task.py 49.1 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
lintangsutawika's avatar
lintangsutawika committed
7
import logging
8
9
10
import evaluate
import random
import itertools
11
import functools
12
from tqdm import tqdm
13
14
15
16

import datasets
import numpy as np

baberabb's avatar
baberabb committed
17
from typing import Union, List, Any, Tuple, Literal
18
from collections.abc import Callable
19

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

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
]

lintangsutawika's avatar
lintangsutawika committed
50

51
eval_logger = logging.getLogger("lm-eval")
52

lintangsutawika's avatar
lintangsutawika committed
53

54
55
@dataclass
class TaskConfig(dict):
56
    # task naming/registry
57
    task: str = None
lintangsutawika's avatar
lintangsutawika committed
58
    task_alias: str = None
59
    group: Union[str, list] = None
lintangsutawika's avatar
lintangsutawika committed
60
    group_alias: Union[str, list] = None
61
62
63
    # HF dataset options.
    # which dataset to use,
    # and what splits for what purpose
64
65
    dataset_path: str = None
    dataset_name: str = None
66
    dataset_kwargs: dict = None
67
68
69
    training_split: str = None
    validation_split: str = None
    test_split: str = None
lintangsutawika's avatar
lintangsutawika committed
70
    fewshot_split: str = None  # TODO: assert that this not None if num_fewshot > 0. (?) assert if this is same split as one evaling (?)
71
72
    # formatting / prompting options.
    # see docs/advanced_task_guide.md for more info
73
    process_docs: Callable = None
74
75
    doc_to_text: Union[Callable, str] = None
    doc_to_target: Union[Callable, str] = None
lintangsutawika's avatar
lintangsutawika committed
76
    doc_to_choice: Union[Callable, str, dict, list] = None
lintangsutawika's avatar
lintangsutawika committed
77
    process_results: Union[Callable, str] = None
78
    use_prompt: str = None
79
    description: str = ""
80
81
    target_delimiter: str = " "
    fewshot_delimiter: str = "\n\n"
haileyschoelkopf's avatar
haileyschoelkopf committed
82
    fewshot_config: dict = None
83
    # runtime configuration options
84
    num_fewshot: int = None
85
    # scoring options
86
    metric_list: list = None
87
    output_type: str = "generate_until"
88
    generation_kwargs: dict = None
89
    repeats: int = 1
lintangsutawika's avatar
lintangsutawika committed
90
    filter_list: Union[str, list] = None
91
92
    should_decontaminate: bool = False
    doc_to_decontamination_query: str = None
93

lintangsutawika's avatar
lintangsutawika committed
94
95
96
    metadata: Union[
        str, list
    ] = None  # by default, not used in the code. allows for users to pass arbitrary info to tasks
97

Ethan Smith's avatar
Ethan Smith committed
98
    def __post_init__(self) -> None:
lintangsutawika's avatar
lintangsutawika committed
99
        if self.dataset_path and ("." in self.dataset_path):
lintangsutawika's avatar
lintangsutawika committed
100
101
            import inspect
            from importlib import import_module
lintangsutawika's avatar
format  
lintangsutawika committed
102

lintangsutawika's avatar
lintangsutawika committed
103
            self.dataset_path = inspect.getfile(import_module(self.dataset_path))
104

Lintang Sutawika's avatar
Lintang Sutawika committed
105
        if self.generation_kwargs is not None:
106
            if self.output_type != "generate_until":
Lintang Sutawika's avatar
Lintang Sutawika committed
107
                eval_logger.warning(
108
                    f"[{self.task}] passed `generation_kwargs`, but not using `output_type: generate_until`!"
Lintang Sutawika's avatar
Lintang Sutawika committed
109
                )
110
                assert self.output_type != "generate_until"
Lintang Sutawika's avatar
Lintang Sutawika committed
111
112
113
114
115
116
117

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

            if "until" not in self.generation_kwargs:
118
                self.generation_kwargs["until"] = [self.fewshot_delimiter]
Lintang Sutawika's avatar
Lintang Sutawika committed
119
        else:
120
            if self.output_type == "generate_until":
Lintang Sutawika's avatar
Lintang Sutawika committed
121
122
                # ensure that we greedily generate in absence of explicit arguments otherwise
                self.generation_kwargs = {
Lintang Sutawika's avatar
Lintang Sutawika committed
123
                    "until": None
124
125
                    if self.fewshot_delimiter is None
                    else [self.fewshot_delimiter],
Lintang Sutawika's avatar
Lintang Sutawika committed
126
127
                    "do_sample": False,
                }
128

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

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

134
135
136
    def __setitem__(self, item, value):
        return setattr(self, item, value)

137
    def to_dict(self):
138
139
        """dumps the current config as a dictionary object, as a printable format.
        null fields will not be printed.
haileyschoelkopf's avatar
haileyschoelkopf committed
140
        Used for dumping results alongside full task configuration
141

haileyschoelkopf's avatar
haileyschoelkopf committed
142
143
144
145
146
147
148
149
150
151
        :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
152
153
154
            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
155
        return cfg_dict
156

157
158
159
160
161
162
163
164
165
166
167
168

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
169

170
171
172
173
174
175
176
177
    # 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
178

179
180
181
182
183
184
    def __init__(
        self,
        data_dir=None,
        cache_dir=None,
        download_mode=None,
        config=None,
Ethan Smith's avatar
Ethan Smith committed
185
    ) -> None:
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
        """
        :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

lintangsutawika's avatar
lintangsutawika committed
212
        self._config = TaskConfig({**config}) if config else TaskConfig()
213

lintangsutawika's avatar
lintangsutawika committed
214
        self._filters = [build_filter_ensemble("none", [["take_first", None]])]
215

Ethan Smith's avatar
Ethan Smith committed
216
    def download(self, data_dir=None, cache_dir=None, download_mode=None) -> None:
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
        """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.
        """
241
242
243
244
245
246
247
        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,
        )
248

249
250
251
252
253
    @property
    def config(self):
        """Returns the TaskConfig associated with this class."""
        return self._config

254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
    @abc.abstractmethod
    def has_training_docs(self):
        """Whether the task has a training set"""
        pass

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

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

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

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

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

290
291
292
293
294
295
296
297
298
299
    def fewshot_docs(self):
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        if self.has_training_docs():
            return self.training_docs()
        elif self.has_validation_docs():
            return self.validation_docs()
        else:
lintangsutawika's avatar
lintangsutawika committed
300
            eval_logger.warning(
301
                "has_training_docs and has_validation_docs are False"
302
                ", using test_docs as fewshot_docs but this is not recommended."
lintangsutawika's avatar
lintangsutawika committed
303
            )
304
305
            return self.test_docs()

306
307
308
309
310
311
312
313
314
315
    def _process_doc(self, doc):
        """
        Override this to process (detokenize, strip, replace, etc.) individual
        documents. This can be used in a map over documents of a data split.
        E.g. `map(self._process_doc, self.dataset["validation"])`

        :return: dict
            The processed version of the specified `doc`.
        """
        return doc
lintangsutawika's avatar
lintangsutawika committed
316

317
318
319
320
321
322
323
324
325
326
327
328
329
    @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
330
    def doc_to_decontamination_query(self, doc) -> None:
331
332
333
334
335
336
337
338
339
340
341
342
343
        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
344
    def build_all_requests(self, limit=None, rank=None, world_size=None) -> None:
345
346
347
348
349
350
351
352
353
354
        """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!"

355
        eval_logger.info(f"Building contexts for task on rank {rank}...")
356

357
        instances = []
358
359
        for doc_id, doc in utils.create_iterator(
            enumerate(docs), rank, world_size, limit
lintangsutawika's avatar
lintangsutawika committed
360
        ):
361
            # sample fewshot context #TODO: need to offset doc_id by rank now!
362
            fewshot_ctx = self.fewshot_context(
363
                doc,
364
                0 if self.config.num_fewshot is None else self.config.num_fewshot,
365
            )
366

367
            # TODO: we should override self.config.repeats if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
368
369
370
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
371
                metadata=(self.config["task"], doc_id, self.config.repeats),
lintangsutawika's avatar
lintangsutawika committed
372
            )
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397

            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
398
            The number of times each instance in a dataset is inferred on. Defaults to 1,
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
            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
434
435
436
437
438
439
440
441
442
443
    @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))

444
    @utils.positional_deprecated
lintangsutawika's avatar
lintangsutawika committed
445
    def fewshot_context(
446
447
448
449
450
        self,
        doc,
        num_fewshot,
        rnd=random.Random(1234),
        description=None,
lintangsutawika's avatar
lintangsutawika committed
451
    ):
452
453
454
455
456
457
458
        """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.
lintangsutawika's avatar
lintangsutawika committed
459
460
461
462
463
        :param rnd: random.Random
            The pseudo-random number generator used to randomly sample examples.
            WARNING: This is currently a required arg although it's optionalized with a default `None`.
        :param description: str
            The task's description that will be prepended to the fewshot examples.
464
465
466
        :returns: str
            The fewshot context.
        """
lintangsutawika's avatar
lintangsutawika committed
467
468
469
470
        assert (
            rnd is not None
        ), "A `random.Random` generator argument must be provided to `rnd`"

471
        description = description if description else ""
472
473

        if num_fewshot == 0:
lintangsutawika's avatar
lintangsutawika committed
474
            labeled_examples = ""
475
        else:
lintangsutawika's avatar
lintangsutawika committed
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
            # for sets with no training docs, draw from other set *but ensure no overlap with current doc*
            if self.has_training_docs():
                fewshotex = self.fewshot_examples(k=num_fewshot, rnd=rnd)
            else:
                if self._fewshot_docs is None:
                    self._fewshot_docs = list(
                        self.validation_docs()
                        if self.has_validation_docs()
                        else self.test_docs()
                    )

                fewshotex = rnd.sample(self._fewshot_docs, num_fewshot + 1)

                # get rid of the doc that's the one we're evaluating, if it's in the fewshot
                fewshotex = [x for x in fewshotex if x != doc][:num_fewshot]

            labeled_examples = (
                "\n\n".join(
                    [
                        self.doc_to_text(doc) + self.doc_to_target(doc)
                        for doc in fewshotex
                    ]
                )
                + "\n\n"
lintangsutawika's avatar
lintangsutawika committed
500
            )
501
502

        example = self.doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
503
        return description + labeled_examples + example
504
505

    def apply_filters(self):
lintangsutawika's avatar
lintangsutawika committed
506
507
        if hasattr(self, "_filters"):
            for f in self._filters:
lintangsutawika's avatar
lintangsutawika committed
508
                f.apply(self._instances, None)
lintangsutawika's avatar
lintangsutawika committed
509
510
511
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances
512

baberabb's avatar
baberabb committed
513
    def dump_config(self) -> dict:
514
        """Returns a dictionary representing the task's config.
515
516
517
518
519

        :returns: str
            The fewshot context.
        """
        # TODO: this should only return the overrides applied to a non-YAML task's configuration.
520
        # (num_fewshot)
521
        return self.config.to_dict()
522

523
524

class ConfigurableTask(Task):
525
    VERSION = "Yaml"
526
    OUTPUT_TYPE = None
527
    CONFIG = None
528
529
530

    def __init__(
        self, data_dir=None, cache_dir=None, download_mode=None, config: dict = None
Ethan Smith's avatar
Ethan Smith committed
531
    ) -> None:  # TODO no super() call here
532
        # Get pre-configured attributes
533
        self._config = self.CONFIG
534

535
        # Use new configurations if there was no preconfiguration
536
        if self.config is None:
537
            self._config = TaskConfig(**config)
538
539
        # Overwrite configs
        else:
lintangsutawika's avatar
lintangsutawika committed
540
            if config is not None:
541
                self._config.__dict__.update(config)
542

543
        if self.config is None:
lintangsutawika's avatar
lintangsutawika committed
544
545
546
            raise ValueError(
                "Must pass a config to ConfigurableTask, either in cls.CONFIG or `config` kwarg"
            )
547

548
549
550
        if self.config.output_type is not None:
            assert self.config.output_type in ALL_OUTPUT_TYPES
            self.OUTPUT_TYPE = self.config.output_type
551

552
553
        if self.config.dataset_path is not None:
            self.DATASET_PATH = self.config.dataset_path
554

555
556
        if self.config.dataset_name is not None:
            self.DATASET_NAME = self.config.dataset_name
557

558
559
560
561
        self._metric_fn_list = {}
        self._metric_fn_kwargs = {}
        self._aggregation_list = {}
        self._higher_is_better = {}
562

563
        if self.config.metric_list is None:
564
            # TODO: handle this in TaskConfig.__post_init__ ?
565
566
            _metric_list = DEFAULT_METRIC_REGISTRY[self.config.output_type]

567
            for metric_name in _metric_list:
haileyschoelkopf's avatar
haileyschoelkopf committed
568
                self._metric_fn_list[metric_name] = get_metric(metric_name)
lintangsutawika's avatar
lintangsutawika committed
569
                self._metric_fn_kwargs[metric_name] = {}
570
571
572
                self._aggregation_list[metric_name] = get_metric_aggregation(
                    metric_name
                )
haileyschoelkopf's avatar
haileyschoelkopf committed
573
                self._higher_is_better[metric_name] = is_higher_better(metric_name)
574
        else:
575
            for metric_config in self.config.metric_list:
576
577
578
579
580
                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
581
582
                    if key
                    not in ["metric", "aggregation", "higher_is_better", "hf_evaluate"]
583
                }
Chris's avatar
Chris committed
584
585
586
587
                hf_evaluate_metric = (
                    "hf_evaluate" in metric_config
                    and metric_config["hf_evaluate"] is True
                )
588

589
                if self.config.process_results is not None:
590
591
                    self._metric_fn_list[metric_name] = None
                    self._metric_fn_kwargs[metric_name] = {}
592
593
594
595
596
597
                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
598
599
600
                    self._metric_fn_list[metric_name] = get_metric(
                        metric_name, hf_evaluate_metric
                    )
601
                    self._metric_fn_kwargs[metric_name] = kwargs
lintangsutawika's avatar
lintangsutawika committed
602

603
                if "aggregation" in metric_config:
604
                    agg_name = metric_config["aggregation"]
605
                    if type(agg_name) == str:
haileyschoelkopf's avatar
haileyschoelkopf committed
606
                        self._aggregation_list[metric_name] = get_aggregation(agg_name)
607
608
609
610
                    elif callable(agg_name):
                        self._aggregation_list[metric_name] = metric_config[
                            "aggregation"
                        ]
611
                else:
612
                    INV_AGG_REGISTRY = {v: k for k, v in AGGREGATION_REGISTRY.items()}
lintangsutawika's avatar
lintangsutawika committed
613
                    metric_agg = get_metric_aggregation(metric_name)
614
                    eval_logger.warning(
baberabb's avatar
baberabb committed
615
                        f"[Task: {self._config.task}] metric {metric_name} is defined, but aggregation is not. "
616
617
                        f"using default "
                        f"aggregation={INV_AGG_REGISTRY[metric_agg]}"
618
                    )
619
                    self._aggregation_list[metric_name] = metric_agg
lintangsutawika's avatar
lintangsutawika committed
620

621
622
623
624
625
626
                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
627
                        f"[Task: {self._config.task}] metric {metric_name} is defined, but higher_is_better is not. "
628
                        f"using default "
haileyschoelkopf's avatar
haileyschoelkopf committed
629
                        f"higher_is_better={is_higher_better(metric_name)}"
630
                    )
haileyschoelkopf's avatar
haileyschoelkopf committed
631
                    self._higher_is_better[metric_name] = is_higher_better(metric_name)
632

633
        self.download(self.config.dataset_kwargs)
634
635
636
        self._training_docs = None
        self._fewshot_docs = None

637
        if self.config.filter_list is not None:
lintangsutawika's avatar
lintangsutawika committed
638
            self._filters = []
639
            for filter_config in self.config.filter_list:
lintangsutawika's avatar
lintangsutawika committed
640
641
642
643
644
645
646
                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
647
648
649
                        }
                        components.append([function["function"], kwargs])
                    filter_pipeline = build_filter_ensemble(filter_name, components)
lintangsutawika's avatar
lintangsutawika committed
650
                self._filters.append(filter_pipeline)
lintangsutawika's avatar
lintangsutawika committed
651
        else:
652
            self._filters = [build_filter_ensemble("none", [["take_first", None]])]
653

654
655
        if self.config.use_prompt is not None:
            eval_logger.info(f"loading prompt {self.config.use_prompt}")
656
            self.prompt = get_prompt(
657
                self.config.use_prompt, self.DATASET_PATH, self.DATASET_NAME
lintangsutawika's avatar
lintangsutawika committed
658
            )
659
660
661
        else:
            self.prompt = None

lintangsutawika's avatar
lintangsutawika committed
662
        if self.fewshot_docs() is not None:
haileyschoelkopf's avatar
haileyschoelkopf committed
663
            self.sampler = samplers.get_sampler(
haileyschoelkopf's avatar
haileyschoelkopf committed
664
665
666
                self.config.fewshot_config.get("sampler", "default")
                if self.config.fewshot_config
                else "default"
haileyschoelkopf's avatar
haileyschoelkopf committed
667
            )(list(self.fewshot_docs()), self, rnd=random.Random(1234))
668

669
        if self.has_test_docs():
670
            self.task_docs = self.test_docs()
671
        elif self.has_validation_docs():
672
            self.task_docs = self.validation_docs()
673
674
675
676
677
        else:
            assert (
                False
            ), f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"

678
        # Test One Doc
679
        self.features = list(self.task_docs.features.keys())
680
681
        self.multiple_input = 0
        self.multiple_target = 0
682
        test_doc = self.task_docs[0]
683
        test_text = self.doc_to_text(test_doc)
684
        test_target = self.doc_to_target(test_doc)
lintangsutawika's avatar
lintangsutawika committed
685

686
        if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
687
688
689
            test_choice = self.doc_to_choice(test_doc)
            if type(test_choice) is not list:
                eval_logger.error("doc_to_choice must return list")
690
691
            else:
                num_choice = len(test_choice)
692

693
694
            if type(test_text) is int:
                self.multiple_input = num_choice
695
696
        else:
            test_choice = None
697

698
        if type(test_target) is list:
699
            self.multiple_target = len(test_target)
700
        else:
lintangsutawika's avatar
lintangsutawika committed
701
            if (type(test_target) is int) and (test_choice is not None):
lintangsutawika's avatar
lintangsutawika committed
702
                test_target = test_choice[test_target]
703
            else:
lintangsutawika's avatar
lintangsutawika committed
704
                test_target = str(test_target)
705

706
707
708
        if test_choice is not None:
            check_choices = test_choice
        else:
lintangsutawika's avatar
lintangsutawika committed
709
            check_choices = [test_target]
710
711
712
713
        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 = (
714
715
                    True
                    if self.config.target_delimiter.rstrip()
716
                    != self.config.target_delimiter
717
                    else False
718
                )
719

720
721
722
723
724
725
                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(
726
                        f'Both target_delimiter "{self.config.target_delimiter}" and target choice: "{choice}" do not have whitespace, ignore if the language you are evaluating on does not require/use whitespace'
727
728
                    )

Ethan Smith's avatar
Ethan Smith committed
729
    def download(self, dataset_kwargs=None) -> None:
730
731
732
733
734
735
        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
736
    def has_training_docs(self) -> bool:
737
        if self.config.training_split is not None:
738
739
740
741
            return True
        else:
            return False

baberabb's avatar
baberabb committed
742
    def has_validation_docs(self) -> bool:
743
        if self.config.validation_split is not None:
744
745
746
747
            return True
        else:
            return False

baberabb's avatar
baberabb committed
748
    def has_test_docs(self) -> bool:
749
        if self.config.test_split is not None:
750
751
752
753
            return True
        else:
            return False

baberabb's avatar
baberabb committed
754
    def training_docs(self) -> datasets.Dataset:
755
        if self.has_training_docs():
756
757
758
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.training_split]
759
                )
760
            return self.dataset[self.config.training_split]
761

baberabb's avatar
baberabb committed
762
    def validation_docs(self) -> datasets.Dataset:
763
        if self.has_validation_docs():
764
765
766
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.validation_split]
767
                )
768
            return self.dataset[self.config.validation_split]
769

baberabb's avatar
baberabb committed
770
    def test_docs(self) -> datasets.Dataset:
771
        if self.has_test_docs():
772
773
774
            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]
775

776
    def fewshot_docs(self):
777
778
        if self.config.fewshot_split is not None:
            return self.dataset[self.config.fewshot_split]
779
        else:
780
            if (self.config.num_fewshot is not None) and (self.config.num_fewshot > 0):
781
                eval_logger.warning(
782
                    f"Task '{self.config.task}': "
783
784
785
786
                    "num_fewshot > 0 but fewshot_split is None. "
                    "using preconfigured rule."
                )
            return super().fewshot_docs()
787

lintangsutawika's avatar
lintangsutawika committed
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
    @utils.positional_deprecated
    def fewshot_context(self, doc, num_fewshot):
        """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:
            # always prepend the (possibly empty) task description
            labeled_examples = self.config.description
        else:
            labeled_examples = self.config.description + self.sampler.get_context(
                doc, num_fewshot
            )

        example = self.doc_to_text(doc)
        if type(example) == str:
            return labeled_examples + example
        elif type(example) == list:
            return [labeled_examples + ex for ex in example]
        elif type(example) == int:
            if self.config.doc_to_choice is not None:
                choices = self.doc_to_choice(doc)
                return labeled_examples + choices[example]
            else:
                return labeled_examples + str(example)

821
822
823
824
825
826
827
828
    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

829
    def should_decontaminate(self):
830
        return self.config.should_decontaminate
831
832

    def doc_to_decontamination_query(self, doc):
833
        if self.config.should_decontaminate:
834
835
            if self.config.doc_to_decontamination_query is None:
                return self.doc_to_text(doc)
836
            else:
837
838
839
840
841
842
843
844
845
846
847
                doc_to_decontamination_query = self.config.doc_to_decontamination_query
                if doc_to_decontamination_query in self.features:
                    return doc[doc_to_decontamination_query]
                elif callable(doc_to_decontamination_query):
                    return doc_to_decontamination_query(doc)
                else:
                    return ast.literal_eval(
                        utils.apply_template(
                            self.config.doc_to_decontamination_query, doc
                        )
                    )
848

849
850
851
852
853
854
855
856
857
858
859
860
    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):
861
862
        if self.prompt is not None:
            doc_to_text = self.prompt
863
        else:
864
            doc_to_text = self.config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
865

866
867
868
        if type(doc_to_text) == int:
            return doc_to_text
        elif type(doc_to_text) == str:
869
            if doc_to_text in self.features:
870
                # if self.config.doc_to_choice is not None:
871
872
                #     return self.doc_to_choice(doc)[doc[doc_to_text]]
                # else:
873
874
                return doc[doc_to_text]
            else:
lintangsutawika's avatar
lintangsutawika committed
875
                text_string = utils.apply_template(doc_to_text, doc)
lintangsutawika's avatar
lintangsutawika committed
876
                if text_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
877
878
879
                    return ast.literal_eval(text_string)
                else:
                    return text_string
880
        elif callable(doc_to_text):
881
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
882
        # Used when applying a Promptsource template
883
        elif hasattr(doc_to_text, "apply"):
884
885
886
887
888
            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")
889
                return self.config.fewshot_delimiter
890
        else:
891
            print(type(doc_to_text))
892
            raise TypeError
893

894
    def doc_to_target(self, doc: dict) -> Union[int, str, list]:
895
896
        if self.prompt is not None:
            doc_to_target = self.prompt
897
        else:
898
            doc_to_target = self.config.doc_to_target
899

900
901
902
        if type(doc_to_target) == int:
            return doc_to_target
        elif type(doc_to_target) == str:
903
            if doc_to_target in self.features:
904
                # if self.config.doc_to_choice is not None:
905
906
907
                #     return self.doc_to_choice(doc)[doc[doc_to_target]]
                # else:
                return doc[doc_to_target]
908
            else:
lintangsutawika's avatar
lintangsutawika committed
909
                target_string = utils.apply_template(doc_to_target, doc)
lintangsutawika's avatar
lintangsutawika committed
910
                if target_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
911
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
912
913
914
915
916
                elif (
                    len(target_string) >= 2
                    and (target_string[0] == "[")
                    and (target_string[-1] == "]")
                ):
baberabb's avatar
baberabb committed
917
918
919
920
                    try:
                        return ast.literal_eval(target_string)
                    except (SyntaxError, ValueError):
                        return target_string
lintangsutawika's avatar
lintangsutawika committed
921
922
                else:
                    return target_string
923
924
        elif type(doc_to_target) == list:
            return doc_to_target
925
        elif callable(doc_to_target):
926
            return doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
927
        # Used when applying a Promptsource template
928
        elif hasattr(doc_to_target, "apply"):
929
            applied_prompt = doc_to_target.apply(doc)
930
931
932
933
            if len(applied_prompt) == 2:
                return applied_prompt[1]
            else:
                eval_logger.warning("Applied prompt returns empty string")
934
                return self.config.fewshot_delimiter
935
936
        else:
            raise TypeError
937

baberabb's avatar
baberabb committed
938
    def doc_to_choice(self, doc: Any) -> List[str]:
939
940
        if self.prompt is not None:
            doc_to_choice = self.prompt
941
        elif self.config.doc_to_choice is None:
942
943
            eval_logger.error("doc_to_choice was called but not set in config")
        else:
944
            doc_to_choice = self.config.doc_to_choice
945
946

        if type(doc_to_choice) == str:
947
948
949
950
            if doc_to_choice in self.features:
                return doc[doc_to_choice]
            else:
                return ast.literal_eval(utils.apply_template(doc_to_choice, doc))
951
952
953
954
955
956
957
958
959
960
        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
961

baberabb's avatar
baberabb committed
962
963
964
    def construct_requests(
        self, doc: dict, ctx: str, **kwargs
    ) -> Union[List[Instance], Instance]:
965
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
966
            arguments = (ctx, self.doc_to_target(doc))
967
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
968
            arguments = (self.doc_to_target(doc),)
969
        elif self.OUTPUT_TYPE == "multiple_choice":
970
            choices = self.doc_to_choice(doc)
971
            target_delimiter = self.config.target_delimiter
972
973
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
974
                cont = self.doc_to_target(doc)
975
                arguments = [(ctx, f"{target_delimiter}{cont}") for ctx in choices]
976
            else:
977
                # Otherwise they are placed in the continuation
978
                arguments = [(ctx, f"{target_delimiter}{cont}") for cont in choices]
979

980
            request_list = [
981
982
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
983
                    doc=doc,
984
                    arguments=arg,
985
                    idx=i,
986
987
                    **kwargs,
                )
988
                for i, arg in enumerate(arguments)
989
            ]
990
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
991
            if "acc_mutual_info" in self._metric_fn_list.keys():
992
993
994
                # if we are calculating multiple choice accuracy
                # using mutual information instead of raw loglikelihood as metric, need unconditional lls.

lintangsutawika's avatar
lintangsutawika committed
995
                # here mutual info refers to calculating
996
997
998
999
1000
1001
                # 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
1002
                            doc=doc,
1003
                            arguments=("", "{}".format(choice)),
1004
1005
1006
                            idx=i,
                            **kwargs,
                        )
lintangsutawika's avatar
lintangsutawika committed
1007
                        for i, choice in enumerate(choices)
1008
1009
1010
                    ]
                )
            return request_list
lintangsutawika's avatar
lintangsutawika committed
1011

1012
        elif self.OUTPUT_TYPE == "generate_until":
1013
            arguments = (ctx, self.config.generation_kwargs)
lintangsutawika's avatar
lintangsutawika committed
1014
1015

        return Instance(
lintangsutawika's avatar
lintangsutawika committed
1016
1017
            request_type=self.OUTPUT_TYPE, doc=doc, arguments=arguments, idx=0, **kwargs
        )
1018
1019

    def process_results(self, doc, results):
1020
1021
        if callable(self.config.process_results):
            return self.config.process_results(doc, results)
lintangsutawika's avatar
lintangsutawika committed
1022

1023
        result_dict = {}
1024
        use_metric = list(self._metric_fn_list.keys())
1025
1026
1027
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
1028
1029
1030
1031
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
1032
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
1033
            (loglikelihood,) = results
1034
1035
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
haileyschoelkopf's avatar
haileyschoelkopf committed
1036
            return {
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
                **(
                    {"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
1052
            }
1053
        elif self.OUTPUT_TYPE == "multiple_choice":
1054
            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
1055

1056
            # retrieve choices in List[str] form, to compute choice lengths, etc.
1057
            choices = self.doc_to_choice(doc)
1058
1059
            completion_len = np.array([float(len(i)) for i in choices])

1060
1061
            if (
                2 * len(choices) == len(lls)
1062
                and "acc_mutual_info" in self._metric_fn_list.keys()
1063
1064
1065
1066
1067
1068
1069
            ):
                # 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]
1070

1071
1072
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
1073

1074
1075
            if self.multiple_input:
                gold = self.doc_to_text(doc)
1076
            else:
1077
                gold = self.doc_to_target(doc)
1078
1079
1080

            gold_index_error = False
            if type(gold) is list:
Lintang Sutawika's avatar
Lintang Sutawika committed
1081
1082
                gold = [i if i < len(choices) else -100 for i in gold]
                if -100 in gold:
1083
1084
1085
                    gold_index_error = True
            else:
                if type(gold) is int:
Lintang Sutawika's avatar
Lintang Sutawika committed
1086
                    gold = gold if gold < len(choices) else -100
1087
                elif type(gold) is str:
Lintang Sutawika's avatar
Lintang Sutawika committed
1088
                    gold = choices.index(gold) if gold in choices else -100
lintangsutawika's avatar
lintangsutawika committed
1089

Lintang Sutawika's avatar
Lintang Sutawika committed
1090
                if gold == -100:
1091
1092
1093
1094
                    gold_index_error = True

            if gold_index_error:
                eval_logger.warning(
lintangsutawika's avatar
lintangsutawika committed
1095
                    f"Label index was not in within range of available choices,"
1096
1097
                    f"Sample:\n\n{doc}\n\n"
                )
lintangsutawika's avatar
lintangsutawika committed
1098

1099
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
1100
1101
                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
1102
                exact_match = int(any([is_greedy[i] if i != -100 else 0 for i in gold]))
lintangsutawika's avatar
lintangsutawika committed
1103
1104
1105
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
1106
                # 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
1107
                exact_match = int(is_greedy[gold]) if gold != -100 else 0
1108
1109

            result_dict = {
1110
                **({"acc": acc} if "acc" in use_metric else {}),
1111
1112
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
1113
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
1114
                **({"exact_match": exact_match} if "exact_match" in use_metric else {}),
1115
1116
            }

1117
            if "acc_mutual_info" in use_metric:
lintangsutawika's avatar
lintangsutawika committed
1118
1119
1120
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
1121
1122
1123
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

1124
        elif self.OUTPUT_TYPE == "generate_until":
1125
            gold = self.doc_to_target(doc)
Chris's avatar
Chris committed
1126
            result = results[0]
1127
            if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1128
                # If you set doc_to_choice,
lintangsutawika's avatar
lintangsutawika committed
1129
                # it assumes that doc_to_target returns a number.
1130
1131
                choices = self.doc_to_choice(doc)
                gold = choices[gold]
1132
1133
            # we expect multiple_targets to be a list.
            elif self.multiple_target:
baberabb's avatar
baberabb committed
1134
                gold = list(gold)
Chris's avatar
Chris committed
1135
1136
1137
            elif type(gold) != type(result):
                # cast gold to the same type as result
                gold = type(result)(gold)
1138

lintangsutawika's avatar
lintangsutawika committed
1139
            for metric in self._metric_fn_list.keys():
haileyschoelkopf's avatar
haileyschoelkopf committed
1140
1141
1142
1143
1144
                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
1145
1146
1147
1148
                    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
1149
                    for gold_option in gold:
1150
                        try:
1151
                            result_score = self._metric_fn_list[metric](
1152
1153
                                references=[gold_option],
                                predictions=[result],
1154
                                **self._metric_fn_kwargs[metric],
1155
                            )
baberabb's avatar
baberabb committed
1156
1157
1158
                        except (
                            TypeError
                        ):  # TODO: this is hacky and I don't want to do it
1159
                            result_score = self._metric_fn_list[metric](
haileyschoelkopf's avatar
haileyschoelkopf committed
1160
1161
1162
                                [gold_option, result]
                            )
                        if isinstance(result_score, dict):
haileyschoelkopf's avatar
haileyschoelkopf committed
1163
                            # TODO: this handles the case where HF evaluate returns a dict.
1164
                            result_score = result_score[metric]
haileyschoelkopf's avatar
haileyschoelkopf committed
1165
                        scores.append(result_score)
haileyschoelkopf's avatar
haileyschoelkopf committed
1166
                    if any(scores):
1167
                        result_score = 1.0
haileyschoelkopf's avatar
haileyschoelkopf committed
1168
                    else:
1169
                        result_score = 0.0
haileyschoelkopf's avatar
haileyschoelkopf committed
1170
                else:
1171
                    try:
1172
                        result_score = self._metric_fn_list[metric](
1173
1174
                            references=[gold],
                            predictions=[result],
1175
                            **self._metric_fn_kwargs[metric],
1176
                        )
baberabb's avatar
baberabb committed
1177
1178
1179
                    except (
                        TypeError
                    ):  # needed for now in order to use a different interface between our own metrics and HF Evaluate metrics
1180
                        result_score = self._metric_fn_list[metric]([gold, result])
1181
1182
1183
1184
                    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
1185
        else:
lintangsutawika's avatar
lintangsutawika committed
1186
1187
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
1188
                "'loglikelihood', 'loglikelihood_rolling', 'generate_until' or 'multiple_choice'",
1189
            )
1190
1191
1192
1193
1194
1195
1196

        return result_dict

    def aggregation(self):
        return self._aggregation_list

    def higher_is_better(self):
haileyschoelkopf's avatar
haileyschoelkopf committed
1197
        return self._higher_is_better
1198
1199
1200
1201
1202


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

baberabb's avatar
baberabb committed
1203
    def doc_to_target(self, doc: dict) -> str:
1204
1205
        return " " + doc["choices"][doc["gold"]]

baberabb's avatar
baberabb committed
1206
    def construct_requests(self, doc: dict, ctx: str, **kwargs) -> List[Instance]:
1207
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1208
1209
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1210
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1211
                doc=doc,
1212
                arguments=(ctx, " {}".format(choice)),
1213
                idx=i,
1214
1215
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1216
1217
            for i, choice in enumerate(doc["choices"])
        ]
1218

baberabb's avatar
baberabb committed
1219
    def process_results(self, doc: dict, results: List[Tuple[float, bool]]) -> dict:
lintangsutawika's avatar
lintangsutawika committed
1220
1221
1222
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
        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
1234
    def higher_is_better(self) -> dict:
1235
1236
1237
1238
1239
        return {
            "acc": True,
            "acc_norm": True,
        }

baberabb's avatar
baberabb committed
1240
    def aggregation(self) -> dict:
1241
1242
1243
1244
1245
1246
        return {
            "acc": mean,
            "acc_norm": mean,
        }


lintangsutawika's avatar
lintangsutawika committed
1247
class PerplexityTask(Task):
1248
1249
    OUTPUT_TYPE = "loglikelihood_rolling"

baberabb's avatar
baberabb committed
1250
    def has_training_docs(self) -> bool:
1251
1252
        return False

baberabb's avatar
baberabb committed
1253
    def fewshot_examples(self, k: int, rnd) -> List:
1254
1255
1256
        assert k == 0
        return []

baberabb's avatar
baberabb committed
1257
    def fewshot_context(self, doc: dict, num_fewshot: int) -> Literal[""]:
1258
1259
1260
1261
1262
1263
        assert (
            num_fewshot == 0
        ), "The number of fewshot examples must be 0 for perplexity tasks."

        return ""

baberabb's avatar
baberabb committed
1264
    def higher_is_better(self) -> dict:
1265
1266
1267
1268
1269
1270
1271
1272
1273
        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
1274
    def doc_to_text(self, doc) -> str:
1275
1276
1277
1278
1279
        return ""

    def doc_to_target(self, doc):
        return doc

baberabb's avatar
baberabb committed
1280
    def construct_requests(self, doc: dict, ctx: Union[str, None], **kwargs):
1281
1282
        assert not ctx

lintangsutawika's avatar
lintangsutawika committed
1283
1284
1285
1286
1287
1288
1289
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1290

baberabb's avatar
baberabb committed
1291
    def process_results(self, doc: dict, results: float) -> dict:
1292
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1293
1294
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1295
1296
1297
1298
1299
1300
        return {
            "word_perplexity": (loglikelihood, words),
            "byte_perplexity": (loglikelihood, bytes_),
            "bits_per_byte": (loglikelihood, bytes_),
        }

baberabb's avatar
baberabb committed
1301
    def aggregation(self) -> dict:
1302
1303
1304
1305
1306
1307
1308
        return {
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
            "bits_per_byte": bits_per_byte,
        }

    @classmethod
baberabb's avatar
baberabb committed
1309
    def count_bytes(cls, doc) -> int:
1310
1311
1312
        return len(doc.encode("utf-8"))

    @classmethod
baberabb's avatar
baberabb committed
1313
    def count_words(cls, doc) -> int:
1314
1315
        """Downstream tasks with custom word boundaries should override this!"""
        return len(re.split(r"\s+", doc))