task.py 47.3 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)
lintangsutawika's avatar
lintangsutawika committed
547
                self._metric_fn_kwargs[metric_name] = {}
548
549
550
                self._aggregation_list[metric_name] = get_metric_aggregation(
                    metric_name
                )
haileyschoelkopf's avatar
haileyschoelkopf committed
551
                self._higher_is_better[metric_name] = is_higher_better(metric_name)
552
        else:
553
            for metric_config in self.config.metric_list:
554
555
556
557
558
                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
559
560
                    if key
                    not in ["metric", "aggregation", "higher_is_better", "hf_evaluate"]
561
                }
Chris's avatar
Chris committed
562
563
564
565
                hf_evaluate_metric = (
                    "hf_evaluate" in metric_config
                    and metric_config["hf_evaluate"] is True
                )
566

567
                if self.config.process_results is not None:
568
569
                    self._metric_fn_list[metric_name] = None
                    self._metric_fn_kwargs[metric_name] = {}
570
571
572
573
574
575
                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
576
577
578
                    self._metric_fn_list[metric_name] = get_metric(
                        metric_name, hf_evaluate_metric
                    )
579
                    self._metric_fn_kwargs[metric_name] = kwargs
lintangsutawika's avatar
lintangsutawika committed
580

581
                if "aggregation" in metric_config:
582
                    agg_name = metric_config["aggregation"]
583
                    if type(agg_name) == str:
haileyschoelkopf's avatar
haileyschoelkopf committed
584
                        self._aggregation_list[metric_name] = get_aggregation(agg_name)
585
586
587
588
                    elif callable(agg_name):
                        self._aggregation_list[metric_name] = metric_config[
                            "aggregation"
                        ]
589
                else:
590
                    INV_AGG_REGISTRY = {v: k for k, v in AGGREGATION_REGISTRY.items()}
lintangsutawika's avatar
lintangsutawika committed
591
                    metric_agg = get_metric_aggregation(metric_name)
592
                    eval_logger.warning(
baberabb's avatar
baberabb committed
593
                        f"[Task: {self._config.task}] metric {metric_name} is defined, but aggregation is not. "
594
595
                        f"using default "
                        f"aggregation={INV_AGG_REGISTRY[metric_agg]}"
596
                    )
597
                    self._aggregation_list[metric_name] = metric_agg
lintangsutawika's avatar
lintangsutawika committed
598

599
600
601
602
603
604
                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
605
                        f"[Task: {self._config.task}] metric {metric_name} is defined, but higher_is_better is not. "
606
                        f"using default "
haileyschoelkopf's avatar
haileyschoelkopf committed
607
                        f"higher_is_better={is_higher_better(metric_name)}"
608
                    )
haileyschoelkopf's avatar
haileyschoelkopf committed
609
                    self._higher_is_better[metric_name] = is_higher_better(metric_name)
610

611
        self.download(self.config.dataset_kwargs)
612
613
614
        self._training_docs = None
        self._fewshot_docs = None

615
        if self.config.filter_list is not None:
lintangsutawika's avatar
lintangsutawika committed
616
            self._filters = []
617
            for filter_config in self.config.filter_list:
lintangsutawika's avatar
lintangsutawika committed
618
619
620
621
622
623
624
                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
625
626
627
                        }
                        components.append([function["function"], kwargs])
                    filter_pipeline = build_filter_ensemble(filter_name, components)
lintangsutawika's avatar
lintangsutawika committed
628
                self._filters.append(filter_pipeline)
lintangsutawika's avatar
lintangsutawika committed
629
        else:
630
            self._filters = [build_filter_ensemble("none", [["take_first", None]])]
631

632
633
        if self.config.use_prompt is not None:
            eval_logger.info(f"loading prompt {self.config.use_prompt}")
634
            self.prompt = get_prompt(
635
                self.config.use_prompt, self.DATASET_PATH, self.DATASET_NAME
lintangsutawika's avatar
lintangsutawika committed
636
            )
637
638
639
        else:
            self.prompt = None

lintangsutawika's avatar
lintangsutawika committed
640
        if self.fewshot_docs() is not None:
haileyschoelkopf's avatar
haileyschoelkopf committed
641
            self.sampler = samplers.get_sampler(
haileyschoelkopf's avatar
haileyschoelkopf committed
642
643
644
                self.config.fewshot_config.get("sampler", "default")
                if self.config.fewshot_config
                else "default"
haileyschoelkopf's avatar
haileyschoelkopf committed
645
            )(list(self.fewshot_docs()), self, rnd=random.Random(1234))
646

647
        if self.has_test_docs():
648
            self.task_docs = self.test_docs()
649
        elif self.has_validation_docs():
650
            self.task_docs = self.validation_docs()
651
652
653
654
655
        else:
            assert (
                False
            ), f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"

656
        # Test One Doc
657
        self.features = list(self.task_docs.features.keys())
658
659
        self.multiple_input = 0
        self.multiple_target = 0
660
        test_doc = self.task_docs[0]
661
        test_text = self.doc_to_text(test_doc)
662
        test_target = self.doc_to_target(test_doc)
lintangsutawika's avatar
lintangsutawika committed
663

664
        if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
665
666
667
            test_choice = self.doc_to_choice(test_doc)
            if type(test_choice) is not list:
                eval_logger.error("doc_to_choice must return list")
668
669
            else:
                num_choice = len(test_choice)
670

671
672
            if type(test_text) is int:
                self.multiple_input = num_choice
673
674
        else:
            test_choice = None
675

676
        if type(test_target) is list:
677
            self.multiple_target = len(test_target)
678
        else:
lintangsutawika's avatar
lintangsutawika committed
679
            if (type(test_target) is int) and (test_choice is not None):
lintangsutawika's avatar
lintangsutawika committed
680
                test_target = test_choice[test_target]
681
            else:
lintangsutawika's avatar
lintangsutawika committed
682
                test_target = str(test_target)
683

684
685
686
        if test_choice is not None:
            check_choices = test_choice
        else:
lintangsutawika's avatar
lintangsutawika committed
687
            check_choices = [test_target]
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 = (
692
693
694
695
                    True
                    if self.config.target_delimiter.rstrip()
                    == self.config.target_delimiter
                    else False
696
                )
697

698
699
700
701
702
703
704
705
706
                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
707
    def download(self, dataset_kwargs=None) -> None:
708
709
710
711
712
713
        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
714
    def has_training_docs(self) -> bool:
715
        if self.config.training_split is not None:
716
717
718
719
            return True
        else:
            return False

baberabb's avatar
baberabb committed
720
    def has_validation_docs(self) -> bool:
721
        if self.config.validation_split is not None:
722
723
724
725
            return True
        else:
            return False

baberabb's avatar
baberabb committed
726
    def has_test_docs(self) -> bool:
727
        if self.config.test_split is not None:
728
729
730
731
            return True
        else:
            return False

baberabb's avatar
baberabb committed
732
    def training_docs(self) -> datasets.Dataset:
733
        if self.has_training_docs():
734
735
736
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.training_split]
737
                )
738
            return self.dataset[self.config.training_split]
739

baberabb's avatar
baberabb committed
740
    def validation_docs(self) -> datasets.Dataset:
741
        if self.has_validation_docs():
742
743
744
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.validation_split]
745
                )
746
            return self.dataset[self.config.validation_split]
747

baberabb's avatar
baberabb committed
748
    def test_docs(self) -> datasets.Dataset:
749
        if self.has_test_docs():
750
751
752
            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]
753

754
    def fewshot_docs(self):
755
756
        if self.config.fewshot_split is not None:
            return self.dataset[self.config.fewshot_split]
757
        else:
758
            if self.config.num_fewshot > 0:
759
                eval_logger.warning(
760
                    f"Task '{self.config.task}': "
761
762
763
764
                    "num_fewshot > 0 but fewshot_split is None. "
                    "using preconfigured rule."
                )
            return super().fewshot_docs()
765

766
767
768
769
770
771
772
773
    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

774
    def should_decontaminate(self):
775
        return self.config.should_decontaminate
776
777

    def doc_to_decontamination_query(self, doc):
778
779
780
        if self.config.should_decontaminate:
            if self.config.doc_to_decontamination_query in self.features:
                return doc[self.config.doc_to_decontamination_query]
781
782
            else:
                return ast.literal_eval(
783
                    utils.apply_template(self.config.doc_to_decontamination_query, doc)
784
                )
785

786
787
788
789
790
791
792
793
794
795
796
797
    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):
798
799
        if self.prompt is not None:
            doc_to_text = self.prompt
800
        else:
801
            doc_to_text = self.config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
802

803
804
805
        if type(doc_to_text) == int:
            return doc_to_text
        elif type(doc_to_text) == str:
806
            if doc_to_text in self.features:
807
                # if self.config.doc_to_choice is not None:
808
809
                #     return self.doc_to_choice(doc)[doc[doc_to_text]]
                # else:
810
811
                return doc[doc_to_text]
            else:
lintangsutawika's avatar
lintangsutawika committed
812
                text_string = utils.apply_template(doc_to_text, doc)
lintangsutawika's avatar
lintangsutawika committed
813
                if text_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
814
815
816
                    return ast.literal_eval(text_string)
                else:
                    return text_string
817
        elif callable(doc_to_text):
818
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
819
        # Used when applying a Promptsource template
820
        elif hasattr(doc_to_text, "apply"):
821
822
823
824
825
            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")
826
                return self.config.fewshot_delimiter
827
        else:
828
            print(type(doc_to_text))
829
            raise TypeError
830

831
    def doc_to_target(self, doc: dict) -> Union[int, str, list]:
832
833
        if self.prompt is not None:
            doc_to_target = self.prompt
834
        else:
835
            doc_to_target = self.config.doc_to_target
836

837
838
839
        if type(doc_to_target) == int:
            return doc_to_target
        elif type(doc_to_target) == str:
840
            if doc_to_target in self.features:
841
                # if self.config.doc_to_choice is not None:
842
843
844
                #     return self.doc_to_choice(doc)[doc[doc_to_target]]
                # else:
                return doc[doc_to_target]
845
            else:
lintangsutawika's avatar
lintangsutawika committed
846
                target_string = utils.apply_template(doc_to_target, doc)
lintangsutawika's avatar
lintangsutawika committed
847
                if target_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
848
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
849
850
851
852
853
                elif (
                    len(target_string) >= 2
                    and (target_string[0] == "[")
                    and (target_string[-1] == "]")
                ):
baberabb's avatar
baberabb committed
854
855
856
857
                    try:
                        return ast.literal_eval(target_string)
                    except (SyntaxError, ValueError):
                        return target_string
lintangsutawika's avatar
lintangsutawika committed
858
859
                else:
                    return target_string
860
861
        elif type(doc_to_target) == list:
            return doc_to_target
862
        elif callable(doc_to_target):
863
            return doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
864
        # Used when applying a Promptsource template
865
        elif hasattr(doc_to_target, "apply"):
866
            applied_prompt = doc_to_target.apply(doc)
867
868
869
870
            if len(applied_prompt) == 2:
                return applied_prompt[1]
            else:
                eval_logger.warning("Applied prompt returns empty string")
871
                return self.config.fewshot_delimiter
872
873
        else:
            raise TypeError
874

baberabb's avatar
baberabb committed
875
    def doc_to_choice(self, doc: Any) -> List[str]:
876
877
        if self.prompt is not None:
            doc_to_choice = self.prompt
878
        elif self.config.doc_to_choice is None:
879
880
            eval_logger.error("doc_to_choice was called but not set in config")
        else:
881
            doc_to_choice = self.config.doc_to_choice
882
883
884
885
886
887
888
889
890
891
892
893
894

        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
895

896
    def gold_alias(self, doc):
897
898
899
900
901
        # 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]}}.
902
903
        if self.config.gold_alias is not None:
            doc_to_target = self.config.gold_alias
904
        else:
lintangsutawika's avatar
lintangsutawika committed
905
            return self.doc_to_target(doc)
906
907
908
909
910
911
912
913
914
915

        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
916
917
918
    def construct_requests(
        self, doc: dict, ctx: str, **kwargs
    ) -> Union[List[Instance], Instance]:
919
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
920
            arguments = (ctx, self.doc_to_target(doc))
921
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
922
            arguments = (self.doc_to_target(doc),)
923
        elif self.OUTPUT_TYPE == "multiple_choice":
924
            choices = self.doc_to_choice(doc)
925
            target_delimiter = self.config.target_delimiter
926
927
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
928
                cont = self.doc_to_target(doc)
929
                arguments = [(ctx, f"{target_delimiter}{cont}") for ctx in choices]
930
            else:
931
                # Otherwise they are placed in the continuation
932
                arguments = [(ctx, f"{target_delimiter}{cont}") for cont in choices]
933

934
            request_list = [
935
936
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
937
                    doc=doc,
938
                    arguments=arg,
939
                    idx=i,
940
941
                    **kwargs,
                )
942
                for i, arg in enumerate(arguments)
943
            ]
944
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
945
            if "acc_mutual_info" in self._metric_fn_list.keys():
946
947
948
                # if we are calculating multiple choice accuracy
                # using mutual information instead of raw loglikelihood as metric, need unconditional lls.

lintangsutawika's avatar
lintangsutawika committed
949
                # here mutual info refers to calculating
950
951
952
953
954
955
                # 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
956
                            doc=doc,
957
                            arguments=("", "{}".format(choice)),
958
959
960
                            idx=i,
                            **kwargs,
                        )
lintangsutawika's avatar
lintangsutawika committed
961
                        for i, choice in enumerate(choices)
962
963
964
                    ]
                )
            return request_list
lintangsutawika's avatar
lintangsutawika committed
965

966
        elif self.OUTPUT_TYPE == "generate_until":
967
            arguments = (ctx, self.config.generation_kwargs)
lintangsutawika's avatar
lintangsutawika committed
968
969

        return Instance(
lintangsutawika's avatar
lintangsutawika committed
970
971
            request_type=self.OUTPUT_TYPE, doc=doc, arguments=arguments, idx=0, **kwargs
        )
972
973

    def process_results(self, doc, results):
974
975
        if callable(self.config.process_results):
            return self.config.process_results(doc, results)
lintangsutawika's avatar
lintangsutawika committed
976

977
        result_dict = {}
978
        use_metric = list(self._metric_fn_list.keys())
979
980
981
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
982
983
984
985
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
986
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
987
            (loglikelihood,) = results
988
989
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
haileyschoelkopf's avatar
haileyschoelkopf committed
990
            return {
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
                **(
                    {"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
1006
            }
1007
        elif self.OUTPUT_TYPE == "multiple_choice":
1008
            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
1009

1010
            # retrieve choices in List[str] form, to compute choice lengths, etc.
1011
            choices = self.doc_to_choice(doc)
1012
1013
            completion_len = np.array([float(len(i)) for i in choices])

1014
1015
            if (
                2 * len(choices) == len(lls)
1016
                and "acc_mutual_info" in self._metric_fn_list.keys()
1017
1018
1019
1020
1021
1022
1023
            ):
                # 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]
1024

1025
1026
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
1027

1028
1029
            if self.multiple_input:
                gold = self.doc_to_text(doc)
1030
            else:
1031
                gold = self.doc_to_target(doc)
1032
1033
1034

            gold_index_error = False
            if type(gold) is list:
Lintang Sutawika's avatar
Lintang Sutawika committed
1035
1036
                gold = [i if i < len(choices) else -100 for i in gold]
                if -100 in gold:
1037
1038
1039
                    gold_index_error = True
            else:
                if type(gold) is int:
Lintang Sutawika's avatar
Lintang Sutawika committed
1040
                    gold = gold if gold < len(choices) else -100
1041
                elif type(gold) is str:
Lintang Sutawika's avatar
Lintang Sutawika committed
1042
                    gold = choices.index(gold) if gold in choices else -100
lintangsutawika's avatar
lintangsutawika committed
1043

Lintang Sutawika's avatar
Lintang Sutawika committed
1044
                if gold == -100:
1045
1046
1047
1048
                    gold_index_error = True

            if gold_index_error:
                eval_logger.warning(
lintangsutawika's avatar
lintangsutawika committed
1049
                    f"Label index was not in within range of available choices,"
1050
1051
                    f"Sample:\n\n{doc}\n\n"
                )
lintangsutawika's avatar
lintangsutawika committed
1052

1053
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
1054
1055
                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
1056
                exact_match = int(any([is_greedy[i] if i != -100 else 0 for i in gold]))
lintangsutawika's avatar
lintangsutawika committed
1057
1058
1059
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
1060
                # 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
1061
                exact_match = int(is_greedy[gold]) if gold != -100 else 0
1062
1063

            result_dict = {
1064
                **({"acc": acc} if "acc" in use_metric else {}),
1065
1066
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
1067
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
1068
                **({"exact_match": exact_match} if "exact_match" in use_metric else {}),
1069
1070
            }

1071
            if "acc_mutual_info" in use_metric:
lintangsutawika's avatar
lintangsutawika committed
1072
1073
1074
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
1075
1076
1077
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

1078
        elif self.OUTPUT_TYPE == "generate_until":
1079
            gold = self.doc_to_target(doc)
Chris's avatar
Chris committed
1080
            result = results[0]
1081
            if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1082
                # If you set doc_to_choice,
lintangsutawika's avatar
lintangsutawika committed
1083
                # it assumes that doc_to_target returns a number.
1084
1085
                choices = self.doc_to_choice(doc)
                gold = choices[gold]
1086
1087
            # we expect multiple_targets to be a list.
            elif self.multiple_target:
baberabb's avatar
baberabb committed
1088
                gold = list(gold)
Chris's avatar
Chris committed
1089
1090
1091
            elif type(gold) != type(result):
                # cast gold to the same type as result
                gold = type(result)(gold)
1092

lintangsutawika's avatar
lintangsutawika committed
1093
            for metric in self._metric_fn_list.keys():
haileyschoelkopf's avatar
haileyschoelkopf committed
1094
1095
1096
1097
1098
                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
1099
1100
1101
1102
                    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
1103
                    for gold_option in gold:
1104
                        try:
1105
                            result_score = self._metric_fn_list[metric](
1106
1107
                                references=[gold_option],
                                predictions=[result],
1108
                                **self._metric_fn_kwargs[metric],
1109
                            )
baberabb's avatar
baberabb committed
1110
1111
1112
                        except (
                            TypeError
                        ):  # TODO: this is hacky and I don't want to do it
1113
                            result_score = self._metric_fn_list[metric](
haileyschoelkopf's avatar
haileyschoelkopf committed
1114
1115
1116
                                [gold_option, result]
                            )
                        if isinstance(result_score, dict):
haileyschoelkopf's avatar
haileyschoelkopf committed
1117
                            # TODO: this handles the case where HF evaluate returns a dict.
1118
                            result_score = result_score[metric]
haileyschoelkopf's avatar
haileyschoelkopf committed
1119
                        scores.append(result_score)
haileyschoelkopf's avatar
haileyschoelkopf committed
1120
                    if any(scores):
1121
                        result_score = 1.0
haileyschoelkopf's avatar
haileyschoelkopf committed
1122
                    else:
1123
                        result_score = 0.0
haileyschoelkopf's avatar
haileyschoelkopf committed
1124
                else:
1125
                    try:
1126
                        result_score = self._metric_fn_list[metric](
1127
1128
                            references=[gold],
                            predictions=[result],
1129
                            **self._metric_fn_kwargs[metric],
1130
                        )
baberabb's avatar
baberabb committed
1131
1132
1133
                    except (
                        TypeError
                    ):  # needed for now in order to use a different interface between our own metrics and HF Evaluate metrics
1134
                        result_score = self._metric_fn_list[metric]([gold, result])
1135
1136
1137
1138
                    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
1139
        else:
lintangsutawika's avatar
lintangsutawika committed
1140
1141
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
1142
                "'loglikelihood', 'loglikelihood_rolling', 'generate_until' or 'multiple_choice'",
1143
            )
1144
1145
1146
1147
1148
1149
1150

        return result_dict

    def aggregation(self):
        return self._aggregation_list

    def higher_is_better(self):
haileyschoelkopf's avatar
haileyschoelkopf committed
1151
        return self._higher_is_better
1152
1153
1154
1155
1156


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

baberabb's avatar
baberabb committed
1157
    def doc_to_target(self, doc: dict) -> str:
1158
1159
        return " " + doc["choices"][doc["gold"]]

baberabb's avatar
baberabb committed
1160
    def construct_requests(self, doc: dict, ctx: str, **kwargs) -> List[Instance]:
1161
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1162
1163
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1164
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1165
                doc=doc,
1166
                arguments=(ctx, " {}".format(choice)),
1167
                idx=i,
1168
1169
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1170
1171
            for i, choice in enumerate(doc["choices"])
        ]
1172

baberabb's avatar
baberabb committed
1173
    def process_results(self, doc: dict, results: List[Tuple[float, bool]]) -> dict:
lintangsutawika's avatar
lintangsutawika committed
1174
1175
1176
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
        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
1188
    def higher_is_better(self) -> dict:
1189
1190
1191
1192
1193
        return {
            "acc": True,
            "acc_norm": True,
        }

baberabb's avatar
baberabb committed
1194
    def aggregation(self) -> dict:
1195
1196
1197
1198
1199
1200
        return {
            "acc": mean,
            "acc_norm": mean,
        }


lintangsutawika's avatar
lintangsutawika committed
1201
class PerplexityTask(Task):
1202
1203
    OUTPUT_TYPE = "loglikelihood_rolling"

baberabb's avatar
baberabb committed
1204
    def has_training_docs(self) -> bool:
1205
1206
        return False

baberabb's avatar
baberabb committed
1207
    def fewshot_examples(self, k: int, rnd) -> List:
1208
1209
1210
        assert k == 0
        return []

baberabb's avatar
baberabb committed
1211
    def fewshot_context(self, doc: dict, num_fewshot: int) -> Literal[""]:
1212
1213
1214
1215
1216
1217
        assert (
            num_fewshot == 0
        ), "The number of fewshot examples must be 0 for perplexity tasks."

        return ""

baberabb's avatar
baberabb committed
1218
    def higher_is_better(self) -> dict:
1219
1220
1221
1222
1223
1224
1225
1226
1227
        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
1228
    def doc_to_text(self, doc) -> str:
1229
1230
1231
1232
1233
        return ""

    def doc_to_target(self, doc):
        return doc

baberabb's avatar
baberabb committed
1234
    def construct_requests(self, doc: dict, ctx: Union[str, None], **kwargs):
1235
1236
        assert not ctx

lintangsutawika's avatar
lintangsutawika committed
1237
1238
1239
1240
1241
1242
1243
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1244

baberabb's avatar
baberabb committed
1245
    def process_results(self, doc: dict, results: float) -> dict:
1246
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1247
1248
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1249
1250
1251
1252
1253
1254
        return {
            "word_perplexity": (loglikelihood, words),
            "byte_perplexity": (loglikelihood, bytes_),
            "bits_per_byte": (loglikelihood, bytes_),
        }

baberabb's avatar
baberabb committed
1255
    def aggregation(self) -> dict:
1256
1257
1258
1259
1260
1261
1262
        return {
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
            "bits_per_byte": bits_per_byte,
        }

    @classmethod
baberabb's avatar
baberabb committed
1263
    def count_bytes(cls, doc) -> int:
1264
1265
1266
        return len(doc.encode("utf-8"))

    @classmethod
baberabb's avatar
baberabb committed
1267
    def count_words(cls, doc) -> int:
1268
1269
        """Downstream tasks with custom word boundaries should override this!"""
        return len(re.split(r"\s+", doc))