiple_input = 0
        self.multiple_target = 0
647
648
        test_doc = docs[0]
        test_text = self.doc_to_text(test_doc)
649
        test_target = self.doc_to_target(test_doc)
lintangsutawika's avatar
lintangsutawika committed
650
651
652
653
654

        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")
655
656
            else:
                num_choice = len(test_choice)
657

658
659
            if type(test_text) is int:
                self.multiple_input = num_choice
660

661
        if type(test_target) is list:
662
663
            self.multiple_target = len(test_target)

664
665
666
667
668
669
670
671
    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 {},
        )

672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
    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]

702
    def fewshot_docs(self):
703
        if self._config.fewshot_split is not None:
704
            return self.dataset[self._config.fewshot_split]
705
706
707
        else:
            if self._config.num_fewshot > 0:
                eval_logger.warning(
haileyschoelkopf's avatar
haileyschoelkopf committed
708
                    f"Task '{self._config.task}': "
709
710
711
712
                    "num_fewshot > 0 but fewshot_split is None. "
                    "using preconfigured rule."
                )
            return super().fewshot_docs()
713

714
715
716
717
718
    def should_decontaminate(self):
        return self._config.should_decontaminate

    def doc_to_decontamination_query(self, doc):
        if self._config.should_decontaminate:
719
720
721
722
723
724
            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)
                )
725

726
727
728
729
730
731
732
733
734
735
736
737
    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):
738
739
740

        if self.prompt is not None:
            doc_to_text = self.prompt
741
742
        else:
            doc_to_text = self._config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
743

744
745
746
        if type(doc_to_text) == int:
            return doc_to_text
        elif type(doc_to_text) == str:
747
            if doc_to_text in self.features:
748
749
750
                # if self._config.doc_to_choice is not None:
                #     return self.doc_to_choice(doc)[doc[doc_to_text]]
                # else:
751
752
                return doc[doc_to_text]
            else:
lintangsutawika's avatar
lintangsutawika committed
753
754
755
756
757
                text_string = utils.apply_template(doc_to_text, doc)
                if text_string.isdigit():
                    return ast.literal_eval(text_string)
                else:
                    return text_string
758
        elif callable(doc_to_text):
759
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
760
        # Used when applying a Promptsource template
761
        elif hasattr(doc_to_text, "apply"):
762
            return doc_to_text.apply(doc)[0]
763
        else:
764
            print(type(doc_to_text))
765
            raise TypeError
766
767

    def doc_to_target(self, doc):
768
769
770

        if self.prompt is not None:
            doc_to_target = self.prompt
771
772
773
        else:
            doc_to_target = self._config.doc_to_target

774
775
776
        if type(doc_to_target) == int:
            return doc_to_target
        elif type(doc_to_target) == str:
777
            if doc_to_target in self.features:
778
779
780
781
                # 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]
782
            else:
lintangsutawika's avatar
lintangsutawika committed
783
784
785
786
787
                target_string = utils.apply_template(doc_to_target, doc)
                if target_string.isdigit():
                    return ast.literal_eval(target_string)
                else:
                    return target_string
788
        elif callable(doc_to_target):
789
            return doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
790
        # Used when applying a Promptsource template
791
792
        elif hasattr(doc_to_target, "apply"):
            return doc_to_target.apply(doc)[1]
793
794
        else:
            raise TypeError
795
796
797
798
799

    def doc_to_choice(self, doc):

        if self.prompt is not None:
            doc_to_choice = self.prompt
lintangsutawika's avatar
lintangsutawika committed
800
        elif self._config.doc_to_choice is None:
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
            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
817

818
    def gold_alias(self, doc):
819
820
821
822
823
        # 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
824
        if self._config.gold_alias is not None:
825
826
            doc_to_target = self._config.gold_alias
        else:
lintangsutawika's avatar
lintangsutawika committed
827
            return self.doc_to_target(doc)
828
829
830
831
832
833
834
835
836
837

        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

838
839
    def construct_requests(self, doc, ctx, **kwargs):

840
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
841
            arguments = (ctx, self.doc_to_target(doc))
842
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
843
            arguments = (self.doc_to_target(doc),)
844
        elif self.OUTPUT_TYPE == "multiple_choice":
845
846
847
848

            choices = self.doc_to_choice(doc)
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
849
                cont = self.doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
850
                arguments = [(ctx, " {}".format(cont)) for ctx in choices]
851
            else:
852
                # Otherwise they are placed in the continuation
lintangsutawika's avatar
lintangsutawika committed
853
                arguments = [(ctx, " {}".format(cont)) for cont in choices]
854

855
            request_list = [
856
857
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
858
                    doc=doc,
859
                    arguments=arg,
860
                    idx=i,
861
862
                    **kwargs,
                )
863
                for i, arg in enumerate(arguments)
864
            ]
865
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
866
            if "acc_mutual_info" in self._metric_fn_list.keys():
867
868
869
                # if we are calculating multiple choice accuracy
                # using mutual information instead of raw loglikelihood as metric, need unconditional lls.

lintangsutawika's avatar
lintangsutawika committed
870
                # here mutual info refers to calculating
871
872
873
874
875
876
                # 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
877
                            doc=doc,
878
                            arguments=("", "{}".format(choice)),
879
880
881
                            idx=i,
                            **kwargs,
                        )
lintangsutawika's avatar
lintangsutawika committed
882
                        for i, choice in enumerate(choices)
883
884
885
                    ]
                )
            return request_list
lintangsutawika's avatar
lintangsutawika committed
886

887
        elif self.OUTPUT_TYPE == "greedy_until":
888
            arguments = (ctx, self._config.generation_kwargs)
lintangsutawika's avatar
lintangsutawika committed
889
890

        return Instance(
lintangsutawika's avatar
lintangsutawika committed
891
892
            request_type=self.OUTPUT_TYPE, doc=doc, arguments=arguments, idx=0, **kwargs
        )
893
894
895

    def process_results(self, doc, results):

lintangsutawika's avatar
lintangsutawika committed
896
897
        if callable(self._config.process_results):
            return self._config.process_results(doc, results)
lintangsutawika's avatar
lintangsutawika committed
898

899
        result_dict = {}
900
        use_metric = list(self._metric_fn_list.keys())
901
902
903
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
904
905
906
907
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
908
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
909
            (loglikelihood,) = results
910
911
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
haileyschoelkopf's avatar
haileyschoelkopf committed
912
            return {
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
                **(
                    {"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
928
            }
929
        elif self.OUTPUT_TYPE == "multiple_choice":
930
931

            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
932

933
            # retrieve choices in List[str] form, to compute choice lengths, etc.
934
            choices = self.doc_to_choice(doc)
935
936
            completion_len = np.array([float(len(i)) for i in choices])

937
938
            if (
                2 * len(choices) == len(lls)
939
                and "acc_mutual_info" in self._metric_fn_list.keys()
940
941
942
943
944
945
946
            ):
                # 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]
947

948
949
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
950

951
952
            if self.multiple_input:
                gold = self.doc_to_text(doc)
953
            else:
954
                gold = self.doc_to_target(doc)
955
956
                if type(gold) is str:
                    gold = choices.index(gold)
lintangsutawika's avatar
lintangsutawika committed
957

958
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
959
960
                acc = 1.0 if pred in gold else 0.0
                acc_norm = 1.0 if pred_norm in gold else 0.0
961
                exact_match = int(any([is_greedy[i] for i in gold]))
lintangsutawika's avatar
lintangsutawika committed
962
963
964
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
965
966
                # 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])
967
968

            result_dict = {
969
                **({"acc": acc} if "acc" in use_metric else {}),
970
971
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
972
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
973
                **({"exact_match": exact_match} if "exact_match" in use_metric else {}),
974
975
            }

976
            if "acc_mutual_info" in use_metric:
lintangsutawika's avatar
lintangsutawika committed
977
978
979
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
980
981
982
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

983
984
        elif self.OUTPUT_TYPE == "greedy_until":

985
            gold = self.doc_to_target(doc)
986
987
988
            if type(gold) == int:
                choices = self.doc_to_choice(doc)
                gold = choices[gold]
989

990
            for key, result in zip(self._metric_fn_list.keys(), results):
haileyschoelkopf's avatar
haileyschoelkopf committed
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
                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],
                    )
1016

haileyschoelkopf's avatar
haileyschoelkopf committed
1017
1018
1019
1020
                if isinstance(result, dict):
                    result_dict.update(result)
                else:
                    result_dict[key] = result
1021
        else:
lintangsutawika's avatar
lintangsutawika committed
1022
1023
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
1024
                "'loglikelihood', 'loglikelihood_rolling', 'greedy_until' or 'multiple_choice'",
1025
            )
1026
1027
1028
1029
1030
1031
1032

        return result_dict

    def aggregation(self):
        return self._aggregation_list

    def higher_is_better(self):
haileyschoelkopf's avatar
haileyschoelkopf committed
1033
        return self._higher_is_better
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043


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):
1044
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1045
1046
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1047
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1048
                doc=doc,
1049
                arguments=(ctx, " {}".format(choice)),
1050
                idx=i,
1051
1052
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1053
1054
            for i, choice in enumerate(doc["choices"])
        ]
1055
1056

    def process_results(self, doc, results):
lintangsutawika's avatar
lintangsutawika committed
1057
1058
1059
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
        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
1084
class PerplexityTask(Task):
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094

    OUTPUT_TYPE = "loglikelihood_rolling"

    def has_training_docs(self):
        return False

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

1095
    def fewshot_context(self, doc, num_fewshot):
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
        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
1121
1122
1123
1124
1125
1126
1127
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1128
1129
1130

    def process_results(self, doc, results):
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1131
1132
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
        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))