task.py 39.9 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
12
13
14

import datasets
import numpy as np

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

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

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

41
42
43
44
45
46
47
ALL_OUTPUT_TYPES = [
    "loglikelihood",
    "multiple_choice",
    "loglikelihood_rolling",
    "greedy_until",
]

48
49
50

@dataclass
class TaskConfig(dict):
51
    # task naming/registry
52
    task: str = None
53
    group: Union[str, list] = None
54
55
56
    # HF dataset options.
    # which dataset to use,
    # and what splits for what purpose
57
58
    dataset_path: str = None
    dataset_name: str = None
59
    dataset_kwargs: dict = None
60
61
62
    training_split: str = None
    validation_split: str = None
    test_split: str = None
lintangsutawika's avatar
lintangsutawika committed
63
    fewshot_split: str = None  # TODO: assert that this not None if num_fewshot > 0. (?) assert if this is same split as one evaling (?)
64
65
    # formatting / prompting options.
    # see docs/advanced_task_guide.md for more info
66
    template_aliases: Union[str, list] = None
67
68
    doc_to_text: Union[Callable, str] = None
    doc_to_target: Union[Callable, str] = None
lintangsutawika's avatar
lintangsutawika committed
69
    doc_to_choice: Union[Callable, str, dict, list] = None
70
    gold_alias: Union[Callable, str] = None
71
    use_prompt: str = None
72
    description: str = ""
73
74
    target_delimiter: str = " "
    fewshot_delimiter: str = "\n\n"
75
    # runtime configuration options
76
    num_fewshot: int = 0
77
    # scoring options
78
79
    metric_list: str = None
    output_type: str = "greedy_until"
80
    generation_kwargs: dict = None
81
    repeats: int = 1
lintangsutawika's avatar
lintangsutawika committed
82
    filter_list: Union[str, list] = None
83
84
    should_decontaminate: bool = False
    doc_to_decontamination_query: str = None
85

lintangsutawika's avatar
lintangsutawika committed
86
    metadata: str = None  # by default, not used in the code. allows for users to pass arbitrary info to tasks
87

88
89
90
91
    def __post_init__(self):
        # allow user-specified aliases so that users can
        # force prompt-compatibility for some prompt regardless of
        # field names in prompt
92
        if self.template_aliases:
93
94
            if type(self.doc_to_text) == str:
                self.doc_to_text = self.template_aliases + self.doc_to_text
95

96
97
            if type(self.doc_to_target) == str:
                self.doc_to_target = self.template_aliases + self.doc_to_target
98

99
            if type(self.gold_alias) == str:
lintangsutawika's avatar
lintangsutawika committed
100
                self.gold_alias = self.template_aliases + self.gold_alias
101

Lintang Sutawika's avatar
Lintang Sutawika committed
102
103
104
105
106
107
108
109
110
111
112
113
        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:
114
                self.generation_kwargs["until"] = [self.fewshot_delimiter]
Lintang Sutawika's avatar
Lintang Sutawika committed
115
116
117
118
        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
119
                    "until": None
120
121
                    if self.fewshot_delimiter is None
                    else [self.fewshot_delimiter],
Lintang Sutawika's avatar
Lintang Sutawika committed
122
123
124
                    "do_sample": False,
                    "temperature": 0.0,
                }
125

haileyschoelkopf's avatar
haileyschoelkopf committed
126
127
        # TODO: how to make TaskConfigs be de- and re-serializable, even when using the !function constructor?

128
129
130
    def __getitem__(self, item):
        return getattr(self, item)

131
    def to_dict(self):
132
133
        """dumps the current config as a dictionary object, as a printable format.
        null fields will not be printed.
haileyschoelkopf's avatar
haileyschoelkopf committed
134
        Used for dumping results alongside full task configuration
135

haileyschoelkopf's avatar
haileyschoelkopf committed
136
137
138
139
140
141
142
143
144
145
        :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
146
147
148
            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
149
        return cfg_dict
150

151
152
153
154
155
156
157
158
159
160
161
162

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
163

164
165
166
167
168
169
170
171
    # 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
172

173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
    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
207
        self._config = TaskConfig(**config) if config else TaskConfig()
208
209
210

        if not hasattr(self, "_filters"):
            self._filters = []
lintangsutawika's avatar
lintangsutawika committed
211
            for name, components in self._config.get(
212
                "filters", [["none", [["take_first", None]]]]
lintangsutawika's avatar
lintangsutawika committed
213
            ):
214
215
216
                filter_pipeline = build_filter_ensemble(name, components)
                self._filters.append(filter_pipeline)

lintangsutawika's avatar
lintangsutawika committed
217
218
219
        self.sampler = samplers.Sampler(
            list(self.fewshot_docs()), self, rnd=random.Random()
        )  # TODO: pass the correct docs in here
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245

    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.
        """
246
247
248
249
250
251
252
        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,
        )
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289

    @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 []

290
291
292
293
294
295
296
297
298
299
    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
300
            eval_logger.warning(
301
                "has_training_docs and has_validation_docs are False"
302
                ", using test_docs as fewshot_docs but this is not recommended."
lintangsutawika's avatar
lintangsutawika committed
303
            )
304
305
            return self.test_docs()

306
307
308
309
310
311
312
313
314
315
    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
316

317
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
    @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

344
    def build_all_requests(self, limit=None, rank=None, world_size=None):
345
346
347
348
349
350
351
352
353
354
355
        """Build a set of Instances for a task, and store them in task.instances"""
        if self.has_test_docs():
            docs = self.test_docs()
        elif self.has_validation_docs():
            docs = self.validation_docs()
        else:
            assert (
                False
            ), f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"

        instances = []
356
357
        for doc_id, doc in utils.create_iterator(
            enumerate(docs), rank, world_size, limit
lintangsutawika's avatar
lintangsutawika committed
358
        ):
359
            # sample fewshot context #TODO: need to offset doc_id by rank now!
360
361
362
            fewshot_ctx = self.fewshot_context(
                doc, self._config.num_fewshot, rnd=random.Random()
            )
363

haileyschoelkopf's avatar
haileyschoelkopf committed
364
            # TODO: we should override self._config.repeats if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
365
366
367
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
368
                metadata=(self._config["task"], doc_id, self._config.repeats),
lintangsutawika's avatar
lintangsutawika committed
369
            )
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394

            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
395
            The number of times each instance in a dataset is inferred on. Defaults to 1,
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
            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
431
432
433
434
435
436
437
438
439
440
    @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))

441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
    @utils.positional_deprecated
    def fewshot_context(self, doc, num_fewshot, rnd=None):
        """Returns a fewshot context string that is made up of a prepended description
        (if provided), the `num_fewshot` number of examples, and an appended prompt example.

        :param doc: str
            The document as returned from training_docs, validation_docs, or test_docs.
        :param num_fewshot: int
            The number of fewshot examples to provide in the returned context string.
        :param rnd: random.Random
            The pseudo-random number generator used to randomly sample examples.
            WARNING: This is currently a required arg although it's optionalized with a default `None`.
        :returns: str
            The fewshot context.
        """
        assert (
            rnd is not None
        ), "A `random.Random` generator argument must be provided to `rnd`"

        if num_fewshot == 0:
461
462
            # always prepend the (possibly empty) task description
            labeled_examples = self._config.description
463
        else:
lintangsutawika's avatar
lintangsutawika committed
464
465
466
            labeled_examples = self._config.description + self.sampler.get_context(
                doc, num_fewshot
            )
467
468

        example = self.doc_to_text(doc)
469
470
471
472
        if type(example) == str:
            return labeled_examples + example
        elif type(example) == list:
            return [labeled_examples + ex for ex in example]
473
474
475

    def apply_filters(self):

lintangsutawika's avatar
lintangsutawika committed
476
477
478
479
480
481
        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
482

483
    def dump_config(self):
484
        """Returns a dictionary representing the task's config.
485
486
487
488
489

        :returns: str
            The fewshot context.
        """
        # TODO: this should only return the overrides applied to a non-YAML task's configuration.
490
        # (num_fewshot)
491
492
        return self._config.to_dict()

493
494
495

class ConfigurableTask(Task):

496
    VERSION = "Yaml"
497
    OUTPUT_TYPE = None
498
    CONFIG = None
499
500
501
502

    def __init__(
        self, data_dir=None, cache_dir=None, download_mode=None, config: dict = None
    ):
503
        # Get pre-configured attributes
504
        self._config = self.CONFIG
505

506
507
        # Use new configurations if there was no preconfiguration
        if self._config is None:
508
            self._config = TaskConfig(**config)
509
510
        # Overwrite configs
        else:
lintangsutawika's avatar
lintangsutawika committed
511
            if config is not None:
512
                self._config.__dict__.update(config)
513

514
        if self._config is None:
lintangsutawika's avatar
lintangsutawika committed
515
516
517
            raise ValueError(
                "Must pass a config to ConfigurableTask, either in cls.CONFIG or `config` kwarg"
            )
518
519

        if self._config.output_type is not None:
520
            assert self._config.output_type in ALL_OUTPUT_TYPES
521
522
            self.OUTPUT_TYPE = self._config.output_type

523
524
525
526
527
528
        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

529
530
531
532
        self._metric_fn_list = {}
        self._metric_fn_kwargs = {}
        self._aggregation_list = {}
        self._higher_is_better = {}
533

534
        _metric_list = DEFAULT_METRIC_REGISTRY[self._config.output_type]
535
        if self._config.metric_list is None:
536
            # TODO: handle this in TaskConfig.__post_init__ ?
537
            for metric_name in _metric_list:
haileyschoelkopf's avatar
haileyschoelkopf committed
538
539
                self._metric_fn_list[metric_name] = get_metric(metric_name)
                self._aggregation_list[metric_name] = get_default_aggregation(
540
                    metric_name
haileyschoelkopf's avatar
haileyschoelkopf committed
541
542
                )
                self._higher_is_better[metric_name] = is_higher_better(metric_name)
543
544
545
546
547
548
549
550
551
        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
552
553
                self._metric_fn_list[metric_name] = get_metric(metric_name)
                self._metric_fn_kwargs[metric_name] = kwargs
lintangsutawika's avatar
lintangsutawika committed
554

555
                if "aggregation" in metric_config:
556
                    agg_name = metric_config["aggregation"]
557
                    if type(agg_name) == str:
haileyschoelkopf's avatar
haileyschoelkopf committed
558
                        self._aggregation_list[metric_name] = get_aggregation(agg_name)
559
560
561
562
                    elif callable(agg_name):
                        self._aggregation_list[metric_name] = metric_config[
                            "aggregation"
                        ]
563
                else:
564
565

                    INV_AGG_REGISTRY = {v: k for k, v in AGGREGATION_REGISTRY.items()}
haileyschoelkopf's avatar
haileyschoelkopf committed
566
                    metric_agg = get_default_aggregation(metric_name)
567
                    eval_logger.warning(
568
569
570
                        f"metric {metric_name} is defined, but aggregation is not. "
                        f"using default "
                        f"aggregation={INV_AGG_REGISTRY[metric_agg]}"
571
                    )
572
                    self._aggregation_list[metric_name] = metric_agg
lintangsutawika's avatar
lintangsutawika committed
573

574
575
576
577
578
579
                if "higher_is_better" in metric_config:
                    self._higher_is_better[metric_name] = metric_config[
                        "higher_is_better"
                    ]
                else:
                    eval_logger.warning(
580
581
                        f"metric {metric_name} is defined, but higher_is_better is not. "
                        f"using default "
haileyschoelkopf's avatar
haileyschoelkopf committed
582
                        f"higher_is_better={is_higher_better(metric_name)}"
583
                    )
haileyschoelkopf's avatar
haileyschoelkopf committed
584
                    self._higher_is_better[metric_name] = is_higher_better(metric_name)
585

586
        self.download(self._config.dataset_kwargs)
587
588
589
        self._training_docs = None
        self._fewshot_docs = None

lintangsutawika's avatar
lintangsutawika committed
590
        if self._config.filter_list is not None:
lintangsutawika's avatar
lintangsutawika committed
591
            self._filters = []
lintangsutawika's avatar
lintangsutawika committed
592
593
594
595
596
597
598
599
            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
600
601
602
                        }
                        components.append([function["function"], kwargs])
                    filter_pipeline = build_filter_ensemble(filter_name, components)
lintangsutawika's avatar
lintangsutawika committed
603
                self._filters.append(filter_pipeline)
lintangsutawika's avatar
lintangsutawika committed
604
        else:
605
            self._filters = [build_filter_ensemble("none", [["take_first", None]])]
606
607

        if self._config.use_prompt is not None:
lintangsutawika's avatar
lintangsutawika committed
608
            eval_logger.info(f"loading prompt {self._config.use_prompt}")
609
            self.prompt = get_prompt(
lintangsutawika's avatar
lintangsutawika committed
610
611
                self._config.use_prompt, self.DATASET_PATH, self.DATASET_NAME
            )
612
613
614
        else:
            self.prompt = None

lintangsutawika's avatar
lintangsutawika committed
615
616
617
        if self.fewshot_docs() is not None:
            self.sampler = samplers.Sampler(
                list(self.fewshot_docs()), self, rnd=random.Random()
618
            )
619

620
621
622
623
        if self._config.template_aliases is not None:
            for key, alias in self._config.template_aliases:
                self.dataset.rename_column(key, alias)

624
625
626
627
628
629
630
631
632
        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!"

633
        # Test One Doc
634
635
636
        self.features = list(docs.features.keys())
        self.multiple_input = 0
        self.multiple_target = 0
637
638
        test_doc = docs[0]
        test_text = self.doc_to_text(test_doc)
639
        test_target = self.doc_to_target(test_doc)
lintangsutawika's avatar
lintangsutawika committed
640
641
642
643
644

        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")
645
646
            else:
                num_choice = len(test_choice)
647

648
649
            if type(test_text) is int:
                self.multiple_input = num_choice
650

651
        if type(test_target) is list:
652
653
654
655
            self.multiple_target = len(test_target)

        eval_logger.info(f" Input choices: {self.multiple_input}")
        eval_logger.info(f"Output choices: {self.multiple_target}")
656

657
658
659
660
661
662
663
664
    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 {},
        )

665
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
    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]

695
    def fewshot_docs(self):
696
        if self._config.fewshot_split is not None:
697
            return self.dataset[self._config.fewshot_split]
698
699
700
        else:
            if self._config.num_fewshot > 0:
                eval_logger.warning(
haileyschoelkopf's avatar
haileyschoelkopf committed
701
                    f"Task '{self._config.task}': "
702
703
704
705
                    "num_fewshot > 0 but fewshot_split is None. "
                    "using preconfigured rule."
                )
            return super().fewshot_docs()
706

707
708
709
710
711
    def should_decontaminate(self):
        return self._config.should_decontaminate

    def doc_to_decontamination_query(self, doc):
        if self._config.should_decontaminate:
712
713
714
715
716
717
            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)
                )
718

719
720
721
722
723
724
725
726
727
728
729
730
    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):
731
732
733

        if self.prompt is not None:
            doc_to_text = self.prompt
734
735
        else:
            doc_to_text = self._config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
736

737
738
739
        if type(doc_to_text) == int:
            return doc_to_text
        elif type(doc_to_text) == str:
740
            if doc_to_text in self.features:
741
742
743
                # if self._config.doc_to_choice is not None:
                #     return self.doc_to_choice(doc)[doc[doc_to_text]]
                # else:
744
745
                return doc[doc_to_text]
            else:
lintangsutawika's avatar
lintangsutawika committed
746
747
748
749
750
                text_string = utils.apply_template(doc_to_text, doc)
                if text_string.isdigit():
                    return ast.literal_eval(text_string)
                else:
                    return text_string
751
        elif callable(doc_to_text):
752
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
753
        # Used when applying a Promptsource template
754
        elif hasattr(doc_to_text, "apply"):
755
            return doc_to_text.apply(doc)[0]
756
        else:
757
            print(type(doc_to_text))
758
            raise TypeError
759
760

    def doc_to_target(self, doc):
761
762
763

        if self.prompt is not None:
            doc_to_target = self.prompt
764
765
766
        else:
            doc_to_target = self._config.doc_to_target

767
768
769
        if type(doc_to_target) == int:
            return doc_to_target
        elif type(doc_to_target) == str:
770
            if doc_to_target in self.features:
771
772
773
774
                # 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]
775
            else:
lintangsutawika's avatar
lintangsutawika committed
776
777
778
779
780
                target_string = utils.apply_template(doc_to_target, doc)
                if target_string.isdigit():
                    return ast.literal_eval(target_string)
                else:
                    return target_string
781
        elif callable(doc_to_target):
782
            return doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
783
        # Used when applying a Promptsource template
784
785
        elif hasattr(doc_to_target, "apply"):
            return doc_to_target.apply(doc)[1]
786
787
        else:
            raise TypeError
788
789
790
791
792

    def doc_to_choice(self, doc):

        if self.prompt is not None:
            doc_to_choice = self.prompt
lintangsutawika's avatar
lintangsutawika committed
793
        elif self._config.doc_to_choice is None:
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
            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
810

811
    def gold_alias(self, doc):
812
813
814
815
816
        # 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
817
        if self._config.gold_alias is not None:
818
819
            doc_to_target = self._config.gold_alias
        else:
lintangsutawika's avatar
lintangsutawika committed
820
            return self.doc_to_target(doc)
821
822
823
824
825
826
827
828
829
830

        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

831
832
    def construct_requests(self, doc, ctx, **kwargs):

833
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
834
            arguments = (ctx, self.doc_to_target(doc))
835
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
836
            arguments = (self.doc_to_target(doc),)
837
        elif self.OUTPUT_TYPE == "multiple_choice":
838
839
840
841

            choices = self.doc_to_choice(doc)
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
842
                cont = self.doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
843
                arguments = [(ctx, " {}".format(cont)) for ctx in choices]
844
            else:
845
                # Otherwise they are placed in the continuation
lintangsutawika's avatar
lintangsutawika committed
846
                arguments = [(ctx, " {}".format(cont)) for cont in choices]
847

848
            request_list = [
849
850
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
851
                    doc=doc,
852
                    arguments=arg,
853
                    idx=i,
854
855
                    **kwargs,
                )
856
                for i, arg in enumerate(arguments)
857
            ]
858
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
859
            if "acc_mutual_info" in self._metric_fn_list.keys():
860
861
862
                # if we are calculating multiple choice accuracy
                # using mutual information instead of raw loglikelihood as metric, need unconditional lls.

lintangsutawika's avatar
lintangsutawika committed
863
                # here mutual info refers to calculating
864
865
866
867
868
869
                # 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
870
                            doc=doc,
871
                            arguments=("", "{}".format(choice)),
872
873
874
                            idx=i,
                            **kwargs,
                        )
lintangsutawika's avatar
lintangsutawika committed
875
                        for i, choice in enumerate(choices)
876
877
878
                    ]
                )
            return request_list
lintangsutawika's avatar
lintangsutawika committed
879

880
        elif self.OUTPUT_TYPE == "greedy_until":
881
            arguments = (ctx, self._config.generation_kwargs)
lintangsutawika's avatar
lintangsutawika committed
882
883

        return Instance(
lintangsutawika's avatar
lintangsutawika committed
884
885
            request_type=self.OUTPUT_TYPE, doc=doc, arguments=arguments, idx=0, **kwargs
        )
886
887
888

    def process_results(self, doc, results):

lintangsutawika's avatar
lintangsutawika committed
889
890
891
        # if callable(self._config.process_results):
        #     return self._config.process_results(doc, results)

892
        result_dict = {}
893
        use_metric = list(self._metric_fn_list.keys())
894
895
896
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
897
898
899
900
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
901
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
902
            (loglikelihood,) = results
903
904
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
haileyschoelkopf's avatar
haileyschoelkopf committed
905
            return {
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
                **(
                    {"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
921
            }
922
        elif self.OUTPUT_TYPE == "multiple_choice":
923
924

            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
925

926
            # retrieve choices in List[str] form, to compute choice lengths, etc.
927
            choices = self.doc_to_choice(doc)
928
929
            completion_len = np.array([float(len(i)) for i in choices])

930
931
            if (
                2 * len(choices) == len(lls)
932
                and "acc_mutual_info" in self._metric_fn_list.keys()
933
934
935
936
937
938
939
            ):
                # 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]
940

941
942
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
943

944
945
            if self.multiple_input:
                gold = self.doc_to_text(doc)
946
            else:
947
                gold = self.doc_to_target(doc)
948
949
                if type(gold) is str:
                    gold = choices.index(gold)
lintangsutawika's avatar
lintangsutawika committed
950

951
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
952
953
954
955
956
                acc = 1.0 if pred in gold else 0.0
                acc_norm = 1.0 if pred_norm in gold else 0.0
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
957
958

            result_dict = {
959
                **({"acc": acc} if "acc" in use_metric else {}),
960
961
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
962
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
963
964
            }

965
            if "exact_match" in self._metric_fn_list.keys():
966
                # TODO: this gets score of 0 on arc_challenge for pythia-70m. need to test that this works properly
lintangsutawika's avatar
lintangsutawika committed
967
                is_greedy = is_greedy[gold]  # take value for the gold answer
968
969
                result_dict["exact_match"] = int(is_greedy)

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
                _dict = self._metric_fn_list[key](
983
                    references=gold if self.multiple_target else [gold],
haileyschoelkopf's avatar
haileyschoelkopf committed
984
985
                    predictions=[result],
                    **self._metric_fn_kwargs[key],
986
                )
987

lintangsutawika's avatar
lintangsutawika committed
988
                result_dict = {**result_dict, **_dict}
989
        else:
lintangsutawika's avatar
lintangsutawika committed
990
991
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
992
                "'loglikelihood', 'loglikelihood_rolling', 'greedy_until' or 'multiple_choice'",
993
            )
994
995
996
997
998
999
1000

        return result_dict

    def aggregation(self):
        return self._aggregation_list

    def higher_is_better(self):
haileyschoelkopf's avatar
haileyschoelkopf committed
1001
        return self._higher_is_better
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011


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):
1012
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1013
1014
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1015
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1016
                doc=doc,
1017
                arguments=(ctx, " {}".format(choice)),
1018
                idx=i,
1019
1020
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1021
1022
            for i, choice in enumerate(doc["choices"])
        ]
1023
1024

    def process_results(self, doc, results):
lintangsutawika's avatar
lintangsutawika committed
1025
1026
1027
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
        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
1052
class PerplexityTask(Task):
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062

    OUTPUT_TYPE = "loglikelihood_rolling"

    def has_training_docs(self):
        return False

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

lintangsutawika's avatar
lintangsutawika committed
1063
    def fewshot_context(self, doc, num_fewshot, rnd=None):
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
        assert (
            num_fewshot == 0
        ), "The number of fewshot examples must be 0 for perplexity tasks."
        assert (
            rnd is not None
        ), "A `random.Random` generator argument must be provided to `rnd`."

        return ""

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

    def doc_to_decontamination_query(self, doc):
        return doc

    def doc_to_text(self, doc):
        return ""

    def doc_to_target(self, doc):
        return doc

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

lintangsutawika's avatar
lintangsutawika committed
1092
1093
1094
1095
1096
1097
1098
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1099
1100
1101

    def process_results(self, doc, results):
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1102
1103
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
        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))