task.py 47.2 KB
Newer Older
1
import abc
2
from dataclasses import dataclass, field, asdict
3
4

import re
5
import ast
lintangsutawika's avatar
lintangsutawika committed
6
import yaml
7
8
9
import evaluate
import random
import itertools
10
import functools
11
from tqdm import tqdm
12
13
14
15

import datasets
import numpy as np

baberabb's avatar
baberabb committed
16
from typing import Union, List, Any, Tuple, Literal
17
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
    get_metric,
    get_aggregation,
36
    get_metric_aggregation,
haileyschoelkopf's avatar
haileyschoelkopf committed
37
    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
ALL_OUTPUT_TYPES = [
    "loglikelihood",
    "multiple_choice",
    "loglikelihood_rolling",
47
    "generate_until",
48
49
]

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
    process_docs: Callable = 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"
haileyschoelkopf's avatar
haileyschoelkopf committed
78
    fewshot_config: dict = None
79
    # runtime configuration options
80
    num_fewshot: int = 0
81
    # scoring options
82
    metric_list: list = None
83
    output_type: str = "generate_until"
84
    generation_kwargs: dict = None
85
    repeats: int = 1
lintangsutawika's avatar
lintangsutawika committed
86
    filter_list: Union[str, list] = None
87
88
    should_decontaminate: bool = False
    doc_to_decontamination_query: str = None
89

lintangsutawika's avatar
lintangsutawika committed
90
    metadata: str = None  # by default, not used in the code. allows for users to pass arbitrary info to tasks
91

Ethan Smith's avatar
Ethan Smith committed
92
    def __post_init__(self) -> None:
lintangsutawika's avatar
lintangsutawika committed
93
94
95
        if "." in self.dataset_path:
            import inspect
            from importlib import import_module
lintangsutawika's avatar
format  
lintangsutawika committed
96

lintangsutawika's avatar
lintangsutawika committed
97
            self.dataset_path = inspect.getfile(import_module(self.dataset_path))
98

Lintang Sutawika's avatar
Lintang Sutawika committed
99
        if self.generation_kwargs is not None:
100
            if self.output_type != "generate_until":
Lintang Sutawika's avatar
Lintang Sutawika committed
101
                eval_logger.warning(
102
                    f"[{self.task}] passed `generation_kwargs`, but not using `output_type: generate_until`!"
Lintang Sutawika's avatar
Lintang Sutawika committed
103
                )
104
                assert self.output_type != "generate_until"
Lintang Sutawika's avatar
Lintang Sutawika committed
105
106
107
108
109
110
111

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

            if "until" not in self.generation_kwargs:
112
                self.generation_kwargs["until"] = [self.fewshot_delimiter]
Lintang Sutawika's avatar
Lintang Sutawika committed
113
        else:
114
            if self.output_type == "generate_until":
Lintang Sutawika's avatar
Lintang Sutawika committed
115
116
                # ensure that we greedily generate in absence of explicit arguments otherwise
                self.generation_kwargs = {
Lintang Sutawika's avatar
Lintang Sutawika committed
117
                    "until": None
118
119
                    if self.fewshot_delimiter is None
                    else [self.fewshot_delimiter],
Lintang Sutawika's avatar
Lintang Sutawika committed
120
121
                    "do_sample": False,
                }
122

haileyschoelkopf's avatar
haileyschoelkopf committed
123
124
        # TODO: how to make TaskConfigs be de- and re-serializable, even when using the !function constructor?

125
126
127
    def __getitem__(self, item):
        return getattr(self, item)

128
129
130
    def __setitem__(self, item, value):
        return setattr(self, item, value)

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
    def __init__(
        self,
        data_dir=None,
        cache_dir=None,
        download_mode=None,
        config=None,
Ethan Smith's avatar
Ethan Smith committed
179
    ) -> None:
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
        """
        :param data_dir: str
            Stores the path to a local folder containing the `Task`'s data files.
            Use this to specify the path to manually downloaded data (usually when
            the dataset is not publicly accessible).
        :param cache_dir: str
            The directory to read/write the `Task` dataset. This follows the
            HuggingFace `datasets` API with the default cache directory located at:
                `~/.cache/huggingface/datasets`
            NOTE: You can change the cache location globally for a given process
            to another directory:
                `export HF_DATASETS_CACHE="/path/to/another/directory"`
        :param download_mode: datasets.DownloadMode
            How to treat pre-existing `Task` downloads and data.
            - `datasets.DownloadMode.REUSE_DATASET_IF_EXISTS`
                Reuse download and reuse dataset.
            - `datasets.DownloadMode.REUSE_CACHE_IF_EXISTS`
                Reuse download with fresh dataset.
            - `datasets.DownloadMode.FORCE_REDOWNLOAD`
                Fresh download and fresh dataset.
        """
        self.download(data_dir, cache_dir, download_mode)
        self._training_docs = None
        self._fewshot_docs = None
        self._instances = None

haileyschoelkopf's avatar
haileyschoelkopf committed
206
        self._config = TaskConfig(**config) if config else TaskConfig()
207
208
209

        if not hasattr(self, "_filters"):
            self._filters = []
lintangsutawika's avatar
lintangsutawika committed
210
            for name, components in self._config.get(
211
                "filters", [["none", [["take_first", None]]]]
lintangsutawika's avatar
lintangsutawika committed
212
            ):
213
214
215
                filter_pipeline = build_filter_ensemble(name, components)
                self._filters.append(filter_pipeline)

lintangsutawika's avatar
lintangsutawika committed
216
        self.sampler = samplers.Sampler(
217
218
            list(self.fewshot_docs()), self, rnd=random.Random(1234)
        )
219

Ethan Smith's avatar
Ethan Smith committed
220
    def download(self, data_dir=None, cache_dir=None, download_mode=None) -> None:
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
        """Downloads and returns the task dataset.
        Override this method to download the dataset from a custom API.

        :param data_dir: str
            Stores the path to a local folder containing the `Task`'s data files.
            Use this to specify the path to manually downloaded data (usually when
            the dataset is not publicly accessible).
        :param cache_dir: str
            The directory to read/write the `Task` dataset. This follows the
            HuggingFace `datasets` API with the default cache directory located at:
                `~/.cache/huggingface/datasets`
            NOTE: You can change the cache location globally for a given process
            by setting the shell environment variable, `HF_DATASETS_CACHE`,
            to another directory:
                `export HF_DATASETS_CACHE="/path/to/another/directory"`
        :param download_mode: datasets.DownloadMode
            How to treat pre-existing `Task` downloads and data.
            - `datasets.DownloadMode.REUSE_DATASET_IF_EXISTS`
                Reuse download and reuse dataset.
            - `datasets.DownloadMode.REUSE_CACHE_IF_EXISTS`
                Reuse download with fresh dataset.
            - `datasets.DownloadMode.FORCE_REDOWNLOAD`
                Fresh download and fresh dataset.
        """
245
246
247
248
249
250
251
        self.dataset = datasets.load_dataset(
            path=self.DATASET_PATH,
            name=self.DATASET_NAME,
            data_dir=data_dir,
            cache_dir=cache_dir,
            download_mode=download_mode,
        )
252

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

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

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

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

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

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

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

294
295
296
297
298
299
300
301
302
303
    def fewshot_docs(self):
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        if self.has_training_docs():
            return self.training_docs()
        elif self.has_validation_docs():
            return self.validation_docs()
        else:
lintangsutawika's avatar
lintangsutawika committed
304
            eval_logger.warning(
305
                "has_training_docs and has_validation_docs are False"
306
                ", using test_docs as fewshot_docs but this is not recommended."
lintangsutawika's avatar
lintangsutawika committed
307
            )
308
309
            return self.test_docs()

310
311
312
313
314
315
316
317
318
319
    def _process_doc(self, doc):
        """
        Override this to process (detokenize, strip, replace, etc.) individual
        documents. This can be used in a map over documents of a data split.
        E.g. `map(self._process_doc, self.dataset["validation"])`

        :return: dict
            The processed version of the specified `doc`.
        """
        return doc
lintangsutawika's avatar
lintangsutawika committed
320

321
322
323
324
325
326
327
328
329
330
331
332
333
    @property
    def instances(self):
        """After calling `task.build_all_requests()`, tasks
        maintain a list of the dataset instances which will be evaluated.
        """
        return self._instances

    def fewshot_examples(self, k, rnd):
        if self._training_docs is None:
            self._training_docs = list(self.training_docs())

        return rnd.sample(self._training_docs, k)

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

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

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

Ethan Smith's avatar
Ethan Smith committed
348
    def build_all_requests(self, limit=None, rank=None, world_size=None) -> None:
349
350
351
352
353
354
355
356
357
358
        """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!"

359
        eval_logger.info(
360
            f"Building contexts for task '{self.config.task}' on rank {rank}..."
361
362
        )

363
        instances = []
364
365
        for doc_id, doc in utils.create_iterator(
            enumerate(docs), rank, world_size, limit
lintangsutawika's avatar
lintangsutawika committed
366
        ):
367
            # sample fewshot context #TODO: need to offset doc_id by rank now!
368
            fewshot_ctx = self.fewshot_context(
369
                doc,
370
                self.config.num_fewshot,
371
            )
372

373
            # TODO: we should override self.config.repeats if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
374
375
376
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
377
                metadata=(self.config["task"], doc_id, self.config.repeats),
lintangsutawika's avatar
lintangsutawika committed
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
403

            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
404
            The number of times each instance in a dataset is inferred on. Defaults to 1,
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
439
            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
440
441
442
443
444
445
446
447
448
449
    @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))

450
    @utils.positional_deprecated
451
    def fewshot_context(self, doc, num_fewshot):
452
453
454
455
456
457
458
459
460
461
462
463
        """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:
464
            # always prepend the (possibly empty) task description
465
            labeled_examples = self.config.description
466
        else:
467
            labeled_examples = self.config.description + self.sampler.get_context(
lintangsutawika's avatar
lintangsutawika committed
468
469
                doc, num_fewshot
            )
470
471

        example = self.doc_to_text(doc)
472
473
474
475
        if type(example) == str:
            return labeled_examples + example
        elif type(example) == list:
            return [labeled_examples + ex for ex in example]
476
        elif type(example) == int:
477
            if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
478
479
480
481
                choices = self.doc_to_choice(doc)
                return labeled_examples + choices[example]
            else:
                return labeled_examples + str(example)
482
483

    def apply_filters(self):
lintangsutawika's avatar
lintangsutawika committed
484
485
486
487
488
489
        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
490

baberabb's avatar
baberabb committed
491
    def dump_config(self) -> dict:
492
        """Returns a dictionary representing the task's config.
493
494
495
496
497

        :returns: str
            The fewshot context.
        """
        # TODO: this should only return the overrides applied to a non-YAML task's configuration.
498
        # (num_fewshot)
499
        return self.config.to_dict()
500

501
502

class ConfigurableTask(Task):
503
    VERSION = "Yaml"
504
    OUTPUT_TYPE = None
505
    CONFIG = None
506
507
508

    def __init__(
        self, data_dir=None, cache_dir=None, download_mode=None, config: dict = None
Ethan Smith's avatar
Ethan Smith committed
509
    ) -> None:  # TODO no super() call here
510
        # Get pre-configured attributes
511
        self._config = self.CONFIG
512

513
        # Use new configurations if there was no preconfiguration
514
        if self.config is None:
515
            self._config = TaskConfig(**config)
516
517
        # Overwrite configs
        else:
lintangsutawika's avatar
lintangsutawika committed
518
            if config is not None:
519
                self._config.__dict__.update(config)
520

521
        if self.config is None:
lintangsutawika's avatar
lintangsutawika committed
522
523
524
            raise ValueError(
                "Must pass a config to ConfigurableTask, either in cls.CONFIG or `config` kwarg"
            )
525

526
527
528
        if self.config.output_type is not None:
            assert self.config.output_type in ALL_OUTPUT_TYPES
            self.OUTPUT_TYPE = self.config.output_type
529

530
531
        if self.config.dataset_path is not None:
            self.DATASET_PATH = self.config.dataset_path
532

533
534
        if self.config.dataset_name is not None:
            self.DATASET_NAME = self.config.dataset_name
535

536
537
538
539
        self._metric_fn_list = {}
        self._metric_fn_kwargs = {}
        self._aggregation_list = {}
        self._higher_is_better = {}
540

541
        if self.config.metric_list is None:
542
            # TODO: handle this in TaskConfig.__post_init__ ?
543
544
            _metric_list = DEFAULT_METRIC_REGISTRY[self.config.output_type]

545
            for metric_name in _metric_list:
haileyschoelkopf's avatar
haileyschoelkopf committed
546
                self._metric_fn_list[metric_name] = get_metric(metric_name)
547
548
549
                self._aggregation_list[metric_name] = get_metric_aggregation(
                    metric_name
                )
haileyschoelkopf's avatar
haileyschoelkopf committed
550
                self._higher_is_better[metric_name] = is_higher_better(metric_name)
551
        else:
552
            for metric_config in self.config.metric_list:
553
554
555
556
557
                assert "metric" in metric_config
                metric_name = metric_config["metric"]
                kwargs = {
                    key: metric_config[key]
                    for key in metric_config
Chris's avatar
Chris committed
558
559
                    if key
                    not in ["metric", "aggregation", "higher_is_better", "hf_evaluate"]
560
                }
Chris's avatar
Chris committed
561
562
563
564
                hf_evaluate_metric = (
                    "hf_evaluate" in metric_config
                    and metric_config["hf_evaluate"] is True
                )
565

566
                if self.config.process_results is not None:
567
568
                    self._metric_fn_list[metric_name] = None
                    self._metric_fn_kwargs[metric_name] = {}
569
570
571
572
573
574
                elif callable(metric_name):
                    metric_fn = metric_name.__call__
                    metric_name = metric_name.__name__
                    self._metric_fn_list[metric_name] = metric_fn
                    self._metric_fn_kwargs[metric_name] = kwargs
                else:
Chris's avatar
Chris committed
575
576
577
                    self._metric_fn_list[metric_name] = get_metric(
                        metric_name, hf_evaluate_metric
                    )
578
                    self._metric_fn_kwargs[metric_name] = kwargs
lintangsutawika's avatar
lintangsutawika committed
579

580
                if "aggregation" in metric_config:
581
                    agg_name = metric_config["aggregation"]
582
                    if type(agg_name) == str:
haileyschoelkopf's avatar
haileyschoelkopf committed
583
                        self._aggregation_list[metric_name] = get_aggregation(agg_name)
584
585
586
587
                    elif callable(agg_name):
                        self._aggregation_list[metric_name] = metric_config[
                            "aggregation"
                        ]
588
                else:
589
                    INV_AGG_REGISTRY = {v: k for k, v in AGGREGATION_REGISTRY.items()}
lintangsutawika's avatar
lintangsutawika committed
590
                    metric_agg = get_metric_aggregation(metric_name)
591
                    eval_logger.warning(
baberabb's avatar
baberabb committed
592
                        f"[Task: {self._config.task}] metric {metric_name} is defined, but aggregation is not. "
593
594
                        f"using default "
                        f"aggregation={INV_AGG_REGISTRY[metric_agg]}"
595
                    )
596
                    self._aggregation_list[metric_name] = metric_agg
lintangsutawika's avatar
lintangsutawika committed
597

598
599
600
601
602
603
                if "higher_is_better" in metric_config:
                    self._higher_is_better[metric_name] = metric_config[
                        "higher_is_better"
                    ]
                else:
                    eval_logger.warning(
baberabb's avatar
baberabb committed
604
                        f"[Task: {self._config.task}] metric {metric_name} is defined, but higher_is_better is not. "
605
                        f"using default "
haileyschoelkopf's avatar
haileyschoelkopf committed
606
                        f"higher_is_better={is_higher_better(metric_name)}"
607
                    )
haileyschoelkopf's avatar
haileyschoelkopf committed
608
                    self._higher_is_better[metric_name] = is_higher_better(metric_name)
609

610
        self.download(self.config.dataset_kwargs)
611
612
613
        self._training_docs = None
        self._fewshot_docs = None

614
        if self.config.filter_list is not None:
lintangsutawika's avatar
lintangsutawika committed
615
            self._filters = []
616
            for filter_config in self.config.filter_list:
lintangsutawika's avatar
lintangsutawika committed
617
618
619
620
621
622
623
                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
624
625
626
                        }
                        components.append([function["function"], kwargs])
                    filter_pipeline = build_filter_ensemble(filter_name, components)
lintangsutawika's avatar
lintangsutawika committed
627
                self._filters.append(filter_pipeline)
lintangsutawika's avatar
lintangsutawika committed
628
        else:
629
            self._filters = [build_filter_ensemble("none", [["take_first", None]])]
630

631
632
        if self.config.use_prompt is not None:
            eval_logger.info(f"loading prompt {self.config.use_prompt}")
633
            self.prompt = get_prompt(
634
                self.config.use_prompt, self.DATASET_PATH, self.DATASET_NAME
lintangsutawika's avatar
lintangsutawika committed
635
            )
636
637
638
        else:
            self.prompt = None

lintangsutawika's avatar
lintangsutawika committed
639
        if self.fewshot_docs() is not None:
haileyschoelkopf's avatar
haileyschoelkopf committed
640
            self.sampler = samplers.get_sampler(
haileyschoelkopf's avatar
haileyschoelkopf committed
641
642
643
                self.config.fewshot_config.get("sampler", "default")
                if self.config.fewshot_config
                else "default"
haileyschoelkopf's avatar
haileyschoelkopf committed
644
            )(list(self.fewshot_docs()), self, rnd=random.Random(1234))
645

646
        if self.has_test_docs():
647
            self.task_docs = self.test_docs()
648
        elif self.has_validation_docs():
649
            self.task_docs = self.validation_docs()
650
651
652
653
654
        else:
            assert (
                False
            ), f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"

655
        # Test One Doc
656
        self.features = list(self.task_docs.features.keys())
657
658
        self.multiple_input = 0
        self.multiple_target = 0
659
        test_doc = self.task_docs[0]
660
        test_text = self.doc_to_text(test_doc)
661
        test_target = self.doc_to_target(test_doc)
lintangsutawika's avatar
lintangsutawika committed
662

663
        if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
664
665
666
            test_choice = self.doc_to_choice(test_doc)
            if type(test_choice) is not list:
                eval_logger.error("doc_to_choice must return list")
667
668
            else:
                num_choice = len(test_choice)
669

670
671
            if type(test_text) is int:
                self.multiple_input = num_choice
672
673
        else:
            test_choice = None
674

675
        if type(test_target) is list:
676
            self.multiple_target = len(test_target)
677
        else:
lintangsutawika's avatar
lintangsutawika committed
678
            if (type(test_target) is int) and (test_choice is not None):
lintangsutawika's avatar
lintangsutawika committed
679
                test_target = test_choice[test_target]
680
            else:
lintangsutawika's avatar
lintangsutawika committed
681
                test_target = str(test_target)
682

683
684
685
        if test_choice is not None:
            check_choices = test_choice
        else:
lintangsutawika's avatar
lintangsutawika committed
686
            check_choices = [test_target]
687
688
689
690
691
        if self.config.doc_to_choice is not None:
            for choice in check_choices:
                choice_has_whitespace = True if choice[0].isspace() else False
                delimiter_has_whitespace = (
                    True if self.config.target_delimiter[-1].isspace() else False
692
                )
693

694
695
696
697
698
699
700
701
702
                if delimiter_has_whitespace and choice_has_whitespace:
                    eval_logger.warning(
                        f'Both target_delimiter and target choice: "{choice}" have whitespace'
                    )
                elif (not delimiter_has_whitespace) and (not choice_has_whitespace):
                    eval_logger.warning(
                        f'Both target_delimiter and target choice: "{choice}" does not have whitespace, ignore if the language you are evaluating on does not require/use whitespace'
                    )

Ethan Smith's avatar
Ethan Smith committed
703
    def download(self, dataset_kwargs=None) -> None:
704
705
706
707
708
709
        self.dataset = datasets.load_dataset(
            path=self.DATASET_PATH,
            name=self.DATASET_NAME,
            **dataset_kwargs if dataset_kwargs is not None else {},
        )

baberabb's avatar
baberabb committed
710
    def has_training_docs(self) -> bool:
711
        if self.config.training_split is not None:
712
713
714
715
            return True
        else:
            return False

baberabb's avatar
baberabb committed
716
    def has_validation_docs(self) -> bool:
717
        if self.config.validation_split is not None:
718
719
720
721
            return True
        else:
            return False

baberabb's avatar
baberabb committed
722
    def has_test_docs(self) -> bool:
723
        if self.config.test_split is not None:
724
725
726
727
            return True
        else:
            return False

baberabb's avatar
baberabb committed
728
    def training_docs(self) -> datasets.Dataset:
729
        if self.has_training_docs():
730
731
732
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.training_split]
733
                )
734
            return self.dataset[self.config.training_split]
735

baberabb's avatar
baberabb committed
736
    def validation_docs(self) -> datasets.Dataset:
737
        if self.has_validation_docs():
738
739
740
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.validation_split]
741
                )
742
            return self.dataset[self.config.validation_split]
743

baberabb's avatar
baberabb committed
744
    def test_docs(self) -> datasets.Dataset:
745
        if self.has_test_docs():
746
747
748
            if self.config.process_docs is not None:
                return self.config.process_docs(self.dataset[self.config.test_split])
            return self.dataset[self.config.test_split]
749

750
    def fewshot_docs(self):
751
752
        if self.config.fewshot_split is not None:
            return self.dataset[self.config.fewshot_split]
753
        else:
754
            if self.config.num_fewshot > 0:
755
                eval_logger.warning(
756
                    f"Task '{self.config.task}': "
757
758
759
760
                    "num_fewshot > 0 but fewshot_split is None. "
                    "using preconfigured rule."
                )
            return super().fewshot_docs()
761

762
763
764
765
766
767
768
769
    def apply_filters(self):
        if hasattr(self, "_filters"):
            for f in self._filters:
                f.apply(self._instances, self.task_docs)
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances

770
    def should_decontaminate(self):
771
        return self.config.should_decontaminate
772
773

    def doc_to_decontamination_query(self, doc):
774
775
776
        if self.config.should_decontaminate:
            if self.config.doc_to_decontamination_query in self.features:
                return doc[self.config.doc_to_decontamination_query]
777
778
            else:
                return ast.literal_eval(
779
                    utils.apply_template(self.config.doc_to_decontamination_query, doc)
780
                )
781

782
783
784
785
786
787
788
789
790
791
792
793
    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):
794
795
        if self.prompt is not None:
            doc_to_text = self.prompt
796
        else:
797
            doc_to_text = self.config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
798

799
800
801
        if type(doc_to_text) == int:
            return doc_to_text
        elif type(doc_to_text) == str:
802
            if doc_to_text in self.features:
803
                # if self.config.doc_to_choice is not None:
804
805
                #     return self.doc_to_choice(doc)[doc[doc_to_text]]
                # else:
806
807
                return doc[doc_to_text]
            else:
lintangsutawika's avatar
lintangsutawika committed
808
                text_string = utils.apply_template(doc_to_text, doc)
lintangsutawika's avatar
lintangsutawika committed
809
                if text_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
810
811
812
                    return ast.literal_eval(text_string)
                else:
                    return text_string
813
        elif callable(doc_to_text):
814
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
815
        # Used when applying a Promptsource template
816
        elif hasattr(doc_to_text, "apply"):
817
818
819
820
821
            applied_prompt = doc_to_text.apply(doc)
            if len(applied_prompt) == 2:
                return applied_prompt[0]
            else:
                eval_logger.warning("Applied prompt returns empty string")
822
                return self.config.fewshot_delimiter
823
        else:
824
            print(type(doc_to_text))
825
            raise TypeError
826

827
    def doc_to_target(self, doc: dict) -> Union[int, str, list]:
828
829
        if self.prompt is not None:
            doc_to_target = self.prompt
830
        else:
831
            doc_to_target = self.config.doc_to_target
832

833
834
835
        if type(doc_to_target) == int:
            return doc_to_target
        elif type(doc_to_target) == str:
836
            if doc_to_target in self.features:
837
                # if self.config.doc_to_choice is not None:
838
839
840
                #     return self.doc_to_choice(doc)[doc[doc_to_target]]
                # else:
                return doc[doc_to_target]
841
            else:
lintangsutawika's avatar
lintangsutawika committed
842
                target_string = utils.apply_template(doc_to_target, doc)
lintangsutawika's avatar
lintangsutawika committed
843
                if target_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
844
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
845
846
847
848
849
                elif (
                    len(target_string) >= 2
                    and (target_string[0] == "[")
                    and (target_string[-1] == "]")
                ):
baberabb's avatar
baberabb committed
850
851
852
853
                    try:
                        return ast.literal_eval(target_string)
                    except (SyntaxError, ValueError):
                        return target_string
lintangsutawika's avatar
lintangsutawika committed
854
855
                else:
                    return target_string
856
857
        elif type(doc_to_target) == list:
            return doc_to_target
858
        elif callable(doc_to_target):
859
            return doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
860
        # Used when applying a Promptsource template
861
        elif hasattr(doc_to_target, "apply"):
862
            applied_prompt = doc_to_target.apply(doc)
863
864
865
866
            if len(applied_prompt) == 2:
                return applied_prompt[1]
            else:
                eval_logger.warning("Applied prompt returns empty string")
867
                return self.config.fewshot_delimiter
868
869
        else:
            raise TypeError
870

baberabb's avatar
baberabb committed
871
    def doc_to_choice(self, doc: Any) -> List[str]:
872
873
        if self.prompt is not None:
            doc_to_choice = self.prompt
874
        elif self.config.doc_to_choice is None:
875
876
            eval_logger.error("doc_to_choice was called but not set in config")
        else:
877
            doc_to_choice = self.config.doc_to_choice
878
879
880
881
882
883
884
885
886
887
888
889
890

        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
891

892
    def gold_alias(self, doc):
893
894
895
896
897
        # 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]}}.
898
899
        if self.config.gold_alias is not None:
            doc_to_target = self.config.gold_alias
900
        else:
lintangsutawika's avatar
lintangsutawika committed
901
            return self.doc_to_target(doc)
902
903
904
905
906
907
908
909
910
911

        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

baberabb's avatar
baberabb committed
912
913
914
    def construct_requests(
        self, doc: dict, ctx: str, **kwargs
    ) -> Union[List[Instance], Instance]:
915
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
916
            arguments = (ctx, self.doc_to_target(doc))
917
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
918
            arguments = (self.doc_to_target(doc),)
919
        elif self.OUTPUT_TYPE == "multiple_choice":
920
            choices = self.doc_to_choice(doc)
921
            target_delimiter = self.config.target_delimiter
922
923
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
924
                cont = self.doc_to_target(doc)
925
                arguments = [(ctx, f"{target_delimiter}{cont}") for ctx in choices]
926
            else:
927
                # Otherwise they are placed in the continuation
928
                arguments = [(ctx, f"{target_delimiter}{cont}") for cont in choices]
929

930
            request_list = [
931
932
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
933
                    doc=doc,
934
                    arguments=arg,
935
                    idx=i,
936
937
                    **kwargs,
                )
938
                for i, arg in enumerate(arguments)
939
            ]
940
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
941
            if "acc_mutual_info" in self._metric_fn_list.keys():
942
943
944
                # if we are calculating multiple choice accuracy
                # using mutual information instead of raw loglikelihood as metric, need unconditional lls.

lintangsutawika's avatar
lintangsutawika committed
945
                # here mutual info refers to calculating
946
947
948
949
950
951
                # 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
952
                            doc=doc,
953
                            arguments=("", "{}".format(choice)),
954
955
956
                            idx=i,
                            **kwargs,
                        )
lintangsutawika's avatar
lintangsutawika committed
957
                        for i, choice in enumerate(choices)
958
959
960
                    ]
                )
            return request_list
lintangsutawika's avatar
lintangsutawika committed
961

962
        elif self.OUTPUT_TYPE == "generate_until":
963
            arguments = (ctx, self.config.generation_kwargs)
lintangsutawika's avatar
lintangsutawika committed
964
965

        return Instance(
lintangsutawika's avatar
lintangsutawika committed
966
967
            request_type=self.OUTPUT_TYPE, doc=doc, arguments=arguments, idx=0, **kwargs
        )
968
969

    def process_results(self, doc, results):
970
971
        if callable(self.config.process_results):
            return self.config.process_results(doc, results)
lintangsutawika's avatar
lintangsutawika committed
972

973
        result_dict = {}
974
        use_metric = list(self._metric_fn_list.keys())
975
976
977
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
978
979
980
981
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
982
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
983
            (loglikelihood,) = results
984
985
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
haileyschoelkopf's avatar
haileyschoelkopf committed
986
            return {
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
                **(
                    {"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
1002
            }
1003
        elif self.OUTPUT_TYPE == "multiple_choice":
1004
            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
1005

1006
            # retrieve choices in List[str] form, to compute choice lengths, etc.
1007
            choices = self.doc_to_choice(doc)
1008
1009
            completion_len = np.array([float(len(i)) for i in choices])

1010
1011
            if (
                2 * len(choices) == len(lls)
1012
                and "acc_mutual_info" in self._metric_fn_list.keys()
1013
1014
1015
1016
1017
1018
1019
            ):
                # 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]
1020

1021
1022
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
1023

1024
1025
            if self.multiple_input:
                gold = self.doc_to_text(doc)
1026
            else:
1027
                gold = self.doc_to_target(doc)
1028
1029
1030

            gold_index_error = False
            if type(gold) is list:
Lintang Sutawika's avatar
Lintang Sutawika committed
1031
1032
                gold = [i if i < len(choices) else -100 for i in gold]
                if -100 in gold:
1033
1034
1035
                    gold_index_error = True
            else:
                if type(gold) is int:
Lintang Sutawika's avatar
Lintang Sutawika committed
1036
                    gold = gold if gold < len(choices) else -100
1037
                elif type(gold) is str:
Lintang Sutawika's avatar
Lintang Sutawika committed
1038
                    gold = choices.index(gold) if gold in choices else -100
lintangsutawika's avatar
lintangsutawika committed
1039

Lintang Sutawika's avatar
Lintang Sutawika committed
1040
                if gold == -100:
1041
1042
1043
1044
                    gold_index_error = True

            if gold_index_error:
                eval_logger.warning(
lintangsutawika's avatar
lintangsutawika committed
1045
                    f"Label index was not in within range of available choices,"
1046
1047
                    f"Sample:\n\n{doc}\n\n"
                )
lintangsutawika's avatar
lintangsutawika committed
1048

1049
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
1050
1051
                acc = 1.0 if pred in gold else 0.0
                acc_norm = 1.0 if pred_norm in gold else 0.0
Lintang Sutawika's avatar
Lintang Sutawika committed
1052
                exact_match = int(any([is_greedy[i] if i != -100 else 0 for i in gold]))
lintangsutawika's avatar
lintangsutawika committed
1053
1054
1055
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
1056
                # TODO: this gets score of 0 on arc_challenge for pythia-70m. need to test that this works properly
Lintang Sutawika's avatar
Lintang Sutawika committed
1057
                exact_match = int(is_greedy[gold]) if gold != -100 else 0
1058
1059

            result_dict = {
1060
                **({"acc": acc} if "acc" in use_metric else {}),
1061
1062
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
1063
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
1064
                **({"exact_match": exact_match} if "exact_match" in use_metric else {}),
1065
1066
            }

1067
            if "acc_mutual_info" in use_metric:
lintangsutawika's avatar
lintangsutawika committed
1068
1069
1070
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
1071
1072
1073
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

1074
        elif self.OUTPUT_TYPE == "generate_until":
1075
            gold = self.doc_to_target(doc)
Chris's avatar
Chris committed
1076
            result = results[0]
1077
            if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1078
                # If you set doc_to_choice,
lintangsutawika's avatar
lintangsutawika committed
1079
                # it assumes that doc_to_target returns a number.
1080
1081
                choices = self.doc_to_choice(doc)
                gold = choices[gold]
1082
1083
            # we expect multiple_targets to be a list.
            elif self.multiple_target:
baberabb's avatar
baberabb committed
1084
                gold = list(gold)
Chris's avatar
Chris committed
1085
1086
1087
            elif type(gold) != type(result):
                # cast gold to the same type as result
                gold = type(result)(gold)
1088

lintangsutawika's avatar
lintangsutawika committed
1089
            for metric in self._metric_fn_list.keys():
haileyschoelkopf's avatar
haileyschoelkopf committed
1090
1091
1092
1093
1094
                if self.multiple_target:
                    # in the case where we have multiple targets,
                    # return true if any are true
                    # TODO: this may break for multipLe_target, non zero-or-1 metrics
                    scores = []
haileyschoelkopf's avatar
haileyschoelkopf committed
1095
1096
1097
1098
                    if not isinstance(gold, list):
                        # sometimes, a multiple_target dataset has exceptions where one doc has only one string answer
                        # print(gold)
                        gold = [gold]
haileyschoelkopf's avatar
haileyschoelkopf committed
1099
                    for gold_option in gold:
1100
                        try:
1101
                            result_score = self._metric_fn_list[metric](
1102
1103
                                references=[gold_option],
                                predictions=[result],
1104
                                **self._metric_fn_kwargs[metric],
1105
                            )
baberabb's avatar
baberabb committed
1106
1107
1108
                        except (
                            TypeError
                        ):  # TODO: this is hacky and I don't want to do it
1109
                            result_score = self._metric_fn_list[metric](
haileyschoelkopf's avatar
haileyschoelkopf committed
1110
1111
1112
                                [gold_option, result]
                            )
                        if isinstance(result_score, dict):
haileyschoelkopf's avatar
haileyschoelkopf committed
1113
                            # TODO: this handles the case where HF evaluate returns a dict.
1114
                            result_score = result_score[metric]
haileyschoelkopf's avatar
haileyschoelkopf committed
1115
                        scores.append(result_score)
haileyschoelkopf's avatar
haileyschoelkopf committed
1116
                    if any(scores):
1117
                        result_score = 1.0
haileyschoelkopf's avatar
haileyschoelkopf committed
1118
                    else:
1119
                        result_score = 0.0
haileyschoelkopf's avatar
haileyschoelkopf committed
1120
                else:
1121
                    try:
1122
                        result_score = self._metric_fn_list[metric](
1123
1124
                            references=[gold],
                            predictions=[result],
1125
                            **self._metric_fn_kwargs[metric],
1126
                        )
baberabb's avatar
baberabb committed
1127
1128
1129
                    except (
                        TypeError
                    ):  # needed for now in order to use a different interface between our own metrics and HF Evaluate metrics
1130
                        result_score = self._metric_fn_list[metric]([gold, result])
1131
1132
1133
1134
                    if isinstance(result_score, dict):
                        # TODO: this handles the case where HF evaluate returns a dict.
                        result_score = result_score[metric]
                result_dict[metric] = result_score
1135
        else:
lintangsutawika's avatar
lintangsutawika committed
1136
1137
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
1138
                "'loglikelihood', 'loglikelihood_rolling', 'generate_until' or 'multiple_choice'",
1139
            )
1140
1141
1142
1143
1144
1145
1146

        return result_dict

    def aggregation(self):
        return self._aggregation_list

    def higher_is_better(self):
haileyschoelkopf's avatar
haileyschoelkopf committed
1147
        return self._higher_is_better
1148
1149
1150
1151
1152


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

baberabb's avatar
baberabb committed
1153
    def doc_to_target(self, doc: dict) -> str:
1154
1155
        return " " + doc["choices"][doc["gold"]]

baberabb's avatar
baberabb committed
1156
    def construct_requests(self, doc: dict, ctx: str, **kwargs) -> List[Instance]:
1157
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1158
1159
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1160
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1161
                doc=doc,
1162
                arguments=(ctx, " {}".format(choice)),
1163
                idx=i,
1164
1165
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1166
1167
            for i, choice in enumerate(doc["choices"])
        ]
1168

baberabb's avatar
baberabb committed
1169
    def process_results(self, doc: dict, results: List[Tuple[float, bool]]) -> dict:
lintangsutawika's avatar
lintangsutawika committed
1170
1171
1172
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
        gold = doc["gold"]

        acc = 1.0 if np.argmax(results) == gold else 0.0
        completion_len = np.array([float(len(i)) for i in doc["choices"]])
        acc_norm = 1.0 if np.argmax(results / completion_len) == gold else 0.0

        return {
            "acc": acc,
            "acc_norm": acc_norm,
        }

baberabb's avatar
baberabb committed
1184
    def higher_is_better(self) -> dict:
1185
1186
1187
1188
1189
        return {
            "acc": True,
            "acc_norm": True,
        }

baberabb's avatar
baberabb committed
1190
    def aggregation(self) -> dict:
1191
1192
1193
1194
1195
1196
        return {
            "acc": mean,
            "acc_norm": mean,
        }


lintangsutawika's avatar
lintangsutawika committed
1197
class PerplexityTask(Task):
1198
1199
    OUTPUT_TYPE = "loglikelihood_rolling"

baberabb's avatar
baberabb committed
1200
    def has_training_docs(self) -> bool:
1201
1202
        return False

baberabb's avatar
baberabb committed
1203
    def fewshot_examples(self, k: int, rnd) -> List:
1204
1205
1206
        assert k == 0
        return []

baberabb's avatar
baberabb committed
1207
    def fewshot_context(self, doc: dict, num_fewshot: int) -> Literal[""]:
1208
1209
1210
1211
1212
1213
        assert (
            num_fewshot == 0
        ), "The number of fewshot examples must be 0 for perplexity tasks."

        return ""

baberabb's avatar
baberabb committed
1214
    def higher_is_better(self) -> dict:
1215
1216
1217
1218
1219
1220
1221
1222
1223
        return {
            "word_perplexity": False,
            "byte_perplexity": False,
            "bits_per_byte": False,
        }

    def doc_to_decontamination_query(self, doc):
        return doc

Ethan Smith's avatar
Ethan Smith committed
1224
    def doc_to_text(self, doc) -> str:
1225
1226
1227
1228
1229
        return ""

    def doc_to_target(self, doc):
        return doc

baberabb's avatar
baberabb committed
1230
    def construct_requests(self, doc: dict, ctx: Union[str, None], **kwargs):
1231
1232
        assert not ctx

lintangsutawika's avatar
lintangsutawika committed
1233
1234
1235
1236
1237
1238
1239
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1240

baberabb's avatar
baberabb committed
1241
    def process_results(self, doc: dict, results: float) -> dict:
1242
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1243
1244
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1245
1246
1247
1248
1249
1250
        return {
            "word_perplexity": (loglikelihood, words),
            "byte_perplexity": (loglikelihood, bytes_),
            "bits_per_byte": (loglikelihood, bytes_),
        }

baberabb's avatar
baberabb committed
1251
    def aggregation(self) -> dict:
1252
1253
1254
1255
1256
1257
1258
        return {
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
            "bits_per_byte": bits_per_byte,
        }

    @classmethod
baberabb's avatar
baberabb committed
1259
    def count_bytes(cls, doc) -> int:
1260
1261
1262
        return len(doc.encode("utf-8"))

    @classmethod
baberabb's avatar
baberabb committed
1263
    def count_words(cls, doc) -> int:
1264
1265
        """Downstream tasks with custom word boundaries should override this!"""
        return len(re.split(r"\s+", doc))